Android 193

README.md

@@ -1,3 +1,19 @@
+| Pull Request | Commit | Title | Author | Merged? |
+|----|----|----|----|----|
+| [12579](https://github.com/yuzu-emu/yuzu-android//pull/12579) | [`66ae60a9e`](https://github.com/yuzu-emu/yuzu-android//pull/12579/files) | Core: Implement Device Mapping & GPU SMMU | [FernandoS27](https://github.com/FernandoS27/) | Yes |
+| [12610](https://github.com/yuzu-emu/yuzu-android//pull/12610) | [`200b371d1`](https://github.com/yuzu-emu/yuzu-android//pull/12610/files) | server_manager: respond to session close correctly | [liamwhite](https://github.com/liamwhite/) | Yes |
+| [12611](https://github.com/yuzu-emu/yuzu-android//pull/12611) | [`2f0b57ca1`](https://github.com/yuzu-emu/yuzu-android//pull/12611/files) | kernel: fix resource management issues | [liamwhite](https://github.com/liamwhite/) | Yes |
+| [12612](https://github.com/yuzu-emu/yuzu-android//pull/12612) | [`76880b84f`](https://github.com/yuzu-emu/yuzu-android//pull/12612/files) | fsp-srv: use program registry for SetCurrentProcess | [liamwhite](https://github.com/liamwhite/) | Yes |
+| [12652](https://github.com/yuzu-emu/yuzu-android//pull/12652) | [`2a0d707ce`](https://github.com/yuzu-emu/yuzu-android//pull/12652/files) | shader_recompiler: emulate 8-bit and 16-bit storage writes with cas loop | [liamwhite](https://github.com/liamwhite/) | Yes |
+| [12659](https://github.com/yuzu-emu/yuzu-android//pull/12659) | [`d94097478`](https://github.com/yuzu-emu/yuzu-android//pull/12659/files) | audio: fetch process object from handle table | [liamwhite](https://github.com/liamwhite/) | Yes |
+| [12665](https://github.com/yuzu-emu/yuzu-android//pull/12665) | [`bee22540a`](https://github.com/yuzu-emu/yuzu-android//pull/12665/files) | service: acc: Only save profiles when profiles have changed | [german77](https://github.com/german77/) | Yes |
+| [12677](https://github.com/yuzu-emu/yuzu-android//pull/12677) | [`d4acdac16`](https://github.com/yuzu-emu/yuzu-android//pull/12677/files) | core: Support multiple modules per patcher | [GPUCode](https://github.com/GPUCode/) | Yes |
+
+
+End of merge log. You can find the original README.md below the break.
+
+-----
+
 <!--
 SPDX-FileCopyrightText: 2018 yuzu Emulator Project
 SPDX-License-Identifier: GPL-2.0-or-later

src/android/app/src/main/jni/native.cpp

@@ -770,8 +770,8 @@ void Java_org_yuzu_yuzu_1emu_NativeLibrary_initializeEmptyUserDirectory(JNIEnv*
     ASSERT(user_id);
 
     const auto user_save_data_path = FileSys::SaveDataFactory::GetFullPath(
-        EmulationSession::GetInstance().System(), vfs_nand_dir, FileSys::SaveDataSpaceId::NandUser,
+        {}, vfs_nand_dir, FileSys::SaveDataSpaceId::NandUser, FileSys::SaveDataType::SaveData, 1,
-        FileSys::SaveDataType::SaveData, 1, user_id->AsU128(), 0);
+        user_id->AsU128(), 0);
 
     const auto full_path = Common::FS::ConcatPathSafe(nand_dir, user_save_data_path);
     if (!Common::FS::CreateParentDirs(full_path)) {
@@ -878,7 +878,7 @@ jstring Java_org_yuzu_yuzu_1emu_NativeLibrary_getSavePath(JNIEnv* env, jobject j
                                                             FileSys::Mode::Read);
 
     const auto user_save_data_path = FileSys::SaveDataFactory::GetFullPath(
-        system, vfsNandDir, FileSys::SaveDataSpaceId::NandUser, FileSys::SaveDataType::SaveData,
+        {}, vfsNandDir, FileSys::SaveDataSpaceId::NandUser, FileSys::SaveDataType::SaveData,
         program_id, user_id->AsU128(), 0);
     return ToJString(env, user_save_data_path);
 }

src/audio_core/device/device_session.cpp

@@ -8,8 +8,11 @@
 #include "audio_core/sink/sink_stream.h"
 #include "core/core.h"
 #include "core/core_timing.h"
+#include "core/guest_memory.h"
 #include "core/memory.h"
 
+#include "core/hle/kernel/k_process.h"
+
 namespace AudioCore {
 
 using namespace std::literals;
@@ -25,7 +28,7 @@ DeviceSession::~DeviceSession() {
 }
 
 Result DeviceSession::Initialize(std::string_view name_, SampleFormat sample_format_,
-                                 u16 channel_count_, size_t session_id_, u32 handle_,
+                                 u16 channel_count_, size_t session_id_, Kernel::KProcess* handle_,
                                  u64 applet_resource_user_id_, Sink::StreamType type_) {
     if (stream) {
         Finalize();
@@ -36,6 +39,7 @@ Result DeviceSession::Initialize(std::string_view name_, SampleFormat sample_for
     channel_count = channel_count_;
     session_id = session_id_;
     handle = handle_;
+    handle->Open();
     applet_resource_user_id = applet_resource_user_id_;
 
     if (type == Sink::StreamType::In) {
@@ -54,6 +58,11 @@ void DeviceSession::Finalize() {
         sink->CloseStream(stream);
         stream = nullptr;
     }
+
+    if (handle) {
+        handle->Close();
+        handle = nullptr;
+    }
 }
 
 void DeviceSession::Start() {
@@ -91,7 +100,7 @@ void DeviceSession::AppendBuffers(std::span<const AudioBuffer> buffers) {
             stream->AppendBuffer(new_buffer, tmp_samples);
         } else {
             Core::Memory::CpuGuestMemory<s16, Core::Memory::GuestMemoryFlags::UnsafeRead> samples(
-                system.ApplicationMemory(), buffer.samples, buffer.size / sizeof(s16));
+                handle->GetMemory(), buffer.samples, buffer.size / sizeof(s16));
             stream->AppendBuffer(new_buffer, samples);
         }
     }
@@ -100,7 +109,7 @@ void DeviceSession::AppendBuffers(std::span<const AudioBuffer> buffers) {
 void DeviceSession::ReleaseBuffer(const AudioBuffer& buffer) const {
     if (type == Sink::StreamType::In) {
         auto samples{stream->ReleaseBuffer(buffer.size / sizeof(s16))};
-        system.ApplicationMemory().WriteBlockUnsafe(buffer.samples, samples.data(), buffer.size);
+        handle->GetMemory().WriteBlockUnsafe(buffer.samples, samples.data(), buffer.size);
     }
 }
 

src/audio_core/device/device_session.h

@@ -20,6 +20,10 @@ struct EventType;
 } // namespace Timing
 } // namespace Core
 
+namespace Kernel {
+class KProcess;
+} // namespace Kernel
+
 namespace AudioCore {
 
 namespace Sink {
@@ -44,13 +48,13 @@ public:
      * @param sample_format           - Sample format for this device's output.
      * @param channel_count           - Number of channels for this device (2 or 6).
      * @param session_id              - This session's id.
-     * @param handle                  - Handle for this device session (unused).
+     * @param handle                  - Process handle for this device session.
      * @param applet_resource_user_id - Applet resource user id for this device session (unused).
      * @param type                    - Type of this stream (Render, In, Out).
      * @return Result code for this call.
      */
     Result Initialize(std::string_view name, SampleFormat sample_format, u16 channel_count,
-                      size_t session_id, u32 handle, u64 applet_resource_user_id,
+                      size_t session_id, Kernel::KProcess* handle, u64 applet_resource_user_id,
                       Sink::StreamType type);
 
     /**
@@ -137,8 +141,8 @@ private:
     u16 channel_count{};
     /// Session id of this device session
     size_t session_id{};
-    /// Handle of this device session
+    /// Process handle of device memory owner
-    u32 handle{};
+    Kernel::KProcess* handle{};
     /// Applet resource user id of this device session
     u64 applet_resource_user_id{};
     /// Total number of samples played by this device session

src/audio_core/in/audio_in_system.cpp

@@ -57,7 +57,7 @@ Result System::IsConfigValid(const std::string_view device_name,
 }
 
 Result System::Initialize(std::string device_name, const AudioInParameter& in_params,
-                          const u32 handle_, const u64 applet_resource_user_id_) {
+                          Kernel::KProcess* handle_, const u64 applet_resource_user_id_) {
     auto result{IsConfigValid(device_name, in_params)};
     if (result.IsError()) {
         return result;

src/audio_core/in/audio_in_system.h

@@ -19,7 +19,8 @@ class System;
 
 namespace Kernel {
 class KEvent;
-}
+class KProcess;
+} // namespace Kernel
 
 namespace AudioCore::AudioIn {
 
@@ -93,12 +94,12 @@ public:
      *
      * @param device_name             - The name of the requested input device.
      * @param in_params               - Input parameters, see AudioInParameter.
-     * @param handle                  - Unused.
+     * @param handle                  - Process handle.
      * @param applet_resource_user_id - Unused.
      * @return Result code.
      */
-    Result Initialize(std::string device_name, const AudioInParameter& in_params, u32 handle,
+    Result Initialize(std::string device_name, const AudioInParameter& in_params,
-                      u64 applet_resource_user_id);
+                      Kernel::KProcess* handle, u64 applet_resource_user_id);
 
     /**
      * Start this system.
@@ -244,8 +245,8 @@ public:
 private:
     /// Core system
     Core::System& system;
-    /// (Unused)
+    /// Process handle
-    u32 handle{};
+    Kernel::KProcess* handle{};
     /// (Unused)
     u64 applet_resource_user_id{};
     /// Buffer event, signalled when a buffer is ready

src/audio_core/out/audio_out_system.cpp

@@ -48,8 +48,8 @@ Result System::IsConfigValid(std::string_view device_name,
     return Service::Audio::ResultInvalidChannelCount;
 }
 
-Result System::Initialize(std::string device_name, const AudioOutParameter& in_params, u32 handle_,
+Result System::Initialize(std::string device_name, const AudioOutParameter& in_params,
-                          u64 applet_resource_user_id_) {
+                          Kernel::KProcess* handle_, u64 applet_resource_user_id_) {
     auto result = IsConfigValid(device_name, in_params);
     if (result.IsError()) {
         return result;

src/audio_core/out/audio_out_system.h

@@ -19,7 +19,8 @@ class System;
 
 namespace Kernel {
 class KEvent;
-}
+class KProcess;
+} // namespace Kernel
 
 namespace AudioCore::AudioOut {
 
@@ -84,12 +85,12 @@ public:
      *
      * @param device_name             - The name of the requested output device.
      * @param in_params               - Input parameters, see AudioOutParameter.
-     * @param handle                  - Unused.
+     * @param handle                  - Process handle.
      * @param applet_resource_user_id - Unused.
      * @return Result code.
      */
-    Result Initialize(std::string device_name, const AudioOutParameter& in_params, u32 handle,
+    Result Initialize(std::string device_name, const AudioOutParameter& in_params,
-                      u64 applet_resource_user_id);
+                      Kernel::KProcess* handle, u64 applet_resource_user_id);
 
     /**
      * Start this system.
@@ -228,8 +229,8 @@ public:
 private:
     /// Core system
     Core::System& system;
-    /// (Unused)
+    /// Process handle
-    u32 handle{};
+    Kernel::KProcess* handle{};
     /// (Unused)
     u64 applet_resource_user_id{};
     /// Buffer event, signalled when a buffer is ready

src/audio_core/renderer/command/data_source/decode.cpp

@@ -9,6 +9,7 @@
 #include "common/fixed_point.h"
 #include "common/logging/log.h"
 #include "common/scratch_buffer.h"
+#include "core/guest_memory.h"
 #include "core/memory.h"
 
 namespace AudioCore::Renderer {

src/common/common_types.h

@@ -45,6 +45,7 @@ using f32 = float;  ///< 32-bit floating point
 using f64 = double; ///< 64-bit floating point
 
 using VAddr = u64;    ///< Represents a pointer in the userspace virtual address space.
+using DAddr = u64;    ///< Represents a pointer in the device specific virtual address space.
 using PAddr = u64;    ///< Represents a pointer in the ARM11 physical address space.
 using GPUVAddr = u64; ///< Represents a pointer in the GPU virtual address space.
 

src/common/page_table.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include "common/page_table.h"
+#include "common/scope_exit.h"
 
 namespace Common {
 
@@ -11,29 +12,10 @@ PageTable::~PageTable() noexcept = default;
 
 bool PageTable::BeginTraversal(TraversalEntry* out_entry, TraversalContext* out_context,
                                Common::ProcessAddress address) const {
-    // Setup invalid defaults.
+    out_context->next_offset = GetInteger(address);
-    out_entry->phys_addr = 0;
+    out_context->next_page = address / page_size;
-    out_entry->block_size = page_size;
-    out_context->next_page = 0;
 
-    // Validate that we can read the actual entry.
+    return this->ContinueTraversal(out_entry, out_context);
-    const auto page = address / page_size;
-    if (page >= backing_addr.size()) {
-        return false;
-    }
-
-    // Validate that the entry is mapped.
-    const auto phys_addr = backing_addr[page];
-    if (phys_addr == 0) {
-        return false;
-    }
-
-    // Populate the results.
-    out_entry->phys_addr = phys_addr + GetInteger(address);
-    out_context->next_page = page + 1;
-    out_context->next_offset = GetInteger(address) + page_size;
-
-    return true;
 }
 
 bool PageTable::ContinueTraversal(TraversalEntry* out_entry, TraversalContext* context) const {
@@ -41,6 +23,12 @@ bool PageTable::ContinueTraversal(TraversalEntry* out_entry, TraversalContext* c
     out_entry->phys_addr = 0;
     out_entry->block_size = page_size;
 
+    // Regardless of whether the page was mapped, advance on exit.
+    SCOPE_EXIT({
+        context->next_page += 1;
+        context->next_offset += page_size;
+    });
+
     // Validate that we can read the actual entry.
     const auto page = context->next_page;
     if (page >= backing_addr.size()) {
@@ -55,8 +43,6 @@ bool PageTable::ContinueTraversal(TraversalEntry* out_entry, TraversalContext* c
 
     // Populate the results.
     out_entry->phys_addr = phys_addr + context->next_offset;
-    context->next_page = page + 1;
-    context->next_offset += page_size;
 
     return true;
 }

src/core/CMakeLists.txt

@@ -37,6 +37,8 @@ add_library(core STATIC
     debugger/gdbstub_arch.h
     debugger/gdbstub.cpp
     debugger/gdbstub.h
+    device_memory_manager.h
+    device_memory_manager.inc
     device_memory.cpp
     device_memory.h
     file_sys/fssystem/fs_i_storage.h
@@ -490,6 +492,10 @@ add_library(core STATIC
     hle/service/filesystem/fsp_pr.h
     hle/service/filesystem/fsp_srv.cpp
     hle/service/filesystem/fsp_srv.h
+    hle/service/filesystem/romfs_controller.cpp
+    hle/service/filesystem/romfs_controller.h
+    hle/service/filesystem/save_data_controller.cpp
+    hle/service/filesystem/save_data_controller.h
     hle/service/fgm/fgm.cpp
     hle/service/fgm/fgm.h
     hle/service/friend/friend.cpp
@@ -605,6 +611,8 @@ add_library(core STATIC
     hle/service/ns/pdm_qry.h
     hle/service/nvdrv/core/container.cpp
     hle/service/nvdrv/core/container.h
+    hle/service/nvdrv/core/heap_mapper.cpp
+    hle/service/nvdrv/core/heap_mapper.h
     hle/service/nvdrv/core/nvmap.cpp
     hle/service/nvdrv/core/nvmap.h
     hle/service/nvdrv/core/syncpoint_manager.cpp

src/core/arm/nce/patcher.cpp

@@ -22,14 +22,10 @@ using NativeExecutionParameters = Kernel::KThread::NativeExecutionParameters;
 constexpr size_t MaxRelativeBranch = 128_MiB;
 constexpr u32 ModuleCodeIndex = 0x24 / sizeof(u32);
 
-Patcher::Patcher() : c(m_patch_instructions) {}
+Patcher::Patcher() : c(m_patch_instructions) {
-
+    // The first word of the patch section is always a branch to the first instruction of the
-Patcher::~Patcher() = default;
+    // module.
-
+    c.dw(0);
-void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
-                        const Kernel::CodeSet::Segment& code) {
-    // Branch to the first instruction of the module.
-    this->BranchToModule(0);
 
     // Write save context helper function.
     c.l(m_save_context);
@@ -38,6 +34,25 @@ void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
     // Write load context helper function.
     c.l(m_load_context);
     WriteLoadContext();
+}
+
+Patcher::~Patcher() = default;
+
+bool Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
+                        const Kernel::CodeSet::Segment& code) {
+    // If we have patched modules but cannot reach the new module, then it needs its own patcher.
+    const size_t image_size = program_image.size();
+    if (total_program_size + image_size > MaxRelativeBranch && total_program_size > 0) {
+        return false;
+    }
+
+    // Add a new module patch to our list
+    modules.emplace_back();
+    curr_patch = &modules.back();
+
+    // The first word of the patch section is always a branch to the first instruction of the
+    // module.
+    curr_patch->m_branch_to_module_relocations.push_back({0, 0});
 
     // Retrieve text segment data.
     const auto text = std::span{program_image}.subspan(code.offset, code.size);
@@ -94,16 +109,17 @@ void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
         }
 
         if (auto exclusive = Exclusive{inst}; exclusive.Verify()) {
-            m_exclusives.push_back(i);
+            curr_patch->m_exclusives.push_back(i);
         }
     }
 
     // Determine patching mode for the final relocation step
-    const size_t image_size = program_image.size();
+    total_program_size += image_size;
     this->mode = image_size > MaxRelativeBranch ? PatchMode::PreText : PatchMode::PostData;
+    return true;
 }
 
-void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
+bool Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
                               const Kernel::CodeSet::Segment& code,
                               Kernel::PhysicalMemory& program_image,
                               EntryTrampolines* out_trampolines) {
@@ -120,7 +136,7 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
         if (mode == PatchMode::PreText) {
             rc.B(rel.patch_offset - patch_size - rel.module_offset);
         } else {
-            rc.B(image_size - rel.module_offset + rel.patch_offset);
+            rc.B(total_program_size - rel.module_offset + rel.patch_offset);
         }
     };
 
@@ -129,7 +145,7 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
         if (mode == PatchMode::PreText) {
             rc.B(patch_size - rel.patch_offset + rel.module_offset);
         } else {
-            rc.B(rel.module_offset - image_size - rel.patch_offset);
+            rc.B(rel.module_offset - total_program_size - rel.patch_offset);
         }
     };
 
@@ -137,7 +153,7 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
         if (mode == PatchMode::PreText) {
             return GetInteger(load_base) + patch_offset;
         } else {
-            return GetInteger(load_base) + image_size + patch_offset;
+            return GetInteger(load_base) + total_program_size + patch_offset;
         }
     };
 
@@ -150,39 +166,50 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
     };
 
     // We are now ready to relocate!
-    for (const Relocation& rel : m_branch_to_patch_relocations) {
+    auto& patch = modules[m_relocate_module_index++];
+    for (const Relocation& rel : patch.m_branch_to_patch_relocations) {
         ApplyBranchToPatchRelocation(text_words.data() + rel.module_offset / sizeof(u32), rel);
     }
-    for (const Relocation& rel : m_branch_to_module_relocations) {
+    for (const Relocation& rel : patch.m_branch_to_module_relocations) {
         ApplyBranchToModuleRelocation(m_patch_instructions.data() + rel.patch_offset / sizeof(u32),
                                       rel);
     }
 
     // Rewrite PC constants and record post trampolines
-    for (const Relocation& rel : m_write_module_pc_relocations) {
+    for (const Relocation& rel : patch.m_write_module_pc_relocations) {
         oaknut::CodeGenerator rc{m_patch_instructions.data() + rel.patch_offset / sizeof(u32)};
         rc.dx(RebasePc(rel.module_offset));
     }
-    for (const Trampoline& rel : m_trampolines) {
+    for (const Trampoline& rel : patch.m_trampolines) {
         out_trampolines->insert({RebasePc(rel.module_offset), RebasePatch(rel.patch_offset)});
     }
 
     // Cortex-A57 seems to treat all exclusives as ordered, but newer processors do not.
     // Convert to ordered to preserve this assumption.
-    for (const ModuleTextAddress i : m_exclusives) {
+    for (const ModuleTextAddress i : patch.m_exclusives) {
         auto exclusive = Exclusive{text_words[i]};
         text_words[i] = exclusive.AsOrdered();
     }
 
-    // Copy to program image
+    // Remove the patched module size from the total. This is done so total_program_size
-    if (this->mode == PatchMode::PreText) {
+    // always represents the distance from the currently patched module to the patch section.
-        std::memcpy(program_image.data(), m_patch_instructions.data(),
+    total_program_size -= image_size;
-                    m_patch_instructions.size() * sizeof(u32));
+
-    } else {
+    // Only copy to the program image of the last module
-        program_image.resize(image_size + patch_size);
+    if (m_relocate_module_index == modules.size()) {
-        std::memcpy(program_image.data() + image_size, m_patch_instructions.data(),
+        if (this->mode == PatchMode::PreText) {
-                    m_patch_instructions.size() * sizeof(u32));
+            ASSERT(image_size == total_program_size);
+            std::memcpy(program_image.data(), m_patch_instructions.data(),
+                        m_patch_instructions.size() * sizeof(u32));
+        } else {
+            program_image.resize(image_size + patch_size);
+            std::memcpy(program_image.data() + image_size, m_patch_instructions.data(),
+                        m_patch_instructions.size() * sizeof(u32));
+        }
+        return true;
     }
+
+    return false;
 }
 
 size_t Patcher::GetSectionSize() const noexcept {
@@ -322,7 +349,7 @@ void Patcher::WriteSvcTrampoline(ModuleDestLabel module_dest, u32 svc_id) {
 
     // Write the post-SVC trampoline address, which will jump back to the guest after restoring its
     // state.
-    m_trampolines.push_back({c.offset(), module_dest});
+    curr_patch->m_trampolines.push_back({c.offset(), module_dest});
 
     // Host called this location. Save the return address so we can
     // unwind the stack properly when jumping back.

src/core/arm/nce/patcher.h

@@ -31,9 +31,9 @@ public:
     explicit Patcher();
     ~Patcher();
 
-    void PatchText(const Kernel::PhysicalMemory& program_image,
+    bool PatchText(const Kernel::PhysicalMemory& program_image,
                    const Kernel::CodeSet::Segment& code);
-    void RelocateAndCopy(Common::ProcessAddress load_base, const Kernel::CodeSet::Segment& code,
+    bool RelocateAndCopy(Common::ProcessAddress load_base, const Kernel::CodeSet::Segment& code,
                          Kernel::PhysicalMemory& program_image, EntryTrampolines* out_trampolines);
     size_t GetSectionSize() const noexcept;
 
@@ -61,16 +61,16 @@ private:
 
 private:
     void BranchToPatch(uintptr_t module_dest) {
-        m_branch_to_patch_relocations.push_back({c.offset(), module_dest});
+        curr_patch->m_branch_to_patch_relocations.push_back({c.offset(), module_dest});
     }
 
     void BranchToModule(uintptr_t module_dest) {
-        m_branch_to_module_relocations.push_back({c.offset(), module_dest});
+        curr_patch->m_branch_to_module_relocations.push_back({c.offset(), module_dest});
         c.dw(0);
     }
 
     void WriteModulePc(uintptr_t module_dest) {
-        m_write_module_pc_relocations.push_back({c.offset(), module_dest});
+        curr_patch->m_write_module_pc_relocations.push_back({c.offset(), module_dest});
         c.dx(0);
     }
 
@@ -84,15 +84,22 @@ private:
         uintptr_t module_offset; ///< Offset in bytes from the start of the text section.
     };
 
+    struct ModulePatch {
+        std::vector<Trampoline> m_trampolines;
+        std::vector<Relocation> m_branch_to_patch_relocations{};
+        std::vector<Relocation> m_branch_to_module_relocations{};
+        std::vector<Relocation> m_write_module_pc_relocations{};
+        std::vector<ModuleTextAddress> m_exclusives{};
+    };
+
     oaknut::VectorCodeGenerator c;
-    std::vector<Trampoline> m_trampolines;
-    std::vector<Relocation> m_branch_to_patch_relocations{};
-    std::vector<Relocation> m_branch_to_module_relocations{};
-    std::vector<Relocation> m_write_module_pc_relocations{};
-    std::vector<ModuleTextAddress> m_exclusives{};
     oaknut::Label m_save_context{};
     oaknut::Label m_load_context{};
     PatchMode mode{PatchMode::None};
+    size_t total_program_size{};
+    size_t m_relocate_module_index{};
+    std::vector<ModulePatch> modules;
+    ModulePatch* curr_patch;
 };
 
 } // namespace Core::NCE

src/core/core.cpp

@@ -28,6 +28,7 @@
 #include "core/file_sys/savedata_factory.h"
 #include "core/file_sys/vfs_concat.h"
 #include "core/file_sys/vfs_real.h"
+#include "core/gpu_dirty_memory_manager.h"
 #include "core/hle/kernel/k_memory_manager.h"
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/kernel/k_resource_limit.h"
@@ -413,6 +414,7 @@ struct System::Impl {
         kernel.ShutdownCores();
         services.reset();
         service_manager.reset();
+        fs_controller.Reset();
         cheat_engine.reset();
         telemetry_session.reset();
         time_manager.Shutdown();
@@ -564,6 +566,9 @@ struct System::Impl {
     std::array<u64, Core::Hardware::NUM_CPU_CORES> dynarmic_ticks{};
     std::array<MicroProfileToken, Core::Hardware::NUM_CPU_CORES> microprofile_cpu{};
 
+    std::array<Core::GPUDirtyMemoryManager, Core::Hardware::NUM_CPU_CORES>
+        gpu_dirty_memory_managers;
+
     std::deque<std::vector<u8>> user_channel;
 };
 
@@ -650,8 +655,14 @@ size_t System::GetCurrentHostThreadID() const {
     return impl->kernel.GetCurrentHostThreadID();
 }
 
-void System::GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback) {
+std::span<GPUDirtyMemoryManager> System::GetGPUDirtyMemoryManager() {
-    return this->ApplicationProcess()->GatherGPUDirtyMemory(callback);
+    return impl->gpu_dirty_memory_managers;
+}
+
+void System::GatherGPUDirtyMemory(std::function<void(PAddr, size_t)>& callback) {
+    for (auto& manager : impl->gpu_dirty_memory_managers) {
+        manager.Gather(callback);
+    }
 }
 
 PerfStatsResults System::GetAndResetPerfStats() {

src/core/core.h

@@ -8,6 +8,7 @@
 #include <functional>
 #include <memory>
 #include <mutex>
+#include <span>
 #include <string>
 #include <vector>
 
@@ -116,6 +117,7 @@ class CpuManager;
 class Debugger;
 class DeviceMemory;
 class ExclusiveMonitor;
+class GPUDirtyMemoryManager;
 class PerfStats;
 class Reporter;
 class SpeedLimiter;
@@ -224,7 +226,9 @@ public:
     /// Prepare the core emulation for a reschedule
     void PrepareReschedule(u32 core_index);
 
-    void GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback);
+    std::span<GPUDirtyMemoryManager> GetGPUDirtyMemoryManager();
+
+    void GatherGPUDirtyMemory(std::function<void(PAddr, size_t)>& callback);
 
     [[nodiscard]] size_t GetCurrentHostThreadID() const;
 

src/core/debugger/debugger.cpp

@@ -114,7 +114,7 @@ public:
     }
 
     Kernel::KThread* GetActiveThread() override {
-        return state->active_thread;
+        return state->active_thread.GetPointerUnsafe();
     }
 
 private:
@@ -147,11 +147,14 @@ private:
 
         std::scoped_lock lk{connection_lock};
 
+        // Find the process we are going to debug.
+        SetDebugProcess();
+
         // Ensure everything is stopped.
         PauseEmulation();
 
         // Set up the new frontend.
-        frontend = std::make_unique<GDBStub>(*this, system);
+        frontend = std::make_unique<GDBStub>(*this, system, debug_process.GetPointerUnsafe());
 
         // Set the new state. This will tear down any existing state.
         state = ConnectionState{
@@ -194,15 +197,20 @@ private:
             UpdateActiveThread();
 
             if (state->info.type == SignalType::Watchpoint) {
-                frontend->Watchpoint(state->active_thread, *state->info.watchpoint);
+                frontend->Watchpoint(std::addressof(*state->active_thread),
+                                     *state->info.watchpoint);
             } else {
-                frontend->Stopped(state->active_thread);
+                frontend->Stopped(std::addressof(*state->active_thread));
             }
 
             break;
         case SignalType::ShuttingDown:
             frontend->ShuttingDown();
 
+            // Release members.
+            state->active_thread.Reset(nullptr);
+            debug_process.Reset(nullptr);
+
             // Wait for emulation to shut down gracefully now.
             state->signal_pipe.close();
             state->client_socket.shutdown(boost::asio::socket_base::shutdown_both);
@@ -222,7 +230,7 @@ private:
                 stopped = true;
                 PauseEmulation();
                 UpdateActiveThread();
-                frontend->Stopped(state->active_thread);
+                frontend->Stopped(state->active_thread.GetPointerUnsafe());
                 break;
             }
             case DebuggerAction::Continue:
@@ -232,7 +240,7 @@ private:
                 MarkResumed([&] {
                     state->active_thread->SetStepState(Kernel::StepState::StepPending);
                     state->active_thread->Resume(Kernel::SuspendType::Debug);
-                    ResumeEmulation(state->active_thread);
+                    ResumeEmulation(state->active_thread.GetPointerUnsafe());
                 });
                 break;
             case DebuggerAction::StepThreadLocked: {
@@ -255,6 +263,7 @@ private:
     }
 
     void PauseEmulation() {
+        Kernel::KScopedLightLock ll{debug_process->GetListLock()};
         Kernel::KScopedSchedulerLock sl{system.Kernel()};
 
         // Put all threads to sleep on next scheduler round.
@@ -264,6 +273,9 @@ private:
     }
 
     void ResumeEmulation(Kernel::KThread* except = nullptr) {
+        Kernel::KScopedLightLock ll{debug_process->GetListLock()};
+        Kernel::KScopedSchedulerLock sl{system.Kernel()};
+
         // Wake up all threads.
         for (auto& thread : ThreadList()) {
             if (std::addressof(thread) == except) {
@@ -277,15 +289,16 @@ private:
 
     template <typename Callback>
     void MarkResumed(Callback&& cb) {
-        Kernel::KScopedSchedulerLock sl{system.Kernel()};
         stopped = false;
         cb();
     }
 
     void UpdateActiveThread() {
+        Kernel::KScopedLightLock ll{debug_process->GetListLock()};
+
         auto& threads{ThreadList()};
         for (auto& thread : threads) {
-            if (std::addressof(thread) == state->active_thread) {
+            if (std::addressof(thread) == state->active_thread.GetPointerUnsafe()) {
                 // Thread is still alive, no need to update.
                 return;
             }
@@ -293,12 +306,18 @@ private:
         state->active_thread = std::addressof(threads.front());
     }
 
+private:
+    void SetDebugProcess() {
+        debug_process = std::move(system.Kernel().GetProcessList().back());
+    }
+
     Kernel::KProcess::ThreadList& ThreadList() {
-        return system.ApplicationProcess()->GetThreadList();
+        return debug_process->GetThreadList();
     }
 
 private:
     System& system;
+    Kernel::KScopedAutoObject<Kernel::KProcess> debug_process;
     std::unique_ptr<DebuggerFrontend> frontend;
 
     boost::asio::io_context io_context;
@@ -310,7 +329,7 @@ private:
         boost::process::async_pipe signal_pipe;
 
         SignalInfo info;
-        Kernel::KThread* active_thread;
+        Kernel::KScopedAutoObject<Kernel::KThread> active_thread;
         std::array<u8, 4096> client_data;
         bool pipe_data;
     };

src/core/debugger/gdbstub.cpp

@@ -108,9 +108,9 @@ static std::string EscapeXML(std::string_view data) {
     return escaped;
 }
 
-GDBStub::GDBStub(DebuggerBackend& backend_, Core::System& system_)
+GDBStub::GDBStub(DebuggerBackend& backend_, Core::System& system_, Kernel::KProcess* debug_process_)
-    : DebuggerFrontend(backend_), system{system_} {
+    : DebuggerFrontend(backend_), system{system_}, debug_process{debug_process_} {
-    if (system.ApplicationProcess()->Is64Bit()) {
+    if (GetProcess()->Is64Bit()) {
         arch = std::make_unique<GDBStubA64>();
     } else {
         arch = std::make_unique<GDBStubA32>();
@@ -276,7 +276,7 @@ void GDBStub::ExecuteCommand(std::string_view packet, std::vector<DebuggerAction
         const size_t size{static_cast<size_t>(strtoll(command.data() + sep, nullptr, 16))};
 
         std::vector<u8> mem(size);
-        if (system.ApplicationMemory().ReadBlock(addr, mem.data(), size)) {
+        if (GetMemory().ReadBlock(addr, mem.data(), size)) {
             // Restore any bytes belonging to replaced instructions.
             auto it = replaced_instructions.lower_bound(addr);
             for (; it != replaced_instructions.end() && it->first < addr + size; it++) {
@@ -310,8 +310,8 @@ void GDBStub::ExecuteCommand(std::string_view packet, std::vector<DebuggerAction
         const auto mem_substr{std::string_view(command).substr(mem_sep)};
         const auto mem{Common::HexStringToVector(mem_substr, false)};
 
-        if (system.ApplicationMemory().WriteBlock(addr, mem.data(), size)) {
+        if (GetMemory().WriteBlock(addr, mem.data(), size)) {
-            Core::InvalidateInstructionCacheRange(system.ApplicationProcess(), addr, size);
+            Core::InvalidateInstructionCacheRange(GetProcess(), addr, size);
             SendReply(GDB_STUB_REPLY_OK);
         } else {
             SendReply(GDB_STUB_REPLY_ERR);
@@ -353,7 +353,7 @@ void GDBStub::HandleBreakpointInsert(std::string_view command) {
     const size_t addr{static_cast<size_t>(strtoll(command.data() + addr_sep, nullptr, 16))};
     const size_t size{static_cast<size_t>(strtoll(command.data() + size_sep, nullptr, 16))};
 
-    if (!system.ApplicationMemory().IsValidVirtualAddressRange(addr, size)) {
+    if (!GetMemory().IsValidVirtualAddressRange(addr, size)) {
         SendReply(GDB_STUB_REPLY_ERR);
         return;
     }
@@ -362,22 +362,20 @@ void GDBStub::HandleBreakpointInsert(std::string_view command) {
 
     switch (type) {
     case BreakpointType::Software:
-        replaced_instructions[addr] = system.ApplicationMemory().Read32(addr);
+        replaced_instructions[addr] = GetMemory().Read32(addr);
-        system.ApplicationMemory().Write32(addr, arch->BreakpointInstruction());
+        GetMemory().Write32(addr, arch->BreakpointInstruction());
-        Core::InvalidateInstructionCacheRange(system.ApplicationProcess(), addr, sizeof(u32));
+        Core::InvalidateInstructionCacheRange(GetProcess(), addr, sizeof(u32));
         success = true;
         break;
     case BreakpointType::WriteWatch:
-        success = system.ApplicationProcess()->InsertWatchpoint(addr, size,
+        success = GetProcess()->InsertWatchpoint(addr, size, Kernel::DebugWatchpointType::Write);
-                                                                Kernel::DebugWatchpointType::Write);
         break;
     case BreakpointType::ReadWatch:
-        success = system.ApplicationProcess()->InsertWatchpoint(addr, size,
+        success = GetProcess()->InsertWatchpoint(addr, size, Kernel::DebugWatchpointType::Read);
-                                                                Kernel::DebugWatchpointType::Read);
         break;
     case BreakpointType::AccessWatch:
-        success = system.ApplicationProcess()->InsertWatchpoint(
+        success =
-            addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
+            GetProcess()->InsertWatchpoint(addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
         break;
     case BreakpointType::Hardware:
     default:
@@ -400,7 +398,7 @@ void GDBStub::HandleBreakpointRemove(std::string_view command) {
     const size_t addr{static_cast<size_t>(strtoll(command.data() + addr_sep, nullptr, 16))};
     const size_t size{static_cast<size_t>(strtoll(command.data() + size_sep, nullptr, 16))};
 
-    if (!system.ApplicationMemory().IsValidVirtualAddressRange(addr, size)) {
+    if (!GetMemory().IsValidVirtualAddressRange(addr, size)) {
         SendReply(GDB_STUB_REPLY_ERR);
         return;
     }
@@ -411,24 +409,22 @@ void GDBStub::HandleBreakpointRemove(std::string_view command) {
     case BreakpointType::Software: {
         const auto orig_insn{replaced_instructions.find(addr)};
         if (orig_insn != replaced_instructions.end()) {
-            system.ApplicationMemory().Write32(addr, orig_insn->second);
+            GetMemory().Write32(addr, orig_insn->second);
-            Core::InvalidateInstructionCacheRange(system.ApplicationProcess(), addr, sizeof(u32));
+            Core::InvalidateInstructionCacheRange(GetProcess(), addr, sizeof(u32));
             replaced_instructions.erase(addr);
             success = true;
         }
         break;
     }
     case BreakpointType::WriteWatch:
-        success = system.ApplicationProcess()->RemoveWatchpoint(addr, size,
+        success = GetProcess()->RemoveWatchpoint(addr, size, Kernel::DebugWatchpointType::Write);
-                                                                Kernel::DebugWatchpointType::Write);
         break;
     case BreakpointType::ReadWatch:
-        success = system.ApplicationProcess()->RemoveWatchpoint(addr, size,
+        success = GetProcess()->RemoveWatchpoint(addr, size, Kernel::DebugWatchpointType::Read);
-                                                                Kernel::DebugWatchpointType::Read);
         break;
     case BreakpointType::AccessWatch:
-        success = system.ApplicationProcess()->RemoveWatchpoint(
+        success =
-            addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
+            GetProcess()->RemoveWatchpoint(addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
         break;
     case BreakpointType::Hardware:
     default:
@@ -466,10 +462,10 @@ void GDBStub::HandleQuery(std::string_view command) {
         const auto target_xml{arch->GetTargetXML()};
         SendReply(PaginateBuffer(target_xml, command.substr(30)));
     } else if (command.starts_with("Offsets")) {
-        const auto main_offset = Core::FindMainModuleEntrypoint(system.ApplicationProcess());
+        const auto main_offset = Core::FindMainModuleEntrypoint(GetProcess());
         SendReply(fmt::format("TextSeg={:x}", GetInteger(main_offset)));
     } else if (command.starts_with("Xfer:libraries:read::")) {
-        auto modules = Core::FindModules(system.ApplicationProcess());
+        auto modules = Core::FindModules(GetProcess());
 
         std::string buffer;
         buffer += R"(<?xml version="1.0"?>)";
@@ -483,7 +479,7 @@ void GDBStub::HandleQuery(std::string_view command) {
         SendReply(PaginateBuffer(buffer, command.substr(21)));
     } else if (command.starts_with("fThreadInfo")) {
         // beginning of list
-        const auto& threads = system.ApplicationProcess()->GetThreadList();
+        const auto& threads = GetProcess()->GetThreadList();
         std::vector<std::string> thread_ids;
         for (const auto& thread : threads) {
             thread_ids.push_back(fmt::format("{:x}", thread.GetThreadId()));
@@ -497,7 +493,7 @@ void GDBStub::HandleQuery(std::string_view command) {
         buffer += R"(<?xml version="1.0"?>)";
         buffer += "<threads>";
 
-        const auto& threads = system.ApplicationProcess()->GetThreadList();
+        const auto& threads = GetProcess()->GetThreadList();
         for (const auto& thread : threads) {
             auto thread_name{Core::GetThreadName(&thread)};
             if (!thread_name) {
@@ -613,7 +609,7 @@ void GDBStub::HandleRcmd(const std::vector<u8>& command) {
     std::string_view command_str{reinterpret_cast<const char*>(&command[0]), command.size()};
     std::string reply;
 
-    auto* process = system.ApplicationProcess();
+    auto* process = GetProcess();
     auto& page_table = process->GetPageTable();
 
     const char* commands = "Commands:\n"
@@ -714,7 +710,7 @@ void GDBStub::HandleRcmd(const std::vector<u8>& command) {
 }
 
 Kernel::KThread* GDBStub::GetThreadByID(u64 thread_id) {
-    auto& threads{system.ApplicationProcess()->GetThreadList()};
+    auto& threads{GetProcess()->GetThreadList()};
     for (auto& thread : threads) {
         if (thread.GetThreadId() == thread_id) {
             return std::addressof(thread);
@@ -783,4 +779,12 @@ void GDBStub::SendStatus(char status) {
     backend.WriteToClient(buf);
 }
 
+Kernel::KProcess* GDBStub::GetProcess() {
+    return debug_process;
+}
+
+Core::Memory::Memory& GDBStub::GetMemory() {
+    return GetProcess()->GetMemory();
+}
+
 } // namespace Core

src/core/debugger/gdbstub.h

@@ -12,13 +12,22 @@
 #include "core/debugger/debugger_interface.h"
 #include "core/debugger/gdbstub_arch.h"
 
+namespace Kernel {
+class KProcess;
+}
+
+namespace Core::Memory {
+class Memory;
+}
+
 namespace Core {
 
 class System;
 
 class GDBStub : public DebuggerFrontend {
 public:
-    explicit GDBStub(DebuggerBackend& backend, Core::System& system);
+    explicit GDBStub(DebuggerBackend& backend, Core::System& system,
+                     Kernel::KProcess* debug_process);
     ~GDBStub() override;
 
     void Connected() override;
@@ -42,8 +51,12 @@ private:
     void SendReply(std::string_view data);
     void SendStatus(char status);
 
+    Kernel::KProcess* GetProcess();
+    Core::Memory::Memory& GetMemory();
+
 private:
     Core::System& system;
+    Kernel::KProcess* debug_process;
     std::unique_ptr<GDBStubArch> arch;
     std::vector<char> current_command;
     std::map<VAddr, u32> replaced_instructions;

src/core/device_memory.h

@@ -31,6 +31,12 @@ public:
                DramMemoryMap::Base;
     }
 
+    template <typename T>
+    PAddr GetRawPhysicalAddr(const T* ptr) const {
+        return static_cast<PAddr>(reinterpret_cast<uintptr_t>(ptr) -
+                                  reinterpret_cast<uintptr_t>(buffer.BackingBasePointer()));
+    }
+
     template <typename T>
     T* GetPointer(Common::PhysicalAddress addr) {
         return reinterpret_cast<T*>(buffer.BackingBasePointer() +
@@ -43,6 +49,16 @@ public:
                                     (GetInteger(addr) - DramMemoryMap::Base));
     }
 
+    template <typename T>
+    T* GetPointerFromRaw(PAddr addr) {
+        return reinterpret_cast<T*>(buffer.BackingBasePointer() + addr);
+    }
+
+    template <typename T>
+    const T* GetPointerFromRaw(PAddr addr) const {
+        return reinterpret_cast<T*>(buffer.BackingBasePointer() + addr);
+    }
+
     Common::HostMemory buffer;
 };
 

src/core/device_memory_manager.h

@@ -0,0 +1,208 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <array>
+#include <atomic>
+#include <deque>
+#include <memory>
+#include <mutex>
+
+#include "common/common_types.h"
+#include "common/scratch_buffer.h"
+#include "common/virtual_buffer.h"
+
+namespace Core {
+
+constexpr size_t DEVICE_PAGEBITS = 12ULL;
+constexpr size_t DEVICE_PAGESIZE = 1ULL << DEVICE_PAGEBITS;
+constexpr size_t DEVICE_PAGEMASK = DEVICE_PAGESIZE - 1ULL;
+
+class DeviceMemory;
+
+namespace Memory {
+class Memory;
+}
+
+template <typename DTraits>
+struct DeviceMemoryManagerAllocator;
+
+template <typename Traits>
+class DeviceMemoryManager {
+    using DeviceInterface = typename Traits::DeviceInterface;
+    using DeviceMethods = typename Traits::DeviceMethods;
+
+public:
+    DeviceMemoryManager(const DeviceMemory& device_memory);
+    ~DeviceMemoryManager();
+
+    void BindInterface(DeviceInterface* device_inter);
+
+    DAddr Allocate(size_t size);
+    void AllocateFixed(DAddr start, size_t size);
+    void Free(DAddr start, size_t size);
+
+    void Map(DAddr address, VAddr virtual_address, size_t size, size_t process_id,
+             bool track = false);
+
+    void Unmap(DAddr address, size_t size);
+
+    void TrackContinuityImpl(DAddr address, VAddr virtual_address, size_t size, size_t process_id);
+    void TrackContinuity(DAddr address, VAddr virtual_address, size_t size, size_t process_id) {
+        std::scoped_lock lk(mapping_guard);
+        TrackContinuityImpl(address, virtual_address, size, process_id);
+    }
+
+    // Write / Read
+    template <typename T>
+    T* GetPointer(DAddr address);
+
+    template <typename T>
+    const T* GetPointer(DAddr address) const;
+
+    template <typename Func>
+    void ApplyOpOnPAddr(PAddr address, Common::ScratchBuffer<u32>& buffer, Func&& operation) {
+        DAddr subbits = static_cast<DAddr>(address & page_mask);
+        const u32 base = compressed_device_addr[(address >> page_bits)];
+        if ((base >> MULTI_FLAG_BITS) == 0) [[likely]] {
+            const DAddr d_address = (static_cast<DAddr>(base) << page_bits) + subbits;
+            operation(d_address);
+            return;
+        }
+        InnerGatherDeviceAddresses(buffer, address);
+        for (u32 value : buffer) {
+            operation((static_cast<DAddr>(value) << page_bits) + subbits);
+        }
+    }
+
+    template <typename Func>
+    void ApplyOpOnPointer(const u8* p, Common::ScratchBuffer<u32>& buffer, Func&& operation) {
+        PAddr address = GetRawPhysicalAddr<u8>(p);
+        ApplyOpOnPAddr(address, buffer, operation);
+    }
+
+    PAddr GetPhysicalRawAddressFromDAddr(DAddr address) const {
+        PAddr subbits = static_cast<PAddr>(address & page_mask);
+        auto paddr = compressed_physical_ptr[(address >> page_bits)];
+        if (paddr == 0) {
+            return 0;
+        }
+        return (static_cast<PAddr>(paddr - 1) << page_bits) + subbits;
+    }
+
+    template <typename T>
+    void Write(DAddr address, T value);
+
+    template <typename T>
+    T Read(DAddr address) const;
+
+    u8* GetSpan(const DAddr src_addr, const std::size_t size);
+    const u8* GetSpan(const DAddr src_addr, const std::size_t size) const;
+
+    void ReadBlock(DAddr address, void* dest_pointer, size_t size);
+    void ReadBlockUnsafe(DAddr address, void* dest_pointer, size_t size);
+    void WriteBlock(DAddr address, const void* src_pointer, size_t size);
+    void WriteBlockUnsafe(DAddr address, const void* src_pointer, size_t size);
+
+    size_t RegisterProcess(Memory::Memory* memory);
+    void UnregisterProcess(size_t id);
+
+    void UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta);
+
+    static constexpr size_t AS_BITS = Traits::device_virtual_bits;
+
+private:
+    static constexpr size_t device_virtual_bits = Traits::device_virtual_bits;
+    static constexpr size_t device_as_size = 1ULL << device_virtual_bits;
+    static constexpr size_t physical_min_bits = 32;
+    static constexpr size_t physical_max_bits = 33;
+    static constexpr size_t page_bits = 12;
+    static constexpr size_t page_size = 1ULL << page_bits;
+    static constexpr size_t page_mask = page_size - 1ULL;
+    static constexpr u32 physical_address_base = 1U << page_bits;
+    static constexpr u32 MULTI_FLAG_BITS = 31;
+    static constexpr u32 MULTI_FLAG = 1U << MULTI_FLAG_BITS;
+    static constexpr u32 MULTI_MASK = ~MULTI_FLAG;
+
+    template <typename T>
+    T* GetPointerFromRaw(PAddr addr) {
+        return reinterpret_cast<T*>(physical_base + addr);
+    }
+
+    template <typename T>
+    const T* GetPointerFromRaw(PAddr addr) const {
+        return reinterpret_cast<T*>(physical_base + addr);
+    }
+
+    template <typename T>
+    PAddr GetRawPhysicalAddr(const T* ptr) const {
+        return static_cast<PAddr>(reinterpret_cast<uintptr_t>(ptr) - physical_base);
+    }
+
+    void WalkBlock(const DAddr addr, const std::size_t size, auto on_unmapped, auto on_memory,
+                   auto increment);
+
+    void InnerGatherDeviceAddresses(Common::ScratchBuffer<u32>& buffer, PAddr address);
+
+    std::unique_ptr<DeviceMemoryManagerAllocator<Traits>> impl;
+
+    const uintptr_t physical_base;
+    DeviceInterface* device_inter;
+    Common::VirtualBuffer<u32> compressed_physical_ptr;
+    Common::VirtualBuffer<u32> compressed_device_addr;
+    Common::VirtualBuffer<u32> continuity_tracker;
+
+    // Process memory interfaces
+
+    std::deque<size_t> id_pool;
+    std::deque<Memory::Memory*> registered_processes;
+
+    // Memory protection management
+
+    static constexpr size_t guest_max_as_bits = 39;
+    static constexpr size_t guest_as_size = 1ULL << guest_max_as_bits;
+    static constexpr size_t guest_mask = guest_as_size - 1ULL;
+    static constexpr size_t process_id_start_bit = guest_max_as_bits;
+
+    std::pair<size_t, VAddr> ExtractCPUBacking(size_t page_index) {
+        auto content = cpu_backing_address[page_index];
+        const VAddr address = content & guest_mask;
+        const size_t process_id = static_cast<size_t>(content >> process_id_start_bit);
+        return std::make_pair(process_id, address);
+    }
+
+    void InsertCPUBacking(size_t page_index, VAddr address, size_t process_id) {
+        cpu_backing_address[page_index] = address | (process_id << process_id_start_bit);
+    }
+
+    Common::VirtualBuffer<VAddr> cpu_backing_address;
+    static constexpr size_t subentries = 8 / sizeof(u8);
+    static constexpr size_t subentries_mask = subentries - 1;
+    class CounterEntry final {
+    public:
+        CounterEntry() = default;
+
+        std::atomic_uint8_t& Count(std::size_t page) {
+            return values[page & subentries_mask];
+        }
+
+        const std::atomic_uint8_t& Count(std::size_t page) const {
+            return values[page & subentries_mask];
+        }
+
+    private:
+        std::array<std::atomic_uint8_t, subentries> values{};
+    };
+    static_assert(sizeof(CounterEntry) == subentries * sizeof(u8),
+                  "CounterEntry should be 8 bytes!");
+
+    static constexpr size_t num_counter_entries =
+        (1ULL << (device_virtual_bits - page_bits)) / subentries;
+    using CachedPages = std::array<CounterEntry, num_counter_entries>;
+    std::unique_ptr<CachedPages> cached_pages;
+    std::mutex counter_guard;
+    std::mutex mapping_guard;
+};
+
+} // namespace Core

src/core/device_memory_manager.inc

@@ -0,0 +1,588 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <atomic>
+#include <limits>
+#include <memory>
+#include <type_traits>
+
+#include "common/address_space.h"
+#include "common/address_space.inc"
+#include "common/alignment.h"
+#include "common/assert.h"
+#include "common/div_ceil.h"
+#include "common/scope_exit.h"
+#include "common/settings.h"
+#include "core/device_memory.h"
+#include "core/device_memory_manager.h"
+#include "core/memory.h"
+
+namespace Core {
+
+namespace {
+
+class MultiAddressContainer {
+public:
+    MultiAddressContainer() = default;
+    ~MultiAddressContainer() = default;
+
+    void GatherValues(u32 start_entry, Common::ScratchBuffer<u32>& buffer) {
+        buffer.resize(8);
+        buffer.resize(0);
+        size_t index = 0;
+        const auto add_value = [&](u32 value) {
+            buffer[index] = value;
+            index++;
+            buffer.resize(index);
+        };
+
+        u32 iter_entry = start_entry;
+        Entry* current = &storage[iter_entry - 1];
+        add_value(current->value);
+        while (current->next_entry != 0) {
+            iter_entry = current->next_entry;
+            current = &storage[iter_entry - 1];
+            add_value(current->value);
+        }
+    }
+
+    u32 Register(u32 value) {
+        return RegisterImplementation(value);
+    }
+
+    void Register(u32 value, u32 start_entry) {
+        auto entry_id = RegisterImplementation(value);
+        u32 iter_entry = start_entry;
+        Entry* current = &storage[iter_entry - 1];
+        while (current->next_entry != 0) {
+            iter_entry = current->next_entry;
+            current = &storage[iter_entry - 1];
+        }
+        current->next_entry = entry_id;
+    }
+
+    std::pair<bool, u32> Unregister(u32 value, u32 start_entry) {
+        u32 iter_entry = start_entry;
+        Entry* previous{};
+        Entry* current = &storage[iter_entry - 1];
+        Entry* next{};
+        bool more_than_one_remaining = false;
+        u32 result_start{start_entry};
+        size_t count = 0;
+        while (current->value != value) {
+            count++;
+            previous = current;
+            iter_entry = current->next_entry;
+            current = &storage[iter_entry - 1];
+        }
+        // Find next
+        u32 next_entry = current->next_entry;
+        if (next_entry != 0) {
+            next = &storage[next_entry - 1];
+            more_than_one_remaining = next->next_entry != 0 || previous != nullptr;
+        }
+        if (previous) {
+            previous->next_entry = next_entry;
+        } else {
+            result_start = next_entry;
+        }
+        free_entries.emplace_back(iter_entry);
+        return std::make_pair(more_than_one_remaining || count > 1, result_start);
+    }
+
+    u32 ReleaseEntry(u32 start_entry) {
+        Entry* current = &storage[start_entry - 1];
+        free_entries.emplace_back(start_entry);
+        return current->value;
+    }
+
+private:
+    u32 RegisterImplementation(u32 value) {
+        auto entry_id = GetNewEntry();
+        auto& entry = storage[entry_id - 1];
+        entry.next_entry = 0;
+        entry.value = value;
+        return entry_id;
+    }
+    u32 GetNewEntry() {
+        if (!free_entries.empty()) {
+            u32 result = free_entries.front();
+            free_entries.pop_front();
+            return result;
+        }
+        storage.emplace_back();
+        u32 new_entry = static_cast<u32>(storage.size());
+        return new_entry;
+    }
+
+    struct Entry {
+        u32 next_entry{};
+        u32 value{};
+    };
+
+    std::deque<Entry> storage;
+    std::deque<u32> free_entries;
+};
+
+struct EmptyAllocator {
+    EmptyAllocator([[maybe_unused]] DAddr address) {}
+};
+
+} // namespace
+
+template <typename DTraits>
+struct DeviceMemoryManagerAllocator {
+    static constexpr size_t device_virtual_bits = DTraits::device_virtual_bits;
+    static constexpr DAddr first_address = 1ULL << Memory::YUZU_PAGEBITS;
+    static constexpr DAddr max_device_area = 1ULL << device_virtual_bits;
+
+    DeviceMemoryManagerAllocator() : main_allocator(first_address) {}
+
+    Common::FlatAllocator<DAddr, 0, device_virtual_bits> main_allocator;
+    MultiAddressContainer multi_dev_address;
+
+    /// Returns true when vaddr -> vaddr+size is fully contained in the buffer
+    template <bool pin_area>
+    [[nodiscard]] bool IsInBounds(VAddr addr, u64 size) const noexcept {
+        return addr >= 0 && addr + size <= max_device_area;
+    }
+
+    DAddr Allocate(size_t size) {
+        return main_allocator.Allocate(size);
+    }
+
+    void AllocateFixed(DAddr b_address, size_t b_size) {
+        main_allocator.AllocateFixed(b_address, b_size);
+    }
+
+    void Free(DAddr b_address, size_t b_size) {
+        main_allocator.Free(b_address, b_size);
+    }
+};
+
+template <typename Traits>
+DeviceMemoryManager<Traits>::DeviceMemoryManager(const DeviceMemory& device_memory_)
+    : physical_base{reinterpret_cast<const uintptr_t>(device_memory_.buffer.BackingBasePointer())},
+      device_inter{nullptr}, compressed_physical_ptr(device_as_size >> Memory::YUZU_PAGEBITS),
+      compressed_device_addr(1ULL << ((Settings::values.memory_layout_mode.GetValue() ==
+                                               Settings::MemoryLayout::Memory_4Gb
+                                           ? physical_min_bits
+                                           : physical_max_bits) -
+                                      Memory::YUZU_PAGEBITS)),
+      continuity_tracker(device_as_size >> Memory::YUZU_PAGEBITS),
+      cpu_backing_address(device_as_size >> Memory::YUZU_PAGEBITS) {
+    impl = std::make_unique<DeviceMemoryManagerAllocator<Traits>>();
+    cached_pages = std::make_unique<CachedPages>();
+
+    const size_t total_virtual = device_as_size >> Memory::YUZU_PAGEBITS;
+    for (size_t i = 0; i < total_virtual; i++) {
+        compressed_physical_ptr[i] = 0;
+        continuity_tracker[i] = 1;
+        cpu_backing_address[i] = 0;
+    }
+    const size_t total_phys = 1ULL << ((Settings::values.memory_layout_mode.GetValue() ==
+                                                Settings::MemoryLayout::Memory_4Gb
+                                            ? physical_min_bits
+                                            : physical_max_bits) -
+                                       Memory::YUZU_PAGEBITS);
+    for (size_t i = 0; i < total_phys; i++) {
+        compressed_device_addr[i] = 0;
+    }
+}
+
+template <typename Traits>
+DeviceMemoryManager<Traits>::~DeviceMemoryManager() = default;
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::BindInterface(DeviceInterface* device_inter_) {
+    device_inter = device_inter_;
+}
+
+template <typename Traits>
+DAddr DeviceMemoryManager<Traits>::Allocate(size_t size) {
+    return impl->Allocate(size);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::AllocateFixed(DAddr start, size_t size) {
+    return impl->AllocateFixed(start, size);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::Free(DAddr start, size_t size) {
+    impl->Free(start, size);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::Map(DAddr address, VAddr virtual_address, size_t size,
+                                      size_t process_id, bool track) {
+    Core::Memory::Memory* process_memory = registered_processes[process_id];
+    size_t start_page_d = address >> Memory::YUZU_PAGEBITS;
+    size_t num_pages = Common::AlignUp(size, Memory::YUZU_PAGESIZE) >> Memory::YUZU_PAGEBITS;
+    std::scoped_lock lk(mapping_guard);
+    for (size_t i = 0; i < num_pages; i++) {
+        const VAddr new_vaddress = virtual_address + i * Memory::YUZU_PAGESIZE;
+        auto* ptr = process_memory->GetPointerSilent(Common::ProcessAddress(new_vaddress));
+        if (ptr == nullptr) [[unlikely]] {
+            compressed_physical_ptr[start_page_d + i] = 0;
+            continue;
+        }
+        auto phys_addr = static_cast<u32>(GetRawPhysicalAddr(ptr) >> Memory::YUZU_PAGEBITS) + 1U;
+        compressed_physical_ptr[start_page_d + i] = phys_addr;
+        InsertCPUBacking(start_page_d + i, new_vaddress, process_id);
+        const u32 base_dev = compressed_device_addr[phys_addr - 1U];
+        const u32 new_dev = static_cast<u32>(start_page_d + i);
+        if (base_dev == 0) [[likely]] {
+            compressed_device_addr[phys_addr - 1U] = new_dev;
+            continue;
+        }
+        u32 start_id = base_dev & MULTI_MASK;
+        if ((base_dev >> MULTI_FLAG_BITS) == 0) {
+            start_id = impl->multi_dev_address.Register(base_dev);
+            compressed_device_addr[phys_addr - 1U] = MULTI_FLAG | start_id;
+        }
+        impl->multi_dev_address.Register(new_dev, start_id);
+    }
+    if (track) {
+        TrackContinuityImpl(address, virtual_address, size, process_id);
+    }
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::Unmap(DAddr address, size_t size) {
+    size_t start_page_d = address >> Memory::YUZU_PAGEBITS;
+    size_t num_pages = Common::AlignUp(size, Memory::YUZU_PAGESIZE) >> Memory::YUZU_PAGEBITS;
+    device_inter->InvalidateRegion(address, size);
+    std::scoped_lock lk(mapping_guard);
+    for (size_t i = 0; i < num_pages; i++) {
+        auto phys_addr = compressed_physical_ptr[start_page_d + i];
+        compressed_physical_ptr[start_page_d + i] = 0;
+        cpu_backing_address[start_page_d + i] = 0;
+        if (phys_addr != 0) [[likely]] {
+            const u32 base_dev = compressed_device_addr[phys_addr - 1U];
+            if ((base_dev >> MULTI_FLAG_BITS) == 0) [[likely]] {
+                compressed_device_addr[phys_addr - 1] = 0;
+                continue;
+            }
+            const auto [more_entries, new_start] = impl->multi_dev_address.Unregister(
+                static_cast<u32>(start_page_d + i), base_dev & MULTI_MASK);
+            if (!more_entries) {
+                compressed_device_addr[phys_addr - 1] =
+                    impl->multi_dev_address.ReleaseEntry(new_start);
+                continue;
+            }
+            compressed_device_addr[phys_addr - 1] = new_start | MULTI_FLAG;
+        }
+    }
+}
+template <typename Traits>
+void DeviceMemoryManager<Traits>::TrackContinuityImpl(DAddr address, VAddr virtual_address,
+                                                      size_t size, size_t process_id) {
+    Core::Memory::Memory* process_memory = registered_processes[process_id];
+    size_t start_page_d = address >> Memory::YUZU_PAGEBITS;
+    size_t num_pages = Common::AlignUp(size, Memory::YUZU_PAGESIZE) >> Memory::YUZU_PAGEBITS;
+    uintptr_t last_ptr = 0;
+    size_t page_count = 1;
+    for (size_t i = num_pages; i > 0; i--) {
+        size_t index = i - 1;
+        const VAddr new_vaddress = virtual_address + index * Memory::YUZU_PAGESIZE;
+        const uintptr_t new_ptr = reinterpret_cast<uintptr_t>(
+            process_memory->GetPointerSilent(Common::ProcessAddress(new_vaddress)));
+        if (new_ptr + page_size == last_ptr) {
+            page_count++;
+        } else {
+            page_count = 1;
+        }
+        last_ptr = new_ptr;
+        continuity_tracker[start_page_d + index] = static_cast<u32>(page_count);
+    }
+}
+template <typename Traits>
+u8* DeviceMemoryManager<Traits>::GetSpan(const DAddr src_addr, const std::size_t size) {
+    size_t page_index = src_addr >> page_bits;
+    size_t subbits = src_addr & page_mask;
+    if ((static_cast<size_t>(continuity_tracker[page_index]) << page_bits) >= size + subbits) {
+        return GetPointer<u8>(src_addr);
+    }
+    return nullptr;
+}
+
+template <typename Traits>
+const u8* DeviceMemoryManager<Traits>::GetSpan(const DAddr src_addr, const std::size_t size) const {
+    size_t page_index = src_addr >> page_bits;
+    size_t subbits = src_addr & page_mask;
+    if ((static_cast<size_t>(continuity_tracker[page_index]) << page_bits) >= size + subbits) {
+        return GetPointer<u8>(src_addr);
+    }
+    return nullptr;
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::InnerGatherDeviceAddresses(Common::ScratchBuffer<u32>& buffer,
+                                                             PAddr address) {
+    size_t phys_addr = address >> page_bits;
+    std::scoped_lock lk(mapping_guard);
+    u32 backing = compressed_device_addr[phys_addr];
+    if ((backing >> MULTI_FLAG_BITS) != 0) {
+        impl->multi_dev_address.GatherValues(backing & MULTI_MASK, buffer);
+        return;
+    }
+    buffer.resize(1);
+    buffer[0] = backing;
+}
+
+template <typename Traits>
+template <typename T>
+T* DeviceMemoryManager<Traits>::GetPointer(DAddr address) {
+    const size_t index = address >> Memory::YUZU_PAGEBITS;
+    const size_t offset = address & Memory::YUZU_PAGEMASK;
+    auto phys_addr = compressed_physical_ptr[index];
+    if (phys_addr == 0) [[unlikely]] {
+        return nullptr;
+    }
+    return GetPointerFromRaw<T>((static_cast<PAddr>(phys_addr - 1) << Memory::YUZU_PAGEBITS) +
+                                offset);
+}
+
+template <typename Traits>
+template <typename T>
+const T* DeviceMemoryManager<Traits>::GetPointer(DAddr address) const {
+    const size_t index = address >> Memory::YUZU_PAGEBITS;
+    const size_t offset = address & Memory::YUZU_PAGEMASK;
+    auto phys_addr = compressed_physical_ptr[index];
+    if (phys_addr == 0) [[unlikely]] {
+        return nullptr;
+    }
+    return GetPointerFromRaw<T>((static_cast<PAddr>(phys_addr - 1) << Memory::YUZU_PAGEBITS) +
+                                offset);
+}
+
+template <typename Traits>
+template <typename T>
+void DeviceMemoryManager<Traits>::Write(DAddr address, T value) {
+    T* ptr = GetPointer<T>(address);
+    if (!ptr) [[unlikely]] {
+        return;
+    }
+    std::memcpy(ptr, &value, sizeof(T));
+}
+
+template <typename Traits>
+template <typename T>
+T DeviceMemoryManager<Traits>::Read(DAddr address) const {
+    const T* ptr = GetPointer<T>(address);
+    T result{};
+    if (!ptr) [[unlikely]] {
+        return result;
+    }
+    std::memcpy(&result, ptr, sizeof(T));
+    return result;
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::WalkBlock(DAddr addr, std::size_t size, auto on_unmapped,
+                                            auto on_memory, auto increment) {
+    std::size_t remaining_size = size;
+    std::size_t page_index = addr >> Memory::YUZU_PAGEBITS;
+    std::size_t page_offset = addr & Memory::YUZU_PAGEMASK;
+
+    while (remaining_size) {
+        const size_t next_pages = static_cast<std::size_t>(continuity_tracker[page_index]);
+        const std::size_t copy_amount =
+            std::min((next_pages << Memory::YUZU_PAGEBITS) - page_offset, remaining_size);
+        const auto current_vaddr =
+            static_cast<u64>((page_index << Memory::YUZU_PAGEBITS) + page_offset);
+        SCOPE_EXIT({
+            page_index += next_pages;
+            page_offset = 0;
+            increment(copy_amount);
+            remaining_size -= copy_amount;
+        });
+
+        auto phys_addr = compressed_physical_ptr[page_index];
+        if (phys_addr == 0) {
+            on_unmapped(copy_amount, current_vaddr);
+            continue;
+        }
+        auto* mem_ptr = GetPointerFromRaw<u8>(
+            (static_cast<PAddr>(phys_addr - 1) << Memory::YUZU_PAGEBITS) + page_offset);
+        on_memory(copy_amount, mem_ptr);
+    }
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::ReadBlock(DAddr address, void* dest_pointer, size_t size) {
+    device_inter->FlushRegion(address, size);
+    WalkBlock(
+        address, size,
+        [&](size_t copy_amount, DAddr current_vaddr) {
+            LOG_ERROR(
+                HW_Memory,
+                "Unmapped Device ReadBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
+                current_vaddr, address, size);
+            std::memset(dest_pointer, 0, copy_amount);
+        },
+        [&](size_t copy_amount, const u8* const src_ptr) {
+            std::memcpy(dest_pointer, src_ptr, copy_amount);
+        },
+        [&](const std::size_t copy_amount) {
+            dest_pointer = static_cast<u8*>(dest_pointer) + copy_amount;
+        });
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::WriteBlock(DAddr address, const void* src_pointer, size_t size) {
+    WalkBlock(
+        address, size,
+        [&](size_t copy_amount, DAddr current_vaddr) {
+            LOG_ERROR(
+                HW_Memory,
+                "Unmapped Device WriteBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
+                current_vaddr, address, size);
+        },
+        [&](size_t copy_amount, u8* const dst_ptr) {
+            std::memcpy(dst_ptr, src_pointer, copy_amount);
+        },
+        [&](const std::size_t copy_amount) {
+            src_pointer = static_cast<const u8*>(src_pointer) + copy_amount;
+        });
+    device_inter->InvalidateRegion(address, size);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::ReadBlockUnsafe(DAddr address, void* dest_pointer, size_t size) {
+    WalkBlock(
+        address, size,
+        [&](size_t copy_amount, DAddr current_vaddr) {
+            LOG_ERROR(
+                HW_Memory,
+                "Unmapped Device ReadBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
+                current_vaddr, address, size);
+            std::memset(dest_pointer, 0, copy_amount);
+        },
+        [&](size_t copy_amount, const u8* const src_ptr) {
+            std::memcpy(dest_pointer, src_ptr, copy_amount);
+        },
+        [&](const std::size_t copy_amount) {
+            dest_pointer = static_cast<u8*>(dest_pointer) + copy_amount;
+        });
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::WriteBlockUnsafe(DAddr address, const void* src_pointer,
+                                                   size_t size) {
+    WalkBlock(
+        address, size,
+        [&](size_t copy_amount, DAddr current_vaddr) {
+            LOG_ERROR(
+                HW_Memory,
+                "Unmapped Device WriteBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
+                current_vaddr, address, size);
+        },
+        [&](size_t copy_amount, u8* const dst_ptr) {
+            std::memcpy(dst_ptr, src_pointer, copy_amount);
+        },
+        [&](const std::size_t copy_amount) {
+            src_pointer = static_cast<const u8*>(src_pointer) + copy_amount;
+        });
+}
+
+template <typename Traits>
+size_t DeviceMemoryManager<Traits>::RegisterProcess(Memory::Memory* memory_device_inter) {
+    size_t new_id;
+    if (!id_pool.empty()) {
+        new_id = id_pool.front();
+        id_pool.pop_front();
+        registered_processes[new_id] = memory_device_inter;
+    } else {
+        registered_processes.emplace_back(memory_device_inter);
+        new_id = registered_processes.size() - 1U;
+    }
+    return new_id;
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UnregisterProcess(size_t id) {
+    registered_processes[id] = nullptr;
+    id_pool.push_front(id);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
+    bool locked = false;
+    auto lock = [&] {
+        if (!locked) {
+            counter_guard.lock();
+            locked = true;
+        }
+    };
+    SCOPE_EXIT({
+        if (locked) {
+            counter_guard.unlock();
+        }
+    });
+    u64 uncache_begin = 0;
+    u64 cache_begin = 0;
+    u64 uncache_bytes = 0;
+    u64 cache_bytes = 0;
+    const auto MarkRegionCaching = &DeviceMemoryManager<Traits>::DeviceMethods::MarkRegionCaching;
+
+    std::atomic_thread_fence(std::memory_order_acquire);
+    const size_t page_end = Common::DivCeil(addr + size, Memory::YUZU_PAGESIZE);
+    size_t page = addr >> Memory::YUZU_PAGEBITS;
+    auto [process_id, base_vaddress] = ExtractCPUBacking(page);
+    size_t vpage = base_vaddress >> Memory::YUZU_PAGEBITS;
+    auto* memory_device_inter = registered_processes[process_id];
+    for (; page != page_end; ++page) {
+        std::atomic_uint8_t& count = cached_pages->at(page >> 3).Count(page);
+
+        if (delta > 0) {
+            ASSERT_MSG(count.load(std::memory_order::relaxed) < std::numeric_limits<u8>::max(),
+                       "Count may overflow!");
+        } else if (delta < 0) {
+            ASSERT_MSG(count.load(std::memory_order::relaxed) > 0, "Count may underflow!");
+        } else {
+            ASSERT_MSG(false, "Delta must be non-zero!");
+        }
+
+        // Adds or subtracts 1, as count is a unsigned 8-bit value
+        count.fetch_add(static_cast<u8>(delta), std::memory_order_release);
+
+        // Assume delta is either -1 or 1
+        if (count.load(std::memory_order::relaxed) == 0) {
+            if (uncache_bytes == 0) {
+                uncache_begin = vpage;
+            }
+            uncache_bytes += Memory::YUZU_PAGESIZE;
+        } else if (uncache_bytes > 0) {
+            lock();
+            MarkRegionCaching(memory_device_inter, uncache_begin << Memory::YUZU_PAGEBITS,
+                              uncache_bytes, false);
+            uncache_bytes = 0;
+        }
+        if (count.load(std::memory_order::relaxed) == 1 && delta > 0) {
+            if (cache_bytes == 0) {
+                cache_begin = vpage;
+            }
+            cache_bytes += Memory::YUZU_PAGESIZE;
+        } else if (cache_bytes > 0) {
+            lock();
+            MarkRegionCaching(memory_device_inter, cache_begin << Memory::YUZU_PAGEBITS, cache_bytes,
+                              true);
+            cache_bytes = 0;
+        }
+        vpage++;
+    }
+    if (uncache_bytes > 0) {
+        lock();
+        MarkRegionCaching(memory_device_inter, uncache_begin << Memory::YUZU_PAGEBITS, uncache_bytes,
+                          false);
+    }
+    if (cache_bytes > 0) {
+        lock();
+        MarkRegionCaching(memory_device_inter, cache_begin << Memory::YUZU_PAGEBITS, cache_bytes,
+                          true);
+    }
+}
+
+} // namespace Core

src/core/file_sys/savedata_factory.cpp

@@ -97,8 +97,9 @@ std::string SaveDataAttribute::DebugInfo() const {
                        static_cast<u8>(rank), index);
 }
 
-SaveDataFactory::SaveDataFactory(Core::System& system_, VirtualDir save_directory_)
+SaveDataFactory::SaveDataFactory(Core::System& system_, ProgramId program_id_,
-    : dir{std::move(save_directory_)}, system{system_} {
+                                 VirtualDir save_directory_)
+    : system{system_}, program_id{program_id_}, dir{std::move(save_directory_)} {
     // Delete all temporary storages
     // On hardware, it is expected that temporary storage be empty at first use.
     dir->DeleteSubdirectoryRecursive("temp");
@@ -110,7 +111,7 @@ VirtualDir SaveDataFactory::Create(SaveDataSpaceId space, const SaveDataAttribut
     PrintSaveDataAttributeWarnings(meta);
 
     const auto save_directory =
-        GetFullPath(system, dir, space, meta.type, meta.title_id, meta.user_id, meta.save_id);
+        GetFullPath(program_id, dir, space, meta.type, meta.title_id, meta.user_id, meta.save_id);
 
     return dir->CreateDirectoryRelative(save_directory);
 }
@@ -118,7 +119,7 @@ VirtualDir SaveDataFactory::Create(SaveDataSpaceId space, const SaveDataAttribut
 VirtualDir SaveDataFactory::Open(SaveDataSpaceId space, const SaveDataAttribute& meta) const {
 
     const auto save_directory =
-        GetFullPath(system, dir, space, meta.type, meta.title_id, meta.user_id, meta.save_id);
+        GetFullPath(program_id, dir, space, meta.type, meta.title_id, meta.user_id, meta.save_id);
 
     auto out = dir->GetDirectoryRelative(save_directory);
 
@@ -147,14 +148,14 @@ std::string SaveDataFactory::GetSaveDataSpaceIdPath(SaveDataSpaceId space) {
     }
 }
 
-std::string SaveDataFactory::GetFullPath(Core::System& system, VirtualDir dir,
+std::string SaveDataFactory::GetFullPath(ProgramId program_id, VirtualDir dir,
                                          SaveDataSpaceId space, SaveDataType type, u64 title_id,
                                          u128 user_id, u64 save_id) {
     // According to switchbrew, if a save is of type SaveData and the title id field is 0, it should
     // be interpreted as the title id of the current process.
     if (type == SaveDataType::SaveData || type == SaveDataType::DeviceSaveData) {
         if (title_id == 0) {
-            title_id = system.GetApplicationProcessProgramID();
+            title_id = program_id;
         }
     }
 
@@ -201,7 +202,7 @@ std::string SaveDataFactory::GetUserGameSaveDataRoot(u128 user_id, bool future)
 SaveDataSize SaveDataFactory::ReadSaveDataSize(SaveDataType type, u64 title_id,
                                                u128 user_id) const {
     const auto path =
-        GetFullPath(system, dir, SaveDataSpaceId::NandUser, type, title_id, user_id, 0);
+        GetFullPath(program_id, dir, SaveDataSpaceId::NandUser, type, title_id, user_id, 0);
     const auto relative_dir = GetOrCreateDirectoryRelative(dir, path);
 
     const auto size_file = relative_dir->GetFile(GetSaveDataSizeFileName());
@@ -220,7 +221,7 @@ SaveDataSize SaveDataFactory::ReadSaveDataSize(SaveDataType type, u64 title_id,
 void SaveDataFactory::WriteSaveDataSize(SaveDataType type, u64 title_id, u128 user_id,
                                         SaveDataSize new_value) const {
     const auto path =
-        GetFullPath(system, dir, SaveDataSpaceId::NandUser, type, title_id, user_id, 0);
+        GetFullPath(program_id, dir, SaveDataSpaceId::NandUser, type, title_id, user_id, 0);
     const auto relative_dir = GetOrCreateDirectoryRelative(dir, path);
 
     const auto size_file = relative_dir->CreateFile(GetSaveDataSizeFileName());

src/core/file_sys/savedata_factory.h

@@ -87,10 +87,13 @@ constexpr const char* GetSaveDataSizeFileName() {
     return ".yuzu_save_size";
 }
 
+using ProgramId = u64;
+
 /// File system interface to the SaveData archive
 class SaveDataFactory {
 public:
-    explicit SaveDataFactory(Core::System& system_, VirtualDir save_directory_);
+    explicit SaveDataFactory(Core::System& system_, ProgramId program_id_,
+                             VirtualDir save_directory_);
     ~SaveDataFactory();
 
     VirtualDir Create(SaveDataSpaceId space, const SaveDataAttribute& meta) const;
@@ -99,7 +102,7 @@ public:
     VirtualDir GetSaveDataSpaceDirectory(SaveDataSpaceId space) const;
 
     static std::string GetSaveDataSpaceIdPath(SaveDataSpaceId space);
-    static std::string GetFullPath(Core::System& system, VirtualDir dir, SaveDataSpaceId space,
+    static std::string GetFullPath(ProgramId program_id, VirtualDir dir, SaveDataSpaceId space,
                                    SaveDataType type, u64 title_id, u128 user_id, u64 save_id);
     static std::string GetUserGameSaveDataRoot(u128 user_id, bool future);
 
@@ -110,8 +113,9 @@ public:
     void SetAutoCreate(bool state);
 
 private:
-    VirtualDir dir;
     Core::System& system;
+    ProgramId program_id;
+    VirtualDir dir;
     bool auto_create{true};
 };
 

src/core/gpu_dirty_memory_manager.h

@@ -10,7 +10,7 @@
 #include <utility>
 #include <vector>
 
-#include "core/memory.h"
+#include "core/device_memory_manager.h"
 
 namespace Core {
 
@@ -23,7 +23,7 @@ public:
 
     ~GPUDirtyMemoryManager() = default;
 
-    void Collect(VAddr address, size_t size) {
+    void Collect(PAddr address, size_t size) {
         TransformAddress t = BuildTransform(address, size);
         TransformAddress tmp, original;
         do {
@@ -47,7 +47,7 @@ public:
                                                 std::memory_order_relaxed));
     }
 
-    void Gather(std::function<void(VAddr, size_t)>& callback) {
+    void Gather(std::function<void(PAddr, size_t)>& callback) {
         {
             std::scoped_lock lk(guard);
             TransformAddress t = current.exchange(default_transform, std::memory_order_relaxed);
@@ -65,7 +65,7 @@ public:
                 mask = mask >> empty_bits;
 
                 const size_t continuous_bits = std::countr_one(mask);
-                callback((static_cast<VAddr>(transform.address) << page_bits) + offset,
+                callback((static_cast<PAddr>(transform.address) << page_bits) + offset,
                          continuous_bits << align_bits);
                 mask = continuous_bits < align_size ? (mask >> continuous_bits) : 0;
                 offset += continuous_bits << align_bits;
@@ -80,7 +80,7 @@ private:
         u32 mask;
     };
 
-    constexpr static size_t page_bits = Memory::YUZU_PAGEBITS - 1;
+    constexpr static size_t page_bits = DEVICE_PAGEBITS - 1;
     constexpr static size_t page_size = 1ULL << page_bits;
     constexpr static size_t page_mask = page_size - 1;
 
@@ -89,7 +89,7 @@ private:
     constexpr static size_t align_mask = align_size - 1;
     constexpr static TransformAddress default_transform = {.address = ~0U, .mask = 0U};
 
-    bool IsValid(VAddr address) {
+    bool IsValid(PAddr address) {
         return address < (1ULL << 39);
     }
 
@@ -103,7 +103,7 @@ private:
         return mask;
     }
 
-    TransformAddress BuildTransform(VAddr address, size_t size) {
+    TransformAddress BuildTransform(PAddr address, size_t size) {
         const size_t minor_address = address & page_mask;
         const size_t minor_bit = minor_address >> align_bits;
         const size_t top_bit = (minor_address + size + align_mask) >> align_bits;

src/core/guest_memory.h

@@ -0,0 +1,214 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <iterator>
+#include <memory>
+#include <optional>
+#include <span>
+#include <vector>
+
+#include "common/assert.h"
+#include "common/scratch_buffer.h"
+
+namespace Core::Memory {
+
+enum GuestMemoryFlags : u32 {
+    Read = 1 << 0,
+    Write = 1 << 1,
+    Safe = 1 << 2,
+    Cached = 1 << 3,
+
+    SafeRead = Read | Safe,
+    SafeWrite = Write | Safe,
+    SafeReadWrite = SafeRead | SafeWrite,
+    SafeReadCachedWrite = SafeReadWrite | Cached,
+
+    UnsafeRead = Read,
+    UnsafeWrite = Write,
+    UnsafeReadWrite = UnsafeRead | UnsafeWrite,
+    UnsafeReadCachedWrite = UnsafeReadWrite | Cached,
+};
+
+namespace {
+template <typename M, typename T, GuestMemoryFlags FLAGS>
+class GuestMemory {
+    using iterator = T*;
+    using const_iterator = const T*;
+    using value_type = T;
+    using element_type = T;
+    using iterator_category = std::contiguous_iterator_tag;
+
+public:
+    GuestMemory() = delete;
+    explicit GuestMemory(M& memory, u64 addr, std::size_t size,
+                         Common::ScratchBuffer<T>* backup = nullptr)
+        : m_memory{memory}, m_addr{addr}, m_size{size} {
+        static_assert(FLAGS & GuestMemoryFlags::Read || FLAGS & GuestMemoryFlags::Write);
+        if constexpr (FLAGS & GuestMemoryFlags::Read) {
+            Read(addr, size, backup);
+        }
+    }
+
+    ~GuestMemory() = default;
+
+    T* data() noexcept {
+        return m_data_span.data();
+    }
+
+    const T* data() const noexcept {
+        return m_data_span.data();
+    }
+
+    size_t size() const noexcept {
+        return m_size;
+    }
+
+    size_t size_bytes() const noexcept {
+        return this->size() * sizeof(T);
+    }
+
+    [[nodiscard]] T* begin() noexcept {
+        return this->data();
+    }
+
+    [[nodiscard]] const T* begin() const noexcept {
+        return this->data();
+    }
+
+    [[nodiscard]] T* end() noexcept {
+        return this->data() + this->size();
+    }
+
+    [[nodiscard]] const T* end() const noexcept {
+        return this->data() + this->size();
+    }
+
+    T& operator[](size_t index) noexcept {
+        return m_data_span[index];
+    }
+
+    const T& operator[](size_t index) const noexcept {
+        return m_data_span[index];
+    }
+
+    void SetAddressAndSize(u64 addr, std::size_t size) noexcept {
+        m_addr = addr;
+        m_size = size;
+        m_addr_changed = true;
+    }
+
+    std::span<T> Read(u64 addr, std::size_t size,
+                      Common::ScratchBuffer<T>* backup = nullptr) noexcept {
+        m_addr = addr;
+        m_size = size;
+        if (m_size == 0) {
+            m_is_data_copy = true;
+            return {};
+        }
+
+        if (this->TrySetSpan()) {
+            if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                m_memory.FlushRegion(m_addr, this->size_bytes());
+            }
+        } else {
+            if (backup) {
+                backup->resize_destructive(this->size());
+                m_data_span = *backup;
+            } else {
+                m_data_copy.resize(this->size());
+                m_data_span = std::span(m_data_copy);
+            }
+            m_is_data_copy = true;
+            m_span_valid = true;
+            if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                m_memory.ReadBlock(m_addr, this->data(), this->size_bytes());
+            } else {
+                m_memory.ReadBlockUnsafe(m_addr, this->data(), this->size_bytes());
+            }
+        }
+        return m_data_span;
+    }
+
+    void Write(std::span<T> write_data) noexcept {
+        if constexpr (FLAGS & GuestMemoryFlags::Cached) {
+            m_memory.WriteBlockCached(m_addr, write_data.data(), this->size_bytes());
+        } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+            m_memory.WriteBlock(m_addr, write_data.data(), this->size_bytes());
+        } else {
+            m_memory.WriteBlockUnsafe(m_addr, write_data.data(), this->size_bytes());
+        }
+    }
+
+    bool TrySetSpan() noexcept {
+        if (u8* ptr = m_memory.GetSpan(m_addr, this->size_bytes()); ptr) {
+            m_data_span = {reinterpret_cast<T*>(ptr), this->size()};
+            m_span_valid = true;
+            return true;
+        }
+        return false;
+    }
+
+protected:
+    bool IsDataCopy() const noexcept {
+        return m_is_data_copy;
+    }
+
+    bool AddressChanged() const noexcept {
+        return m_addr_changed;
+    }
+
+    M& m_memory;
+    u64 m_addr{};
+    size_t m_size{};
+    std::span<T> m_data_span{};
+    std::vector<T> m_data_copy{};
+    bool m_span_valid{false};
+    bool m_is_data_copy{false};
+    bool m_addr_changed{false};
+};
+
+template <typename M, typename T, GuestMemoryFlags FLAGS>
+class GuestMemoryScoped : public GuestMemory<M, T, FLAGS> {
+public:
+    GuestMemoryScoped() = delete;
+    explicit GuestMemoryScoped(M& memory, u64 addr, std::size_t size,
+                               Common::ScratchBuffer<T>* backup = nullptr)
+        : GuestMemory<M, T, FLAGS>(memory, addr, size, backup) {
+        if constexpr (!(FLAGS & GuestMemoryFlags::Read)) {
+            if (!this->TrySetSpan()) {
+                if (backup) {
+                    this->m_data_span = *backup;
+                    this->m_span_valid = true;
+                    this->m_is_data_copy = true;
+                }
+            }
+        }
+    }
+
+    ~GuestMemoryScoped() {
+        if constexpr (FLAGS & GuestMemoryFlags::Write) {
+            if (this->size() == 0) [[unlikely]] {
+                return;
+            }
+
+            if (this->AddressChanged() || this->IsDataCopy()) {
+                ASSERT(this->m_span_valid);
+                if constexpr (FLAGS & GuestMemoryFlags::Cached) {
+                    this->m_memory.WriteBlockCached(this->m_addr, this->data(), this->size_bytes());
+                } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
+                    this->m_memory.WriteBlock(this->m_addr, this->data(), this->size_bytes());
+                } else {
+                    this->m_memory.WriteBlockUnsafe(this->m_addr, this->data(), this->size_bytes());
+                }
+            } else if constexpr ((FLAGS & GuestMemoryFlags::Safe) ||
+                                 (FLAGS & GuestMemoryFlags::Cached)) {
+                this->m_memory.InvalidateRegion(this->m_addr, this->size_bytes());
+            }
+        }
+    }
+};
+} // namespace
+
+} // namespace Core::Memory

src/core/hle/kernel/k_memory_block_manager.cpp

@@ -28,14 +28,14 @@ Result KMemoryBlockManager::Initialize(KProcessAddress st, KProcessAddress nd,
 }
 
 void KMemoryBlockManager::Finalize(KMemoryBlockSlabManager* slab_manager,
-                                   HostUnmapCallback&& host_unmap_callback) {
+                                   BlockCallback&& block_callback) {
     // Erase every block until we have none left.
     auto it = m_memory_block_tree.begin();
     while (it != m_memory_block_tree.end()) {
         KMemoryBlock* block = std::addressof(*it);
         it = m_memory_block_tree.erase(it);
+        block_callback(block->GetAddress(), block->GetSize());
         slab_manager->Free(block);
-        host_unmap_callback(block->GetAddress(), block->GetSize());
     }
 
     ASSERT(m_memory_block_tree.empty());

src/core/hle/kernel/k_memory_block_manager.h

@@ -85,11 +85,11 @@ public:
 public:
     KMemoryBlockManager();
 
-    using HostUnmapCallback = std::function<void(Common::ProcessAddress, u64)>;
+    using BlockCallback = std::function<void(Common::ProcessAddress, u64)>;
 
     Result Initialize(KProcessAddress st, KProcessAddress nd,
                       KMemoryBlockSlabManager* slab_manager);
-    void Finalize(KMemoryBlockSlabManager* slab_manager, HostUnmapCallback&& host_unmap_callback);
+    void Finalize(KMemoryBlockSlabManager* slab_manager, BlockCallback&& block_callback);
 
     iterator end() {
         return m_memory_block_tree.end();

src/core/hle/kernel/k_page_table_base.cpp

@@ -431,15 +431,43 @@ Result KPageTableBase::InitializeForProcess(Svc::CreateProcessFlag as_type, bool
                                                m_memory_block_slab_manager));
 }
 
+Result KPageTableBase::FinalizeProcess() {
+    // Only process tables should be finalized.
+    ASSERT(!this->IsKernel());
+
+    // NOTE: Here Nintendo calls an unknown OnFinalize function.
+    // this->OnFinalize();
+
+    // NOTE: Here Nintendo calls a second unknown OnFinalize function.
+    // this->OnFinalize2();
+
+    // NOTE: Here Nintendo does a page table walk to discover heap pages to free.
+    // We will use the block manager finalization below to free them.
+
+    R_SUCCEED();
+}
+
 void KPageTableBase::Finalize() {
-    auto HostUnmapCallback = [&](KProcessAddress addr, u64 size) {
+    this->FinalizeProcess();
-        if (Settings::IsFastmemEnabled()) {
+
+    auto BlockCallback = [&](KProcessAddress addr, u64 size) {
+        if (m_impl->fastmem_arena) {
             m_system.DeviceMemory().buffer.Unmap(GetInteger(addr), size, false);
         }
+
+        // Get physical pages.
+        KPageGroup pg(m_kernel, m_block_info_manager);
+        this->MakePageGroup(pg, addr, size / PageSize);
+
+        // Free the pages.
+        pg.CloseAndReset();
     };
 
     // Finalize memory blocks.
-    m_memory_block_manager.Finalize(m_memory_block_slab_manager, std::move(HostUnmapCallback));
+    {
+        KScopedLightLock lk(m_general_lock);
+        m_memory_block_manager.Finalize(m_memory_block_slab_manager, std::move(BlockCallback));
+    }
 
     // Free any unsafe mapped memory.
     if (m_mapped_unsafe_physical_memory) {

src/core/hle/kernel/k_page_table_base.h

@@ -241,6 +241,7 @@ public:
                                 KResourceLimit* resource_limit, Core::Memory::Memory& memory,
                                 KProcessAddress aslr_space_start);
 
+    Result FinalizeProcess();
     void Finalize();
 
     bool IsKernel() const {

src/core/hle/kernel/k_process.cpp

@@ -5,6 +5,7 @@
 #include "common/scope_exit.h"
 #include "common/settings.h"
 #include "core/core.h"
+#include "core/gpu_dirty_memory_manager.h"
 #include "core/hle/kernel/k_process.h"
 #include "core/hle/kernel/k_scoped_resource_reservation.h"
 #include "core/hle/kernel/k_shared_memory.h"
@@ -171,6 +172,12 @@ void KProcess::Finalize() {
         m_resource_limit->Close();
     }
 
+    // Clear expensive resources, as the destructor is not called for guest objects.
+    for (auto& interface : m_arm_interfaces) {
+        interface.reset();
+    }
+    m_exclusive_monitor.reset();
+
     // Perform inherited finalization.
     KSynchronizationObject::Finalize();
 }
@@ -314,7 +321,7 @@ Result KProcess::Initialize(const Svc::CreateProcessParameter& params, const KPa
 
     // Ensure our memory is initialized.
     m_memory.SetCurrentPageTable(*this);
-    m_memory.SetGPUDirtyManagers(m_dirty_memory_managers);
+    m_memory.SetGPUDirtyManagers(m_kernel.System().GetGPUDirtyMemoryManager());
 
     // Ensure we can insert the code region.
     R_UNLESS(m_page_table.CanContain(params.code_address, params.code_num_pages * PageSize,
@@ -411,7 +418,7 @@ Result KProcess::Initialize(const Svc::CreateProcessParameter& params,
 
     // Ensure our memory is initialized.
     m_memory.SetCurrentPageTable(*this);
-    m_memory.SetGPUDirtyManagers(m_dirty_memory_managers);
+    m_memory.SetGPUDirtyManagers(m_kernel.System().GetGPUDirtyMemoryManager());
 
     // Ensure we can insert the code region.
     R_UNLESS(m_page_table.CanContain(params.code_address, code_size, KMemoryState::Code),
@@ -1135,8 +1142,7 @@ void KProcess::Switch(KProcess* cur_process, KProcess* next_process) {}
 KProcess::KProcess(KernelCore& kernel)
     : KAutoObjectWithSlabHeapAndContainer(kernel), m_page_table{kernel}, m_state_lock{kernel},
       m_list_lock{kernel}, m_cond_var{kernel.System()}, m_address_arbiter{kernel.System()},
-      m_handle_table{kernel}, m_dirty_memory_managers{},
+      m_handle_table{kernel}, m_exclusive_monitor{}, m_memory{kernel.System()} {}
-      m_exclusive_monitor{}, m_memory{kernel.System()} {}
 KProcess::~KProcess() = default;
 
 Result KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std::size_t code_size,
@@ -1233,10 +1239,10 @@ void KProcess::LoadModule(CodeSet code_set, KProcessAddress base_addr) {
     ReprotectSegment(code_set.DataSegment(), Svc::MemoryPermission::ReadWrite);
 
 #ifdef HAS_NCE
-    if (this->IsApplication() && Settings::IsNceEnabled()) {
+    const auto& patch = code_set.PatchSegment();
+    if (this->IsApplication() && Settings::IsNceEnabled() && patch.size != 0) {
         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,
                        Common::MemoryPermission::Read | Common::MemoryPermission::Execute);
         buffer.Protect(GetInteger(base_addr + patch.addr), patch.size,
@@ -1318,10 +1324,4 @@ bool KProcess::RemoveWatchpoint(KProcessAddress addr, u64 size, DebugWatchpointT
     return true;
 }
 
-void KProcess::GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback) {
-    for (auto& manager : m_dirty_memory_managers) {
-        manager.Gather(callback);
-    }
-}
-
 } // namespace Kernel

src/core/hle/kernel/k_process.h

@@ -7,7 +7,6 @@
 
 #include "core/arm/arm_interface.h"
 #include "core/file_sys/program_metadata.h"
-#include "core/gpu_dirty_memory_manager.h"
 #include "core/hle/kernel/code_set.h"
 #include "core/hle/kernel/k_address_arbiter.h"
 #include "core/hle/kernel/k_capabilities.h"
@@ -128,7 +127,6 @@ private:
 #ifdef HAS_NCE
     std::unordered_map<u64, u64> m_post_handlers{};
 #endif
-    std::array<Core::GPUDirtyMemoryManager, Core::Hardware::NUM_CPU_CORES> m_dirty_memory_managers;
     std::unique_ptr<Core::ExclusiveMonitor> m_exclusive_monitor;
     Core::Memory::Memory m_memory;
 
@@ -511,8 +509,6 @@ public:
         return m_memory;
     }
 
-    void GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback);
-
     Core::ExclusiveMonitor& GetExclusiveMonitor() const {
         return *m_exclusive_monitor;
     }

src/core/hle/kernel/kernel.cpp

@@ -112,7 +112,14 @@ struct KernelCore::Impl {
             old_process->Close();
         }
 
-        process_list.clear();
+        {
+            std::scoped_lock lk{process_list_lock};
+            for (auto* const process : process_list) {
+                process->Terminate();
+                process->Close();
+            }
+            process_list.clear();
+        }
 
         next_object_id = 0;
         next_kernel_process_id = KProcess::InitialProcessIdMin;
@@ -770,6 +777,7 @@ struct KernelCore::Impl {
     std::atomic<u64> next_thread_id{1};
 
     // Lists all processes that exist in the current session.
+    std::mutex process_list_lock;
     std::vector<KProcess*> process_list;
     std::atomic<KProcess*> application_process{};
     std::unique_ptr<Kernel::GlobalSchedulerContext> global_scheduler_context;
@@ -869,9 +877,19 @@ KResourceLimit* KernelCore::GetSystemResourceLimit() {
 }
 
 void KernelCore::AppendNewProcess(KProcess* process) {
+    process->Open();
+
+    std::scoped_lock lk{impl->process_list_lock};
     impl->process_list.push_back(process);
 }
 
+void KernelCore::RemoveProcess(KProcess* process) {
+    std::scoped_lock lk{impl->process_list_lock};
+    if (std::erase(impl->process_list, process)) {
+        process->Close();
+    }
+}
+
 void KernelCore::MakeApplicationProcess(KProcess* process) {
     impl->MakeApplicationProcess(process);
 }
@@ -884,8 +902,15 @@ const KProcess* KernelCore::ApplicationProcess() const {
     return impl->application_process;
 }
 
-const std::vector<KProcess*>& KernelCore::GetProcessList() const {
+std::list<KScopedAutoObject<KProcess>> KernelCore::GetProcessList() {
-    return impl->process_list;
+    std::list<KScopedAutoObject<KProcess>> processes;
+    std::scoped_lock lk{impl->process_list_lock};
+
+    for (auto* const process : impl->process_list) {
+        processes.emplace_back(process);
+    }
+
+    return processes;
 }
 
 Kernel::GlobalSchedulerContext& KernelCore::GlobalSchedulerContext() {

src/core/hle/kernel/kernel.h

@@ -5,6 +5,7 @@
 
 #include <array>
 #include <functional>
+#include <list>
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -116,8 +117,9 @@ public:
     /// Retrieves a shared pointer to the system resource limit instance.
     KResourceLimit* GetSystemResourceLimit();
 
-    /// Adds the given shared pointer to an internal list of active processes.
+    /// Adds/removes the given pointer to an internal list of active processes.
     void AppendNewProcess(KProcess* process);
+    void RemoveProcess(KProcess* process);
 
     /// Makes the given process the new application process.
     void MakeApplicationProcess(KProcess* process);
@@ -129,7 +131,7 @@ public:
     const KProcess* ApplicationProcess() const;
 
     /// Retrieves the list of processes.
-    const std::vector<KProcess*>& GetProcessList() const;
+    std::list<KScopedAutoObject<KProcess>> GetProcessList();
 
     /// Gets the sole instance of the global scheduler
     Kernel::GlobalSchedulerContext& GlobalSchedulerContext();

src/core/hle/kernel/svc/svc_process.cpp

@@ -74,13 +74,15 @@ Result GetProcessList(Core::System& system, s32* out_num_processes, u64 out_proc
     }
 
     auto& memory = GetCurrentMemory(kernel);
-    const auto& process_list = kernel.GetProcessList();
+    auto process_list = kernel.GetProcessList();
+    auto it = process_list.begin();
+
     const auto num_processes = process_list.size();
     const auto copy_amount =
         std::min(static_cast<std::size_t>(out_process_ids_size), num_processes);
 
-    for (std::size_t i = 0; i < copy_amount; ++i) {
+    for (std::size_t i = 0; i < copy_amount && it != process_list.end(); ++i, ++it) {
-        memory.Write64(out_process_ids, process_list[i]->GetProcessId());
+        memory.Write64(out_process_ids, (*it)->GetProcessId());
         out_process_ids += sizeof(u64);
     }
 

src/core/hle/service/acc/profile_manager.cpp

@@ -61,9 +61,7 @@ ProfileManager::ProfileManager() {
     OpenUser(*GetUser(current));
 }
 
-ProfileManager::~ProfileManager() {
+ProfileManager::~ProfileManager() = default;
-    WriteUserSaveFile();
-}
 
 /// After a users creation it needs to be "registered" to the system. AddToProfiles handles the
 /// internal management of the users profiles
@@ -113,6 +111,8 @@ Result ProfileManager::CreateNewUser(UUID uuid, const ProfileUsername& username)
         return ERROR_USER_ALREADY_EXISTS;
     }
 
+    is_save_needed = true;
+
     return AddUser({
         .user_uuid = uuid,
         .username = username,
@@ -326,6 +326,9 @@ bool ProfileManager::RemoveUser(UUID uuid) {
     profiles[*index] = ProfileInfo{};
     std::stable_partition(profiles.begin(), profiles.end(),
                           [](const ProfileInfo& profile) { return profile.user_uuid.IsValid(); });
+
+    is_save_needed = true;
+
     return true;
 }
 
@@ -340,6 +343,8 @@ bool ProfileManager::SetProfileBase(UUID uuid, const ProfileBase& profile_new) {
     profile.username = profile_new.username;
     profile.creation_time = profile_new.timestamp;
 
+    is_save_needed = true;
+
     return true;
 }
 
@@ -348,6 +353,7 @@ bool ProfileManager::SetProfileBaseAndData(Common::UUID uuid, const ProfileBase&
     const auto index = GetUserIndex(uuid);
     if (index.has_value() && SetProfileBase(uuid, profile_new)) {
         profiles[*index].data = data_new;
+        is_save_needed = true;
         return true;
     }
 
@@ -391,6 +397,10 @@ void ProfileManager::ParseUserSaveFile() {
 }
 
 void ProfileManager::WriteUserSaveFile() {
+    if (!is_save_needed) {
+        return;
+    }
+
     ProfileDataRaw raw{};
 
     for (std::size_t i = 0; i < MAX_USERS; ++i) {
@@ -423,7 +433,10 @@ void ProfileManager::WriteUserSaveFile() {
     if (!save.IsOpen() || !save.SetSize(sizeof(ProfileDataRaw)) || !save.WriteObject(raw)) {
         LOG_WARNING(Service_ACC, "Failed to write save data to file... No changes to user data "
                                  "made in current session will be saved.");
+        return;
     }
+
+    is_save_needed = false;
 }
 
 }; // namespace Service::Account

src/core/hle/service/acc/profile_manager.h

@@ -103,6 +103,7 @@ private:
     std::optional<std::size_t> AddToProfiles(const ProfileInfo& profile);
     bool RemoveProfileAtIndex(std::size_t index);
 
+    bool is_save_needed{};
     std::array<ProfileInfo, MAX_USERS> profiles{};
     std::array<ProfileInfo, MAX_USERS> stored_opened_profiles{};
     std::size_t user_count{};

src/core/hle/service/am/am.cpp

@@ -36,6 +36,7 @@
 #include "core/hle/service/caps/caps_su.h"
 #include "core/hle/service/caps/caps_types.h"
 #include "core/hle/service/filesystem/filesystem.h"
+#include "core/hle/service/filesystem/save_data_controller.h"
 #include "core/hle/service/ipc_helpers.h"
 #include "core/hle/service/ns/ns.h"
 #include "core/hle/service/nvnflinger/fb_share_buffer_manager.h"
@@ -2178,7 +2179,7 @@ void IApplicationFunctions::EnsureSaveData(HLERequestContext& ctx) {
     attribute.type = FileSys::SaveDataType::SaveData;
 
     FileSys::VirtualDir save_data{};
-    const auto res = system.GetFileSystemController().CreateSaveData(
+    const auto res = system.GetFileSystemController().OpenSaveDataController()->CreateSaveData(
         &save_data, FileSys::SaveDataSpaceId::NandUser, attribute);
 
     IPC::ResponseBuilder rb{ctx, 4};
@@ -2353,7 +2354,7 @@ void IApplicationFunctions::ExtendSaveData(HLERequestContext& ctx) {
               "new_journal={:016X}",
               static_cast<u8>(type), user_id[1], user_id[0], new_normal_size, new_journal_size);
 
-    system.GetFileSystemController().WriteSaveDataSize(
+    system.GetFileSystemController().OpenSaveDataController()->WriteSaveDataSize(
         type, system.GetApplicationProcessProgramID(), user_id,
         {new_normal_size, new_journal_size});
 
@@ -2378,7 +2379,7 @@ void IApplicationFunctions::GetSaveDataSize(HLERequestContext& ctx) {
     LOG_DEBUG(Service_AM, "called with type={:02X}, user_id={:016X}{:016X}", type, user_id[1],
               user_id[0]);
 
-    const auto size = system.GetFileSystemController().ReadSaveDataSize(
+    const auto size = system.GetFileSystemController().OpenSaveDataController()->ReadSaveDataSize(
         type, system.GetApplicationProcessProgramID(), user_id);
 
     IPC::ResponseBuilder rb{ctx, 6};

src/core/hle/service/audio/audin_u.cpp

@@ -18,11 +18,11 @@ using namespace AudioCore::AudioIn;
 class IAudioIn final : public ServiceFramework<IAudioIn> {
 public:
     explicit IAudioIn(Core::System& system_, Manager& manager, size_t session_id,
-                      const std::string& device_name, const AudioInParameter& in_params, u32 handle,
+                      const std::string& device_name, const AudioInParameter& in_params,
-                      u64 applet_resource_user_id)
+                      Kernel::KProcess* handle, u64 applet_resource_user_id)
         : ServiceFramework{system_, "IAudioIn"},
           service_context{system_, "IAudioIn"}, event{service_context.CreateEvent("AudioInEvent")},
-          impl{std::make_shared<In>(system_, manager, event, session_id)} {
+          process{handle}, impl{std::make_shared<In>(system_, manager, event, session_id)} {
         // clang-format off
         static const FunctionInfo functions[] = {
             {0, &IAudioIn::GetAudioInState, "GetAudioInState"},
@@ -45,6 +45,8 @@ public:
 
         RegisterHandlers(functions);
 
+        process->Open();
+
         if (impl->GetSystem()
                 .Initialize(device_name, in_params, handle, applet_resource_user_id)
                 .IsError()) {
@@ -55,6 +57,7 @@ public:
     ~IAudioIn() override {
         impl->Free();
         service_context.CloseEvent(event);
+        process->Close();
     }
 
     [[nodiscard]] std::shared_ptr<In> GetImpl() {
@@ -196,6 +199,7 @@ private:
 
     KernelHelpers::ServiceContext service_context;
     Kernel::KEvent* event;
+    Kernel::KProcess* process;
     std::shared_ptr<AudioCore::AudioIn::In> impl;
     Common::ScratchBuffer<u64> released_buffer;
 };
@@ -267,6 +271,14 @@ void AudInU::OpenAudioIn(HLERequestContext& ctx) {
     auto device_name = Common::StringFromBuffer(device_name_data);
     auto handle{ctx.GetCopyHandle(0)};
 
+    auto process{ctx.GetObjectFromHandle<Kernel::KProcess>(handle)};
+    if (process.IsNull()) {
+        LOG_ERROR(Service_Audio, "Failed to get process handle");
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(ResultUnknown);
+        return;
+    }
+
     std::scoped_lock l{impl->mutex};
     auto link{impl->LinkToManager()};
     if (link.IsError()) {
@@ -287,8 +299,9 @@ void AudInU::OpenAudioIn(HLERequestContext& ctx) {
     LOG_DEBUG(Service_Audio, "Opening new AudioIn, sessionid={}, free sessions={}", new_session_id,
               impl->num_free_sessions);
 
-    auto audio_in = std::make_shared<IAudioIn>(system, *impl, new_session_id, device_name,
+    auto audio_in =
-                                               in_params, handle, applet_resource_user_id);
+        std::make_shared<IAudioIn>(system, *impl, new_session_id, device_name, in_params,
+                                   process.GetPointerUnsafe(), applet_resource_user_id);
     impl->sessions[new_session_id] = audio_in->GetImpl();
     impl->applet_resource_user_ids[new_session_id] = applet_resource_user_id;
 
@@ -318,6 +331,14 @@ void AudInU::OpenAudioInProtocolSpecified(HLERequestContext& ctx) {
     auto device_name = Common::StringFromBuffer(device_name_data);
     auto handle{ctx.GetCopyHandle(0)};
 
+    auto process{ctx.GetObjectFromHandle<Kernel::KProcess>(handle)};
+    if (process.IsNull()) {
+        LOG_ERROR(Service_Audio, "Failed to get process handle");
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(ResultUnknown);
+        return;
+    }
+
     std::scoped_lock l{impl->mutex};
     auto link{impl->LinkToManager()};
     if (link.IsError()) {
@@ -338,8 +359,9 @@ void AudInU::OpenAudioInProtocolSpecified(HLERequestContext& ctx) {
     LOG_DEBUG(Service_Audio, "Opening new AudioIn, sessionid={}, free sessions={}", new_session_id,
               impl->num_free_sessions);
 
-    auto audio_in = std::make_shared<IAudioIn>(system, *impl, new_session_id, device_name,
+    auto audio_in =
-                                               in_params, handle, applet_resource_user_id);
+        std::make_shared<IAudioIn>(system, *impl, new_session_id, device_name, in_params,
+                                   process.GetPointerUnsafe(), applet_resource_user_id);
     impl->sessions[new_session_id] = audio_in->GetImpl();
     impl->applet_resource_user_ids[new_session_id] = applet_resource_user_id;
 

src/core/hle/service/audio/audout_u.cpp

@@ -26,9 +26,10 @@ class IAudioOut final : public ServiceFramework<IAudioOut> {
 public:
     explicit IAudioOut(Core::System& system_, AudioCore::AudioOut::Manager& manager,
                        size_t session_id, const std::string& device_name,
-                       const AudioOutParameter& in_params, u32 handle, u64 applet_resource_user_id)
+                       const AudioOutParameter& in_params, Kernel::KProcess* handle,
+                       u64 applet_resource_user_id)
         : ServiceFramework{system_, "IAudioOut"}, service_context{system_, "IAudioOut"},
-          event{service_context.CreateEvent("AudioOutEvent")},
+          event{service_context.CreateEvent("AudioOutEvent")}, process{handle},
           impl{std::make_shared<AudioCore::AudioOut::Out>(system_, manager, event, session_id)} {
 
         // clang-format off
@@ -50,11 +51,14 @@ public:
         };
         // clang-format on
         RegisterHandlers(functions);
+
+        process->Open();
     }
 
     ~IAudioOut() override {
         impl->Free();
         service_context.CloseEvent(event);
+        process->Close();
     }
 
     [[nodiscard]] std::shared_ptr<AudioCore::AudioOut::Out> GetImpl() {
@@ -206,6 +210,7 @@ private:
 
     KernelHelpers::ServiceContext service_context;
     Kernel::KEvent* event;
+    Kernel::KProcess* process;
     std::shared_ptr<AudioCore::AudioOut::Out> impl;
     Common::ScratchBuffer<u64> released_buffer;
 };
@@ -257,6 +262,14 @@ void AudOutU::OpenAudioOut(HLERequestContext& ctx) {
     auto device_name = Common::StringFromBuffer(device_name_data);
     auto handle{ctx.GetCopyHandle(0)};
 
+    auto process{ctx.GetObjectFromHandle<Kernel::KProcess>(handle)};
+    if (process.IsNull()) {
+        LOG_ERROR(Service_Audio, "Failed to get process handle");
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(ResultUnknown);
+        return;
+    }
+
     auto link{impl->LinkToManager()};
     if (link.IsError()) {
         LOG_ERROR(Service_Audio, "Failed to link Audio Out to Audio Manager");
@@ -276,10 +289,11 @@ void AudOutU::OpenAudioOut(HLERequestContext& ctx) {
     LOG_DEBUG(Service_Audio, "Opening new AudioOut, sessionid={}, free sessions={}", new_session_id,
               impl->num_free_sessions);
 
-    auto audio_out = std::make_shared<IAudioOut>(system, *impl, new_session_id, device_name,
+    auto audio_out =
-                                                 in_params, handle, applet_resource_user_id);
+        std::make_shared<IAudioOut>(system, *impl, new_session_id, device_name, in_params,
-    result = audio_out->GetImpl()->GetSystem().Initialize(device_name, in_params, handle,
+                                    process.GetPointerUnsafe(), applet_resource_user_id);
-                                                          applet_resource_user_id);
+    result = audio_out->GetImpl()->GetSystem().Initialize(
+        device_name, in_params, process.GetPointerUnsafe(), applet_resource_user_id);
     if (result.IsError()) {
         LOG_ERROR(Service_Audio, "Failed to initialize the AudioOut System!");
         IPC::ResponseBuilder rb{ctx, 2};

src/core/hle/service/filesystem/filesystem.cpp

@@ -24,15 +24,13 @@
 #include "core/hle/service/filesystem/fsp_ldr.h"
 #include "core/hle/service/filesystem/fsp_pr.h"
 #include "core/hle/service/filesystem/fsp_srv.h"
+#include "core/hle/service/filesystem/romfs_controller.h"
+#include "core/hle/service/filesystem/save_data_controller.h"
 #include "core/hle/service/server_manager.h"
 #include "core/loader/loader.h"
 
 namespace Service::FileSystem {
 
-// A default size for normal/journal save data size if application control metadata cannot be found.
-// This should be large enough to satisfy even the most extreme requirements (~4.2GB)
-constexpr u64 SUFFICIENT_SAVE_DATA_SIZE = 0xF0000000;
-
 static FileSys::VirtualDir GetDirectoryRelativeWrapped(FileSys::VirtualDir base,
                                                        std::string_view dir_name_) {
     std::string dir_name(Common::FS::SanitizePath(dir_name_));
@@ -297,145 +295,65 @@ FileSystemController::FileSystemController(Core::System& system_) : system{syste
 
 FileSystemController::~FileSystemController() = default;
 
-Result FileSystemController::RegisterRomFS(std::unique_ptr<FileSys::RomFSFactory>&& factory) {
+Result FileSystemController::RegisterProcess(
-    romfs_factory = std::move(factory);
+    ProcessId process_id, ProgramId program_id,
-    LOG_DEBUG(Service_FS, "Registered RomFS");
+    std::shared_ptr<FileSys::RomFSFactory>&& romfs_factory) {
+    std::scoped_lock lk{registration_lock};
+
+    registrations.emplace(process_id, Registration{
+                                          .program_id = program_id,
+                                          .romfs_factory = std::move(romfs_factory),
+                                          .save_data_factory = CreateSaveDataFactory(program_id),
+                                      });
+
+    LOG_DEBUG(Service_FS, "Registered for process {}", process_id);
     return ResultSuccess;
 }
 
-Result FileSystemController::RegisterSaveData(std::unique_ptr<FileSys::SaveDataFactory>&& factory) {
+Result FileSystemController::OpenProcess(
-    ASSERT_MSG(save_data_factory == nullptr, "Tried to register a second save data");
+    ProgramId* out_program_id, std::shared_ptr<SaveDataController>* out_save_data_controller,
-    save_data_factory = std::move(factory);
+    std::shared_ptr<RomFsController>* out_romfs_controller, ProcessId process_id) {
-    LOG_DEBUG(Service_FS, "Registered save data");
+    std::scoped_lock lk{registration_lock};
+
+    const auto it = registrations.find(process_id);
+    if (it == registrations.end()) {
+        return FileSys::ERROR_ENTITY_NOT_FOUND;
+    }
+
+    *out_program_id = it->second.program_id;
+    *out_save_data_controller =
+        std::make_shared<SaveDataController>(system, it->second.save_data_factory);
+    *out_romfs_controller =
+        std::make_shared<RomFsController>(it->second.romfs_factory, it->second.program_id);
     return ResultSuccess;
 }
 
-Result FileSystemController::RegisterSDMC(std::unique_ptr<FileSys::SDMCFactory>&& factory) {
+void FileSystemController::SetPackedUpdate(ProcessId process_id, FileSys::VirtualFile update_raw) {
-    ASSERT_MSG(sdmc_factory == nullptr, "Tried to register a second SDMC");
-    sdmc_factory = std::move(factory);
-    LOG_DEBUG(Service_FS, "Registered SDMC");
-    return ResultSuccess;
-}
-
-Result FileSystemController::RegisterBIS(std::unique_ptr<FileSys::BISFactory>&& factory) {
-    ASSERT_MSG(bis_factory == nullptr, "Tried to register a second BIS");
-    bis_factory = std::move(factory);
-    LOG_DEBUG(Service_FS, "Registered BIS");
-    return ResultSuccess;
-}
-
-void FileSystemController::SetPackedUpdate(FileSys::VirtualFile update_raw) {
     LOG_TRACE(Service_FS, "Setting packed update for romfs");
 
-    if (romfs_factory == nullptr)
+    std::scoped_lock lk{registration_lock};
+    const auto it = registrations.find(process_id);
+    if (it == registrations.end()) {
         return;
+    }
 
-    romfs_factory->SetPackedUpdate(std::move(update_raw));
+    it->second.romfs_factory->SetPackedUpdate(std::move(update_raw));
 }
 
-FileSys::VirtualFile FileSystemController::OpenRomFSCurrentProcess() const {
+std::shared_ptr<SaveDataController> FileSystemController::OpenSaveDataController() {
-    LOG_TRACE(Service_FS, "Opening RomFS for current process");
+    return std::make_shared<SaveDataController>(system, CreateSaveDataFactory(ProgramId{}));
-
-    if (romfs_factory == nullptr) {
-        return nullptr;
-    }
-
-    return romfs_factory->OpenCurrentProcess(system.GetApplicationProcessProgramID());
 }
 
-FileSys::VirtualFile FileSystemController::OpenPatchedRomFS(u64 title_id,
+std::shared_ptr<FileSys::SaveDataFactory> FileSystemController::CreateSaveDataFactory(
-                                                            FileSys::ContentRecordType type) const {
+    ProgramId program_id) {
-    LOG_TRACE(Service_FS, "Opening patched RomFS for title_id={:016X}", title_id);
+    using YuzuPath = Common::FS::YuzuPath;
+    const auto rw_mode = FileSys::Mode::ReadWrite;
 
-    if (romfs_factory == nullptr) {
+    auto vfs = system.GetFilesystem();
-        return nullptr;
+    const auto nand_directory =
-    }
+        vfs->OpenDirectory(Common::FS::GetYuzuPathString(YuzuPath::NANDDir), rw_mode);
-
+    return std::make_shared<FileSys::SaveDataFactory>(system, program_id,
-    return romfs_factory->OpenPatchedRomFS(title_id, type);
+                                                      std::move(nand_directory));
-}
-
-FileSys::VirtualFile FileSystemController::OpenPatchedRomFSWithProgramIndex(
-    u64 title_id, u8 program_index, FileSys::ContentRecordType type) const {
-    LOG_TRACE(Service_FS, "Opening patched RomFS for title_id={:016X}, program_index={}", title_id,
-              program_index);
-
-    if (romfs_factory == nullptr) {
-        return nullptr;
-    }
-
-    return romfs_factory->OpenPatchedRomFSWithProgramIndex(title_id, program_index, type);
-}
-
-FileSys::VirtualFile FileSystemController::OpenRomFS(u64 title_id, FileSys::StorageId storage_id,
-                                                     FileSys::ContentRecordType type) const {
-    LOG_TRACE(Service_FS, "Opening RomFS for title_id={:016X}, storage_id={:02X}, type={:02X}",
-              title_id, storage_id, type);
-
-    if (romfs_factory == nullptr) {
-        return nullptr;
-    }
-
-    return romfs_factory->Open(title_id, storage_id, type);
-}
-
-std::shared_ptr<FileSys::NCA> FileSystemController::OpenBaseNca(
-    u64 title_id, FileSys::StorageId storage_id, FileSys::ContentRecordType type) const {
-    return romfs_factory->GetEntry(title_id, storage_id, type);
-}
-
-Result FileSystemController::CreateSaveData(FileSys::VirtualDir* out_save_data,
-                                            FileSys::SaveDataSpaceId space,
-                                            const FileSys::SaveDataAttribute& save_struct) const {
-    LOG_TRACE(Service_FS, "Creating Save Data for space_id={:01X}, save_struct={}", space,
-              save_struct.DebugInfo());
-
-    if (save_data_factory == nullptr) {
-        return FileSys::ERROR_ENTITY_NOT_FOUND;
-    }
-
-    auto save_data = save_data_factory->Create(space, save_struct);
-    if (save_data == nullptr) {
-        return FileSys::ERROR_ENTITY_NOT_FOUND;
-    }
-
-    *out_save_data = save_data;
-    return ResultSuccess;
-}
-
-Result FileSystemController::OpenSaveData(FileSys::VirtualDir* out_save_data,
-                                          FileSys::SaveDataSpaceId space,
-                                          const FileSys::SaveDataAttribute& attribute) const {
-    LOG_TRACE(Service_FS, "Opening Save Data for space_id={:01X}, save_struct={}", space,
-              attribute.DebugInfo());
-
-    if (save_data_factory == nullptr) {
-        return FileSys::ERROR_ENTITY_NOT_FOUND;
-    }
-
-    auto save_data = save_data_factory->Open(space, attribute);
-    if (save_data == nullptr) {
-        return FileSys::ERROR_ENTITY_NOT_FOUND;
-    }
-
-    *out_save_data = save_data;
-    return ResultSuccess;
-}
-
-Result FileSystemController::OpenSaveDataSpace(FileSys::VirtualDir* out_save_data_space,
-                                               FileSys::SaveDataSpaceId space) const {
-    LOG_TRACE(Service_FS, "Opening Save Data Space for space_id={:01X}", space);
-
-    if (save_data_factory == nullptr) {
-        return FileSys::ERROR_ENTITY_NOT_FOUND;
-    }
-
-    auto save_data_space = save_data_factory->GetSaveDataSpaceDirectory(space);
-    if (save_data_space == nullptr) {
-        return FileSys::ERROR_ENTITY_NOT_FOUND;
-    }
-
-    *out_save_data_space = save_data_space;
-    return ResultSuccess;
 }
 
 Result FileSystemController::OpenSDMC(FileSys::VirtualDir* out_sdmc) const {
@@ -540,48 +458,6 @@ u64 FileSystemController::GetTotalSpaceSize(FileSys::StorageId id) const {
     return 0;
 }
 
-FileSys::SaveDataSize FileSystemController::ReadSaveDataSize(FileSys::SaveDataType type,
-                                                             u64 title_id, u128 user_id) const {
-    if (save_data_factory == nullptr) {
-        return {0, 0};
-    }
-
-    const auto value = save_data_factory->ReadSaveDataSize(type, title_id, user_id);
-
-    if (value.normal == 0 && value.journal == 0) {
-        FileSys::SaveDataSize new_size{SUFFICIENT_SAVE_DATA_SIZE, SUFFICIENT_SAVE_DATA_SIZE};
-
-        FileSys::NACP nacp;
-        const auto res = system.GetAppLoader().ReadControlData(nacp);
-
-        if (res != Loader::ResultStatus::Success) {
-            const FileSys::PatchManager pm{system.GetApplicationProcessProgramID(),
-                                           system.GetFileSystemController(),
-                                           system.GetContentProvider()};
-            const auto metadata = pm.GetControlMetadata();
-            const auto& nacp_unique = metadata.first;
-
-            if (nacp_unique != nullptr) {
-                new_size = {nacp_unique->GetDefaultNormalSaveSize(),
-                            nacp_unique->GetDefaultJournalSaveSize()};
-            }
-        } else {
-            new_size = {nacp.GetDefaultNormalSaveSize(), nacp.GetDefaultJournalSaveSize()};
-        }
-
-        WriteSaveDataSize(type, title_id, user_id, new_size);
-        return new_size;
-    }
-
-    return value;
-}
-
-void FileSystemController::WriteSaveDataSize(FileSys::SaveDataType type, u64 title_id, u128 user_id,
-                                             FileSys::SaveDataSize new_value) const {
-    if (save_data_factory != nullptr)
-        save_data_factory->WriteSaveDataSize(type, title_id, user_id, new_value);
-}
-
 void FileSystemController::SetGameCard(FileSys::VirtualFile file) {
     gamecard = std::make_unique<FileSys::XCI>(file);
     const auto dir = gamecard->ConcatenatedPseudoDirectory();
@@ -801,14 +677,9 @@ FileSys::VirtualDir FileSystemController::GetBCATDirectory(u64 title_id) const {
     return bis_factory->GetBCATDirectory(title_id);
 }
 
-void FileSystemController::SetAutoSaveDataCreation(bool enable) {
-    save_data_factory->SetAutoCreate(enable);
-}
-
 void FileSystemController::CreateFactories(FileSys::VfsFilesystem& vfs, bool overwrite) {
     if (overwrite) {
         bis_factory = nullptr;
-        save_data_factory = nullptr;
         sdmc_factory = nullptr;
     }
 
@@ -836,11 +707,6 @@ void FileSystemController::CreateFactories(FileSys::VfsFilesystem& vfs, bool ove
                                        bis_factory->GetUserNANDContents());
     }
 
-    if (save_data_factory == nullptr) {
-        save_data_factory =
-            std::make_unique<FileSys::SaveDataFactory>(system, std::move(nand_directory));
-    }
-
     if (sdmc_factory == nullptr) {
         sdmc_factory = std::make_unique<FileSys::SDMCFactory>(std::move(sd_directory),
                                                               std::move(sd_load_directory));
@@ -849,12 +715,19 @@ void FileSystemController::CreateFactories(FileSys::VfsFilesystem& vfs, bool ove
     }
 }
 
+void FileSystemController::Reset() {
+    std::scoped_lock lk{registration_lock};
+    registrations.clear();
+}
+
 void LoopProcess(Core::System& system) {
     auto server_manager = std::make_unique<ServerManager>(system);
 
+    const auto FileSystemProxyFactory = [&] { return std::make_shared<FSP_SRV>(system); };
+
     server_manager->RegisterNamedService("fsp-ldr", std::make_shared<FSP_LDR>(system));
     server_manager->RegisterNamedService("fsp:pr", std::make_shared<FSP_PR>(system));
-    server_manager->RegisterNamedService("fsp-srv", std::make_shared<FSP_SRV>(system));
+    server_manager->RegisterNamedService("fsp-srv", std::move(FileSystemProxyFactory));
     ServerManager::RunServer(std::move(server_manager));
 }
 

src/core/hle/service/filesystem/filesystem.h

@@ -43,6 +43,9 @@ class ServiceManager;
 
 namespace FileSystem {
 
+class RomFsController;
+class SaveDataController;
+
 enum class ContentStorageId : u32 {
     System,
     User,
@@ -61,32 +64,24 @@ enum class OpenDirectoryMode : u64 {
 };
 DECLARE_ENUM_FLAG_OPERATORS(OpenDirectoryMode);
 
+using ProcessId = u64;
+using ProgramId = u64;
+
 class FileSystemController {
 public:
     explicit FileSystemController(Core::System& system_);
     ~FileSystemController();
 
-    Result RegisterRomFS(std::unique_ptr<FileSys::RomFSFactory>&& factory);
+    Result RegisterProcess(ProcessId process_id, ProgramId program_id,
-    Result RegisterSaveData(std::unique_ptr<FileSys::SaveDataFactory>&& factory);
+                           std::shared_ptr<FileSys::RomFSFactory>&& factory);
-    Result RegisterSDMC(std::unique_ptr<FileSys::SDMCFactory>&& factory);
+    Result OpenProcess(ProgramId* out_program_id,
-    Result RegisterBIS(std::unique_ptr<FileSys::BISFactory>&& factory);
+                       std::shared_ptr<SaveDataController>* out_save_data_controller,
+                       std::shared_ptr<RomFsController>* out_romfs_controller,
+                       ProcessId process_id);
+    void SetPackedUpdate(ProcessId process_id, FileSys::VirtualFile update_raw);
 
-    void SetPackedUpdate(FileSys::VirtualFile update_raw);
+    std::shared_ptr<SaveDataController> OpenSaveDataController();
-    FileSys::VirtualFile OpenRomFSCurrentProcess() const;
-    FileSys::VirtualFile OpenPatchedRomFS(u64 title_id, FileSys::ContentRecordType type) const;
-    FileSys::VirtualFile OpenPatchedRomFSWithProgramIndex(u64 title_id, u8 program_index,
-                                                          FileSys::ContentRecordType type) const;
-    FileSys::VirtualFile OpenRomFS(u64 title_id, FileSys::StorageId storage_id,
-                                   FileSys::ContentRecordType type) const;
-    std::shared_ptr<FileSys::NCA> OpenBaseNca(u64 title_id, FileSys::StorageId storage_id,
-                                              FileSys::ContentRecordType type) const;
 
-    Result CreateSaveData(FileSys::VirtualDir* out_save_data, FileSys::SaveDataSpaceId space,
-                          const FileSys::SaveDataAttribute& save_struct) const;
-    Result OpenSaveData(FileSys::VirtualDir* out_save_data, FileSys::SaveDataSpaceId space,
-                        const FileSys::SaveDataAttribute& save_struct) const;
-    Result OpenSaveDataSpace(FileSys::VirtualDir* out_save_data_space,
-                             FileSys::SaveDataSpaceId space) const;
     Result OpenSDMC(FileSys::VirtualDir* out_sdmc) const;
     Result OpenBISPartition(FileSys::VirtualDir* out_bis_partition,
                             FileSys::BisPartitionId id) const;
@@ -96,11 +91,6 @@ public:
     u64 GetFreeSpaceSize(FileSys::StorageId id) const;
     u64 GetTotalSpaceSize(FileSys::StorageId id) const;
 
-    FileSys::SaveDataSize ReadSaveDataSize(FileSys::SaveDataType type, u64 title_id,
-                                           u128 user_id) const;
-    void WriteSaveDataSize(FileSys::SaveDataType type, u64 title_id, u128 user_id,
-                           FileSys::SaveDataSize new_value) const;
-
     void SetGameCard(FileSys::VirtualFile file);
     FileSys::XCI* GetGameCard() const;
 
@@ -133,15 +123,24 @@ public:
 
     FileSys::VirtualDir GetBCATDirectory(u64 title_id) const;
 
-    void SetAutoSaveDataCreation(bool enable);
-
     // Creates the SaveData, SDMC, and BIS Factories. Should be called once and before any function
     // above is called.
     void CreateFactories(FileSys::VfsFilesystem& vfs, bool overwrite = true);
 
+    void Reset();
+
 private:
-    std::unique_ptr<FileSys::RomFSFactory> romfs_factory;
+    std::shared_ptr<FileSys::SaveDataFactory> CreateSaveDataFactory(ProgramId program_id);
-    std::unique_ptr<FileSys::SaveDataFactory> save_data_factory;
+
+    struct Registration {
+        ProgramId program_id;
+        std::shared_ptr<FileSys::RomFSFactory> romfs_factory;
+        std::shared_ptr<FileSys::SaveDataFactory> save_data_factory;
+    };
+
+    std::mutex registration_lock;
+    std::map<ProcessId, Registration> registrations;
+
     std::unique_ptr<FileSys::SDMCFactory> sdmc_factory;
     std::unique_ptr<FileSys::BISFactory> bis_factory;
 

src/core/hle/service/filesystem/fsp_srv.cpp

@@ -27,6 +27,8 @@
 #include "core/hle/result.h"
 #include "core/hle/service/filesystem/filesystem.h"
 #include "core/hle/service/filesystem/fsp_srv.h"
+#include "core/hle/service/filesystem/romfs_controller.h"
+#include "core/hle/service/filesystem/save_data_controller.h"
 #include "core/hle/service/hle_ipc.h"
 #include "core/hle/service/ipc_helpers.h"
 #include "core/reporter.h"
@@ -577,9 +579,11 @@ private:
 
 class ISaveDataInfoReader final : public ServiceFramework<ISaveDataInfoReader> {
 public:
-    explicit ISaveDataInfoReader(Core::System& system_, FileSys::SaveDataSpaceId space,
+    explicit ISaveDataInfoReader(Core::System& system_,
-                                 FileSystemController& fsc_)
+                                 std::shared_ptr<SaveDataController> save_data_controller_,
-        : ServiceFramework{system_, "ISaveDataInfoReader"}, fsc{fsc_} {
+                                 FileSys::SaveDataSpaceId space)
+        : ServiceFramework{system_, "ISaveDataInfoReader"}, save_data_controller{
+                                                                save_data_controller_} {
         static const FunctionInfo functions[] = {
             {0, &ISaveDataInfoReader::ReadSaveDataInfo, "ReadSaveDataInfo"},
         };
@@ -626,7 +630,7 @@ private:
 
     void FindAllSaves(FileSys::SaveDataSpaceId space) {
         FileSys::VirtualDir save_root{};
-        const auto result = fsc.OpenSaveDataSpace(&save_root, space);
+        const auto result = save_data_controller->OpenSaveDataSpace(&save_root, space);
 
         if (result != ResultSuccess || save_root == nullptr) {
             LOG_ERROR(Service_FS, "The save root for the space_id={:02X} was invalid!", space);
@@ -723,7 +727,8 @@ private:
     };
     static_assert(sizeof(SaveDataInfo) == 0x60, "SaveDataInfo has incorrect size.");
 
-    FileSystemController& fsc;
+    ProcessId process_id = 0;
+    std::shared_ptr<SaveDataController> save_data_controller;
     std::vector<SaveDataInfo> info;
     u64 next_entry_index = 0;
 };
@@ -863,21 +868,20 @@ FSP_SRV::FSP_SRV(Core::System& system_)
     if (Settings::values.enable_fs_access_log) {
         access_log_mode = AccessLogMode::SdCard;
     }
-
-    // This should be true on creation
-    fsc.SetAutoSaveDataCreation(true);
 }
 
 FSP_SRV::~FSP_SRV() = default;
 
 void FSP_SRV::SetCurrentProcess(HLERequestContext& ctx) {
-    IPC::RequestParser rp{ctx};
+    current_process_id = ctx.GetPID();
-    current_process_id = rp.Pop<u64>();
 
     LOG_DEBUG(Service_FS, "called. current_process_id=0x{:016X}", current_process_id);
 
+    const auto res =
+        fsc.OpenProcess(&program_id, &save_data_controller, &romfs_controller, current_process_id);
+
     IPC::ResponseBuilder rb{ctx, 2};
-    rb.Push(ResultSuccess);
+    rb.Push(res);
 }
 
 void FSP_SRV::OpenFileSystemWithPatch(HLERequestContext& ctx) {
@@ -916,7 +920,8 @@ void FSP_SRV::CreateSaveDataFileSystem(HLERequestContext& ctx) {
               uid[1], uid[0]);
 
     FileSys::VirtualDir save_data_dir{};
-    fsc.CreateSaveData(&save_data_dir, FileSys::SaveDataSpaceId::NandUser, save_struct);
+    save_data_controller->CreateSaveData(&save_data_dir, FileSys::SaveDataSpaceId::NandUser,
+                                         save_struct);
 
     IPC::ResponseBuilder rb{ctx, 2};
     rb.Push(ResultSuccess);
@@ -931,7 +936,8 @@ void FSP_SRV::CreateSaveDataFileSystemBySystemSaveDataId(HLERequestContext& ctx)
     LOG_DEBUG(Service_FS, "called save_struct = {}", save_struct.DebugInfo());
 
     FileSys::VirtualDir save_data_dir{};
-    fsc.CreateSaveData(&save_data_dir, FileSys::SaveDataSpaceId::NandSystem, save_struct);
+    save_data_controller->CreateSaveData(&save_data_dir, FileSys::SaveDataSpaceId::NandSystem,
+                                         save_struct);
 
     IPC::ResponseBuilder rb{ctx, 2};
     rb.Push(ResultSuccess);
@@ -950,7 +956,8 @@ void FSP_SRV::OpenSaveDataFileSystem(HLERequestContext& ctx) {
     LOG_INFO(Service_FS, "called.");
 
     FileSys::VirtualDir dir{};
-    auto result = fsc.OpenSaveData(&dir, parameters.space_id, parameters.attribute);
+    auto result =
+        save_data_controller->OpenSaveData(&dir, parameters.space_id, parameters.attribute);
     if (result != ResultSuccess) {
         IPC::ResponseBuilder rb{ctx, 2, 0, 0};
         rb.Push(FileSys::ERROR_ENTITY_NOT_FOUND);
@@ -1001,7 +1008,7 @@ void FSP_SRV::OpenSaveDataInfoReaderBySaveDataSpaceId(HLERequestContext& ctx) {
     IPC::ResponseBuilder rb{ctx, 2, 0, 1};
     rb.Push(ResultSuccess);
     rb.PushIpcInterface<ISaveDataInfoReader>(
-        std::make_shared<ISaveDataInfoReader>(system, space, fsc));
+        std::make_shared<ISaveDataInfoReader>(system, save_data_controller, space));
 }
 
 void FSP_SRV::OpenSaveDataInfoReaderOnlyCacheStorage(HLERequestContext& ctx) {
@@ -1009,8 +1016,8 @@ void FSP_SRV::OpenSaveDataInfoReaderOnlyCacheStorage(HLERequestContext& ctx) {
 
     IPC::ResponseBuilder rb{ctx, 2, 0, 1};
     rb.Push(ResultSuccess);
-    rb.PushIpcInterface<ISaveDataInfoReader>(system, FileSys::SaveDataSpaceId::TemporaryStorage,
+    rb.PushIpcInterface<ISaveDataInfoReader>(system, save_data_controller,
-                                             fsc);
+                                             FileSys::SaveDataSpaceId::TemporaryStorage);
 }
 
 void FSP_SRV::WriteSaveDataFileSystemExtraDataBySaveDataAttribute(HLERequestContext& ctx) {
@@ -1050,7 +1057,7 @@ void FSP_SRV::OpenDataStorageByCurrentProcess(HLERequestContext& ctx) {
     LOG_DEBUG(Service_FS, "called");
 
     if (!romfs) {
-        auto current_romfs = fsc.OpenRomFSCurrentProcess();
+        auto current_romfs = romfs_controller->OpenRomFSCurrentProcess();
         if (!current_romfs) {
             // TODO (bunnei): Find the right error code to use here
             LOG_CRITICAL(Service_FS, "no file system interface available!");
@@ -1078,7 +1085,7 @@ void FSP_SRV::OpenDataStorageByDataId(HLERequestContext& ctx) {
     LOG_DEBUG(Service_FS, "called with storage_id={:02X}, unknown={:08X}, title_id={:016X}",
               storage_id, unknown, title_id);
 
-    auto data = fsc.OpenRomFS(title_id, storage_id, FileSys::ContentRecordType::Data);
+    auto data = romfs_controller->OpenRomFS(title_id, storage_id, FileSys::ContentRecordType::Data);
 
     if (!data) {
         const auto archive = FileSys::SystemArchive::SynthesizeSystemArchive(title_id);
@@ -1101,7 +1108,8 @@ void FSP_SRV::OpenDataStorageByDataId(HLERequestContext& ctx) {
 
     const FileSys::PatchManager pm{title_id, fsc, content_provider};
 
-    auto base = fsc.OpenBaseNca(title_id, storage_id, FileSys::ContentRecordType::Data);
+    auto base =
+        romfs_controller->OpenBaseNca(title_id, storage_id, FileSys::ContentRecordType::Data);
     auto storage = std::make_shared<IStorage>(
         system, pm.PatchRomFS(base.get(), std::move(data), FileSys::ContentRecordType::Data));
 
@@ -1129,9 +1137,8 @@ void FSP_SRV::OpenDataStorageWithProgramIndex(HLERequestContext& ctx) {
 
     LOG_DEBUG(Service_FS, "called, program_index={}", program_index);
 
-    auto patched_romfs =
+    auto patched_romfs = romfs_controller->OpenPatchedRomFSWithProgramIndex(
-        fsc.OpenPatchedRomFSWithProgramIndex(system.GetApplicationProcessProgramID(), program_index,
+        program_id, program_index, FileSys::ContentRecordType::Program);
-                                             FileSys::ContentRecordType::Program);
 
     if (!patched_romfs) {
         // TODO: Find the right error code to use here
@@ -1152,7 +1159,7 @@ void FSP_SRV::OpenDataStorageWithProgramIndex(HLERequestContext& ctx) {
 void FSP_SRV::DisableAutoSaveDataCreation(HLERequestContext& ctx) {
     LOG_DEBUG(Service_FS, "called");
 
-    fsc.SetAutoSaveDataCreation(false);
+    save_data_controller->SetAutoCreate(false);
 
     IPC::ResponseBuilder rb{ctx, 2};
     rb.Push(ResultSuccess);

src/core/hle/service/filesystem/fsp_srv.h

@@ -17,6 +17,9 @@ class FileSystemBackend;
 
 namespace Service::FileSystem {
 
+class RomFsController;
+class SaveDataController;
+
 enum class AccessLogVersion : u32 {
     V7_0_0 = 2,
 
@@ -67,6 +70,9 @@ private:
     u64 current_process_id = 0;
     u32 access_log_program_index = 0;
     AccessLogMode access_log_mode = AccessLogMode::None;
+    u64 program_id = 0;
+    std::shared_ptr<SaveDataController> save_data_controller;
+    std::shared_ptr<RomFsController> romfs_controller;
 };
 
 } // namespace Service::FileSystem

src/core/hle/service/filesystem/romfs_controller.cpp

@@ -0,0 +1,37 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "core/hle/service/filesystem/romfs_controller.h"
+
+namespace Service::FileSystem {
+
+RomFsController::RomFsController(std::shared_ptr<FileSys::RomFSFactory> factory_, u64 program_id_)
+    : factory{std::move(factory_)}, program_id{program_id_} {}
+RomFsController::~RomFsController() = default;
+
+FileSys::VirtualFile RomFsController::OpenRomFSCurrentProcess() {
+    return factory->OpenCurrentProcess(program_id);
+}
+
+FileSys::VirtualFile RomFsController::OpenPatchedRomFS(u64 title_id,
+                                                       FileSys::ContentRecordType type) {
+    return factory->OpenPatchedRomFS(title_id, type);
+}
+
+FileSys::VirtualFile RomFsController::OpenPatchedRomFSWithProgramIndex(
+    u64 title_id, u8 program_index, FileSys::ContentRecordType type) {
+    return factory->OpenPatchedRomFSWithProgramIndex(title_id, program_index, type);
+}
+
+FileSys::VirtualFile RomFsController::OpenRomFS(u64 title_id, FileSys::StorageId storage_id,
+                                                FileSys::ContentRecordType type) {
+    return factory->Open(title_id, storage_id, type);
+}
+
+std::shared_ptr<FileSys::NCA> RomFsController::OpenBaseNca(u64 title_id,
+                                                           FileSys::StorageId storage_id,
+                                                           FileSys::ContentRecordType type) {
+    return factory->GetEntry(title_id, storage_id, type);
+}
+
+} // namespace Service::FileSystem

src/core/hle/service/filesystem/romfs_controller.h

@@ -0,0 +1,31 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "core/file_sys/nca_metadata.h"
+#include "core/file_sys/romfs_factory.h"
+#include "core/file_sys/vfs_types.h"
+
+namespace Service::FileSystem {
+
+class RomFsController {
+public:
+    explicit RomFsController(std::shared_ptr<FileSys::RomFSFactory> factory_, u64 program_id_);
+    ~RomFsController();
+
+    FileSys::VirtualFile OpenRomFSCurrentProcess();
+    FileSys::VirtualFile OpenPatchedRomFS(u64 title_id, FileSys::ContentRecordType type);
+    FileSys::VirtualFile OpenPatchedRomFSWithProgramIndex(u64 title_id, u8 program_index,
+                                                          FileSys::ContentRecordType type);
+    FileSys::VirtualFile OpenRomFS(u64 title_id, FileSys::StorageId storage_id,
+                                   FileSys::ContentRecordType type);
+    std::shared_ptr<FileSys::NCA> OpenBaseNca(u64 title_id, FileSys::StorageId storage_id,
+                                              FileSys::ContentRecordType type);
+
+private:
+    const std::shared_ptr<FileSys::RomFSFactory> factory;
+    const u64 program_id;
+};
+
+} // namespace Service::FileSystem

src/core/hle/service/filesystem/save_data_controller.cpp

@@ -0,0 +1,99 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "core/core.h"
+#include "core/file_sys/control_metadata.h"
+#include "core/file_sys/errors.h"
+#include "core/file_sys/patch_manager.h"
+#include "core/hle/service/filesystem/save_data_controller.h"
+#include "core/loader/loader.h"
+
+namespace Service::FileSystem {
+
+namespace {
+
+// A default size for normal/journal save data size if application control metadata cannot be found.
+// This should be large enough to satisfy even the most extreme requirements (~4.2GB)
+constexpr u64 SufficientSaveDataSize = 0xF0000000;
+
+FileSys::SaveDataSize GetDefaultSaveDataSize(Core::System& system, u64 program_id) {
+    const FileSys::PatchManager pm{program_id, system.GetFileSystemController(),
+                                   system.GetContentProvider()};
+    const auto metadata = pm.GetControlMetadata();
+    const auto& nacp = metadata.first;
+
+    if (nacp != nullptr) {
+        return {nacp->GetDefaultNormalSaveSize(), nacp->GetDefaultJournalSaveSize()};
+    }
+
+    return {SufficientSaveDataSize, SufficientSaveDataSize};
+}
+
+} // namespace
+
+SaveDataController::SaveDataController(Core::System& system_,
+                                       std::shared_ptr<FileSys::SaveDataFactory> factory_)
+    : system{system_}, factory{std::move(factory_)} {}
+SaveDataController::~SaveDataController() = default;
+
+Result SaveDataController::CreateSaveData(FileSys::VirtualDir* out_save_data,
+                                          FileSys::SaveDataSpaceId space,
+                                          const FileSys::SaveDataAttribute& attribute) {
+    LOG_TRACE(Service_FS, "Creating Save Data for space_id={:01X}, save_struct={}", space,
+              attribute.DebugInfo());
+
+    auto save_data = factory->Create(space, attribute);
+    if (save_data == nullptr) {
+        return FileSys::ERROR_ENTITY_NOT_FOUND;
+    }
+
+    *out_save_data = save_data;
+    return ResultSuccess;
+}
+
+Result SaveDataController::OpenSaveData(FileSys::VirtualDir* out_save_data,
+                                        FileSys::SaveDataSpaceId space,
+                                        const FileSys::SaveDataAttribute& attribute) {
+    auto save_data = factory->Open(space, attribute);
+    if (save_data == nullptr) {
+        return FileSys::ERROR_ENTITY_NOT_FOUND;
+    }
+
+    *out_save_data = save_data;
+    return ResultSuccess;
+}
+
+Result SaveDataController::OpenSaveDataSpace(FileSys::VirtualDir* out_save_data_space,
+                                             FileSys::SaveDataSpaceId space) {
+    auto save_data_space = factory->GetSaveDataSpaceDirectory(space);
+    if (save_data_space == nullptr) {
+        return FileSys::ERROR_ENTITY_NOT_FOUND;
+    }
+
+    *out_save_data_space = save_data_space;
+    return ResultSuccess;
+}
+
+FileSys::SaveDataSize SaveDataController::ReadSaveDataSize(FileSys::SaveDataType type, u64 title_id,
+                                                           u128 user_id) {
+    const auto value = factory->ReadSaveDataSize(type, title_id, user_id);
+
+    if (value.normal == 0 && value.journal == 0) {
+        const auto size = GetDefaultSaveDataSize(system, title_id);
+        factory->WriteSaveDataSize(type, title_id, user_id, size);
+        return size;
+    }
+
+    return value;
+}
+
+void SaveDataController::WriteSaveDataSize(FileSys::SaveDataType type, u64 title_id, u128 user_id,
+                                           FileSys::SaveDataSize new_value) {
+    factory->WriteSaveDataSize(type, title_id, user_id, new_value);
+}
+
+void SaveDataController::SetAutoCreate(bool state) {
+    factory->SetAutoCreate(state);
+}
+
+} // namespace Service::FileSystem

src/core/hle/service/filesystem/save_data_controller.h

@@ -0,0 +1,35 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "core/file_sys/nca_metadata.h"
+#include "core/file_sys/savedata_factory.h"
+#include "core/file_sys/vfs_types.h"
+
+namespace Service::FileSystem {
+
+class SaveDataController {
+public:
+    explicit SaveDataController(Core::System& system,
+                                std::shared_ptr<FileSys::SaveDataFactory> factory_);
+    ~SaveDataController();
+
+    Result CreateSaveData(FileSys::VirtualDir* out_save_data, FileSys::SaveDataSpaceId space,
+                          const FileSys::SaveDataAttribute& attribute);
+    Result OpenSaveData(FileSys::VirtualDir* out_save_data, FileSys::SaveDataSpaceId space,
+                        const FileSys::SaveDataAttribute& attribute);
+    Result OpenSaveDataSpace(FileSys::VirtualDir* out_save_data_space,
+                             FileSys::SaveDataSpaceId space);
+
+    FileSys::SaveDataSize ReadSaveDataSize(FileSys::SaveDataType type, u64 title_id, u128 user_id);
+    void WriteSaveDataSize(FileSys::SaveDataType type, u64 title_id, u128 user_id,
+                           FileSys::SaveDataSize new_value);
+    void SetAutoCreate(bool state);
+
+private:
+    Core::System& system;
+    const std::shared_ptr<FileSys::SaveDataFactory> factory;
+};
+
+} // namespace Service::FileSystem

src/core/hle/service/glue/arp.cpp

@@ -15,9 +15,10 @@
 namespace Service::Glue {
 
 namespace {
-std::optional<u64> GetTitleIDForProcessID(const Core::System& system, u64 process_id) {
+std::optional<u64> GetTitleIDForProcessID(Core::System& system, u64 process_id) {
-    const auto& list = system.Kernel().GetProcessList();
+    auto list = system.Kernel().GetProcessList();
-    const auto iter = std::find_if(list.begin(), list.end(), [&process_id](const auto& process) {
+
+    const auto iter = std::find_if(list.begin(), list.end(), [&process_id](auto& process) {
         return process->GetProcessId() == process_id;
     });
 

src/core/hle/service/hid/hid.cpp

@@ -22,12 +22,10 @@ void LoopProcess(Core::System& system) {
     std::shared_ptr<HidFirmwareSettings> firmware_settings =
         std::make_shared<HidFirmwareSettings>();
 
-    // TODO: Remove this hack until this service is emulated properly.
+    // TODO: Remove this hack when am is emulated properly.
-    const auto process_list = system.Kernel().GetProcessList();
+    resource_manager->Initialize();
-    if (!process_list.empty()) {
+    resource_manager->RegisterAppletResourceUserId(system.ApplicationProcess()->GetProcessId(),
-        resource_manager->Initialize();
+                                                   true);
-        resource_manager->RegisterAppletResourceUserId(process_list[0]->GetId(), true);
-    }
 
     server_manager->RegisterNamedService(
         "hid", std::make_shared<IHidServer>(system, resource_manager, firmware_settings));

src/core/hle/service/hle_ipc.cpp

@@ -12,6 +12,7 @@
 #include "common/common_types.h"
 #include "common/logging/log.h"
 #include "common/scratch_buffer.h"
+#include "core/guest_memory.h"
 #include "core/hle/kernel/k_auto_object.h"
 #include "core/hle/kernel/k_handle_table.h"
 #include "core/hle/kernel/k_process.h"
@@ -344,9 +345,9 @@ std::vector<u8> HLERequestContext::ReadBufferCopy(std::size_t buffer_index) cons
 
 std::span<const u8> HLERequestContext::ReadBufferA(std::size_t buffer_index) const {
     static thread_local std::array read_buffer_a{
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
     };
 
     ASSERT_OR_EXECUTE_MSG(
@@ -360,9 +361,9 @@ std::span<const u8> HLERequestContext::ReadBufferA(std::size_t buffer_index) con
 
 std::span<const u8> HLERequestContext::ReadBufferX(std::size_t buffer_index) const {
     static thread_local std::array read_buffer_x{
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
     };
 
     ASSERT_OR_EXECUTE_MSG(
@@ -376,14 +377,14 @@ std::span<const u8> HLERequestContext::ReadBufferX(std::size_t buffer_index) con
 
 std::span<const u8> HLERequestContext::ReadBuffer(std::size_t buffer_index) const {
     static thread_local std::array read_buffer_a{
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
     };
     static thread_local std::array read_buffer_x{
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
+        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead>(memory, 0, 0),
     };
 
     const bool is_buffer_a{BufferDescriptorA().size() > buffer_index &&

src/core/hle/service/hle_ipc.h

@@ -41,6 +41,8 @@ class KernelCore;
 class KHandleTable;
 class KProcess;
 class KServerSession;
+template <typename T>
+class KScopedAutoObject;
 class KThread;
 } // namespace Kernel
 

src/core/hle/service/nvdrv/core/container.cpp

@@ -2,27 +2,134 @@
 // SPDX-FileCopyrightText: 2022 Skyline Team and Contributors
 // SPDX-License-Identifier: GPL-3.0-or-later
 
+#include <atomic>
+#include <deque>
+#include <mutex>
+
+#include "core/hle/kernel/k_process.h"
 #include "core/hle/service/nvdrv/core/container.h"
+#include "core/hle/service/nvdrv/core/heap_mapper.h"
 #include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/hle/service/nvdrv/core/syncpoint_manager.h"
+#include "core/memory.h"
 #include "video_core/host1x/host1x.h"
 
 namespace Service::Nvidia::NvCore {
 
+Session::Session(size_t id_, Kernel::KProcess* process_, size_t smmu_id_)
+    : id{id_}, process{process_}, smmu_id{smmu_id_},
+      has_preallocated_area{}, mapper{}, is_active{} {}
+
+Session::~Session() = default;
+
 struct ContainerImpl {
-    explicit ContainerImpl(Tegra::Host1x::Host1x& host1x_)
+    explicit ContainerImpl(Container& core, Tegra::Host1x::Host1x& host1x_)
-        : file{host1x_}, manager{host1x_}, device_file_data{} {}
+        : host1x{host1x_}, file{core, host1x_}, manager{host1x_}, device_file_data{} {}
+    Tegra::Host1x::Host1x& host1x;
     NvMap file;
     SyncpointManager manager;
     Container::Host1xDeviceFileData device_file_data;
+    std::deque<Session> sessions;
+    size_t new_ids{};
+    std::deque<size_t> id_pool;
+    std::mutex session_guard;
 };
 
 Container::Container(Tegra::Host1x::Host1x& host1x_) {
-    impl = std::make_unique<ContainerImpl>(host1x_);
+    impl = std::make_unique<ContainerImpl>(*this, host1x_);
 }
 
 Container::~Container() = default;
 
+size_t Container::OpenSession(Kernel::KProcess* process) {
+    std::scoped_lock lk(impl->session_guard);
+    for (auto& session : impl->sessions) {
+        if (!session.is_active) {
+            continue;
+        }
+        if (session.process == process) {
+            return session.id;
+        }
+    }
+    size_t new_id{};
+    auto* memory_interface = &process->GetMemory();
+    auto& smmu = impl->host1x.MemoryManager();
+    auto smmu_id = smmu.RegisterProcess(memory_interface);
+    if (!impl->id_pool.empty()) {
+        new_id = impl->id_pool.front();
+        impl->id_pool.pop_front();
+        impl->sessions[new_id] = Session{new_id, process, smmu_id};
+    } else {
+        new_id = impl->new_ids++;
+        impl->sessions.emplace_back(new_id, process, smmu_id);
+    }
+    auto& session = impl->sessions[new_id];
+    session.is_active = true;
+    // Optimization
+    if (process->IsApplication()) {
+        auto& page_table = process->GetPageTable().GetBasePageTable();
+        auto heap_start = page_table.GetHeapRegionStart();
+
+        Kernel::KProcessAddress cur_addr = heap_start;
+        size_t region_size = 0;
+        VAddr region_start = 0;
+        while (true) {
+            Kernel::KMemoryInfo mem_info{};
+            Kernel::Svc::PageInfo page_info{};
+            R_ASSERT(page_table.QueryInfo(std::addressof(mem_info), std::addressof(page_info),
+                                          cur_addr));
+            auto svc_mem_info = mem_info.GetSvcMemoryInfo();
+
+            // check if this memory block is heap
+            if (svc_mem_info.state == Kernel::Svc::MemoryState::Normal) {
+                if (svc_mem_info.size > region_size) {
+                    region_size = svc_mem_info.size;
+                    region_start = svc_mem_info.base_address;
+                }
+            }
+
+            // Check if we're done.
+            const uintptr_t next_address = svc_mem_info.base_address + svc_mem_info.size;
+            if (next_address <= GetInteger(cur_addr)) {
+                break;
+            }
+
+            cur_addr = next_address;
+        }
+        session.has_preallocated_area = false;
+        auto start_region = (region_size >> 15) >= 1024 ? smmu.Allocate(region_size) : 0;
+        if (start_region != 0) {
+            session.mapper = std::make_unique<HeapMapper>(region_start, start_region, region_size,
+                                                          smmu_id, impl->host1x);
+            smmu.TrackContinuity(start_region, region_start, region_size, smmu_id);
+            session.has_preallocated_area = true;
+            LOG_CRITICAL(Debug, "Preallocation created!");
+        }
+    }
+    return new_id;
+}
+
+void Container::CloseSession(size_t id) {
+    std::scoped_lock lk(impl->session_guard);
+    auto& session = impl->sessions[id];
+    auto& smmu = impl->host1x.MemoryManager();
+    if (session.has_preallocated_area) {
+        const DAddr region_start = session.mapper->GetRegionStart();
+        const size_t region_size = session.mapper->GetRegionSize();
+        session.mapper.reset();
+        smmu.Free(region_start, region_size);
+        session.has_preallocated_area = false;
+    }
+    session.is_active = false;
+    smmu.UnregisterProcess(impl->sessions[id].smmu_id);
+    impl->id_pool.emplace_front(id);
+}
+
+Session* Container::GetSession(size_t id) {
+    std::atomic_thread_fence(std::memory_order_acquire);
+    return &impl->sessions[id];
+}
+
 NvMap& Container::GetNvMapFile() {
     return impl->file;
 }

src/core/hle/service/nvdrv/core/container.h

@@ -10,22 +10,49 @@
 
 #include "core/hle/service/nvdrv/nvdata.h"
 
+namespace Kernel {
+class KProcess;
+}
+
 namespace Tegra::Host1x {
 class Host1x;
 } // namespace Tegra::Host1x
 
 namespace Service::Nvidia::NvCore {
 
+class HeapMapper;
 class NvMap;
 class SyncpointManager;
 
 struct ContainerImpl;
 
+struct Session {
+    Session(size_t id_, Kernel::KProcess* process_, size_t smmu_id_);
+    ~Session();
+
+    Session(const Session&) = delete;
+    Session& operator=(const Session&) = delete;
+    Session(Session&&) = default;
+    Session& operator=(Session&&) = default;
+
+    size_t id;
+    Kernel::KProcess* process;
+    size_t smmu_id;
+    bool has_preallocated_area{};
+    std::unique_ptr<HeapMapper> mapper{};
+    bool is_active{};
+};
+
 class Container {
 public:
     explicit Container(Tegra::Host1x::Host1x& host1x);
     ~Container();
 
+    size_t OpenSession(Kernel::KProcess* process);
+    void CloseSession(size_t id);
+
+    Session* GetSession(size_t id);
+
     NvMap& GetNvMapFile();
 
     const NvMap& GetNvMapFile() const;

src/core/hle/service/nvdrv/core/heap_mapper.cpp

@@ -0,0 +1,175 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <mutex>
+
+#include <boost/container/small_vector.hpp>
+#define BOOST_NO_MT
+#include <boost/pool/detail/mutex.hpp>
+#undef BOOST_NO_MT
+#include <boost/icl/interval.hpp>
+#include <boost/icl/interval_base_set.hpp>
+#include <boost/icl/interval_set.hpp>
+#include <boost/icl/split_interval_map.hpp>
+#include <boost/pool/pool.hpp>
+#include <boost/pool/pool_alloc.hpp>
+#include <boost/pool/poolfwd.hpp>
+
+#include "core/hle/service/nvdrv/core/heap_mapper.h"
+#include "video_core/host1x/host1x.h"
+
+namespace boost {
+template <typename T>
+class fast_pool_allocator<T, default_user_allocator_new_delete, details::pool::null_mutex, 4096, 0>;
+}
+
+namespace Service::Nvidia::NvCore {
+
+using IntervalCompare = std::less<DAddr>;
+using IntervalInstance = boost::icl::interval_type_default<DAddr, std::less>;
+using IntervalAllocator = boost::fast_pool_allocator<DAddr>;
+using IntervalSet = boost::icl::interval_set<DAddr>;
+using IntervalType = typename IntervalSet::interval_type;
+
+template <typename Type>
+struct counter_add_functor : public boost::icl::identity_based_inplace_combine<Type> {
+    // types
+    typedef counter_add_functor<Type> type;
+    typedef boost::icl::identity_based_inplace_combine<Type> base_type;
+
+    // public member functions
+    void operator()(Type& current, const Type& added) const {
+        current += added;
+        if (current < base_type::identity_element()) {
+            current = base_type::identity_element();
+        }
+    }
+
+    // public static functions
+    static void version(Type&){};
+};
+
+using OverlapCombine = counter_add_functor<int>;
+using OverlapSection = boost::icl::inter_section<int>;
+using OverlapCounter = boost::icl::split_interval_map<DAddr, int>;
+
+struct HeapMapper::HeapMapperInternal {
+    HeapMapperInternal(Tegra::Host1x::Host1x& host1x) : device_memory{host1x.MemoryManager()} {}
+    ~HeapMapperInternal() = default;
+
+    template <typename Func>
+    void ForEachInOverlapCounter(OverlapCounter& current_range, VAddr cpu_addr, u64 size,
+                                 Func&& func) {
+        const DAddr start_address = cpu_addr;
+        const DAddr end_address = start_address + size;
+        const IntervalType search_interval{start_address, end_address};
+        auto it = current_range.lower_bound(search_interval);
+        if (it == current_range.end()) {
+            return;
+        }
+        auto end_it = current_range.upper_bound(search_interval);
+        for (; it != end_it; it++) {
+            auto& inter = it->first;
+            DAddr inter_addr_end = inter.upper();
+            DAddr inter_addr = inter.lower();
+            if (inter_addr_end > end_address) {
+                inter_addr_end = end_address;
+            }
+            if (inter_addr < start_address) {
+                inter_addr = start_address;
+            }
+            func(inter_addr, inter_addr_end, it->second);
+        }
+    }
+
+    void RemoveEachInOverlapCounter(OverlapCounter& current_range,
+                                    const IntervalType search_interval, int subtract_value) {
+        bool any_removals = false;
+        current_range.add(std::make_pair(search_interval, subtract_value));
+        do {
+            any_removals = false;
+            auto it = current_range.lower_bound(search_interval);
+            if (it == current_range.end()) {
+                return;
+            }
+            auto end_it = current_range.upper_bound(search_interval);
+            for (; it != end_it; it++) {
+                if (it->second <= 0) {
+                    any_removals = true;
+                    current_range.erase(it);
+                    break;
+                }
+            }
+        } while (any_removals);
+    }
+
+    IntervalSet base_set;
+    OverlapCounter mapping_overlaps;
+    Tegra::MaxwellDeviceMemoryManager& device_memory;
+    std::mutex guard;
+};
+
+HeapMapper::HeapMapper(VAddr start_vaddress, DAddr start_daddress, size_t size, size_t smmu_id,
+                       Tegra::Host1x::Host1x& host1x)
+    : m_vaddress{start_vaddress}, m_daddress{start_daddress}, m_size{size}, m_smmu_id{smmu_id} {
+    m_internal = std::make_unique<HeapMapperInternal>(host1x);
+}
+
+HeapMapper::~HeapMapper() {
+    m_internal->device_memory.Unmap(m_daddress, m_size);
+}
+
+DAddr HeapMapper::Map(VAddr start, size_t size) {
+    std::scoped_lock lk(m_internal->guard);
+    m_internal->base_set.clear();
+    const IntervalType interval{start, start + size};
+    m_internal->base_set.insert(interval);
+    m_internal->ForEachInOverlapCounter(m_internal->mapping_overlaps, start, size,
+                                        [this](VAddr start_addr, VAddr end_addr, int) {
+                                            const IntervalType other{start_addr, end_addr};
+                                            m_internal->base_set.subtract(other);
+                                        });
+    if (!m_internal->base_set.empty()) {
+        auto it = m_internal->base_set.begin();
+        auto end_it = m_internal->base_set.end();
+        for (; it != end_it; it++) {
+            const VAddr inter_addr_end = it->upper();
+            const VAddr inter_addr = it->lower();
+            const size_t offset = inter_addr - m_vaddress;
+            const size_t sub_size = inter_addr_end - inter_addr;
+            m_internal->device_memory.Map(m_daddress + offset, m_vaddress + offset, sub_size,
+                                          m_smmu_id);
+        }
+    }
+    m_internal->mapping_overlaps += std::make_pair(interval, 1);
+    m_internal->base_set.clear();
+    return m_daddress + (start - m_vaddress);
+}
+
+void HeapMapper::Unmap(VAddr start, size_t size) {
+    std::scoped_lock lk(m_internal->guard);
+    m_internal->base_set.clear();
+    m_internal->ForEachInOverlapCounter(m_internal->mapping_overlaps, start, size,
+                                        [this](VAddr start_addr, VAddr end_addr, int value) {
+                                            if (value <= 1) {
+                                                const IntervalType other{start_addr, end_addr};
+                                                m_internal->base_set.insert(other);
+                                            }
+                                        });
+    if (!m_internal->base_set.empty()) {
+        auto it = m_internal->base_set.begin();
+        auto end_it = m_internal->base_set.end();
+        for (; it != end_it; it++) {
+            const VAddr inter_addr_end = it->upper();
+            const VAddr inter_addr = it->lower();
+            const size_t offset = inter_addr - m_vaddress;
+            const size_t sub_size = inter_addr_end - inter_addr;
+            m_internal->device_memory.Unmap(m_daddress + offset, sub_size);
+        }
+    }
+    const IntervalType to_remove{start, start + size};
+    m_internal->RemoveEachInOverlapCounter(m_internal->mapping_overlaps, to_remove, -1);
+    m_internal->base_set.clear();
+}
+
+} // namespace Service::Nvidia::NvCore

src/core/hle/service/nvdrv/core/heap_mapper.h

@@ -0,0 +1,48 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <memory>
+
+#include "common/common_types.h"
+
+namespace Tegra::Host1x {
+class Host1x;
+} // namespace Tegra::Host1x
+
+namespace Service::Nvidia::NvCore {
+
+class HeapMapper {
+public:
+    HeapMapper(VAddr start_vaddress, DAddr start_daddress, size_t size, size_t smmu_id,
+               Tegra::Host1x::Host1x& host1x);
+    ~HeapMapper();
+
+    bool IsInBounds(VAddr start, size_t size) const {
+        VAddr end = start + size;
+        return start >= m_vaddress && end <= (m_vaddress + m_size);
+    }
+
+    DAddr Map(VAddr start, size_t size);
+
+    void Unmap(VAddr start, size_t size);
+
+    DAddr GetRegionStart() const {
+        return m_daddress;
+    }
+
+    size_t GetRegionSize() const {
+        return m_size;
+    }
+
+private:
+    struct HeapMapperInternal;
+    VAddr m_vaddress;
+    DAddr m_daddress;
+    size_t m_size;
+    size_t m_smmu_id;
+    std::unique_ptr<HeapMapperInternal> m_internal;
+};
+
+} // namespace Service::Nvidia::NvCore

src/core/hle/service/nvdrv/core/nvmap.cpp

@@ -2,14 +2,19 @@
 // SPDX-FileCopyrightText: 2022 Skyline Team and Contributors
 // SPDX-License-Identifier: GPL-3.0-or-later
 
+#include <functional>
+
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "core/hle/service/nvdrv/core/container.h"
+#include "core/hle/service/nvdrv/core/heap_mapper.h"
 #include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/memory.h"
 #include "video_core/host1x/host1x.h"
 
 using Core::Memory::YUZU_PAGESIZE;
+constexpr size_t BIG_PAGE_SIZE = YUZU_PAGESIZE * 16;
 
 namespace Service::Nvidia::NvCore {
 NvMap::Handle::Handle(u64 size_, Id id_)
@@ -17,9 +22,8 @@ NvMap::Handle::Handle(u64 size_, Id id_)
     flags.raw = 0;
 }
 
-NvResult NvMap::Handle::Alloc(Flags pFlags, u32 pAlign, u8 pKind, u64 pAddress) {
+NvResult NvMap::Handle::Alloc(Flags pFlags, u32 pAlign, u8 pKind, u64 pAddress, size_t pSessionId) {
     std::scoped_lock lock(mutex);
-
     // Handles cannot be allocated twice
     if (allocated) {
         return NvResult::AccessDenied;
@@ -28,6 +32,7 @@ NvResult NvMap::Handle::Alloc(Flags pFlags, u32 pAlign, u8 pKind, u64 pAddress)
     flags = pFlags;
     kind = pKind;
     align = pAlign < YUZU_PAGESIZE ? YUZU_PAGESIZE : pAlign;
+    session_id = pSessionId;
 
     // This flag is only applicable for handles with an address passed
     if (pAddress) {
@@ -63,7 +68,7 @@ NvResult NvMap::Handle::Duplicate(bool internal_session) {
     return NvResult::Success;
 }
 
-NvMap::NvMap(Tegra::Host1x::Host1x& host1x_) : host1x{host1x_} {}
+NvMap::NvMap(Container& core_, Tegra::Host1x::Host1x& host1x_) : host1x{host1x_}, core{core_} {}
 
 void NvMap::AddHandle(std::shared_ptr<Handle> handle_description) {
     std::scoped_lock lock(handles_lock);
@@ -78,12 +83,30 @@ void NvMap::UnmapHandle(Handle& handle_description) {
         handle_description.unmap_queue_entry.reset();
     }
 
+    // Free and unmap the handle from Host1x GMMU
+    if (handle_description.pin_virt_address) {
+        host1x.GMMU().Unmap(static_cast<GPUVAddr>(handle_description.pin_virt_address),
+                            handle_description.aligned_size);
+        host1x.Allocator().Free(handle_description.pin_virt_address,
+                                static_cast<u32>(handle_description.aligned_size));
+        handle_description.pin_virt_address = 0;
+    }
+
     // Free and unmap the handle from the SMMU
-    host1x.MemoryManager().Unmap(static_cast<GPUVAddr>(handle_description.pin_virt_address),
+    const size_t map_size = handle_description.aligned_size;
-                                 handle_description.aligned_size);
+    if (!handle_description.in_heap) {
-    host1x.Allocator().Free(handle_description.pin_virt_address,
+        auto& smmu = host1x.MemoryManager();
-                            static_cast<u32>(handle_description.aligned_size));
+        size_t aligned_up = Common::AlignUp(map_size, BIG_PAGE_SIZE);
-    handle_description.pin_virt_address = 0;
+        smmu.Unmap(handle_description.d_address, map_size);
+        smmu.Free(handle_description.d_address, static_cast<size_t>(aligned_up));
+        handle_description.d_address = 0;
+        return;
+    }
+    const VAddr vaddress = handle_description.address;
+    auto* session = core.GetSession(handle_description.session_id);
+    session->mapper->Unmap(vaddress, map_size);
+    handle_description.d_address = 0;
+    handle_description.in_heap = false;
 }
 
 bool NvMap::TryRemoveHandle(const Handle& handle_description) {
@@ -124,22 +147,33 @@ std::shared_ptr<NvMap::Handle> NvMap::GetHandle(Handle::Id handle) {
     }
 }
 
-VAddr NvMap::GetHandleAddress(Handle::Id handle) {
+DAddr NvMap::GetHandleAddress(Handle::Id handle) {
     std::scoped_lock lock(handles_lock);
     try {
-        return handles.at(handle)->address;
+        return handles.at(handle)->d_address;
     } catch (std::out_of_range&) {
         return 0;
     }
 }
 
-u32 NvMap::PinHandle(NvMap::Handle::Id handle) {
+DAddr NvMap::PinHandle(NvMap::Handle::Id handle, bool low_area_pin) {
     auto handle_description{GetHandle(handle)};
     if (!handle_description) [[unlikely]] {
         return 0;
     }
 
     std::scoped_lock lock(handle_description->mutex);
+    const auto map_low_area = [&] {
+        if (handle_description->pin_virt_address == 0) {
+            auto& gmmu_allocator = host1x.Allocator();
+            auto& gmmu = host1x.GMMU();
+            u32 address =
+                gmmu_allocator.Allocate(static_cast<u32>(handle_description->aligned_size));
+            gmmu.Map(static_cast<GPUVAddr>(address), handle_description->d_address,
+                     handle_description->aligned_size);
+            handle_description->pin_virt_address = address;
+        }
+    };
     if (!handle_description->pins) {
         // If we're in the unmap queue we can just remove ourselves and return since we're already
         // mapped
@@ -151,37 +185,58 @@ u32 NvMap::PinHandle(NvMap::Handle::Id handle) {
                 unmap_queue.erase(*handle_description->unmap_queue_entry);
                 handle_description->unmap_queue_entry.reset();
 
+                if (low_area_pin) {
+                    map_low_area();
+                    handle_description->pins++;
+                    return static_cast<DAddr>(handle_description->pin_virt_address);
+                }
+
                 handle_description->pins++;
-                return handle_description->pin_virt_address;
+                return handle_description->d_address;
             }
         }
 
+        using namespace std::placeholders;
         // If not then allocate some space and map it
-        u32 address{};
+        DAddr address{};
-        auto& smmu_allocator = host1x.Allocator();
+        auto& smmu = host1x.MemoryManager();
-        auto& smmu_memory_manager = host1x.MemoryManager();
+        auto* session = core.GetSession(handle_description->session_id);
-        while ((address = smmu_allocator.Allocate(
+        const VAddr vaddress = handle_description->address;
-                    static_cast<u32>(handle_description->aligned_size))) == 0) {
+        const size_t map_size = handle_description->aligned_size;
-            // Free handles until the allocation succeeds
+        if (session->has_preallocated_area && session->mapper->IsInBounds(vaddress, map_size)) {
-            std::scoped_lock queueLock(unmap_queue_lock);
+            handle_description->d_address = session->mapper->Map(vaddress, map_size);
-            if (auto freeHandleDesc{unmap_queue.front()}) {
+            handle_description->in_heap = true;
-                // Handles in the unmap queue are guaranteed not to be pinned so don't bother
+        } else {
-                // checking if they are before unmapping
+            size_t aligned_up = Common::AlignUp(map_size, BIG_PAGE_SIZE);
-                std::scoped_lock freeLock(freeHandleDesc->mutex);
+            while ((address = smmu.Allocate(aligned_up)) == 0) {
-                if (handle_description->pin_virt_address)
+                // Free handles until the allocation succeeds
-                    UnmapHandle(*freeHandleDesc);
+                std::scoped_lock queueLock(unmap_queue_lock);
-            } else {
+                if (auto freeHandleDesc{unmap_queue.front()}) {
-                LOG_CRITICAL(Service_NVDRV, "Ran out of SMMU address space!");
+                    // Handles in the unmap queue are guaranteed not to be pinned so don't bother
+                    // checking if they are before unmapping
+                    std::scoped_lock freeLock(freeHandleDesc->mutex);
+                    if (handle_description->d_address)
+                        UnmapHandle(*freeHandleDesc);
+                } else {
+                    LOG_CRITICAL(Service_NVDRV, "Ran out of SMMU address space!");
+                }
             }
-        }
 
-        smmu_memory_manager.Map(static_cast<GPUVAddr>(address), handle_description->address,
+            handle_description->d_address = address;
-                                handle_description->aligned_size);
+            smmu.Map(address, vaddress, map_size, session->smmu_id, true);
-        handle_description->pin_virt_address = address;
+            handle_description->in_heap = false;
+        }
+    }
+
+    if (low_area_pin) {
+        map_low_area();
     }
 
     handle_description->pins++;
-    return handle_description->pin_virt_address;
+    if (low_area_pin) {
+        return static_cast<DAddr>(handle_description->pin_virt_address);
+    }
+    return handle_description->d_address;
 }
 
 void NvMap::UnpinHandle(Handle::Id handle) {
@@ -232,7 +287,7 @@ std::optional<NvMap::FreeInfo> NvMap::FreeHandle(Handle::Id handle, bool interna
                 LOG_WARNING(Service_NVDRV, "User duplicate count imbalance detected!");
             } else if (handle_description->dupes == 0) {
                 // Force unmap the handle
-                if (handle_description->pin_virt_address) {
+                if (handle_description->d_address) {
                     std::scoped_lock queueLock(unmap_queue_lock);
                     UnmapHandle(*handle_description);
                 }

src/core/hle/service/nvdrv/core/nvmap.h

@@ -25,6 +25,8 @@ class Host1x;
 } // namespace Tegra
 
 namespace Service::Nvidia::NvCore {
+
+class Container;
 /**
  * @brief The nvmap core class holds the global state for nvmap and provides methods to manage
  * handles
@@ -48,7 +50,7 @@ public:
         using Id = u32;
         Id id; //!< A globally unique identifier for this handle
 
-        s32 pins{};
+        s64 pins{};
         u32 pin_virt_address{};
         std::optional<typename std::list<std::shared_ptr<Handle>>::iterator> unmap_queue_entry{};
 
@@ -61,15 +63,18 @@ public:
         } flags{};
         static_assert(sizeof(Flags) == sizeof(u32));
 
-        u64 address{}; //!< The memory location in the guest's AS that this handle corresponds to,
+        VAddr address{}; //!< The memory location in the guest's AS that this handle corresponds to,
-                       //!< this can also be in the nvdrv tmem
+                         //!< this can also be in the nvdrv tmem
         bool is_shared_mem_mapped{}; //!< If this nvmap has been mapped with the MapSharedMem IPC
                                      //!< call
 
         u8 kind{};        //!< Used for memory compression
         bool allocated{}; //!< If the handle has been allocated with `Alloc`
+        bool in_heap{};
+        size_t session_id{};
 
-        u64 dma_map_addr{}; //! remove me after implementing pinning.
+        DAddr d_address{}; //!< The memory location in the device's AS that this handle corresponds
+                           //!< to, this can also be in the nvdrv tmem
 
         Handle(u64 size, Id id);
 
@@ -77,7 +82,8 @@ public:
          * @brief Sets up the handle with the given memory config, can allocate memory from the tmem
          * if a 0 address is passed
          */
-        [[nodiscard]] NvResult Alloc(Flags pFlags, u32 pAlign, u8 pKind, u64 pAddress);
+        [[nodiscard]] NvResult Alloc(Flags pFlags, u32 pAlign, u8 pKind, u64 pAddress,
+                                     size_t pSessionId);
 
         /**
          * @brief Increases the dupe counter of the handle for the given session
@@ -108,7 +114,7 @@ public:
         bool can_unlock;   //!< If the address region is ready to be unlocked
     };
 
-    explicit NvMap(Tegra::Host1x::Host1x& host1x);
+    explicit NvMap(Container& core, Tegra::Host1x::Host1x& host1x);
 
     /**
      * @brief Creates an unallocated handle of the given size
@@ -117,7 +123,7 @@ public:
 
     std::shared_ptr<Handle> GetHandle(Handle::Id handle);
 
-    VAddr GetHandleAddress(Handle::Id handle);
+    DAddr GetHandleAddress(Handle::Id handle);
 
     /**
      * @brief Maps a handle into the SMMU address space
@@ -125,7 +131,7 @@ public:
      * number of calls to `UnpinHandle`
      * @return The SMMU virtual address that the handle has been mapped to
      */
-    u32 PinHandle(Handle::Id handle);
+    DAddr PinHandle(Handle::Id handle, bool low_area_pin);
 
     /**
      * @brief When this has been called an equal number of times to `PinHandle` for the supplied
@@ -172,5 +178,7 @@ private:
      * @return If the handle was removed from the map
      */
     bool TryRemoveHandle(const Handle& handle_description);
+
+    Container& core;
 };
 } // namespace Service::Nvidia::NvCore

src/core/hle/service/nvdrv/devices/nvdevice.h

@@ -62,7 +62,7 @@ public:
      * Called once a device is opened
      * @param fd The device fd
      */
-    virtual void OnOpen(DeviceFD fd) = 0;
+    virtual void OnOpen(size_t session_id, DeviceFD fd) = 0;
 
     /**
      * Called once a device is closed

src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp

@@ -35,14 +35,14 @@ NvResult nvdisp_disp0::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> in
     return NvResult::NotImplemented;
 }
 
-void nvdisp_disp0::OnOpen(DeviceFD fd) {}
+void nvdisp_disp0::OnOpen(size_t session_id, DeviceFD fd) {}
 void nvdisp_disp0::OnClose(DeviceFD fd) {}
 
 void nvdisp_disp0::flip(u32 buffer_handle, u32 offset, android::PixelFormat format, u32 width,
                         u32 height, u32 stride, android::BufferTransformFlags transform,
                         const Common::Rectangle<int>& crop_rect,
                         std::array<Service::Nvidia::NvFence, 4>& fences, u32 num_fences) {
-    const VAddr addr = nvmap.GetHandleAddress(buffer_handle);
+    const DAddr addr = nvmap.GetHandleAddress(buffer_handle);
     LOG_TRACE(Service,
               "Drawing from address {:X} offset {:08X} Width {} Height {} Stride {} Format {}",
               addr, offset, width, height, stride, format);

src/core/hle/service/nvdrv/devices/nvdisp_disp0.h

@@ -32,7 +32,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
     /// Performs a screen flip, drawing the buffer pointed to by the handle.

src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp

@@ -86,7 +86,7 @@ NvResult nvhost_as_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> i
     return NvResult::NotImplemented;
 }
 
-void nvhost_as_gpu::OnOpen(DeviceFD fd) {}
+void nvhost_as_gpu::OnOpen(size_t session_id, DeviceFD fd) {}
 void nvhost_as_gpu::OnClose(DeviceFD fd) {}
 
 NvResult nvhost_as_gpu::AllocAsEx(IoctlAllocAsEx& params) {
@@ -206,6 +206,8 @@ void nvhost_as_gpu::FreeMappingLocked(u64 offset) {
                        static_cast<u32>(aligned_size >> page_size_bits));
     }
 
+    nvmap.UnpinHandle(mapping->handle);
+
     // Sparse mappings shouldn't be fully unmapped, just returned to their sparse state
     // Only FreeSpace can unmap them fully
     if (mapping->sparse_alloc) {
@@ -293,12 +295,12 @@ NvResult nvhost_as_gpu::Remap(std::span<IoctlRemapEntry> entries) {
                 return NvResult::BadValue;
             }
 
-            VAddr cpu_address{static_cast<VAddr>(
+            DAddr base = nvmap.PinHandle(entry.handle, false);
-                handle->address +
+            DAddr device_address{static_cast<DAddr>(
-                (static_cast<u64>(entry.handle_offset_big_pages) << vm.big_page_size_bits))};
+                base + (static_cast<u64>(entry.handle_offset_big_pages) << vm.big_page_size_bits))};
 
-            gmmu->Map(virtual_address, cpu_address, size, static_cast<Tegra::PTEKind>(entry.kind),
+            gmmu->Map(virtual_address, device_address, size,
-                      use_big_pages);
+                      static_cast<Tegra::PTEKind>(entry.kind), use_big_pages);
         }
     }
 
@@ -331,9 +333,9 @@ NvResult nvhost_as_gpu::MapBufferEx(IoctlMapBufferEx& params) {
             }
 
             u64 gpu_address{static_cast<u64>(params.offset + params.buffer_offset)};
-            VAddr cpu_address{mapping->ptr + params.buffer_offset};
+            VAddr device_address{mapping->ptr + params.buffer_offset};
 
-            gmmu->Map(gpu_address, cpu_address, params.mapping_size,
+            gmmu->Map(gpu_address, device_address, params.mapping_size,
                       static_cast<Tegra::PTEKind>(params.kind), mapping->big_page);
 
             return NvResult::Success;
@@ -349,7 +351,8 @@ NvResult nvhost_as_gpu::MapBufferEx(IoctlMapBufferEx& params) {
         return NvResult::BadValue;
     }
 
-    VAddr cpu_address{static_cast<VAddr>(handle->address + params.buffer_offset)};
+    DAddr device_address{
+        static_cast<DAddr>(nvmap.PinHandle(params.handle, false) + params.buffer_offset)};
     u64 size{params.mapping_size ? params.mapping_size : handle->orig_size};
 
     bool big_page{[&]() {
@@ -373,15 +376,14 @@ NvResult nvhost_as_gpu::MapBufferEx(IoctlMapBufferEx& params) {
         }
 
         const bool use_big_pages = alloc->second.big_pages && big_page;
-        gmmu->Map(params.offset, cpu_address, size, static_cast<Tegra::PTEKind>(params.kind),
+        gmmu->Map(params.offset, device_address, size, static_cast<Tegra::PTEKind>(params.kind),
                   use_big_pages);
 
-        auto mapping{std::make_shared<Mapping>(cpu_address, params.offset, size, true,
+        auto mapping{std::make_shared<Mapping>(params.handle, device_address, params.offset, size,
-                                               use_big_pages, alloc->second.sparse)};
+                                               true, use_big_pages, alloc->second.sparse)};
         alloc->second.mappings.push_back(mapping);
         mapping_map[params.offset] = mapping;
     } else {
-
         auto& allocator{big_page ? *vm.big_page_allocator : *vm.small_page_allocator};
         u32 page_size{big_page ? vm.big_page_size : VM::YUZU_PAGESIZE};
         u32 page_size_bits{big_page ? vm.big_page_size_bits : VM::PAGE_SIZE_BITS};
@@ -394,11 +396,11 @@ NvResult nvhost_as_gpu::MapBufferEx(IoctlMapBufferEx& params) {
             return NvResult::InsufficientMemory;
         }
 
-        gmmu->Map(params.offset, cpu_address, Common::AlignUp(size, page_size),
+        gmmu->Map(params.offset, device_address, Common::AlignUp(size, page_size),
                   static_cast<Tegra::PTEKind>(params.kind), big_page);
 
-        auto mapping{
+        auto mapping{std::make_shared<Mapping>(params.handle, device_address, params.offset, size,
-            std::make_shared<Mapping>(cpu_address, params.offset, size, false, big_page, false)};
+                                               false, big_page, false)};
         mapping_map[params.offset] = mapping;
     }
 
@@ -433,6 +435,8 @@ NvResult nvhost_as_gpu::UnmapBuffer(IoctlUnmapBuffer& params) {
             gmmu->Unmap(params.offset, mapping->size);
         }
 
+        nvmap.UnpinHandle(mapping->handle);
+
         mapping_map.erase(params.offset);
     } catch (const std::out_of_range&) {
         LOG_WARNING(Service_NVDRV, "Couldn't find region to unmap at 0x{:X}", params.offset);

src/core/hle/service/nvdrv/devices/nvhost_as_gpu.h

@@ -55,7 +55,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;
@@ -159,16 +159,18 @@ private:
     NvCore::NvMap& nvmap;
 
     struct Mapping {
-        VAddr ptr;
+        NvCore::NvMap::Handle::Id handle;
+        DAddr ptr;
         u64 offset;
         u64 size;
         bool fixed;
         bool big_page; // Only valid if fixed == false
         bool sparse_alloc;
 
-        Mapping(VAddr ptr_, u64 offset_, u64 size_, bool fixed_, bool big_page_, bool sparse_alloc_)
+        Mapping(NvCore::NvMap::Handle::Id handle_, DAddr ptr_, u64 offset_, u64 size_, bool fixed_,
-            : ptr(ptr_), offset(offset_), size(size_), fixed(fixed_), big_page(big_page_),
+                bool big_page_, bool sparse_alloc_)
-              sparse_alloc(sparse_alloc_) {}
+            : handle(handle_), ptr(ptr_), offset(offset_), size(size_), fixed(fixed_),
+              big_page(big_page_), sparse_alloc(sparse_alloc_) {}
     };
 
     struct Allocation {
@@ -212,9 +214,6 @@ private:
         bool initialised{};
     } vm;
     std::shared_ptr<Tegra::MemoryManager> gmmu;
-
-    // s32 channel{};
-    // u32 big_page_size{VM::DEFAULT_BIG_PAGE_SIZE};
 };
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_ctrl.cpp

@@ -76,7 +76,7 @@ NvResult nvhost_ctrl::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> inp
     return NvResult::NotImplemented;
 }
 
-void nvhost_ctrl::OnOpen(DeviceFD fd) {}
+void nvhost_ctrl::OnOpen(size_t session_id, DeviceFD fd) {}
 
 void nvhost_ctrl::OnClose(DeviceFD fd) {}
 

src/core/hle/service/nvdrv/devices/nvhost_ctrl.h

@@ -32,7 +32,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;

src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.cpp

@@ -82,7 +82,7 @@ NvResult nvhost_ctrl_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8>
     return NvResult::NotImplemented;
 }
 
-void nvhost_ctrl_gpu::OnOpen(DeviceFD fd) {}
+void nvhost_ctrl_gpu::OnOpen(size_t session_id, DeviceFD fd) {}
 void nvhost_ctrl_gpu::OnClose(DeviceFD fd) {}
 
 NvResult nvhost_ctrl_gpu::GetCharacteristics1(IoctlCharacteristics& params) {

src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.h

@@ -28,7 +28,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;

src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp

@@ -120,7 +120,7 @@ NvResult nvhost_gpu::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> inpu
     return NvResult::NotImplemented;
 }
 
-void nvhost_gpu::OnOpen(DeviceFD fd) {}
+void nvhost_gpu::OnOpen(size_t session_id, DeviceFD fd) {}
 void nvhost_gpu::OnClose(DeviceFD fd) {}
 
 NvResult nvhost_gpu::SetNVMAPfd(IoctlSetNvmapFD& params) {

src/core/hle/service/nvdrv/devices/nvhost_gpu.h

@@ -47,7 +47,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;

src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp

@@ -35,7 +35,7 @@ NvResult nvhost_nvdec::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> in
         case 0x7:
             return WrapFixed(this, &nvhost_nvdec::SetSubmitTimeout, input, output);
         case 0x9:
-            return WrapFixedVariable(this, &nvhost_nvdec::MapBuffer, input, output);
+            return WrapFixedVariable(this, &nvhost_nvdec::MapBuffer, input, output, fd);
         case 0xa:
             return WrapFixedVariable(this, &nvhost_nvdec::UnmapBuffer, input, output);
         default:
@@ -68,9 +68,10 @@ NvResult nvhost_nvdec::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> in
     return NvResult::NotImplemented;
 }
 
-void nvhost_nvdec::OnOpen(DeviceFD fd) {
+void nvhost_nvdec::OnOpen(size_t session_id, DeviceFD fd) {
     LOG_INFO(Service_NVDRV, "NVDEC video stream started");
     system.SetNVDECActive(true);
+    sessions[fd] = session_id;
 }
 
 void nvhost_nvdec::OnClose(DeviceFD fd) {
@@ -81,6 +82,10 @@ void nvhost_nvdec::OnClose(DeviceFD fd) {
         system.GPU().ClearCdmaInstance(iter->second);
     }
     system.SetNVDECActive(false);
+    auto it = sessions.find(fd);
+    if (it != sessions.end()) {
+        sessions.erase(it);
+    }
 }
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_nvdec.h

@@ -20,7 +20,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 };
 

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp

@@ -8,6 +8,7 @@
 #include "common/common_types.h"
 #include "common/logging/log.h"
 #include "core/core.h"
+#include "core/hle/kernel/k_process.h"
 #include "core/hle/service/nvdrv/core/container.h"
 #include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/hle/service/nvdrv/core/syncpoint_manager.h"
@@ -95,6 +96,8 @@ NvResult nvhost_nvdec_common::Submit(IoctlSubmit& params, std::span<u8> data, De
     offset += SliceVectors(data, fence_thresholds, params.fence_count, offset);
 
     auto& gpu = system.GPU();
+    auto* session = core.GetSession(sessions[fd]);
+
     if (gpu.UseNvdec()) {
         for (std::size_t i = 0; i < syncpt_increments.size(); i++) {
             const SyncptIncr& syncpt_incr = syncpt_increments[i];
@@ -106,8 +109,8 @@ NvResult nvhost_nvdec_common::Submit(IoctlSubmit& params, std::span<u8> data, De
         const auto object = nvmap.GetHandle(cmd_buffer.memory_id);
         ASSERT_OR_EXECUTE(object, return NvResult::InvalidState;);
         Tegra::ChCommandHeaderList cmdlist(cmd_buffer.word_count);
-        system.ApplicationMemory().ReadBlock(object->address + cmd_buffer.offset, cmdlist.data(),
+        session->process->GetMemory().ReadBlock(object->address + cmd_buffer.offset, cmdlist.data(),
-                                             cmdlist.size() * sizeof(u32));
+                                                cmdlist.size() * sizeof(u32));
         gpu.PushCommandBuffer(core.Host1xDeviceFile().fd_to_id[fd], cmdlist);
     }
     // Some games expect command_buffers to be written back
@@ -133,10 +136,12 @@ NvResult nvhost_nvdec_common::GetWaitbase(IoctlGetWaitbase& params) {
     return NvResult::Success;
 }
 
-NvResult nvhost_nvdec_common::MapBuffer(IoctlMapBuffer& params, std::span<MapBufferEntry> entries) {
+NvResult nvhost_nvdec_common::MapBuffer(IoctlMapBuffer& params, std::span<MapBufferEntry> entries,
+                                        DeviceFD fd) {
     const size_t num_entries = std::min(params.num_entries, static_cast<u32>(entries.size()));
     for (size_t i = 0; i < num_entries; i++) {
-        entries[i].map_address = nvmap.PinHandle(entries[i].map_handle);
+        DAddr pin_address = nvmap.PinHandle(entries[i].map_handle, true);
+        entries[i].map_address = static_cast<u32>(pin_address);
     }
 
     return NvResult::Success;

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h

@@ -4,7 +4,9 @@
 #pragma once
 
 #include <deque>
+#include <unordered_map>
 #include <vector>
+
 #include "common/common_types.h"
 #include "common/swap.h"
 #include "core/hle/service/nvdrv/core/syncpoint_manager.h"
@@ -111,7 +113,7 @@ protected:
     NvResult Submit(IoctlSubmit& params, std::span<u8> input, DeviceFD fd);
     NvResult GetSyncpoint(IoctlGetSyncpoint& params);
     NvResult GetWaitbase(IoctlGetWaitbase& params);
-    NvResult MapBuffer(IoctlMapBuffer& params, std::span<MapBufferEntry> entries);
+    NvResult MapBuffer(IoctlMapBuffer& params, std::span<MapBufferEntry> entries, DeviceFD fd);
     NvResult UnmapBuffer(IoctlMapBuffer& params, std::span<MapBufferEntry> entries);
     NvResult SetSubmitTimeout(u32 timeout);
 
@@ -125,6 +127,7 @@ protected:
     NvCore::NvMap& nvmap;
     NvCore::ChannelType channel_type;
     std::array<u32, MaxSyncPoints> device_syncpoints{};
+    std::unordered_map<DeviceFD, size_t> sessions;
 };
 }; // namespace Devices
 } // namespace Service::Nvidia

src/core/hle/service/nvdrv/devices/nvhost_nvjpg.cpp

@@ -44,7 +44,7 @@ NvResult nvhost_nvjpg::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> in
     return NvResult::NotImplemented;
 }
 
-void nvhost_nvjpg::OnOpen(DeviceFD fd) {}
+void nvhost_nvjpg::OnOpen(size_t session_id, DeviceFD fd) {}
 void nvhost_nvjpg::OnClose(DeviceFD fd) {}
 
 NvResult nvhost_nvjpg::SetNVMAPfd(IoctlSetNvmapFD& params) {

src/core/hle/service/nvdrv/devices/nvhost_nvjpg.h

@@ -22,7 +22,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
 private:

src/core/hle/service/nvdrv/devices/nvhost_vic.cpp

@@ -33,7 +33,7 @@ NvResult nvhost_vic::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> inpu
         case 0x3:
             return WrapFixed(this, &nvhost_vic::GetWaitbase, input, output);
         case 0x9:
-            return WrapFixedVariable(this, &nvhost_vic::MapBuffer, input, output);
+            return WrapFixedVariable(this, &nvhost_vic::MapBuffer, input, output, fd);
         case 0xa:
             return WrapFixedVariable(this, &nvhost_vic::UnmapBuffer, input, output);
         default:
@@ -68,7 +68,9 @@ NvResult nvhost_vic::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> inpu
     return NvResult::NotImplemented;
 }
 
-void nvhost_vic::OnOpen(DeviceFD fd) {}
+void nvhost_vic::OnOpen(size_t session_id, DeviceFD fd) {
+    sessions[fd] = session_id;
+}
 
 void nvhost_vic::OnClose(DeviceFD fd) {
     auto& host1x_file = core.Host1xDeviceFile();
@@ -76,6 +78,10 @@ void nvhost_vic::OnClose(DeviceFD fd) {
     if (iter != host1x_file.fd_to_id.end()) {
         system.GPU().ClearCdmaInstance(iter->second);
     }
+    auto it = sessions.find(fd);
+    if (it != sessions.end()) {
+        sessions.erase(it);
+    }
 }
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_vic.h

@@ -19,7 +19,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 };
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvmap.cpp

@@ -36,9 +36,9 @@ NvResult nvmap::Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input,
         case 0x3:
             return WrapFixed(this, &nvmap::IocFromId, input, output);
         case 0x4:
-            return WrapFixed(this, &nvmap::IocAlloc, input, output);
+            return WrapFixed(this, &nvmap::IocAlloc, input, output, fd);
         case 0x5:
-            return WrapFixed(this, &nvmap::IocFree, input, output);
+            return WrapFixed(this, &nvmap::IocFree, input, output, fd);
         case 0x9:
             return WrapFixed(this, &nvmap::IocParam, input, output);
         case 0xe:
@@ -67,8 +67,15 @@ NvResult nvmap::Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, st
     return NvResult::NotImplemented;
 }
 
-void nvmap::OnOpen(DeviceFD fd) {}
+void nvmap::OnOpen(size_t session_id, DeviceFD fd) {
-void nvmap::OnClose(DeviceFD fd) {}
+    sessions[fd] = session_id;
+}
+void nvmap::OnClose(DeviceFD fd) {
+    auto it = sessions.find(fd);
+    if (it != sessions.end()) {
+        sessions.erase(it);
+    }
+}
 
 NvResult nvmap::IocCreate(IocCreateParams& params) {
     LOG_DEBUG(Service_NVDRV, "called, size=0x{:08X}", params.size);
@@ -87,7 +94,7 @@ NvResult nvmap::IocCreate(IocCreateParams& params) {
     return NvResult::Success;
 }
 
-NvResult nvmap::IocAlloc(IocAllocParams& params) {
+NvResult nvmap::IocAlloc(IocAllocParams& params, DeviceFD fd) {
     LOG_DEBUG(Service_NVDRV, "called, addr={:X}", params.address);
 
     if (!params.handle) {
@@ -116,15 +123,15 @@ NvResult nvmap::IocAlloc(IocAllocParams& params) {
         return NvResult::InsufficientMemory;
     }
 
-    const auto result =
+    const auto result = handle_description->Alloc(params.flags, params.align, params.kind,
-        handle_description->Alloc(params.flags, params.align, params.kind, params.address);
+                                                  params.address, sessions[fd]);
     if (result != NvResult::Success) {
         LOG_CRITICAL(Service_NVDRV, "Object failed to allocate, handle={:08X}", params.handle);
         return result;
     }
     bool is_out_io{};
-    ASSERT(system.ApplicationProcess()
+    auto process = container.GetSession(sessions[fd])->process;
-               ->GetPageTable()
+    ASSERT(process->GetPageTable()
                .LockForMapDeviceAddressSpace(&is_out_io, handle_description->address,
                                              handle_description->size,
                                              Kernel::KMemoryPermission::None, true, false)
@@ -224,7 +231,7 @@ NvResult nvmap::IocParam(IocParamParams& params) {
     return NvResult::Success;
 }
 
-NvResult nvmap::IocFree(IocFreeParams& params) {
+NvResult nvmap::IocFree(IocFreeParams& params, DeviceFD fd) {
     LOG_DEBUG(Service_NVDRV, "called");
 
     if (!params.handle) {
@@ -233,9 +240,9 @@ NvResult nvmap::IocFree(IocFreeParams& params) {
     }
 
     if (auto freeInfo{file.FreeHandle(params.handle, false)}) {
+        auto process = container.GetSession(sessions[fd])->process;
         if (freeInfo->can_unlock) {
-            ASSERT(system.ApplicationProcess()
+            ASSERT(process->GetPageTable()
-                       ->GetPageTable()
                        .UnlockForDeviceAddressSpace(freeInfo->address, freeInfo->size)
                        .IsSuccess());
         }

src/core/hle/service/nvdrv/devices/nvmap.h

@@ -33,7 +33,7 @@ public:
     NvResult Ioctl3(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output,
                     std::span<u8> inline_output) override;
 
-    void OnOpen(DeviceFD fd) override;
+    void OnOpen(size_t session_id, DeviceFD fd) override;
     void OnClose(DeviceFD fd) override;
 
     enum class HandleParameterType : u32_le {
@@ -100,11 +100,11 @@ public:
     static_assert(sizeof(IocGetIdParams) == 8, "IocGetIdParams has wrong size");
 
     NvResult IocCreate(IocCreateParams& params);
-    NvResult IocAlloc(IocAllocParams& params);
+    NvResult IocAlloc(IocAllocParams& params, DeviceFD fd);
     NvResult IocGetId(IocGetIdParams& params);
     NvResult IocFromId(IocFromIdParams& params);
     NvResult IocParam(IocParamParams& params);
-    NvResult IocFree(IocFreeParams& params);
+    NvResult IocFree(IocFreeParams& params, DeviceFD fd);
 
 private:
     /// Id to use for the next handle that is created.
@@ -115,6 +115,7 @@ private:
 
     NvCore::Container& container;
     NvCore::NvMap& file;
+    std::unordered_map<DeviceFD, size_t> sessions;
 };
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/nvdrv.cpp

@@ -45,13 +45,22 @@ void EventInterface::FreeEvent(Kernel::KEvent* event) {
 void LoopProcess(Nvnflinger::Nvnflinger& nvnflinger, Core::System& system) {
     auto server_manager = std::make_unique<ServerManager>(system);
     auto module = std::make_shared<Module>(system);
-    server_manager->RegisterNamedService("nvdrv", std::make_shared<NVDRV>(system, module, "nvdrv"));
+    const auto NvdrvInterfaceFactoryForApplication = [&, module] {
-    server_manager->RegisterNamedService("nvdrv:a",
+        return std::make_shared<NVDRV>(system, module, "nvdrv");
-                                         std::make_shared<NVDRV>(system, module, "nvdrv:a"));
+    };
-    server_manager->RegisterNamedService("nvdrv:s",
+    const auto NvdrvInterfaceFactoryForApplets = [&, module] {
-                                         std::make_shared<NVDRV>(system, module, "nvdrv:s"));
+        return std::make_shared<NVDRV>(system, module, "nvdrv:a");
-    server_manager->RegisterNamedService("nvdrv:t",
+    };
-                                         std::make_shared<NVDRV>(system, module, "nvdrv:t"));
+    const auto NvdrvInterfaceFactoryForSysmodules = [&, module] {
+        return std::make_shared<NVDRV>(system, module, "nvdrv:a");
+    };
+    const auto NvdrvInterfaceFactory = [&, module] {
+        return std::make_shared<NVDRV>(system, module, "nvdrv:t");
+    };
+    server_manager->RegisterNamedService("nvdrv", NvdrvInterfaceFactoryForApplication);
+    server_manager->RegisterNamedService("nvdrv:a", NvdrvInterfaceFactoryForApplets);
+    server_manager->RegisterNamedService("nvdrv:s", NvdrvInterfaceFactoryForSysmodules);
+    server_manager->RegisterNamedService("nvdrv:t", NvdrvInterfaceFactory);
     server_manager->RegisterNamedService("nvmemp", std::make_shared<NVMEMP>(system));
     nvnflinger.SetNVDrvInstance(module);
     ServerManager::RunServer(std::move(server_manager));
@@ -113,7 +122,7 @@ NvResult Module::VerifyFD(DeviceFD fd) const {
     return NvResult::Success;
 }
 
-DeviceFD Module::Open(const std::string& device_name) {
+DeviceFD Module::Open(const std::string& device_name, size_t session_id) {
     auto it = builders.find(device_name);
     if (it == builders.end()) {
         LOG_ERROR(Service_NVDRV, "Trying to open unknown device {}", device_name);
@@ -124,7 +133,7 @@ DeviceFD Module::Open(const std::string& device_name) {
     auto& builder = it->second;
     auto device = builder(fd)->second;
 
-    device->OnOpen(fd);
+    device->OnOpen(session_id, fd);
 
     return fd;
 }

src/core/hle/service/nvdrv/nvdrv.h

@@ -77,7 +77,7 @@ public:
     NvResult VerifyFD(DeviceFD fd) const;
 
     /// Opens a device node and returns a file descriptor to it.
-    DeviceFD Open(const std::string& device_name);
+    DeviceFD Open(const std::string& device_name, size_t session_id);
 
     /// Sends an ioctl command to the specified file descriptor.
     NvResult Ioctl1(DeviceFD fd, Ioctl command, std::span<const u8> input, std::span<u8> output);
@@ -93,6 +93,10 @@ public:
 
     NvResult QueryEvent(DeviceFD fd, u32 event_id, Kernel::KEvent*& event);
 
+    NvCore::Container& GetContainer() {
+        return container;
+    }
+
 private:
     friend class EventInterface;
     friend class Service::Nvnflinger::Nvnflinger;

src/core/hle/service/nvdrv/nvdrv_interface.cpp

@@ -3,8 +3,10 @@
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 #include "common/logging/log.h"
+#include "common/scope_exit.h"
 #include "core/core.h"
 #include "core/hle/kernel/k_event.h"
+#include "core/hle/kernel/k_process.h"
 #include "core/hle/kernel/k_readable_event.h"
 #include "core/hle/service/ipc_helpers.h"
 #include "core/hle/service/nvdrv/nvdata.h"
@@ -37,7 +39,7 @@ void NVDRV::Open(HLERequestContext& ctx) {
         return;
     }
 
-    DeviceFD fd = nvdrv->Open(device_name);
+    DeviceFD fd = nvdrv->Open(device_name, session_id);
 
     rb.Push<DeviceFD>(fd);
     rb.PushEnum(fd != INVALID_NVDRV_FD ? NvResult::Success : NvResult::FileOperationFailed);
@@ -150,12 +152,29 @@ void NVDRV::Close(HLERequestContext& ctx) {
 
 void NVDRV::Initialize(HLERequestContext& ctx) {
     LOG_WARNING(Service_NVDRV, "(STUBBED) called");
+    IPC::ResponseBuilder rb{ctx, 3};
+    SCOPE_EXIT({
+        rb.Push(ResultSuccess);
+        rb.PushEnum(NvResult::Success);
+    });
+
+    if (is_initialized) {
+        // No need to initialize again
+        return;
+    }
+
+    IPC::RequestParser rp{ctx};
+    const auto process_handle{ctx.GetCopyHandle(0)};
+    // The transfer memory is lent to nvdrv as a work buffer since nvdrv is
+    // unable to allocate as much memory on its own. For HLE it's unnecessary to handle it
+    [[maybe_unused]] const auto transfer_memory_handle{ctx.GetCopyHandle(1)};
+    [[maybe_unused]] const auto transfer_memory_size = rp.Pop<u32>();
+
+    auto& container = nvdrv->GetContainer();
+    auto process = ctx.GetObjectFromHandle<Kernel::KProcess>(process_handle);
+    session_id = container.OpenSession(process.GetPointerUnsafe());
 
     is_initialized = true;
-
-    IPC::ResponseBuilder rb{ctx, 3};
-    rb.Push(ResultSuccess);
-    rb.PushEnum(NvResult::Success);
 }
 
 void NVDRV::QueryEvent(HLERequestContext& ctx) {
@@ -242,6 +261,9 @@ NVDRV::NVDRV(Core::System& system_, std::shared_ptr<Module> nvdrv_, const char*
     RegisterHandlers(functions);
 }
 
-NVDRV::~NVDRV() = default;
+NVDRV::~NVDRV() {
+    auto& container = nvdrv->GetContainer();
+    container.CloseSession(session_id);
+}
 
 } // namespace Service::Nvidia

src/core/hle/service/nvdrv/nvdrv_interface.h

@@ -35,6 +35,7 @@ private:
 
     u64 pid{};
     bool is_initialized{};
+    size_t session_id{};
     Common::ScratchBuffer<u8> output_buffer;
     Common::ScratchBuffer<u8> inline_output_buffer;
 };

src/core/hle/service/nvnflinger/fb_share_buffer_manager.cpp

@@ -87,19 +87,20 @@ Result CreateNvMapHandle(u32* out_nv_map_handle, Nvidia::Devices::nvmap& nvmap,
     R_SUCCEED();
 }
 
-Result FreeNvMapHandle(Nvidia::Devices::nvmap& nvmap, u32 handle) {
+Result FreeNvMapHandle(Nvidia::Devices::nvmap& nvmap, u32 handle, Nvidia::DeviceFD nvmap_fd) {
     // Free the handle.
     Nvidia::Devices::nvmap::IocFreeParams free_params{
         .handle = handle,
     };
-    R_UNLESS(nvmap.IocFree(free_params) == Nvidia::NvResult::Success, VI::ResultOperationFailed);
+    R_UNLESS(nvmap.IocFree(free_params, nvmap_fd) == Nvidia::NvResult::Success,
+             VI::ResultOperationFailed);
 
     // We succeeded.
     R_SUCCEED();
 }
 
 Result AllocNvMapHandle(Nvidia::Devices::nvmap& nvmap, u32 handle, Common::ProcessAddress buffer,
-                        u32 size) {
+                        u32 size, Nvidia::DeviceFD nvmap_fd) {
     // Assign the allocated memory to the handle.
     Nvidia::Devices::nvmap::IocAllocParams alloc_params{
         .handle = handle,
@@ -109,16 +110,16 @@ Result AllocNvMapHandle(Nvidia::Devices::nvmap& nvmap, u32 handle, Common::Proce
         .kind = 0,
         .address = GetInteger(buffer),
     };
-    R_UNLESS(nvmap.IocAlloc(alloc_params) == Nvidia::NvResult::Success, VI::ResultOperationFailed);
+    R_UNLESS(nvmap.IocAlloc(alloc_params, nvmap_fd) == Nvidia::NvResult::Success,
+             VI::ResultOperationFailed);
 
     // We succeeded.
     R_SUCCEED();
 }
 
-Result AllocateHandleForBuffer(u32* out_handle, Nvidia::Module& nvdrv,
+Result AllocateHandleForBuffer(u32* out_handle, Nvidia::Module& nvdrv, Nvidia::DeviceFD nvmap_fd,
                                Common::ProcessAddress buffer, u32 size) {
     // Get the nvmap device.
-    auto nvmap_fd = nvdrv.Open("/dev/nvmap");
     auto nvmap = nvdrv.GetDevice<Nvidia::Devices::nvmap>(nvmap_fd);
     ASSERT(nvmap != nullptr);
 
@@ -127,11 +128,11 @@ Result AllocateHandleForBuffer(u32* out_handle, Nvidia::Module& nvdrv,
 
     // Ensure we maintain a clean state on failure.
     ON_RESULT_FAILURE {
-        ASSERT(R_SUCCEEDED(FreeNvMapHandle(*nvmap, *out_handle)));
+        ASSERT(R_SUCCEEDED(FreeNvMapHandle(*nvmap, *out_handle, nvmap_fd)));
     };
 
     // Assign the allocated memory to the handle.
-    R_RETURN(AllocNvMapHandle(*nvmap, *out_handle, buffer, size));
+    R_RETURN(AllocNvMapHandle(*nvmap, *out_handle, buffer, size, nvmap_fd));
 }
 
 constexpr auto SharedBufferBlockLinearFormat = android::PixelFormat::Rgba8888;
@@ -197,9 +198,13 @@ Result FbShareBufferManager::Initialize(u64* out_buffer_id, u64* out_layer_id, u
                                            std::addressof(m_buffer_page_group), m_system,
                                            SharedBufferSize));
 
+    auto& container = m_nvdrv->GetContainer();
+    m_session_id = container.OpenSession(m_system.ApplicationProcess());
+    m_nvmap_fd = m_nvdrv->Open("/dev/nvmap", m_session_id);
+
     // Create an nvmap handle for the buffer and assign the memory to it.
-    R_TRY(AllocateHandleForBuffer(std::addressof(m_buffer_nvmap_handle), *m_nvdrv, map_address,
+    R_TRY(AllocateHandleForBuffer(std::addressof(m_buffer_nvmap_handle), *m_nvdrv, m_nvmap_fd,
-                                  SharedBufferSize));
+                                  map_address, SharedBufferSize));
 
     // Record the display id.
     m_display_id = display_id;

src/core/hle/service/nvnflinger/fb_share_buffer_manager.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "common/math_util.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 #include "core/hle/service/nvnflinger/nvnflinger.h"
 #include "core/hle/service/nvnflinger/ui/fence.h"
 
@@ -53,7 +54,8 @@ private:
     u64 m_layer_id = 0;
     u32 m_buffer_nvmap_handle = 0;
     SharedMemoryPoolLayout m_pool_layout = {};
-
+    Nvidia::DeviceFD m_nvmap_fd = {};
+    size_t m_session_id = {};
     std::unique_ptr<Kernel::KPageGroup> m_buffer_page_group;
 
     std::mutex m_guard;

src/core/hle/service/nvnflinger/nvnflinger.cpp

@@ -126,7 +126,7 @@ void Nvnflinger::ShutdownLayers() {
 
 void Nvnflinger::SetNVDrvInstance(std::shared_ptr<Nvidia::Module> instance) {
     nvdrv = std::move(instance);
-    disp_fd = nvdrv->Open("/dev/nvdisp_disp0");
+    disp_fd = nvdrv->Open("/dev/nvdisp_disp0", 0);
 }
 
 std::optional<u64> Nvnflinger::OpenDisplay(std::string_view name) {

src/core/hle/service/nvnflinger/ui/graphic_buffer.cpp

@@ -22,11 +22,13 @@ GraphicBuffer::GraphicBuffer(Service::Nvidia::NvCore::NvMap& nvmap,
     : NvGraphicBuffer(GetBuffer(buffer)), m_nvmap(std::addressof(nvmap)) {
     if (this->BufferId() > 0) {
         m_nvmap->DuplicateHandle(this->BufferId(), true);
+        m_nvmap->PinHandle(this->BufferId(), false);
     }
 }
 
 GraphicBuffer::~GraphicBuffer() {
     if (m_nvmap != nullptr && this->BufferId() > 0) {
+        m_nvmap->UnpinHandle(this->BufferId());
         m_nvmap->FreeHandle(this->BufferId(), true);
     }
 }

src/core/hle/service/pm/pm.cpp

@@ -22,27 +22,26 @@ constexpr Result ResultProcessNotFound{ErrorModule::PM, 1};
 
 constexpr u64 NO_PROCESS_FOUND_PID{0};
 
-std::optional<Kernel::KProcess*> SearchProcessList(
+using ProcessList = std::list<Kernel::KScopedAutoObject<Kernel::KProcess>>;
-    const std::vector<Kernel::KProcess*>& process_list,
+
-    std::function<bool(Kernel::KProcess*)> predicate) {
+template <typename F>
+Kernel::KScopedAutoObject<Kernel::KProcess> SearchProcessList(ProcessList& process_list,
+                                                              F&& predicate) {
     const auto iter = std::find_if(process_list.begin(), process_list.end(), predicate);
 
     if (iter == process_list.end()) {
-        return std::nullopt;
+        return nullptr;
     }
 
-    return *iter;
+    return iter->GetPointerUnsafe();
 }
 
-void GetApplicationPidGeneric(HLERequestContext& ctx,
+void GetApplicationPidGeneric(HLERequestContext& ctx, ProcessList& process_list) {
-                              const std::vector<Kernel::KProcess*>& process_list) {
+    auto process = SearchProcessList(process_list, [](auto& p) { return p->IsApplication(); });
-    const auto process = SearchProcessList(process_list, [](const auto& proc) {
-        return proc->GetProcessId() == Kernel::KProcess::ProcessIdMin;
-    });
 
     IPC::ResponseBuilder rb{ctx, 4};
     rb.Push(ResultSuccess);
-    rb.Push(process.has_value() ? (*process)->GetProcessId() : NO_PROCESS_FOUND_PID);
+    rb.Push(process.IsNull() ? NO_PROCESS_FOUND_PID : process->GetProcessId());
 }
 
 } // Anonymous namespace
@@ -80,8 +79,7 @@ private:
 
 class DebugMonitor final : public ServiceFramework<DebugMonitor> {
 public:
-    explicit DebugMonitor(Core::System& system_)
+    explicit DebugMonitor(Core::System& system_) : ServiceFramework{system_, "pm:dmnt"} {
-        : ServiceFramework{system_, "pm:dmnt"}, kernel{system_.Kernel()} {
         // clang-format off
         static const FunctionInfo functions[] = {
             {0, nullptr, "GetJitDebugProcessIdList"},
@@ -106,12 +104,11 @@ private:
 
         LOG_DEBUG(Service_PM, "called, program_id={:016X}", program_id);
 
-        const auto process =
+        auto list = kernel.GetProcessList();
-            SearchProcessList(kernel.GetProcessList(), [program_id](const auto& proc) {
+        auto process = SearchProcessList(
-                return proc->GetProgramId() == program_id;
+            list, [program_id](auto& p) { return p->GetProgramId() == program_id; });
-            });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -119,12 +116,13 @@ private:
 
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(ResultSuccess);
-        rb.Push((*process)->GetProcessId());
+        rb.Push(process->GetProcessId());
     }
 
     void GetApplicationProcessId(HLERequestContext& ctx) {
         LOG_DEBUG(Service_PM, "called");
-        GetApplicationPidGeneric(ctx, kernel.GetProcessList());
+        auto list = kernel.GetProcessList();
+        GetApplicationPidGeneric(ctx, list);
     }
 
     void AtmosphereGetProcessInfo(HLERequestContext& ctx) {
@@ -135,11 +133,10 @@ private:
 
         LOG_WARNING(Service_PM, "(Partial Implementation) called, pid={:016X}", pid);
 
-        const auto process = SearchProcessList(kernel.GetProcessList(), [pid](const auto& proc) {
+        auto list = kernel.GetProcessList();
-            return proc->GetProcessId() == pid;
+        auto process = SearchProcessList(list, [pid](auto& p) { return p->GetProcessId() == pid; });
-        });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -159,7 +156,7 @@ private:
 
         OverrideStatus override_status{};
         ProgramLocation program_location{
-            .program_id = (*process)->GetProgramId(),
+            .program_id = process->GetProgramId(),
             .storage_id = 0,
         };
 
@@ -169,14 +166,11 @@ private:
         rb.PushRaw(program_location);
         rb.PushRaw(override_status);
     }
-
-    const Kernel::KernelCore& kernel;
 };
 
 class Info final : public ServiceFramework<Info> {
 public:
-    explicit Info(Core::System& system_, const std::vector<Kernel::KProcess*>& process_list_)
+    explicit Info(Core::System& system_) : ServiceFramework{system_, "pm:info"} {
-        : ServiceFramework{system_, "pm:info"}, process_list{process_list_} {
         static const FunctionInfo functions[] = {
             {0, &Info::GetProgramId, "GetProgramId"},
             {65000, &Info::AtmosphereGetProcessId, "AtmosphereGetProcessId"},
@@ -193,11 +187,11 @@ private:
 
         LOG_DEBUG(Service_PM, "called, process_id={:016X}", process_id);
 
-        const auto process = SearchProcessList(process_list, [process_id](const auto& proc) {
+        auto list = kernel.GetProcessList();
-            return proc->GetProcessId() == process_id;
+        auto process = SearchProcessList(
-        });
+            list, [process_id](auto& p) { return p->GetProcessId() == process_id; });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -205,7 +199,7 @@ private:
 
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(ResultSuccess);
-        rb.Push((*process)->GetProgramId());
+        rb.Push(process->GetProgramId());
     }
 
     void AtmosphereGetProcessId(HLERequestContext& ctx) {
@@ -214,11 +208,11 @@ private:
 
         LOG_DEBUG(Service_PM, "called, program_id={:016X}", program_id);
 
-        const auto process = SearchProcessList(process_list, [program_id](const auto& proc) {
+        auto list = system.Kernel().GetProcessList();
-            return proc->GetProgramId() == program_id;
+        auto process = SearchProcessList(
-        });
+            list, [program_id](auto& p) { return p->GetProgramId() == program_id; });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -226,16 +220,13 @@ private:
 
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(ResultSuccess);
-        rb.Push((*process)->GetProcessId());
+        rb.Push(process->GetProcessId());
     }
-
-    const std::vector<Kernel::KProcess*>& process_list;
 };
 
 class Shell final : public ServiceFramework<Shell> {
 public:
-    explicit Shell(Core::System& system_)
+    explicit Shell(Core::System& system_) : ServiceFramework{system_, "pm:shell"} {
-        : ServiceFramework{system_, "pm:shell"}, kernel{system_.Kernel()} {
         // clang-format off
         static const FunctionInfo functions[] = {
             {0, nullptr, "LaunchProgram"},
@@ -257,10 +248,9 @@ public:
 private:
     void GetApplicationProcessIdForShell(HLERequestContext& ctx) {
         LOG_DEBUG(Service_PM, "called");
-        GetApplicationPidGeneric(ctx, kernel.GetProcessList());
+        auto list = kernel.GetProcessList();
+        GetApplicationPidGeneric(ctx, list);
     }
-
-    const Kernel::KernelCore& kernel;
 };
 
 void LoopProcess(Core::System& system) {
@@ -268,8 +258,7 @@ void LoopProcess(Core::System& system) {
 
     server_manager->RegisterNamedService("pm:bm", std::make_shared<BootMode>(system));
     server_manager->RegisterNamedService("pm:dmnt", std::make_shared<DebugMonitor>(system));
-    server_manager->RegisterNamedService(
+    server_manager->RegisterNamedService("pm:info", std::make_shared<Info>(system));
-        "pm:info", std::make_shared<Info>(system, system.Kernel().GetProcessList()));
     server_manager->RegisterNamedService("pm:shell", std::make_shared<Shell>(system));
     ServerManager::RunServer(std::move(server_manager));
 }

src/core/hle/service/server_manager.cpp

@@ -256,8 +256,13 @@ Result ServerManager::WaitAndProcessImpl() {
 
         // Wait for a signal.
         s32 out_index{-1};
-        R_TRY(Kernel::KSynchronizationObject::Wait(m_system.Kernel(), &out_index, wait_objs.data(),
+        R_TRY_CATCH(Kernel::KSynchronizationObject::Wait(m_system.Kernel(), &out_index,
-                                                   num_objs, -1));
+                                                         wait_objs.data(), num_objs, -1)) {
+            R_CATCH(Kernel::ResultSessionClosed) {
+                // On session closed, index is updated and we don't want to return an error.
+            }
+        }
+        R_END_TRY_CATCH;
         ASSERT(out_index >= 0 && out_index < num_objs);
 
         // Set the output index.

src/core/loader/deconstructed_rom_directory.cpp

@@ -19,8 +19,54 @@
 #include "core/arm/nce/patcher.h"
 #endif
 
+#ifndef HAS_NCE
+namespace Core::NCE {
+class Patcher {};
+} // namespace Core::NCE
+#endif
+
 namespace Loader {
 
+struct PatchCollection {
+    explicit PatchCollection(bool is_application_) : is_application{is_application_} {
+        module_patcher_indices.fill(-1);
+        patchers.emplace_back();
+    }
+
+    std::vector<Core::NCE::Patcher>* GetPatchers() {
+        if (is_application && Settings::IsNceEnabled()) {
+            return &patchers;
+        }
+        return nullptr;
+    }
+
+    size_t GetTotalPatchSize() const {
+        size_t total_size{};
+#ifdef HAS_NCE
+        for (auto& patcher : patchers) {
+            total_size += patcher.GetSectionSize();
+        }
+#endif
+        return total_size;
+    }
+
+    void SaveIndex(size_t module) {
+        module_patcher_indices[module] = static_cast<s32>(patchers.size() - 1);
+    }
+
+    s32 GetIndex(size_t module) const {
+        return module_patcher_indices[module];
+    }
+
+    s32 GetLastIndex() const {
+        return static_cast<s32>(patchers.size()) - 1;
+    }
+
+    bool is_application;
+    std::vector<Core::NCE::Patcher> patchers;
+    std::array<s32, 13> module_patcher_indices{};
+};
+
 AppLoader_DeconstructedRomDirectory::AppLoader_DeconstructedRomDirectory(FileSys::VirtualFile file_,
                                                                          bool override_update_)
     : AppLoader(std::move(file_)), override_update(override_update_), is_hbl(false) {
@@ -142,18 +188,7 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
     std::size_t code_size{};
 
     // Define an nce patch context for each potential module.
-#ifdef HAS_NCE
+    PatchCollection patch_ctx{is_application};
-    std::array<Core::NCE::Patcher, 13> module_patchers;
-#endif
-
-    const auto GetPatcher = [&](size_t i) -> Core::NCE::Patcher* {
-#ifdef HAS_NCE
-        if (is_application && Settings::IsNceEnabled()) {
-            return &module_patchers[i];
-        }
-#endif
-        return nullptr;
-    };
 
     // Use the NSO module loader to figure out the code layout
     for (size_t i = 0; i < static_modules.size(); i++) {
@@ -164,13 +199,14 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
         }
 
         const bool should_pass_arguments = std::strcmp(module, "rtld") == 0;
-        const auto tentative_next_load_addr =
+        const auto tentative_next_load_addr = AppLoader_NSO::LoadModule(
-            AppLoader_NSO::LoadModule(process, system, *module_file, code_size,
+            process, system, *module_file, code_size, should_pass_arguments, false, {},
-                                      should_pass_arguments, false, {}, GetPatcher(i));
+            patch_ctx.GetPatchers(), patch_ctx.GetLastIndex());
         if (!tentative_next_load_addr) {
             return {ResultStatus::ErrorLoadingNSO, {}};
         }
 
+        patch_ctx.SaveIndex(i);
         code_size = *tentative_next_load_addr;
     }
 
@@ -184,6 +220,9 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
         return 0;
     }();
 
+    // Add patch size to the total module size
+    code_size += patch_ctx.GetTotalPatchSize();
+
     // Setup the process code layout
     if (process.LoadFromMetadata(metadata, code_size, fastmem_base, is_hbl).IsError()) {
         return {ResultStatus::ErrorUnableToParseKernelMetadata, {}};
@@ -204,9 +243,9 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
 
         const VAddr load_addr{next_load_addr};
         const bool should_pass_arguments = std::strcmp(module, "rtld") == 0;
-        const auto tentative_next_load_addr =
+        const auto tentative_next_load_addr = AppLoader_NSO::LoadModule(
-            AppLoader_NSO::LoadModule(process, system, *module_file, load_addr,
+            process, system, *module_file, load_addr, should_pass_arguments, true, pm,
-                                      should_pass_arguments, true, pm, GetPatcher(i));
+            patch_ctx.GetPatchers(), patch_ctx.GetIndex(i));
         if (!tentative_next_load_addr) {
             return {ResultStatus::ErrorLoadingNSO, {}};
         }
@@ -216,20 +255,6 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
         LOG_DEBUG(Loader, "loaded module {} @ {:#X}", module, load_addr);
     }
 
-    // Find the RomFS by searching for a ".romfs" file in this directory
-    const auto& files = dir->GetFiles();
-    const auto romfs_iter =
-        std::find_if(files.begin(), files.end(), [](const FileSys::VirtualFile& f) {
-            return f->GetName().find(".romfs") != std::string::npos;
-        });
-
-    // Register the RomFS if a ".romfs" file was found
-    if (romfs_iter != files.end() && *romfs_iter != nullptr) {
-        romfs = *romfs_iter;
-        system.GetFileSystemController().RegisterRomFS(std::make_unique<FileSys::RomFSFactory>(
-            *this, system.GetContentProvider(), system.GetFileSystemController()));
-    }
-
     is_loaded = true;
     return {ResultStatus::Success,
             LoadParameters{metadata.GetMainThreadPriority(), metadata.GetMainThreadStackSize()}};

src/core/loader/nca.cpp

@@ -74,8 +74,10 @@ AppLoader_NCA::LoadResult AppLoader_NCA::Load(Kernel::KProcess& process, Core::S
         return load_result;
     }
 
-    system.GetFileSystemController().RegisterRomFS(std::make_unique<FileSys::RomFSFactory>(
+    system.GetFileSystemController().RegisterProcess(
-        *this, system.GetContentProvider(), system.GetFileSystemController()));
+        process.GetProcessId(), nca->GetTitleId(),
+        std::make_shared<FileSys::RomFSFactory>(*this, system.GetContentProvider(),
+                                                system.GetFileSystemController()));
 
     is_loaded = true;
     return load_result;

src/core/loader/nro.cpp

@@ -275,10 +275,12 @@ AppLoader_NRO::LoadResult AppLoader_NRO::Load(Kernel::KProcess& process, Core::S
         return {ResultStatus::ErrorLoadingNRO, {}};
     }
 
-    if (romfs != nullptr) {
+    u64 program_id{};
-        system.GetFileSystemController().RegisterRomFS(std::make_unique<FileSys::RomFSFactory>(
+    ReadProgramId(program_id);
-            *this, system.GetContentProvider(), system.GetFileSystemController()));
+    system.GetFileSystemController().RegisterProcess(
-    }
+        process.GetProcessId(), program_id,
+        std::make_unique<FileSys::RomFSFactory>(*this, system.GetContentProvider(),
+                                                system.GetFileSystemController()));
 
     is_loaded = true;
     return {ResultStatus::Success, LoadParameters{Kernel::KThread::DefaultThreadPriority,

src/core/loader/nso.cpp

@@ -77,7 +77,8 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
                                                const FileSys::VfsFile& nso_file, VAddr load_base,
                                                bool should_pass_arguments, bool load_into_process,
                                                std::optional<FileSys::PatchManager> pm,
-                                               Core::NCE::Patcher* patch) {
+                                               std::vector<Core::NCE::Patcher>* patches,
+                                               s32 patch_index) {
     if (nso_file.GetSize() < sizeof(NSOHeader)) {
         return std::nullopt;
     }
@@ -94,8 +95,11 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
     // Allocate some space at the beginning if we are patching in PreText mode.
     const size_t module_start = [&]() -> size_t {
 #ifdef HAS_NCE
-        if (patch && patch->GetPatchMode() == Core::NCE::PatchMode::PreText) {
+        if (patches && load_into_process) {
-            return patch->GetSectionSize();
+            auto* patch = &patches->operator[](patch_index);
+            if (patch->GetPatchMode() == Core::NCE::PatchMode::PreText) {
+                return patch->GetSectionSize();
+            }
         }
 #endif
         return 0;
@@ -160,27 +164,24 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
 #ifdef HAS_NCE
     // If we are computing the process code layout and using nce backend, patch.
     const auto& code = codeset.CodeSegment();
-    if (patch && patch->GetPatchMode() == Core::NCE::PatchMode::None) {
+    auto* patch = patches ? &patches->operator[](patch_index) : nullptr;
+    if (patch && !load_into_process) {
         // Patch SVCs and MRS calls in the guest code
-        patch->PatchText(program_image, code);
+        while (!patch->PatchText(program_image, code)) {
-
+            patch = &patches->emplace_back();
-        // Add patch section size to the module size.
+        }
-        image_size += static_cast<u32>(patch->GetSectionSize());
     } else if (patch) {
         // Relocate code patch and copy to the program_image.
-        patch->RelocateAndCopy(load_base, code, program_image, &process.GetPostHandlers());
+        if (patch->RelocateAndCopy(load_base, code, program_image, &process.GetPostHandlers())) {
-
+            // Update patch section.
-        // Update patch section.
+            auto& patch_segment = codeset.PatchSegment();
-        auto& patch_segment = codeset.PatchSegment();
+            patch_segment.addr =
-        patch_segment.addr =
+                patch->GetPatchMode() == Core::NCE::PatchMode::PreText ? 0 : image_size;
-            patch->GetPatchMode() == Core::NCE::PatchMode::PreText ? 0 : image_size;
+            patch_segment.size = static_cast<u32>(patch->GetSectionSize());
-        patch_segment.size = static_cast<u32>(patch->GetSectionSize());
-
-        // Add patch section size to the module size. In PreText mode image_size
-        // already contains the patch segment as part of module_start.
-        if (patch->GetPatchMode() == Core::NCE::PatchMode::PostData) {
-            image_size += patch_segment.size;
         }
+
+        // Refresh image_size to take account the patch section if it was added by RelocateAndCopy
+        image_size = static_cast<u32>(program_image.size());
     }
 #endif
 

src/core/loader/nso.h

@@ -93,7 +93,8 @@ public:
                                            const FileSys::VfsFile& nso_file, VAddr load_base,
                                            bool should_pass_arguments, bool load_into_process,
                                            std::optional<FileSys::PatchManager> pm = {},
-                                           Core::NCE::Patcher* patch = nullptr);
+                                           std::vector<Core::NCE::Patcher>* patches = nullptr,
+                                           s32 patch_index = -1);
 
     LoadResult Load(Kernel::KProcess& process, Core::System& system) override;
 

src/core/loader/nsp.cpp

@@ -111,7 +111,8 @@ AppLoader_NSP::LoadResult AppLoader_NSP::Load(Kernel::KProcess& process, Core::S
 
     FileSys::VirtualFile update_raw;
     if (ReadUpdateRaw(update_raw) == ResultStatus::Success && update_raw != nullptr) {
-        system.GetFileSystemController().SetPackedUpdate(std::move(update_raw));
+        system.GetFileSystemController().SetPackedUpdate(process.GetProcessId(),
+                                                         std::move(update_raw));
     }
 
     is_loaded = true;

src/core/loader/xci.cpp

@@ -78,7 +78,8 @@ AppLoader_XCI::LoadResult AppLoader_XCI::Load(Kernel::KProcess& process, Core::S
 
     FileSys::VirtualFile update_raw;
     if (ReadUpdateRaw(update_raw) == ResultStatus::Success && update_raw != nullptr) {
-        system.GetFileSystemController().SetPackedUpdate(std::move(update_raw));
+        system.GetFileSystemController().SetPackedUpdate(process.GetProcessId(),
+                                                         std::move(update_raw));
     }
 
     is_loaded = true;

src/core/memory.cpp

@@ -24,6 +24,8 @@
 #include "core/hle/kernel/k_process.h"
 #include "core/memory.h"
 #include "video_core/gpu.h"
+#include "video_core/host1x/gpu_device_memory_manager.h"
+#include "video_core/host1x/host1x.h"
 #include "video_core/rasterizer_download_area.h"
 
 namespace Core::Memory {
@@ -637,17 +639,6 @@ struct Memory::Impl {
         LOG_DEBUG(HW_Memory, "Mapping {:016X} onto {:016X}-{:016X}", GetInteger(target),
                   base * YUZU_PAGESIZE, (base + size) * YUZU_PAGESIZE);
 
-        // During boot, current_page_table might not be set yet, in which case we need not flush
-        if (system.IsPoweredOn()) {
-            auto& gpu = system.GPU();
-            for (u64 i = 0; i < size; i++) {
-                const auto page = base + i;
-                if (page_table.pointers[page].Type() == Common::PageType::RasterizerCachedMemory) {
-                    gpu.FlushAndInvalidateRegion(page << YUZU_PAGEBITS, YUZU_PAGESIZE);
-                }
-            }
-        }
-
         const auto end = base + size;
         ASSERT_MSG(end <= page_table.pointers.size(), "out of range mapping at {:016X}",
                    base + page_table.pointers.size());
@@ -811,21 +802,33 @@ struct Memory::Impl {
         return true;
     }
 
-    void HandleRasterizerDownload(VAddr address, size_t size) {
+    void HandleRasterizerDownload(VAddr v_address, size_t size) {
+        const auto* p = GetPointerImpl(
+            v_address, []() {}, []() {});
+        if (!gpu_device_memory) [[unlikely]] {
+            gpu_device_memory = &system.Host1x().MemoryManager();
+        }
         const size_t core = system.GetCurrentHostThreadID();
         auto& current_area = rasterizer_read_areas[core];
-        const VAddr end_address = address + size;
+        gpu_device_memory->ApplyOpOnPointer(p, scratch_buffers[core], [&](DAddr address) {
-        if (current_area.start_address <= address && end_address <= current_area.end_address)
+            const DAddr end_address = address + size;
-            [[likely]] {
+            if (current_area.start_address <= address && end_address <= current_area.end_address)
-            return;
+                [[likely]] {
-        }
+                return;
-        current_area = system.GPU().OnCPURead(address, size);
+            }
+            current_area = system.GPU().OnCPURead(address, size);
+        });
     }
 
-    void HandleRasterizerWrite(VAddr address, size_t size) {
+    void HandleRasterizerWrite(VAddr v_address, size_t size) {
+        const auto* p = GetPointerImpl(
+            v_address, []() {}, []() {});
         constexpr size_t sys_core = Core::Hardware::NUM_CPU_CORES - 1;
         const size_t core = std::min(system.GetCurrentHostThreadID(),
                                      sys_core); // any other calls threads go to syscore.
+        if (!gpu_device_memory) [[unlikely]] {
+            gpu_device_memory = &system.Host1x().MemoryManager();
+        }
         // Guard on sys_core;
         if (core == sys_core) [[unlikely]] {
             sys_core_guard.lock();
@@ -835,36 +838,53 @@ struct Memory::Impl {
                 sys_core_guard.unlock();
             }
         });
-        auto& current_area = rasterizer_write_areas[core];
+        gpu_device_memory->ApplyOpOnPointer(p, scratch_buffers[core], [&](DAddr address) {
-        VAddr subaddress = address >> YUZU_PAGEBITS;
+            auto& current_area = rasterizer_write_areas[core];
-        bool do_collection = current_area.last_address == subaddress;
+            PAddr subaddress = address >> YUZU_PAGEBITS;
-        if (!do_collection) [[unlikely]] {
+            bool do_collection = current_area.last_address == subaddress;
-            do_collection = system.GPU().OnCPUWrite(address, size);
+            if (!do_collection) [[unlikely]] {
-            if (!do_collection) {
+                do_collection = system.GPU().OnCPUWrite(address, size);
-                return;
+                if (!do_collection) {
+                    return;
+                }
+                current_area.last_address = subaddress;
             }
-            current_area.last_address = subaddress;
+            gpu_dirty_managers[core].Collect(address, size);
-        }
+        });
-        gpu_dirty_managers[core].Collect(address, size);
     }
 
     struct GPUDirtyState {
-        VAddr last_address;
+        PAddr last_address;
     };
 
-    void InvalidateRegion(Common::ProcessAddress dest_addr, size_t size) {
+    void InvalidateGPUMemory(u8* p, size_t size) {
-        system.GPU().InvalidateRegion(GetInteger(dest_addr), size);
+        constexpr size_t sys_core = Core::Hardware::NUM_CPU_CORES - 1;
-    }
+        const size_t core = std::min(system.GetCurrentHostThreadID(),
-
+                                     sys_core); // any other calls threads go to syscore.
-    void FlushRegion(Common::ProcessAddress dest_addr, size_t size) {
+        if (!gpu_device_memory) [[unlikely]] {
-        system.GPU().FlushRegion(GetInteger(dest_addr), size);
+            gpu_device_memory = &system.Host1x().MemoryManager();
+        }
+        // Guard on sys_core;
+        if (core == sys_core) [[unlikely]] {
+            sys_core_guard.lock();
+        }
+        SCOPE_EXIT({
+            if (core == sys_core) [[unlikely]] {
+                sys_core_guard.unlock();
+            }
+        });
+        auto& gpu = system.GPU();
+        gpu_device_memory->ApplyOpOnPointer(
+            p, scratch_buffers[core], [&](DAddr address) { gpu.InvalidateRegion(address, size); });
     }
 
     Core::System& system;
+    Tegra::MaxwellDeviceMemoryManager* gpu_device_memory{};
     Common::PageTable* current_page_table = nullptr;
     std::array<VideoCore::RasterizerDownloadArea, Core::Hardware::NUM_CPU_CORES>
         rasterizer_read_areas{};
     std::array<GPUDirtyState, Core::Hardware::NUM_CPU_CORES> rasterizer_write_areas{};
+    std::array<Common::ScratchBuffer<u32>, Core::Hardware::NUM_CPU_CORES> scratch_buffers{};
     std::span<Core::GPUDirtyMemoryManager> gpu_dirty_managers;
     std::mutex sys_core_guard;
 
@@ -1059,14 +1079,6 @@ void Memory::MarkRegionDebug(Common::ProcessAddress vaddr, u64 size, bool debug)
     impl->MarkRegionDebug(GetInteger(vaddr), size, debug);
 }
 
-void Memory::InvalidateRegion(Common::ProcessAddress dest_addr, size_t size) {
-    impl->InvalidateRegion(dest_addr, size);
-}
-
-void Memory::FlushRegion(Common::ProcessAddress dest_addr, size_t size) {
-    impl->FlushRegion(dest_addr, size);
-}
-
 bool Memory::InvalidateNCE(Common::ProcessAddress vaddr, size_t size) {
     [[maybe_unused]] bool mapped = true;
     [[maybe_unused]] bool rasterizer = false;
@@ -1078,10 +1090,10 @@ bool Memory::InvalidateNCE(Common::ProcessAddress vaddr, size_t size) {
                       GetInteger(vaddr));
             mapped = false;
         },
-        [&] {
+        [&] { rasterizer = true; });
-            impl->system.GPU().InvalidateRegion(GetInteger(vaddr), size);
+    if (rasterizer) {
-            rasterizer = true;
+        impl->InvalidateGPUMemory(ptr, size);
-        });
+    }
 
 #ifdef __linux__
     if (!rasterizer && mapped) {

src/core/memory.h

@@ -12,6 +12,7 @@
 
 #include "common/scratch_buffer.h"
 #include "common/typed_address.h"
+#include "core/guest_memory.h"
 #include "core/hle/result.h"
 
 namespace Common {
@@ -486,10 +487,10 @@ public:
     void MarkRegionDebug(Common::ProcessAddress vaddr, u64 size, bool debug);
 
     void SetGPUDirtyManagers(std::span<Core::GPUDirtyMemoryManager> managers);
-    void InvalidateRegion(Common::ProcessAddress dest_addr, size_t size);
+
     bool InvalidateNCE(Common::ProcessAddress vaddr, size_t size);
+
     bool InvalidateSeparateHeap(void* fault_address);
-    void FlushRegion(Common::ProcessAddress dest_addr, size_t size);
 
 private:
     Core::System& system;
@@ -498,209 +499,9 @@ private:
     std::unique_ptr<Impl> impl;
 };
 
-enum GuestMemoryFlags : u32 {
-    Read = 1 << 0,
-    Write = 1 << 1,
-    Safe = 1 << 2,
-    Cached = 1 << 3,
-
-    SafeRead = Read | Safe,
-    SafeWrite = Write | Safe,
-    SafeReadWrite = SafeRead | SafeWrite,
-    SafeReadCachedWrite = SafeReadWrite | Cached,
-
-    UnsafeRead = Read,
-    UnsafeWrite = Write,
-    UnsafeReadWrite = UnsafeRead | UnsafeWrite,
-    UnsafeReadCachedWrite = UnsafeReadWrite | Cached,
-};
-
-namespace {
-template <typename M, typename T, GuestMemoryFlags FLAGS>
-class GuestMemory {
-    using iterator = T*;
-    using const_iterator = const T*;
-    using value_type = T;
-    using element_type = T;
-    using iterator_category = std::contiguous_iterator_tag;
-
-public:
-    GuestMemory() = delete;
-    explicit GuestMemory(M& memory, u64 addr, std::size_t size,
-                         Common::ScratchBuffer<T>* backup = nullptr)
-        : m_memory{memory}, m_addr{addr}, m_size{size} {
-        static_assert(FLAGS & GuestMemoryFlags::Read || FLAGS & GuestMemoryFlags::Write);
-        if constexpr (FLAGS & GuestMemoryFlags::Read) {
-            Read(addr, size, backup);
-        }
-    }
-
-    ~GuestMemory() = default;
-
-    T* data() noexcept {
-        return m_data_span.data();
-    }
-
-    const T* data() const noexcept {
-        return m_data_span.data();
-    }
-
-    size_t size() const noexcept {
-        return m_size;
-    }
-
-    size_t size_bytes() const noexcept {
-        return this->size() * sizeof(T);
-    }
-
-    [[nodiscard]] T* begin() noexcept {
-        return this->data();
-    }
-
-    [[nodiscard]] const T* begin() const noexcept {
-        return this->data();
-    }
-
-    [[nodiscard]] T* end() noexcept {
-        return this->data() + this->size();
-    }
-
-    [[nodiscard]] const T* end() const noexcept {
-        return this->data() + this->size();
-    }
-
-    T& operator[](size_t index) noexcept {
-        return m_data_span[index];
-    }
-
-    const T& operator[](size_t index) const noexcept {
-        return m_data_span[index];
-    }
-
-    void SetAddressAndSize(u64 addr, std::size_t size) noexcept {
-        m_addr = addr;
-        m_size = size;
-        m_addr_changed = true;
-    }
-
-    std::span<T> Read(u64 addr, std::size_t size,
-                      Common::ScratchBuffer<T>* backup = nullptr) noexcept {
-        m_addr = addr;
-        m_size = size;
-        if (m_size == 0) {
-            m_is_data_copy = true;
-            return {};
-        }
-
-        if (this->TrySetSpan()) {
-            if constexpr (FLAGS & GuestMemoryFlags::Safe) {
-                m_memory.FlushRegion(m_addr, this->size_bytes());
-            }
-        } else {
-            if (backup) {
-                backup->resize_destructive(this->size());
-                m_data_span = *backup;
-            } else {
-                m_data_copy.resize(this->size());
-                m_data_span = std::span(m_data_copy);
-            }
-            m_is_data_copy = true;
-            m_span_valid = true;
-            if constexpr (FLAGS & GuestMemoryFlags::Safe) {
-                m_memory.ReadBlock(m_addr, this->data(), this->size_bytes());
-            } else {
-                m_memory.ReadBlockUnsafe(m_addr, this->data(), this->size_bytes());
-            }
-        }
-        return m_data_span;
-    }
-
-    void Write(std::span<T> write_data) noexcept {
-        if constexpr (FLAGS & GuestMemoryFlags::Cached) {
-            m_memory.WriteBlockCached(m_addr, write_data.data(), this->size_bytes());
-        } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
-            m_memory.WriteBlock(m_addr, write_data.data(), this->size_bytes());
-        } else {
-            m_memory.WriteBlockUnsafe(m_addr, write_data.data(), this->size_bytes());
-        }
-    }
-
-    bool TrySetSpan() noexcept {
-        if (u8* ptr = m_memory.GetSpan(m_addr, this->size_bytes()); ptr) {
-            m_data_span = {reinterpret_cast<T*>(ptr), this->size()};
-            m_span_valid = true;
-            return true;
-        }
-        return false;
-    }
-
-protected:
-    bool IsDataCopy() const noexcept {
-        return m_is_data_copy;
-    }
-
-    bool AddressChanged() const noexcept {
-        return m_addr_changed;
-    }
-
-    M& m_memory;
-    u64 m_addr{};
-    size_t m_size{};
-    std::span<T> m_data_span{};
-    std::vector<T> m_data_copy{};
-    bool m_span_valid{false};
-    bool m_is_data_copy{false};
-    bool m_addr_changed{false};
-};
-
-template <typename M, typename T, GuestMemoryFlags FLAGS>
-class GuestMemoryScoped : public GuestMemory<M, T, FLAGS> {
-public:
-    GuestMemoryScoped() = delete;
-    explicit GuestMemoryScoped(M& memory, u64 addr, std::size_t size,
-                               Common::ScratchBuffer<T>* backup = nullptr)
-        : GuestMemory<M, T, FLAGS>(memory, addr, size, backup) {
-        if constexpr (!(FLAGS & GuestMemoryFlags::Read)) {
-            if (!this->TrySetSpan()) {
-                if (backup) {
-                    this->m_data_span = *backup;
-                    this->m_span_valid = true;
-                    this->m_is_data_copy = true;
-                }
-            }
-        }
-    }
-
-    ~GuestMemoryScoped() {
-        if constexpr (FLAGS & GuestMemoryFlags::Write) {
-            if (this->size() == 0) [[unlikely]] {
-                return;
-            }
-
-            if (this->AddressChanged() || this->IsDataCopy()) {
-                ASSERT(this->m_span_valid);
-                if constexpr (FLAGS & GuestMemoryFlags::Cached) {
-                    this->m_memory.WriteBlockCached(this->m_addr, this->data(), this->size_bytes());
-                } else if constexpr (FLAGS & GuestMemoryFlags::Safe) {
-                    this->m_memory.WriteBlock(this->m_addr, this->data(), this->size_bytes());
-                } else {
-                    this->m_memory.WriteBlockUnsafe(this->m_addr, this->data(), this->size_bytes());
-                }
-            } else if constexpr ((FLAGS & GuestMemoryFlags::Safe) ||
-                                 (FLAGS & GuestMemoryFlags::Cached)) {
-                this->m_memory.InvalidateRegion(this->m_addr, this->size_bytes());
-            }
-        }
-    }
-};
-} // namespace
-
 template <typename T, GuestMemoryFlags FLAGS>
-using CpuGuestMemory = GuestMemory<Memory, T, FLAGS>;
+using CpuGuestMemory = GuestMemory<Core::Memory::Memory, T, FLAGS>;
 template <typename T, GuestMemoryFlags FLAGS>
-using CpuGuestMemoryScoped = GuestMemoryScoped<Memory, T, FLAGS>;
+using CpuGuestMemoryScoped = GuestMemoryScoped<Core::Memory::Memory, T, FLAGS>;
-template <typename T, GuestMemoryFlags FLAGS>
+
-using GpuGuestMemory = GuestMemory<Tegra::MemoryManager, T, FLAGS>;
-template <typename T, GuestMemoryFlags FLAGS>
-using GpuGuestMemoryScoped = GuestMemoryScoped<Tegra::MemoryManager, T, FLAGS>;
 } // namespace Core::Memory

src/shader_recompiler/backend/spirv/emit_spirv_memory.cpp

@@ -65,6 +65,14 @@ void WriteStorage32(EmitContext& ctx, const IR::Value& binding, const IR::Value&
     WriteStorage(ctx, binding, offset, value, ctx.storage_types.U32, sizeof(u32),
                  &StorageDefinitions::U32, index_offset);
 }
+
+void WriteStorageByCasLoop(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
+                           Id value, Id bit_offset, Id bit_count) {
+    const Id pointer{StoragePointer(ctx, binding, offset, ctx.storage_types.U32, sizeof(u32),
+                                    &StorageDefinitions::U32)};
+    ctx.OpFunctionCall(ctx.TypeVoid(), ctx.write_storage_cas_loop_func, pointer, value, bit_offset,
+                       bit_count);
+}
 } // Anonymous namespace
 
 void EmitLoadGlobalU8(EmitContext&) {
@@ -219,26 +227,42 @@ Id EmitLoadStorage128(EmitContext& ctx, const IR::Value& binding, const IR::Valu
 
 void EmitWriteStorageU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
                         Id value) {
-    WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.U8, value), ctx.storage_types.U8,
+    if (ctx.profile.support_int8) {
-                 sizeof(u8), &StorageDefinitions::U8);
+        WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.U8, value), ctx.storage_types.U8,
+                     sizeof(u8), &StorageDefinitions::U8);
+    } else {
+        WriteStorageByCasLoop(ctx, binding, offset, value, ctx.BitOffset8(offset), ctx.Const(8u));
+    }
 }
 
 void EmitWriteStorageS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
                         Id value) {
-    WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.S8, value), ctx.storage_types.S8,
+    if (ctx.profile.support_int8) {
-                 sizeof(s8), &StorageDefinitions::S8);
+        WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.S8, value), ctx.storage_types.S8,
+                     sizeof(s8), &StorageDefinitions::S8);
+    } else {
+        WriteStorageByCasLoop(ctx, binding, offset, value, ctx.BitOffset8(offset), ctx.Const(8u));
+    }
 }
 
 void EmitWriteStorageU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
                          Id value) {
-    WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.U16, value), ctx.storage_types.U16,
+    if (ctx.profile.support_int16) {
-                 sizeof(u16), &StorageDefinitions::U16);
+        WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.U16, value), ctx.storage_types.U16,
+                     sizeof(u16), &StorageDefinitions::U16);
+    } else {
+        WriteStorageByCasLoop(ctx, binding, offset, value, ctx.BitOffset16(offset), ctx.Const(16u));
+    }
 }
 
 void EmitWriteStorageS16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
                          Id value) {
-    WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.S16, value), ctx.storage_types.S16,
+    if (ctx.profile.support_int16) {
-                 sizeof(s16), &StorageDefinitions::S16);
+        WriteStorage(ctx, binding, offset, ctx.OpSConvert(ctx.S16, value), ctx.storage_types.S16,
+                     sizeof(s16), &StorageDefinitions::S16);
+    } else {
+        WriteStorageByCasLoop(ctx, binding, offset, value, ctx.BitOffset16(offset), ctx.Const(16u));
+    }
 }
 
 void EmitWriteStorage32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -480,6 +480,7 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
     DefineTextures(program.info, texture_binding, bindings.texture_scaling_index);
     DefineImages(program.info, image_binding, bindings.image_scaling_index);
     DefineAttributeMemAccess(program.info);
+    DefineWriteStorageCasLoopFunction(program.info);
     DefineGlobalMemoryFunctions(program.info);
     DefineRescalingInput(program.info);
     DefineRenderArea(program.info);
@@ -877,6 +878,56 @@ void EmitContext::DefineAttributeMemAccess(const Info& info) {
     }
 }
 
+void EmitContext::DefineWriteStorageCasLoopFunction(const Info& info) {
+    if (profile.support_int8 && profile.support_int16) {
+        return;
+    }
+    if (!info.uses_int8 && !info.uses_int16) {
+        return;
+    }
+
+    AddCapability(spv::Capability::VariablePointersStorageBuffer);
+
+    const Id ptr_type{TypePointer(spv::StorageClass::StorageBuffer, U32[1])};
+    const Id func_type{TypeFunction(void_id, ptr_type, U32[1], U32[1], U32[1])};
+    const Id func{OpFunction(void_id, spv::FunctionControlMask::MaskNone, func_type)};
+    const Id pointer{OpFunctionParameter(ptr_type)};
+    const Id value{OpFunctionParameter(U32[1])};
+    const Id bit_offset{OpFunctionParameter(U32[1])};
+    const Id bit_count{OpFunctionParameter(U32[1])};
+
+    AddLabel();
+    const Id scope_device{Const(1u)};
+    const Id ordering_relaxed{u32_zero_value};
+    const Id body_label{OpLabel()};
+    const Id continue_label{OpLabel()};
+    const Id endloop_label{OpLabel()};
+    const Id beginloop_label{OpLabel()};
+    OpBranch(beginloop_label);
+
+    AddLabel(beginloop_label);
+    OpLoopMerge(endloop_label, continue_label, spv::LoopControlMask::MaskNone);
+    OpBranch(body_label);
+
+    AddLabel(body_label);
+    const Id expected_value{OpLoad(U32[1], pointer)};
+    const Id desired_value{OpBitFieldInsert(U32[1], expected_value, value, bit_offset, bit_count)};
+    const Id actual_value{OpAtomicCompareExchange(U32[1], pointer, scope_device, ordering_relaxed,
+                                                  ordering_relaxed, desired_value, expected_value)};
+    const Id store_successful{OpIEqual(U1, expected_value, actual_value)};
+    OpBranchConditional(store_successful, endloop_label, continue_label);
+
+    AddLabel(endloop_label);
+    OpReturn();
+
+    AddLabel(continue_label);
+    OpBranch(beginloop_label);
+
+    OpFunctionEnd();
+
+    write_storage_cas_loop_func = func;
+}
+
 void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
     if (!info.uses_global_memory || !profile.support_int64) {
         return;