Merge pull request #12579 from FernandoS27/smmu

Core: Implement Device Mapping & GPU SMMU

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
@@ -609,6 +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/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"
@@ -565,6 +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;
 };
 
@@ -651,8 +655,14 @@ size_t System::GetCurrentHostThreadID() const {
     return impl->kernel.GetCurrentHostThreadID();
 }
 
-void System::GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback) {
-    return this->ApplicationProcess()->GatherGPUDirtyMemory(callback);
+std::span<GPUDirtyMemoryManager> System::GetGPUDirtyMemoryManager() {
+    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/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 +1,211 @@
+// 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;
+
+struct Asid {
+    size_t id;
+};
+
+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, Asid asid, bool track = false);
+
+    void Unmap(DAddr address, size_t size);
+
+    void TrackContinuityImpl(DAddr address, VAddr virtual_address, size_t size, Asid asid);
+    void TrackContinuity(DAddr address, VAddr virtual_address, size_t size, Asid asid) {
+        std::scoped_lock lk(mapping_guard);
+        TrackContinuityImpl(address, virtual_address, size, asid);
+    }
+
+    // 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);
+
+    Asid RegisterProcess(Memory::Memory* memory);
+    void UnregisterProcess(Asid 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 asid_start_bit = guest_max_as_bits;
+
+    std::pair<Asid, VAddr> ExtractCPUBacking(size_t page_index) {
+        auto content = cpu_backing_address[page_index];
+        const VAddr address = content & guest_mask;
+        const Asid asid{static_cast<size_t>(content >> asid_start_bit)};
+        return std::make_pair(asid, address);
+    }
+
+    void InsertCPUBacking(size_t page_index, VAddr address, Asid asid) {
+        cpu_backing_address[page_index] = address | (asid.id << asid_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 +1,582 @@
+// 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,
+                                      Asid asid, bool track) {
+    Core::Memory::Memory* process_memory = registered_processes[asid.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, asid);
+        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, asid);
+    }
+}
+
+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, Asid asid) {
+    Core::Memory::Memory* process_memory = registered_processes[asid.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>
+Asid 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 Asid{new_id};
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UnregisterProcess(Asid asid) {
+    registered_processes[asid.id] = nullptr;
+    id_pool.push_front(asid.id);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
+    std::unique_lock<std::mutex> lk(counter_guard, std::defer_lock);
+    const auto Lock = [&] {
+        if (!lk) {
+            lk.lock();
+        }
+    };
+    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 [asid, base_vaddress] = ExtractCPUBacking(page);
+    size_t vpage = base_vaddress >> Memory::YUZU_PAGEBITS;
+    auto* memory_device_inter = registered_processes[asid.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/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_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"
@@ -320,7 +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,
@@ -417,7 +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),
@@ -1141,8 +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_exclusive_monitor{}, m_memory{kernel.System()} {}
+      m_handle_table{kernel}, m_exclusive_monitor{}, m_memory{kernel.System()} {}
 KProcess::~KProcess() = default;
 
 Result KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std::size_t code_size,
@@ -1324,10 +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/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"
@@ -23,19 +24,6 @@
 #include "core/hle/service/ipc_helpers.h"
 #include "core/memory.h"
 
-namespace {
-static thread_local std::array read_buffer_data_a{
-    Common::ScratchBuffer<u8>(),
-    Common::ScratchBuffer<u8>(),
-    Common::ScratchBuffer<u8>(),
-};
-static thread_local std::array read_buffer_data_x{
-    Common::ScratchBuffer<u8>(),
-    Common::ScratchBuffer<u8>(),
-    Common::ScratchBuffer<u8>(),
-};
-} // Anonymous namespace
-
 namespace Service {
 
 SessionRequestHandler::SessionRequestHandler(Kernel::KernelCore& kernel_, const char* service_name_)
@@ -343,48 +331,27 @@ 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::SafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
-    };
+    Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> gm(memory, 0, 0);
 
     ASSERT_OR_EXECUTE_MSG(
         BufferDescriptorA().size() > buffer_index, { return {}; },
         "BufferDescriptorA invalid buffer_index {}", buffer_index);
-    auto& read_buffer = read_buffer_a[buffer_index];
-    return read_buffer.Read(BufferDescriptorA()[buffer_index].Address(),
-                            BufferDescriptorA()[buffer_index].Size(),
-                            &read_buffer_data_a[buffer_index]);
+    return gm.Read(BufferDescriptorA()[buffer_index].Address(),
+                   BufferDescriptorA()[buffer_index].Size(), &read_buffer_data_a[buffer_index]);
 }
 
 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::SafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
-    };
+    Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> gm(memory, 0, 0);
 
     ASSERT_OR_EXECUTE_MSG(
         BufferDescriptorX().size() > buffer_index, { return {}; },
         "BufferDescriptorX invalid buffer_index {}", buffer_index);
-    auto& read_buffer = read_buffer_x[buffer_index];
-    return read_buffer.Read(BufferDescriptorX()[buffer_index].Address(),
-                            BufferDescriptorX()[buffer_index].Size(),
-                            &read_buffer_data_x[buffer_index]);
+    return gm.Read(BufferDescriptorX()[buffer_index].Address(),
+                   BufferDescriptorX()[buffer_index].Size(), &read_buffer_data_x[buffer_index]);
 }
 
 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::SafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(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::SafeRead>(memory, 0, 0),
-        Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead>(memory, 0, 0),
-    };
+    Core::Memory::CpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> gm(memory, 0, 0);
 
     const bool is_buffer_a{BufferDescriptorA().size() > buffer_index &&
                            BufferDescriptorA()[buffer_index].Size()};
@@ -401,18 +368,14 @@ std::span<const u8> HLERequestContext::ReadBuffer(std::size_t buffer_index) cons
         ASSERT_OR_EXECUTE_MSG(
             BufferDescriptorA().size() > buffer_index, { return {}; },
             "BufferDescriptorA invalid buffer_index {}", buffer_index);
-        auto& read_buffer = read_buffer_a[buffer_index];
-        return read_buffer.Read(BufferDescriptorA()[buffer_index].Address(),
-                                BufferDescriptorA()[buffer_index].Size(),
-                                &read_buffer_data_a[buffer_index]);
+        return gm.Read(BufferDescriptorA()[buffer_index].Address(),
+                       BufferDescriptorA()[buffer_index].Size(), &read_buffer_data_a[buffer_index]);
     } else {
         ASSERT_OR_EXECUTE_MSG(
             BufferDescriptorX().size() > buffer_index, { return {}; },
             "BufferDescriptorX invalid buffer_index {}", buffer_index);
-        auto& read_buffer = read_buffer_x[buffer_index];
-        return read_buffer.Read(BufferDescriptorX()[buffer_index].Address(),
-                                BufferDescriptorX()[buffer_index].Size(),
-                                &read_buffer_data_x[buffer_index]);
+        return gm.Read(BufferDescriptorX()[buffer_index].Address(),
+                       BufferDescriptorX()[buffer_index].Size(), &read_buffer_data_x[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
 
@@ -424,6 +426,9 @@ private:
 
     Kernel::KernelCore& kernel;
     Core::Memory::Memory& memory;
+
+    mutable std::array<Common::ScratchBuffer<u8>, 3> read_buffer_data_a{};
+    mutable std::array<Common::ScratchBuffer<u8>, 3> read_buffer_data_x{};
 };
 
 } // namespace Service

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

@@ -2,27 +2,135 @@
 // 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(SessionId id_, Kernel::KProcess* process_, Core::Asid asid_)
+    : id{id_}, process{process_}, asid{asid_}, has_preallocated_area{}, mapper{}, is_active{} {}
+
+Session::~Session() = default;
+
 struct ContainerImpl {
-    explicit ContainerImpl(Tegra::Host1x::Host1x& host1x_)
-        : file{host1x_}, manager{host1x_}, device_file_data{} {}
+    explicit ContainerImpl(Container& core, Tegra::Host1x::Host1x& host1x_)
+        : 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;
 
+SessionId Container::OpenSession(Kernel::KProcess* process) {
+    using namespace Common::Literals;
+
+    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 asid = 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{SessionId{new_id}, process, asid};
+    } else {
+        new_id = impl->new_ids++;
+        impl->sessions.emplace_back(SessionId{new_id}, process, asid);
+    }
+    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 >= 32_MiB ? smmu.Allocate(region_size) : 0;
+        if (start_region != 0) {
+            session.mapper = std::make_unique<HeapMapper>(region_start, start_region, region_size,
+                                                          asid, impl->host1x);
+            smmu.TrackContinuity(start_region, region_start, region_size, asid);
+            session.has_preallocated_area = true;
+            LOG_DEBUG(Debug, "Preallocation created!");
+        }
+    }
+    return SessionId{new_id};
+}
+
+void Container::CloseSession(SessionId session_id) {
+    std::scoped_lock lk(impl->session_guard);
+    auto& session = impl->sessions[session_id.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[session_id.id].asid);
+    impl->id_pool.emplace_front(session_id.id);
+}
+
+Session* Container::GetSession(SessionId session_id) {
+    std::atomic_thread_fence(std::memory_order_acquire);
+    return &impl->sessions[session_id.id];
+}
+
 NvMap& Container::GetNvMapFile() {
     return impl->file;
 }

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

@@ -8,24 +8,56 @@
 #include <memory>
 #include <unordered_map>
 
+#include "core/device_memory_manager.h"
 #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 SessionId {
+    size_t id;
+};
+
+struct Session {
+    Session(SessionId id_, Kernel::KProcess* process_, Core::Asid asid_);
+    ~Session();
+
+    Session(const Session&) = delete;
+    Session& operator=(const Session&) = delete;
+    Session(Session&&) = default;
+    Session& operator=(Session&&) = default;
+
+    SessionId id;
+    Kernel::KProcess* process;
+    Core::Asid asid;
+    bool has_preallocated_area{};
+    std::unique_ptr<HeapMapper> mapper{};
+    bool is_active{};
+};
+
 class Container {
 public:
     explicit Container(Tegra::Host1x::Host1x& host1x);
     ~Container();
 
+    SessionId OpenSession(Kernel::KProcess* process);
+    void CloseSession(SessionId id);
+
+    Session* GetSession(SessionId 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, Core::Asid asid,
+                       Tegra::Host1x::Host1x& host1x)
+    : m_vaddress{start_vaddress}, m_daddress{start_daddress}, m_size{size}, m_asid{asid} {
+    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_asid);
+        }
+    }
+    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 +1,49 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <memory>
+
+#include "common/common_types.h"
+#include "core/device_memory_manager.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, Core::Asid asid,
+               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;
+    Core::Asid m_asid;
+    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 +22,9 @@ 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,
+                              NvCore::SessionId pSessionId) {
     std::scoped_lock lock(mutex);
-
     // Handles cannot be allocated twice
     if (allocated) {
         return NvResult::AccessDenied;
@@ -28,6 +33,7 @@ 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 +69,7 @@ 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 +84,30 @@ 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),
-                                 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;
+    const size_t map_size = handle_description.aligned_size;
+    if (!handle_description.in_heap) {
+        auto& smmu = host1x.MemoryManager();
+        size_t aligned_up = Common::AlignUp(map_size, BIG_PAGE_SIZE);
+        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 +148,33 @@ 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 +186,58 @@ 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{};
-        auto& smmu_allocator = host1x.Allocator();
-        auto& smmu_memory_manager = host1x.MemoryManager();
-        while ((address = smmu_allocator.Allocate(
-                    static_cast<u32>(handle_description->aligned_size))) == 0) {
-            // Free handles until the allocation succeeds
-            std::scoped_lock queueLock(unmap_queue_lock);
-            if (auto freeHandleDesc{unmap_queue.front()}) {
-                // 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->pin_virt_address)
-                    UnmapHandle(*freeHandleDesc);
-            } else {
-                LOG_CRITICAL(Service_NVDRV, "Ran out of SMMU address space!");
+        DAddr address{};
+        auto& smmu = host1x.MemoryManager();
+        auto* session = core.GetSession(handle_description->session_id);
+        const VAddr vaddress = handle_description->address;
+        const size_t map_size = handle_description->aligned_size;
+        if (session->has_preallocated_area && session->mapper->IsInBounds(vaddress, map_size)) {
+            handle_description->d_address = session->mapper->Map(vaddress, map_size);
+            handle_description->in_heap = true;
+        } else {
+            size_t aligned_up = Common::AlignUp(map_size, BIG_PAGE_SIZE);
+            while ((address = smmu.Allocate(aligned_up)) == 0) {
+                // Free handles until the allocation succeeds
+                std::scoped_lock queueLock(unmap_queue_lock);
+                if (auto freeHandleDesc{unmap_queue.front()}) {
+                    // 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->aligned_size);
-        handle_description->pin_virt_address = address;
+            handle_description->d_address = address;
+            smmu.Map(address, vaddress, map_size, session->asid, true);
+            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 +288,7 @@ 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

@@ -14,6 +14,7 @@
 
 #include "common/bit_field.h"
 #include "common/common_types.h"
+#include "core/hle/service/nvdrv/core/container.h"
 #include "core/hle/service/nvdrv/nvdata.h"
 
 namespace Tegra {
@@ -25,6 +26,8 @@ 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 +51,7 @@ 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 +64,18 @@ public:
         } flags{};
         static_assert(sizeof(Flags) == sizeof(u32));
 
-        u64 address{}; //!< The memory location in the guest's AS that this handle corresponds to,
-                       //!< this can also be in the nvdrv tmem
+        VAddr address{}; //!< The memory location in the guest's AS that this handle corresponds to,
+                         //!< 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{};
+        NvCore::SessionId 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 +83,8 @@ 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,
+                                     NvCore::SessionId pSessionId);
 
         /**
          * @brief Increases the dupe counter of the handle for the given session
@@ -108,7 +115,7 @@ 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 +124,7 @@ 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 +132,7 @@ 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 +179,7 @@ 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

@@ -7,6 +7,7 @@
 #include <vector>
 
 #include "common/common_types.h"
+#include "core/hle/service/nvdrv/core/container.h"
 #include "core/hle/service/nvdrv/nvdata.h"
 
 namespace Core {
@@ -62,7 +63,7 @@ public:
      * Called once a device is opened
      * @param fd The device fd
      */
-    virtual void OnOpen(DeviceFD fd) = 0;
+    virtual void OnOpen(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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>(
-                handle->address +
-                (static_cast<u64>(entry.handle_offset_big_pages) << vm.big_page_size_bits))};
+            DAddr base = nvmap.PinHandle(entry.handle, false);
+            DAddr device_address{static_cast<DAddr>(
+                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),
-                      use_big_pages);
+            gmmu->Map(virtual_address, device_address, size,
+                      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,
-                                               use_big_pages, alloc->second.sparse)};
+        auto mapping{std::make_shared<Mapping>(params.handle, device_address, params.offset, size,
+                                               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{
-            std::make_shared<Mapping>(cpu_address, params.offset, size, false, big_page, false)};
+        auto mapping{std::make_shared<Mapping>(params.handle, device_address, params.offset, size,
+                                               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(NvCore::SessionId 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_)
-            : ptr(ptr_), offset(offset_), size(size_), fixed(fixed_), big_page(big_page_),
-              sparse_alloc(sparse_alloc_) {}
+        Mapping(NvCore::NvMap::Handle::Id handle_, DAddr ptr_, u64 offset_, u64 size_, bool fixed_,
+                bool big_page_, bool 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId 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(),
-                                             cmdlist.size() * sizeof(u32));
+        session->process->GetMemory().ReadBlock(object->address + cmd_buffer.offset, cmdlist.data(),
+                                                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, NvCore::SessionId> 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(NvCore::SessionId 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(NvCore::SessionId 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(NvCore::SessionId session_id, DeviceFD fd) {
+    sessions[fd] = session_id;
+}
 
 void nvhost_vic::OnClose(DeviceFD fd) {
     auto& host1x_file = core.Host1xDeviceFile();
@@ -76,6 +78,7 @@ void nvhost_vic::OnClose(DeviceFD fd) {
     if (iter != host1x_file.fd_to_id.end()) {
         system.GPU().ClearCdmaInstance(iter->second);
     }
+    sessions.erase(fd);
 }
 
 } // 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(NvCore::SessionId 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::OnClose(DeviceFD fd) {}
+void nvmap::OnOpen(NvCore::SessionId session_id, 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 =
-        handle_description->Alloc(params.flags, params.align, params.kind, params.address);
+    const auto result = handle_description->Alloc(params.flags, params.align, params.kind,
+                                                  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()
-               ->GetPageTable()
+    auto process = container.GetSession(sessions[fd])->process;
+    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()
-                       ->GetPageTable()
+            ASSERT(process->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(NvCore::SessionId 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, NvCore::SessionId> 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"));
-    server_manager->RegisterNamedService("nvdrv:a",
-                                         std::make_shared<NVDRV>(system, module, "nvdrv:a"));
-    server_manager->RegisterNamedService("nvdrv:s",
-                                         std::make_shared<NVDRV>(system, module, "nvdrv:s"));
-    server_manager->RegisterNamedService("nvdrv:t",
-                                         std::make_shared<NVDRV>(system, module, "nvdrv:t"));
+    const auto NvdrvInterfaceFactoryForApplication = [&, module] {
+        return std::make_shared<NVDRV>(system, module, "nvdrv");
+    };
+    const auto NvdrvInterfaceFactoryForApplets = [&, module] {
+        return std::make_shared<NVDRV>(system, module, "nvdrv:a");
+    };
+    const auto NvdrvInterfaceFactoryForSysmodules = [&, module] {
+        return std::make_shared<NVDRV>(system, module, "nvdrv:s");
+    };
+    const auto NvdrvInterfaceFactoryForTesting = [&, 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", NvdrvInterfaceFactoryForTesting);
     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, NvCore::SessionId 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, NvCore::SessionId 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{};
+    NvCore::SessionId 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,
-                                  SharedBufferSize));
+    R_TRY(AllocateHandleForBuffer(std::addressof(m_buffer_nvmap_handle), *m_nvdrv, m_nvmap_fd,
+                                  map_address, SharedBufferSize));
 
     // Record the display id.
     m_display_id = display_id;

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

@@ -4,6 +4,8 @@
 #pragma once
 
 #include "common/math_util.h"
+#include "core/hle/service/nvdrv/core/container.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 +55,8 @@ private:
     u64 m_layer_id = 0;
     u32 m_buffer_nvmap_handle = 0;
     SharedMemoryPoolLayout m_pool_layout = {};
-
+    Nvidia::DeviceFD m_nvmap_fd = {};
+    Nvidia::NvCore::SessionId m_session_id = {};
     std::unique_ptr<Kernel::KPageGroup> m_buffer_page_group;
 
     std::mutex m_guard;

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

@@ -124,7 +124,7 @@ 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", {});
 }
 
 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/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;
-        if (current_area.start_address <= address && end_address <= current_area.end_address)
-            [[likely]] {
-            return;
-        }
-        current_area = system.GPU().OnCPURead(address, size);
+        gpu_device_memory->ApplyOpOnPointer(p, scratch_buffers[core], [&](DAddr address) {
+            const DAddr end_address = address + size;
+            if (current_area.start_address <= address && end_address <= current_area.end_address)
+                [[likely]] {
+                return;
+            }
+            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];
-        VAddr subaddress = address >> YUZU_PAGEBITS;
-        bool do_collection = current_area.last_address == subaddress;
-        if (!do_collection) [[unlikely]] {
-            do_collection = system.GPU().OnCPUWrite(address, size);
-            if (!do_collection) {
-                return;
+        gpu_device_memory->ApplyOpOnPointer(p, scratch_buffers[core], [&](DAddr address) {
+            auto& current_area = rasterizer_write_areas[core];
+            PAddr subaddress = address >> YUZU_PAGEBITS;
+            bool do_collection = current_area.last_address == subaddress;
+            if (!do_collection) [[unlikely]] {
+                do_collection = system.GPU().OnCPUWrite(address, size);
+                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) {
-        system.GPU().InvalidateRegion(GetInteger(dest_addr), size);
-    }
-
-    void FlushRegion(Common::ProcessAddress dest_addr, size_t size) {
-        system.GPU().FlushRegion(GetInteger(dest_addr), size);
+    void InvalidateGPUMemory(u8* p, size_t 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.
+        if (!gpu_device_memory) [[unlikely]] {
+            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;
         },
-        [&] {
-            impl->system.GPU().InvalidateRegion(GetInteger(vaddr), size);
-            rasterizer = true;
-        });
+        [&] { rasterizer = true; });
+    if (rasterizer) {
+        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>;
-template <typename T, GuestMemoryFlags FLAGS>
-using GpuGuestMemory = GuestMemory<Tegra::MemoryManager, T, FLAGS>;
-template <typename T, GuestMemoryFlags FLAGS>
-using GpuGuestMemoryScoped = GuestMemoryScoped<Tegra::MemoryManager, T, FLAGS>;
+using CpuGuestMemoryScoped = GuestMemoryScoped<Core::Memory::Memory, T, FLAGS>;
+
 } // namespace Core::Memory