Core: Initial implementation of device memory mapping

2 years ago · 04b4af6044
6 changed files with 459 additions and 0 deletions
--- a/src/common/common_types.h
+++ b/src/common/common_types.h
@ -45,6 +45,7 @@ using f32 = float;  ///< 32-bit floating point
 using f64 = double; ///< 64-bit floating point

 using VAddr = u64;    ///< Represents a pointer in the userspace virtual address space.
+using DAddr = u64;    ///< Represents a pointer in the device specific virtual address space.
 using PAddr = u64;    ///< Represents a pointer in the ARM11 physical address space.
 using GPUVAddr = u64; ///< Represents a pointer in the GPU virtual address space.

--- a/src/core/device_memory.h
+++ b/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;
 };

--- a/src/core/device_memory_manager.h
+++ b/src/core/device_memory_manager.h
@ -0,0 +1,97 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <deque>
+#include <memory>
+
+#include "common/common_types.h"
+#include "common/virtual_buffer.h"
+
+namespace Core {
+
+class DeviceMemory;
+
+namespace Memory {
+class Memory;
+}
+
+template <typename DTraits>
+struct DeviceMemoryManagerAllocator;
+
+template <typename Traits>
+class DeviceMemoryManager {
+    using DeviceInterface = typename Traits::DeviceInterface;
+
+public:
+    DeviceMemoryManager(const DeviceMemory& device_memory);
+    ~DeviceMemoryManager();
+
+    void BindInterface(DeviceInterface* interface);
+
+    DAddr Allocate(size_t size);
+    void AllocateFixed(DAddr start, size_t size);
+    DAddr AllocatePinned(size_t size);
+    void Free(DAddr start, size_t size);
+
+    void Map(DAddr address, VAddr virtual_address, size_t size, size_t p_id);
+    void Unmap(DAddr address, size_t size);
+
+    // Write / Read
+    template <typename T>
+    T* GetPointer(DAddr address);
+
+    template <typename T>
+    const T* GetPointer(DAddr address) const;
+
+    template <typename T>
+    void Write(DAddr address, T value);
+
+    template <typename T>
+    T Read(DAddr address) const;
+
+    void ReadBlock(DAddr address, void* dest_pointer, size_t size);
+    void WriteBlock(DAddr address, void* src_pointer, size_t size);
+
+    size_t RegisterProcess(Memory::Memory* memory);
+    void UnregisterProcess(size_t id);
+
+private:
+    static constexpr bool supports_pinning = Traits::supports_pinning;
+    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_max_bits = 33;
+    static constexpr size_t page_bits = 12;
+    static constexpr u32 physical_address_base = 1U << page_bits;
+
+    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);
+
+    std::unique_ptr<DeviceMemoryManagerAllocator<Traits>> impl;
+
+    const uintptr_t physical_base;
+    DeviceInterface* interface;
+    Common::VirtualBuffer<u32> compressed_physical_ptr;
+    Common::VirtualBuffer<u32> compressed_device_addr;
+
+    std::deque<size_t> id_pool;
+    std::deque<Memory::Memory*> registered_processes;
+};
+
+} // namespace Core
--- a/src/core/device_memory_manager.inc
+++ b/src/core/device_memory_manager.inc
@ -0,0 +1,304 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <atomic>
+#include <memory>
+#include <type_traits>
+
+#include "common/address_space.h"
+#include "common/address_space.inc"
+#include "common/alignment.h"
+#include "common/scope_exit.h"
+#include "core/device_memory.h"
+#include "core/device_memory_manager.h"
+#include "core/memory.h"
+
+namespace Core {
+
+struct EmptyAllocator {
+    EmptyAllocator([[maybe_unused]] DAddr address) {}
+};
+
+template <typename DTraits>
+struct DeviceMemoryManagerAllocator {
+    static constexpr bool supports_pinning = DTraits::supports_pinning;
+    static constexpr size_t device_virtual_bits = DTraits::device_virtual_bits;
+    static constexpr size_t pin_bits = 32;
+    static constexpr DAddr first_address = 1ULL << Memory::YUZU_PAGEBITS;
+    static constexpr DAddr max_pin_area = supports_pinning ? 1ULL << pin_bits : first_address;
+    static constexpr DAddr max_device_area = 1ULL << device_virtual_bits;
+
+    DeviceMemoryManagerAllocator()
+        : pin_allocator(first_address),
+          main_allocator(supports_pinning ? 1ULL << pin_bits : first_address) {}
+
+    std::conditional_t<supports_pinning, Common::FlatAllocator<DAddr, 0, pin_bits>, EmptyAllocator>
+        pin_allocator;
+    Common::FlatAllocator<DAddr, 0, device_virtual_bits> main_allocator;
+
+    /// 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 {
+        if constexpr (pin_area) {
+            return addr >= 0 && addr + size <= max_pin_area;
+        } else {
+            return addr >= max_pin_area && addr + size <= max_device_area;
+        }
+    }
+
+    DAddr Allocate(size_t size) {
+        return main_allocator.Allocate(size);
+    }
+
+    DAddr AllocatePinned(size_t size) {
+        return pin_allocator.Allocate(size);
+    }
+
+    void DoInRange(DAddr address, size_t size, auto pin_func, auto main_func) {
+        if (IsInBounds<true>(address, size)) {
+            pin_func(address, size);
+            return;
+        }
+        if (IsInBounds<false>(address, size)) {
+            main_func(address, size);
+            return;
+        }
+        DAddr end_size = address + size - max_pin_area;
+        DAddr end_size2 = max_pin_area - address;
+        pin_func(address, end_size2);
+        main_func(max_pin_area, end_size);
+    }
+
+    void AllocateFixed(DAddr b_address, size_t b_size) {
+        if constexpr (supports_pinning) {
+            DoInRange(
+                b_address, b_size,
+                [this](DAddr address, size_t size) { pin_allocator.AllocateFixed(address, size); },
+                [this](DAddr address, size_t size) {
+                    main_allocator.AllocateFixed(address, size);
+                });
+        } else {
+            main_allocator.AllocateFixed(b_address, b_size);
+        }
+    }
+
+    void Free(DAddr b_address, size_t b_size) {
+        if constexpr (supports_pinning) {
+            DoInRange(
+                b_address, b_size,
+                [this](DAddr address, size_t size) { pin_allocator.Free(address, size); },
+                [this](DAddr address, size_t size) { main_allocator.Free(address, size); });
+        } else {
+            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())},
+      interface{nullptr}, compressed_physical_ptr(device_as_size >> Memory::YUZU_PAGEBITS),
+      compressed_device_addr(1ULL << (physical_max_bits - Memory::YUZU_PAGEBITS)) {
+    impl = std::make_unique<DeviceMemoryManagerAllocator<Traits>>();
+}
+
+template <typename Traits>
+DeviceMemoryManager<Traits>::~DeviceMemoryManager() = default;
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::BindInterface(DeviceInterface* interface_) {
+    interface = interface_;
+}
+
+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>
+DAddr DeviceMemoryManager<Traits>::AllocatePinned(size_t size) {
+    return impl->AllocatePinned(size);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::Free(DAddr start, size_t size) {
+    impl->Free(start, size);
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::Map(DAddr address, VAddr virtual_address, size_t size,
+                                      size_t p_id) {
+    Core::Memory::Memory* process_memory = registered_processes[p_id];
+    size_t start_page_d = address >> Memory::YUZU_PAGEBITS;
+    size_t num_pages = Common::AlignUp(size, Memory::YUZU_PAGESIZE) >> Memory::YUZU_PAGEBITS;
+    std::atomic_thread_fence(std::memory_order_acquire);
+    for (size_t i = 0; i < num_pages; i++) {
+        auto* ptr = process_memory->GetPointer(
+            Common::ProcessAddress(virtual_address + i * Memory::YUZU_PAGESIZE));
+        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;
+        compressed_device_addr[phys_addr - 1U] = static_cast<u32>(start_page_d + i);
+    }
+    std::atomic_thread_fence(std::memory_order_release);
+}
+
+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;
+    std::atomic_thread_fence(std::memory_order_acquire);
+    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;
+        if (phys_addr != 0) {
+            compressed_device_addr[phys_addr - 1] = 0;
+        }
+    }
+    std::atomic_thread_fence(std::memory_order_release);
+}
+
+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 std::size_t copy_amount =
+            std::min(static_cast<std::size_t>(Memory::YUZU_PAGESIZE) - page_offset, remaining_size);
+        const auto current_vaddr =
+            static_cast<u64>((page_index << Memory::YUZU_PAGEBITS) + page_offset);
+        SCOPE_EXIT({
+            page_index++;
+            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) {
+    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, 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<u8*>(src_pointer) + copy_amount;
+        });
+}
+
+template <typename Traits>
+size_t DeviceMemoryManager<Traits>::RegisterProcess(Memory::Memory* memory_interface) {
+    size_t new_id;
+    if (!id_pool.empty()) {
+        new_id = id_pool.front();
+        id_pool.pop_front();
+        registered_processes[new_id] = memory_interface;
+    } else {
+        registered_processes.emplace_back(memory_interface);
+        new_id = registered_processes.size() - 1U;
+    }
+    return new_id;
+}
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UnregisterProcess(size_t id) {
+    registered_processes[id] = nullptr;
+    id_pool.push_front(id);
+}
+
+} // namespace Core
--- a/src/video_core/host1x/gpu_device_memory_manager.cpp
+++ b/src/video_core/host1x/gpu_device_memory_manager.cpp
@ -0,0 +1,21 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "core/device_memory_manager.inc"
+#include "video_core/host1x/gpu_device_memory_manager.h"
+#include "video_core/rasterizer_interface.h"
+
+template struct Core::DeviceMemoryManagerAllocator<Tegra::MaxwellDeviceTraits>;
+template class Core::DeviceMemoryManager<Tegra::MaxwellDeviceTraits>;
+
+template const u8* Tegra::MaxwellDeviceMemoryManager::GetPointer<u8>(DAddr addr) const;
+template u8* Tegra::MaxwellDeviceMemoryManager::GetPointer<u8>(DAddr addr);
+
+template u8 Tegra::MaxwellDeviceMemoryManager::Read<u8>(DAddr addr) const;
+template u16 Tegra::MaxwellDeviceMemoryManager::Read<u16>(DAddr addr) const;
+template u32 Tegra::MaxwellDeviceMemoryManager::Read<u32>(DAddr addr) const;
+template u64 Tegra::MaxwellDeviceMemoryManager::Read<u64>(DAddr addr) const;
+template void Tegra::MaxwellDeviceMemoryManager::Write<u8>(DAddr addr, u8 data);
+template void Tegra::MaxwellDeviceMemoryManager::Write<u16>(DAddr addr, u16 data);
+template void Tegra::MaxwellDeviceMemoryManager::Write<u32>(DAddr addr, u32 data);
+template void Tegra::MaxwellDeviceMemoryManager::Write<u64>(DAddr addr, u64 data);
--- a/src/video_core/host1x/gpu_device_memory_manager.h
+++ b/src/video_core/host1x/gpu_device_memory_manager.h
@ -0,0 +1,20 @@
+// SPDX-FileCopyrightText: 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "core/device_memory_manager.h"
+
+namespace VideoCore {
+class RasterizerInterface;
+}
+
+namespace Tegra {
+
+struct MaxwellDeviceTraits {
+    static constexpr bool supports_pinning = true;
+    static constexpr size_t device_virtual_bits = 34;
+    using DeviceInterface = typename VideoCore::RasterizerInterface;
+};
+
+using MaxwellDeviceMemoryManager = Core::DeviceMemoryManager<MaxwellDeviceTraits>;
+
+} // namespace Tegra