From 419055e484f0f0073d5832f7ded5fd3a3e5ad7de Mon Sep 17 00:00:00 2001
From: Liam <byteslice@airmail.cc>
Date: Mon, 11 Dec 2023 20:21:23 -0500
Subject: [PATCH] kernel: instantiate memory separately for each guest process

---
 src/core/arm/arm_interface.cpp               |  2 +-
 src/core/arm/arm_interface.h                 |  2 +-
 src/core/arm/debug.cpp                       | 14 ++--
 src/core/arm/debug.h                         |  6 +-
 src/core/arm/dynarmic/arm_dynarmic_32.cpp    |  6 +-
 src/core/arm/dynarmic/arm_dynarmic_32.h      |  2 +-
 src/core/arm/dynarmic/arm_dynarmic_64.cpp    |  6 +-
 src/core/arm/dynarmic/arm_dynarmic_64.h      |  2 +-
 src/core/core.cpp                            | 73 ++++++--------------
 src/core/core.h                              | 13 ----
 src/core/hle/kernel/k_address_arbiter.cpp    | 19 ++---
 src/core/hle/kernel/k_condition_variable.cpp |  8 +--
 src/core/hle/kernel/k_process.cpp            | 18 +++--
 src/core/hle/kernel/k_process.h              | 15 +++-
 src/core/hle/kernel/k_thread.h               |  6 +-
 src/core/hle/kernel/kernel.cpp               | 13 ----
 src/core/hle/kernel/kernel.h                 |  4 --
 17 files changed, 82 insertions(+), 127 deletions(-)

diff --git a/src/core/arm/arm_interface.cpp b/src/core/arm/arm_interface.cpp
index 698c9c8ad..5dc7e5d59 100644
--- a/src/core/arm/arm_interface.cpp
+++ b/src/core/arm/arm_interface.cpp
@@ -9,7 +9,7 @@
 
 namespace Core {
 
-void ArmInterface::LogBacktrace(const Kernel::KProcess* process) const {
+void ArmInterface::LogBacktrace(Kernel::KProcess* process) const {
     Kernel::Svc::ThreadContext ctx;
     this->GetContext(ctx);
 
diff --git a/src/core/arm/arm_interface.h b/src/core/arm/arm_interface.h
index 806c7c9e9..495963eef 100644
--- a/src/core/arm/arm_interface.h
+++ b/src/core/arm/arm_interface.h
@@ -95,7 +95,7 @@ public:
     virtual void SignalInterrupt(Kernel::KThread* thread) = 0;
 
     // Stack trace generation.
-    void LogBacktrace(const Kernel::KProcess* process) const;
+    void LogBacktrace(Kernel::KProcess* process) const;
 
     // Debug functionality.
     virtual const Kernel::DebugWatchpoint* HaltedWatchpoint() const = 0;
diff --git a/src/core/arm/debug.cpp b/src/core/arm/debug.cpp
index af1c34bc3..854509463 100644
--- a/src/core/arm/debug.cpp
+++ b/src/core/arm/debug.cpp
@@ -79,7 +79,7 @@ constexpr std::array<u64, 2> SegmentBases{
     0x7100000000ULL,
 };
 
-void SymbolicateBacktrace(const Kernel::KProcess* process, std::vector<BacktraceEntry>& out) {
+void SymbolicateBacktrace(Kernel::KProcess* process, std::vector<BacktraceEntry>& out) {
     auto modules = FindModules(process);
 
     const bool is_64 = process->Is64Bit();
@@ -118,7 +118,7 @@ void SymbolicateBacktrace(const Kernel::KProcess* process, std::vector<Backtrace
     }
 }
 
-std::vector<BacktraceEntry> GetAArch64Backtrace(const Kernel::KProcess* process,
+std::vector<BacktraceEntry> GetAArch64Backtrace(Kernel::KProcess* process,
                                                 const Kernel::Svc::ThreadContext& ctx) {
     std::vector<BacktraceEntry> out;
     auto& memory = process->GetMemory();
@@ -144,7 +144,7 @@ std::vector<BacktraceEntry> GetAArch64Backtrace(const Kernel::KProcess* process,
     return out;
 }
 
-std::vector<BacktraceEntry> GetAArch32Backtrace(const Kernel::KProcess* process,
+std::vector<BacktraceEntry> GetAArch32Backtrace(Kernel::KProcess* process,
                                                 const Kernel::Svc::ThreadContext& ctx) {
     std::vector<BacktraceEntry> out;
     auto& memory = process->GetMemory();
@@ -173,7 +173,7 @@ std::vector<BacktraceEntry> GetAArch32Backtrace(const Kernel::KProcess* process,
 } // namespace
 
 std::optional<std::string> GetThreadName(const Kernel::KThread* thread) {
-    const auto* process = thread->GetOwnerProcess();
+    auto* process = thread->GetOwnerProcess();
     if (process->Is64Bit()) {
         return GetNameFromThreadType64(process->GetMemory(), *thread);
     } else {
@@ -248,7 +248,7 @@ Kernel::KProcessAddress GetModuleEnd(const Kernel::KProcess* process,
     return cur_addr - 1;
 }
 
-Loader::AppLoader::Modules FindModules(const Kernel::KProcess* process) {
+Loader::AppLoader::Modules FindModules(Kernel::KProcess* process) {
     Loader::AppLoader::Modules modules;
 
     auto& page_table = process->GetPageTable();
@@ -312,7 +312,7 @@ Loader::AppLoader::Modules FindModules(const Kernel::KProcess* process) {
     return modules;
 }
 
-Kernel::KProcessAddress FindMainModuleEntrypoint(const Kernel::KProcess* process) {
+Kernel::KProcessAddress FindMainModuleEntrypoint(Kernel::KProcess* process) {
     // Do we have any loaded executable sections?
     auto modules = FindModules(process);
 
@@ -337,7 +337,7 @@ void InvalidateInstructionCacheRange(const Kernel::KProcess* process, u64 addres
     }
 }
 
-std::vector<BacktraceEntry> GetBacktraceFromContext(const Kernel::KProcess* process,
+std::vector<BacktraceEntry> GetBacktraceFromContext(Kernel::KProcess* process,
                                                     const Kernel::Svc::ThreadContext& ctx) {
     if (process->Is64Bit()) {
         return GetAArch64Backtrace(process, ctx);
diff --git a/src/core/arm/debug.h b/src/core/arm/debug.h
index c542633db..3cd671365 100644
--- a/src/core/arm/debug.h
+++ b/src/core/arm/debug.h
@@ -14,9 +14,9 @@ std::optional<std::string> GetThreadName(const Kernel::KThread* thread);
 std::string_view GetThreadWaitReason(const Kernel::KThread* thread);
 std::string GetThreadState(const Kernel::KThread* thread);
 
-Loader::AppLoader::Modules FindModules(const Kernel::KProcess* process);
+Loader::AppLoader::Modules FindModules(Kernel::KProcess* process);
 Kernel::KProcessAddress GetModuleEnd(const Kernel::KProcess* process, Kernel::KProcessAddress base);
-Kernel::KProcessAddress FindMainModuleEntrypoint(const Kernel::KProcess* process);
+Kernel::KProcessAddress FindMainModuleEntrypoint(Kernel::KProcess* process);
 
 void InvalidateInstructionCacheRange(const Kernel::KProcess* process, u64 address, u64 size);
 
@@ -28,7 +28,7 @@ struct BacktraceEntry {
     std::string name;
 };
 
-std::vector<BacktraceEntry> GetBacktraceFromContext(const Kernel::KProcess* process,
+std::vector<BacktraceEntry> GetBacktraceFromContext(Kernel::KProcess* process,
                                                     const Kernel::Svc::ThreadContext& ctx);
 std::vector<BacktraceEntry> GetBacktrace(const Kernel::KThread* thread);
 
diff --git a/src/core/arm/dynarmic/arm_dynarmic_32.cpp b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
index f34865e26..c78cfd528 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_32.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
@@ -15,7 +15,7 @@ using namespace Common::Literals;
 
 class DynarmicCallbacks32 : public Dynarmic::A32::UserCallbacks {
 public:
-    explicit DynarmicCallbacks32(ArmDynarmic32& parent, const Kernel::KProcess* process)
+    explicit DynarmicCallbacks32(ArmDynarmic32& parent, Kernel::KProcess* process)
         : m_parent{parent}, m_memory(process->GetMemory()),
           m_process(process), m_debugger_enabled{parent.m_system.DebuggerEnabled()},
           m_check_memory_access{m_debugger_enabled ||
@@ -169,7 +169,7 @@ public:
 
     ArmDynarmic32& m_parent;
     Core::Memory::Memory& m_memory;
-    const Kernel::KProcess* m_process{};
+    Kernel::KProcess* m_process{};
     const bool m_debugger_enabled{};
     const bool m_check_memory_access{};
     static constexpr u64 MinimumRunCycles = 10000U;
@@ -370,7 +370,7 @@ void ArmDynarmic32::RewindBreakpointInstruction() {
     this->SetContext(m_breakpoint_context);
 }
 
-ArmDynarmic32::ArmDynarmic32(System& system, bool uses_wall_clock, const Kernel::KProcess* process,
+ArmDynarmic32::ArmDynarmic32(System& system, bool uses_wall_clock, Kernel::KProcess* process,
                              DynarmicExclusiveMonitor& exclusive_monitor, std::size_t core_index)
     : ArmInterface{uses_wall_clock}, m_system{system}, m_exclusive_monitor{exclusive_monitor},
       m_cb(std::make_unique<DynarmicCallbacks32>(*this, process)),
diff --git a/src/core/arm/dynarmic/arm_dynarmic_32.h b/src/core/arm/dynarmic/arm_dynarmic_32.h
index 185ac7cbf..b580efe61 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_32.h
+++ b/src/core/arm/dynarmic/arm_dynarmic_32.h
@@ -20,7 +20,7 @@ class System;
 
 class ArmDynarmic32 final : public ArmInterface {
 public:
-    ArmDynarmic32(System& system, bool uses_wall_clock, const Kernel::KProcess* process,
+    ArmDynarmic32(System& system, bool uses_wall_clock, Kernel::KProcess* process,
                   DynarmicExclusiveMonitor& exclusive_monitor, std::size_t core_index);
     ~ArmDynarmic32() override;
 
diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.cpp b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
index dff14756e..f351b13d9 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
@@ -15,7 +15,7 @@ using namespace Common::Literals;
 
 class DynarmicCallbacks64 : public Dynarmic::A64::UserCallbacks {
 public:
-    explicit DynarmicCallbacks64(ArmDynarmic64& parent, const Kernel::KProcess* process)
+    explicit DynarmicCallbacks64(ArmDynarmic64& parent, Kernel::KProcess* process)
         : m_parent{parent}, m_memory(process->GetMemory()),
           m_process(process), m_debugger_enabled{parent.m_system.DebuggerEnabled()},
           m_check_memory_access{m_debugger_enabled ||
@@ -216,7 +216,7 @@ public:
     Core::Memory::Memory& m_memory;
     u64 m_tpidrro_el0{};
     u64 m_tpidr_el0{};
-    const Kernel::KProcess* m_process{};
+    Kernel::KProcess* m_process{};
     const bool m_debugger_enabled{};
     const bool m_check_memory_access{};
     static constexpr u64 MinimumRunCycles = 10000U;
@@ -399,7 +399,7 @@ void ArmDynarmic64::RewindBreakpointInstruction() {
     this->SetContext(m_breakpoint_context);
 }
 
-ArmDynarmic64::ArmDynarmic64(System& system, bool uses_wall_clock, const Kernel::KProcess* process,
+ArmDynarmic64::ArmDynarmic64(System& system, bool uses_wall_clock, Kernel::KProcess* process,
                              DynarmicExclusiveMonitor& exclusive_monitor, std::size_t core_index)
     : ArmInterface{uses_wall_clock}, m_system{system}, m_exclusive_monitor{exclusive_monitor},
       m_cb(std::make_unique<DynarmicCallbacks64>(*this, process)), m_core_index{core_index} {
diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.h b/src/core/arm/dynarmic/arm_dynarmic_64.h
index 4f3dd026f..08cd982b3 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.h
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.h
@@ -25,7 +25,7 @@ class System;
 
 class ArmDynarmic64 final : public ArmInterface {
 public:
-    ArmDynarmic64(System& system, bool uses_wall_clock, const Kernel::KProcess* process,
+    ArmDynarmic64(System& system, bool uses_wall_clock, Kernel::KProcess* process,
                   DynarmicExclusiveMonitor& exclusive_monitor, std::size_t core_index);
     ~ArmDynarmic64() override;
 
diff --git a/src/core/core.cpp b/src/core/core.cpp
index b14f74976..66f444d39 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -28,7 +28,6 @@
 #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/hid/hid_core.h"
 #include "core/hle/kernel/k_memory_manager.h"
 #include "core/hle/kernel/k_process.h"
@@ -130,11 +129,8 @@ FileSys::VirtualFile GetGameFileFromPath(const FileSys::VirtualFilesystem& vfs,
 
 struct System::Impl {
     explicit Impl(System& system)
-        : kernel{system}, fs_controller{system}, memory{system}, hid_core{}, room_network{},
-          cpu_manager{system}, reporter{system}, applet_manager{system}, profile_manager{},
-          time_manager{system}, gpu_dirty_memory_write_manager{} {
-        memory.SetGPUDirtyManagers(gpu_dirty_memory_write_manager);
-    }
+        : kernel{system}, fs_controller{system}, hid_core{}, room_network{}, cpu_manager{system},
+          reporter{system}, applet_manager{system}, profile_manager{}, time_manager{system} {}
 
     void Initialize(System& system) {
         device_memory = std::make_unique<Core::DeviceMemory>();
@@ -241,17 +237,17 @@ struct System::Impl {
         debugger = std::make_unique<Debugger>(system, port);
     }
 
-    SystemResultStatus SetupForApplicationProcess(System& system, Frontend::EmuWindow& emu_window) {
+    void InitializeKernel(System& system) {
         LOG_DEBUG(Core, "initialized OK");
 
         // Setting changes may require a full system reinitialization (e.g., disabling multicore).
         ReinitializeIfNecessary(system);
 
-        memory.SetGPUDirtyManagers(gpu_dirty_memory_write_manager);
-
         kernel.Initialize();
         cpu_manager.Initialize();
+    }
 
+    SystemResultStatus SetupForApplicationProcess(System& system, Frontend::EmuWindow& emu_window) {
         /// Reset all glue registrations
         arp_manager.ResetAll();
 
@@ -300,17 +296,9 @@ struct System::Impl {
             return SystemResultStatus::ErrorGetLoader;
         }
 
-        SystemResultStatus init_result{SetupForApplicationProcess(system, emu_window)};
-        if (init_result != SystemResultStatus::Success) {
-            LOG_CRITICAL(Core, "Failed to initialize system (Error {})!",
-                         static_cast<int>(init_result));
-            ShutdownMainProcess();
-            return init_result;
-        }
+        InitializeKernel(system);
 
-        telemetry_session->AddInitialInfo(*app_loader, fs_controller, *content_provider);
-
-        // Create the process.
+        // Create the application process.
         auto main_process = Kernel::KProcess::Create(system.Kernel());
         Kernel::KProcess::Register(system.Kernel(), main_process);
         kernel.AppendNewProcess(main_process);
@@ -323,7 +311,18 @@ struct System::Impl {
             return static_cast<SystemResultStatus>(
                 static_cast<u32>(SystemResultStatus::ErrorLoader) + static_cast<u32>(load_result));
         }
+
+        // Set up the rest of the system.
+        SystemResultStatus init_result{SetupForApplicationProcess(system, emu_window)};
+        if (init_result != SystemResultStatus::Success) {
+            LOG_CRITICAL(Core, "Failed to initialize system (Error {})!",
+                         static_cast<int>(init_result));
+            ShutdownMainProcess();
+            return init_result;
+        }
+
         AddGlueRegistrationForProcess(*app_loader, *main_process);
+        telemetry_session->AddInitialInfo(*app_loader, fs_controller, *content_provider);
 
         // Initialize cheat engine
         if (cheat_engine) {
@@ -426,7 +425,6 @@ struct System::Impl {
         cpu_manager.Shutdown();
         debugger.reset();
         kernel.Shutdown();
-        memory.Reset();
         Network::RestartSocketOperations();
 
         if (auto room_member = room_network.GetRoomMember().lock()) {
@@ -507,7 +505,6 @@ struct System::Impl {
     std::unique_ptr<Tegra::Host1x::Host1x> host1x_core;
     std::unique_ptr<Core::DeviceMemory> device_memory;
     std::unique_ptr<AudioCore::AudioCore> audio_core;
-    Core::Memory::Memory memory;
     Core::HID::HIDCore hid_core;
     Network::RoomNetwork room_network;
 
@@ -567,9 +564,6 @@ 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_write_manager{};
-
     std::deque<std::vector<u8>> user_channel;
 };
 
@@ -652,29 +646,12 @@ void System::PrepareReschedule(const u32 core_index) {
     impl->kernel.PrepareReschedule(core_index);
 }
 
-Core::GPUDirtyMemoryManager& System::CurrentGPUDirtyMemoryManager() {
-    const std::size_t core = impl->kernel.GetCurrentHostThreadID();
-    return impl->gpu_dirty_memory_write_manager[core < Core::Hardware::NUM_CPU_CORES
-                                                    ? core
-                                                    : Core::Hardware::NUM_CPU_CORES - 1];
-}
-
-/// Provides a constant reference to the current gou dirty memory manager.
-const Core::GPUDirtyMemoryManager& System::CurrentGPUDirtyMemoryManager() const {
-    const std::size_t core = impl->kernel.GetCurrentHostThreadID();
-    return impl->gpu_dirty_memory_write_manager[core < Core::Hardware::NUM_CPU_CORES
-                                                    ? core
-                                                    : Core::Hardware::NUM_CPU_CORES - 1];
-}
-
 size_t System::GetCurrentHostThreadID() const {
     return impl->kernel.GetCurrentHostThreadID();
 }
 
 void System::GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback) {
-    for (auto& manager : impl->gpu_dirty_memory_write_manager) {
-        manager.Gather(callback);
-    }
+    return this->ApplicationProcess()->GatherGPUDirtyMemory(callback);
 }
 
 PerfStatsResults System::GetAndResetPerfStats() {
@@ -723,20 +700,12 @@ const Kernel::KProcess* System::ApplicationProcess() const {
     return impl->kernel.ApplicationProcess();
 }
 
-ExclusiveMonitor& System::Monitor() {
-    return impl->kernel.GetExclusiveMonitor();
-}
-
-const ExclusiveMonitor& System::Monitor() const {
-    return impl->kernel.GetExclusiveMonitor();
-}
-
 Memory::Memory& System::ApplicationMemory() {
-    return impl->memory;
+    return impl->kernel.ApplicationProcess()->GetMemory();
 }
 
 const Core::Memory::Memory& System::ApplicationMemory() const {
-    return impl->memory;
+    return impl->kernel.ApplicationProcess()->GetMemory();
 }
 
 Tegra::GPU& System::GPU() {
diff --git a/src/core/core.h b/src/core/core.h
index 473204db7..ba5add0dc 100644
--- a/src/core/core.h
+++ b/src/core/core.h
@@ -116,7 +116,6 @@ class CpuManager;
 class Debugger;
 class DeviceMemory;
 class ExclusiveMonitor;
-class GPUDirtyMemoryManager;
 class PerfStats;
 class Reporter;
 class SpeedLimiter;
@@ -225,12 +224,6 @@ public:
     /// Prepare the core emulation for a reschedule
     void PrepareReschedule(u32 core_index);
 
-    /// Provides a reference to the gou dirty memory manager.
-    [[nodiscard]] Core::GPUDirtyMemoryManager& CurrentGPUDirtyMemoryManager();
-
-    /// Provides a constant reference to the current gou dirty memory manager.
-    [[nodiscard]] const Core::GPUDirtyMemoryManager& CurrentGPUDirtyMemoryManager() const;
-
     void GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback);
 
     [[nodiscard]] size_t GetCurrentHostThreadID() const;
@@ -250,12 +243,6 @@ public:
     /// Gets a const reference to the underlying CPU manager
     [[nodiscard]] const CpuManager& GetCpuManager() const;
 
-    /// Gets a reference to the exclusive monitor
-    [[nodiscard]] ExclusiveMonitor& Monitor();
-
-    /// Gets a constant reference to the exclusive monitor
-    [[nodiscard]] const ExclusiveMonitor& Monitor() const;
-
     /// Gets a mutable reference to the system memory instance.
     [[nodiscard]] Core::Memory::Memory& ApplicationMemory();
 
diff --git a/src/core/hle/kernel/k_address_arbiter.cpp b/src/core/hle/kernel/k_address_arbiter.cpp
index 78d43d729..48889253d 100644
--- a/src/core/hle/kernel/k_address_arbiter.cpp
+++ b/src/core/hle/kernel/k_address_arbiter.cpp
@@ -4,6 +4,7 @@
 #include "core/arm/exclusive_monitor.h"
 #include "core/core.h"
 #include "core/hle/kernel/k_address_arbiter.h"
+#include "core/hle/kernel/k_process.h"
 #include "core/hle/kernel/k_scheduler.h"
 #include "core/hle/kernel/k_scoped_scheduler_lock_and_sleep.h"
 #include "core/hle/kernel/k_thread.h"
@@ -26,9 +27,9 @@ bool ReadFromUser(KernelCore& kernel, s32* out, KProcessAddress address) {
     return true;
 }
 
-bool DecrementIfLessThan(Core::System& system, s32* out, KProcessAddress address, s32 value) {
-    auto& monitor = system.Monitor();
-    const auto current_core = system.Kernel().CurrentPhysicalCoreIndex();
+bool DecrementIfLessThan(KernelCore& kernel, s32* out, KProcessAddress address, s32 value) {
+    auto& monitor = GetCurrentProcess(kernel).GetExclusiveMonitor();
+    const auto current_core = kernel.CurrentPhysicalCoreIndex();
 
     // NOTE: If scheduler lock is not held here, interrupt disable is required.
     // KScopedInterruptDisable di;
@@ -66,10 +67,10 @@ bool DecrementIfLessThan(Core::System& system, s32* out, KProcessAddress address
     return true;
 }
 
-bool UpdateIfEqual(Core::System& system, s32* out, KProcessAddress address, s32 value,
+bool UpdateIfEqual(KernelCore& kernel, s32* out, KProcessAddress address, s32 value,
                    s32 new_value) {
-    auto& monitor = system.Monitor();
-    const auto current_core = system.Kernel().CurrentPhysicalCoreIndex();
+    auto& monitor = GetCurrentProcess(kernel).GetExclusiveMonitor();
+    const auto current_core = kernel.CurrentPhysicalCoreIndex();
 
     // NOTE: If scheduler lock is not held here, interrupt disable is required.
     // KScopedInterruptDisable di;
@@ -159,7 +160,7 @@ Result KAddressArbiter::SignalAndIncrementIfEqual(uint64_t addr, s32 value, s32
 
         // Check the userspace value.
         s32 user_value{};
-        R_UNLESS(UpdateIfEqual(m_system, std::addressof(user_value), addr, value, value + 1),
+        R_UNLESS(UpdateIfEqual(m_kernel, std::addressof(user_value), addr, value, value + 1),
                  ResultInvalidCurrentMemory);
         R_UNLESS(user_value == value, ResultInvalidState);
 
@@ -219,7 +220,7 @@ Result KAddressArbiter::SignalAndModifyByWaitingCountIfEqual(uint64_t addr, s32
         s32 user_value{};
         bool succeeded{};
         if (value != new_value) {
-            succeeded = UpdateIfEqual(m_system, std::addressof(user_value), addr, value, new_value);
+            succeeded = UpdateIfEqual(m_kernel, std::addressof(user_value), addr, value, new_value);
         } else {
             succeeded = ReadFromUser(m_kernel, std::addressof(user_value), addr);
         }
@@ -262,7 +263,7 @@ Result KAddressArbiter::WaitIfLessThan(uint64_t addr, s32 value, bool decrement,
         s32 user_value{};
         bool succeeded{};
         if (decrement) {
-            succeeded = DecrementIfLessThan(m_system, std::addressof(user_value), addr, value);
+            succeeded = DecrementIfLessThan(m_kernel, std::addressof(user_value), addr, value);
         } else {
             succeeded = ReadFromUser(m_kernel, std::addressof(user_value), addr);
         }
diff --git a/src/core/hle/kernel/k_condition_variable.cpp b/src/core/hle/kernel/k_condition_variable.cpp
index 7633a51fb..94ea3527a 100644
--- a/src/core/hle/kernel/k_condition_variable.cpp
+++ b/src/core/hle/kernel/k_condition_variable.cpp
@@ -28,10 +28,10 @@ bool WriteToUser(KernelCore& kernel, KProcessAddress address, const u32* p) {
     return true;
 }
 
-bool UpdateLockAtomic(Core::System& system, u32* out, KProcessAddress address, u32 if_zero,
+bool UpdateLockAtomic(KernelCore& kernel, u32* out, KProcessAddress address, u32 if_zero,
                       u32 new_orr_mask) {
-    auto& monitor = system.Monitor();
-    const auto current_core = system.Kernel().CurrentPhysicalCoreIndex();
+    auto& monitor = GetCurrentProcess(kernel).GetExclusiveMonitor();
+    const auto current_core = kernel.CurrentPhysicalCoreIndex();
 
     u32 expected{};
 
@@ -208,7 +208,7 @@ void KConditionVariable::SignalImpl(KThread* thread) {
         // TODO(bunnei): We should call CanAccessAtomic(..) here.
         can_access = true;
         if (can_access) [[likely]] {
-            UpdateLockAtomic(m_system, std::addressof(prev_tag), address, own_tag,
+            UpdateLockAtomic(m_kernel, std::addressof(prev_tag), address, own_tag,
                              Svc::HandleWaitMask);
         }
     }
diff --git a/src/core/hle/kernel/k_process.cpp b/src/core/hle/kernel/k_process.cpp
index 3a2635e1f..905d141ea 100644
--- a/src/core/hle/kernel/k_process.cpp
+++ b/src/core/hle/kernel/k_process.cpp
@@ -1128,7 +1128,8 @@ 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_handle_table{kernel}, m_dirty_memory_managers{}, m_exclusive_monitor{},
+      m_memory{kernel.System()} {}
 KProcess::~KProcess() = default;
 
 Result KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std::size_t code_size,
@@ -1235,7 +1236,11 @@ void KProcess::LoadModule(CodeSet code_set, KProcessAddress base_addr) {
 }
 
 void KProcess::InitializeInterfaces() {
+    m_exclusive_monitor =
+        Core::MakeExclusiveMonitor(this->GetMemory(), Core::Hardware::NUM_CPU_CORES);
+
     this->GetMemory().SetCurrentPageTable(*this);
+    this->GetMemory().SetGPUDirtyManagers(m_dirty_memory_managers);
 
 #ifdef HAS_NCE
     if (this->Is64Bit() && Settings::IsNceEnabled()) {
@@ -1248,13 +1253,13 @@ void KProcess::InitializeInterfaces() {
         for (size_t i = 0; i < Core::Hardware::NUM_CPU_CORES; i++) {
             m_arm_interfaces[i] = std::make_unique<Core::ArmDynarmic64>(
                 m_kernel.System(), m_kernel.IsMulticore(), this,
-                static_cast<Core::DynarmicExclusiveMonitor&>(m_kernel.GetExclusiveMonitor()), i);
+                static_cast<Core::DynarmicExclusiveMonitor&>(*m_exclusive_monitor), i);
         }
     } else {
         for (size_t i = 0; i < Core::Hardware::NUM_CPU_CORES; i++) {
             m_arm_interfaces[i] = std::make_unique<Core::ArmDynarmic32>(
                 m_kernel.System(), m_kernel.IsMulticore(), this,
-                static_cast<Core::DynarmicExclusiveMonitor&>(m_kernel.GetExclusiveMonitor()), i);
+                static_cast<Core::DynarmicExclusiveMonitor&>(*m_exclusive_monitor), i);
         }
     }
 }
@@ -1305,9 +1310,10 @@ bool KProcess::RemoveWatchpoint(KProcessAddress addr, u64 size, DebugWatchpointT
     return true;
 }
 
-Core::Memory::Memory& KProcess::GetMemory() const {
-    // TODO: per-process memory
-    return m_kernel.System().ApplicationMemory();
+void KProcess::GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback) {
+    for (auto& manager : m_dirty_memory_managers) {
+        manager.Gather(callback);
+    }
 }
 
 } // namespace Kernel
diff --git a/src/core/hle/kernel/k_process.h b/src/core/hle/kernel/k_process.h
index 4b114e39b..53c0e3316 100644
--- a/src/core/hle/kernel/k_process.h
+++ b/src/core/hle/kernel/k_process.h
@@ -7,6 +7,7 @@
 
 #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"
@@ -17,6 +18,7 @@
 #include "core/hle/kernel/k_system_resource.h"
 #include "core/hle/kernel/k_thread.h"
 #include "core/hle/kernel/k_thread_local_page.h"
+#include "core/memory.h"
 
 namespace Kernel {
 
@@ -126,6 +128,9 @@ 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;
 
 private:
     Result StartTermination();
@@ -502,7 +507,15 @@ public:
 
     void InitializeInterfaces();
 
-    Core::Memory::Memory& GetMemory() const;
+    Core::Memory::Memory& GetMemory() {
+        return m_memory;
+    }
+
+    void GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback);
+
+    Core::ExclusiveMonitor& GetExclusiveMonitor() const {
+        return *m_exclusive_monitor;
+    }
 
 public:
     // Overridden parent functions.
diff --git a/src/core/hle/kernel/k_thread.h b/src/core/hle/kernel/k_thread.h
index e9925d231..f13e232b2 100644
--- a/src/core/hle/kernel/k_thread.h
+++ b/src/core/hle/kernel/k_thread.h
@@ -314,11 +314,7 @@ public:
         m_current_core_id = core;
     }
 
-    KProcess* GetOwnerProcess() {
-        return m_parent;
-    }
-
-    const KProcess* GetOwnerProcess() const {
+    KProcess* GetOwnerProcess() const {
         return m_parent;
     }
 
diff --git a/src/core/hle/kernel/kernel.cpp b/src/core/hle/kernel/kernel.cpp
index e479dacde..2efca27c2 100644
--- a/src/core/hle/kernel/kernel.cpp
+++ b/src/core/hle/kernel/kernel.cpp
@@ -126,8 +126,6 @@ struct KernelCore::Impl {
 
         preemption_event = nullptr;
 
-        exclusive_monitor.reset();
-
         // Cleanup persistent kernel objects
         auto CleanupObject = [](KAutoObject* obj) {
             if (obj) {
@@ -191,8 +189,6 @@ struct KernelCore::Impl {
     }
 
     void InitializePhysicalCores() {
-        exclusive_monitor =
-            Core::MakeExclusiveMonitor(system.ApplicationMemory(), Core::Hardware::NUM_CPU_CORES);
         for (u32 i = 0; i < Core::Hardware::NUM_CPU_CORES; i++) {
             const s32 core{static_cast<s32>(i)};
 
@@ -805,7 +801,6 @@ struct KernelCore::Impl {
     std::mutex server_lock;
     std::vector<std::unique_ptr<Service::ServerManager>> server_managers;
 
-    std::unique_ptr<Core::ExclusiveMonitor> exclusive_monitor;
     std::array<std::unique_ptr<Kernel::PhysicalCore>, Core::Hardware::NUM_CPU_CORES> cores;
 
     // Next host thead ID to use, 0-3 IDs represent core threads, >3 represent others
@@ -959,14 +954,6 @@ Kernel::KHardwareTimer& KernelCore::HardwareTimer() {
     return *impl->hardware_timer;
 }
 
-Core::ExclusiveMonitor& KernelCore::GetExclusiveMonitor() {
-    return *impl->exclusive_monitor;
-}
-
-const Core::ExclusiveMonitor& KernelCore::GetExclusiveMonitor() const {
-    return *impl->exclusive_monitor;
-}
-
 KAutoObjectWithListContainer& KernelCore::ObjectListContainer() {
     return *impl->global_object_list_container;
 }
diff --git a/src/core/hle/kernel/kernel.h b/src/core/hle/kernel/kernel.h
index 78c88902c..fefd7aaba 100644
--- a/src/core/hle/kernel/kernel.h
+++ b/src/core/hle/kernel/kernel.h
@@ -170,10 +170,6 @@ public:
     /// Stops execution of 'id' core, in order to reschedule a new thread.
     void PrepareReschedule(std::size_t id);
 
-    Core::ExclusiveMonitor& GetExclusiveMonitor();
-
-    const Core::ExclusiveMonitor& GetExclusiveMonitor() const;
-
     KAutoObjectWithListContainer& ObjectListContainer();
 
     const KAutoObjectWithListContainer& ObjectListContainer() const;