MemoryManager: initial multi paging system implementation.

src/common/multi_level_page_table.inc

@@ -19,6 +19,9 @@ MultiLevelPageTable<BaseAddr>::MultiLevelPageTable(std::size_t address_space_bit
                                                    std::size_t page_bits_)
     : address_space_bits{address_space_bits_},
       first_level_bits{first_level_bits_}, page_bits{page_bits_} {
+    if (page_bits == 0) {
+      return;
+    }
     first_level_shift = address_space_bits - first_level_bits;
     first_level_chunk_size = (1ULL << (first_level_shift - page_bits)) * sizeof(BaseAddr);
     alloc_size = (1ULL << (address_space_bits - page_bits)) * sizeof(BaseAddr);

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

@@ -133,7 +133,8 @@ NvResult nvhost_as_gpu::AllocAsEx(const std::vector<u8>& input, std::vector<u8>&
     const u64 end_big_pages{(vm.va_range_end - vm.va_range_split) >> vm.big_page_size_bits};
     vm.big_page_allocator = std::make_unique<VM::Allocator>(start_big_pages, end_big_pages);
 
-    gmmu = std::make_shared<Tegra::MemoryManager>(system, 40, VM::PAGE_SIZE_BITS);
+    gmmu = std::make_shared<Tegra::MemoryManager>(system, 40, vm.big_page_size_bits,
+                                                  VM::PAGE_SIZE_BITS);
     system.GPU().InitAddressSpace(*gmmu);
     vm.initialised = true;
 
@@ -189,6 +190,7 @@ NvResult nvhost_as_gpu::AllocateSpace(const std::vector<u8>& input, std::vector<
         .size = size,
         .page_size = params.page_size,
         .sparse = (params.flags & MappingFlags::Sparse) != MappingFlags::None,
+        .big_pages = params.page_size != VM::YUZU_PAGESIZE,
     };
 
     std::memcpy(output.data(), &params, output.size());
@@ -209,7 +211,7 @@ void nvhost_as_gpu::FreeMappingLocked(u64 offset) {
     // Sparse mappings shouldn't be fully unmapped, just returned to their sparse state
     // Only FreeSpace can unmap them fully
     if (mapping->sparse_alloc)
-        gmmu->MapSparse(offset, mapping->size);
+        gmmu->MapSparse(offset, mapping->size, mapping->big_page);
     else
         gmmu->Unmap(offset, mapping->size);
 
@@ -294,8 +296,9 @@ NvResult nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& out
             return NvResult::BadValue;
         }
 
+        const bool use_big_pages = alloc->second.big_pages;
         if (!entry.handle) {
-            gmmu->MapSparse(virtual_address, size);
+            gmmu->MapSparse(virtual_address, size, use_big_pages);
         } else {
             auto handle{nvmap.GetHandle(entry.handle)};
             if (!handle) {
@@ -306,7 +309,7 @@ NvResult nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& out
                 handle->address +
                 (static_cast<u64>(entry.handle_offset_big_pages) << vm.big_page_size_bits))};
 
-            gmmu->Map(virtual_address, cpu_address, size);
+            gmmu->Map(virtual_address, cpu_address, size, use_big_pages);
         }
     }
 
@@ -345,7 +348,7 @@ NvResult nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8
             u64 gpu_address{static_cast<u64>(params.offset + params.buffer_offset)};
             VAddr cpu_address{mapping->ptr + params.buffer_offset};
 
-            gmmu->Map(gpu_address, cpu_address, params.mapping_size);
+            gmmu->Map(gpu_address, cpu_address, params.mapping_size, mapping->big_page);
 
             return NvResult::Success;
         } catch ([[maybe_unused]] const std::out_of_range& e) {
@@ -363,6 +366,17 @@ NvResult nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8
     VAddr cpu_address{static_cast<VAddr>(handle->address + params.buffer_offset)};
     u64 size{params.mapping_size ? params.mapping_size : handle->orig_size};
 
+    bool big_page{[&]() {
+        if (Common::IsAligned(handle->align, vm.big_page_size))
+            return true;
+        else if (Common::IsAligned(handle->align, VM::YUZU_PAGESIZE))
+            return false;
+        else {
+            UNREACHABLE();
+            return false;
+        }
+    }()};
+
     if ((params.flags & MappingFlags::Fixed) != MappingFlags::None) {
         auto alloc{allocation_map.upper_bound(params.offset)};
 
@@ -372,23 +386,14 @@ NvResult nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8
             return NvResult::BadValue;
         }
 
-        gmmu->Map(params.offset, cpu_address, size);
+        const bool use_big_pages = alloc->second.big_pages && big_page;
+        gmmu->Map(params.offset, cpu_address, size, use_big_pages);
 
-        auto mapping{std::make_shared<Mapping>(cpu_address, params.offset, size, true, false,
-                                               alloc->second.sparse)};
+        auto mapping{std::make_shared<Mapping>(cpu_address, params.offset, size, true,
+                                               use_big_pages, alloc->second.sparse)};
         alloc->second.mappings.push_back(mapping);
         mapping_map[params.offset] = mapping;
     } else {
-        bool big_page{[&]() {
-            if (Common::IsAligned(handle->align, vm.big_page_size))
-                return true;
-            else if (Common::IsAligned(handle->align, VM::YUZU_PAGESIZE))
-                return false;
-            else {
-                UNREACHABLE();
-                return false;
-            }
-        }()};
 
         auto& allocator{big_page ? *vm.big_page_allocator : *vm.small_page_allocator};
         u32 page_size{big_page ? vm.big_page_size : VM::YUZU_PAGESIZE};
@@ -402,7 +407,7 @@ NvResult nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8
             return NvResult::InsufficientMemory;
         }
 
-        gmmu->Map(params.offset, cpu_address, size);
+        gmmu->Map(params.offset, cpu_address, Common::AlignUp(size, page_size), big_page);
 
         auto mapping{
             std::make_shared<Mapping>(cpu_address, params.offset, size, false, big_page, false)};
@@ -439,7 +444,7 @@ NvResult nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8
         // Sparse mappings shouldn't be fully unmapped, just returned to their sparse state
         // Only FreeSpace can unmap them fully
         if (mapping->sparse_alloc) {
-            gmmu->MapSparse(params.offset, mapping->size);
+            gmmu->MapSparse(params.offset, mapping->size, mapping->big_page);
         } else {
             gmmu->Unmap(params.offset, mapping->size);
         }

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

@@ -177,6 +177,7 @@ private:
         std::list<std::shared_ptr<Mapping>> mappings;
         u32 page_size;
         bool sparse;
+        bool big_pages;
     };
 
     std::map<u64, std::shared_ptr<Mapping>>

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

@@ -9,6 +9,8 @@
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "core/core.h"
+#include "core/hle/kernel/k_page_table.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/devices/nvmap.h"
@@ -136,6 +138,10 @@ NvResult nvmap::IocAlloc(const std::vector<u8>& input, std::vector<u8>& output)
         LOG_CRITICAL(Service_NVDRV, "Object failed to allocate, handle={:08X}", params.handle);
         return result;
     }
+    ASSERT(system.CurrentProcess()
+               ->PageTable()
+               .LockForDeviceAddressSpace(handle_description->address, handle_description->size)
+               .IsSuccess());
     std::memcpy(output.data(), &params, sizeof(params));
     return result;
 }
@@ -256,6 +262,10 @@ NvResult nvmap::IocFree(const std::vector<u8>& input, std::vector<u8>& output) {
     }
 
     if (auto freeInfo{file.FreeHandle(params.handle, false)}) {
+        ASSERT(system.CurrentProcess()
+                   ->PageTable()
+                   .UnlockForDeviceAddressSpace(freeInfo->address, freeInfo->size)
+                   .IsSuccess());
         params.address = freeInfo->address;
         params.size = static_cast<u32>(freeInfo->size);
         params.flags.raw = 0;

src/video_core/memory_manager.cpp

@@ -7,6 +7,7 @@
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "core/core.h"
+#include "core/device_memory.h"
 #include "core/hle/kernel/k_page_table.h"
 #include "core/hle/kernel/k_process.h"
 #include "core/memory.h"
@@ -14,40 +15,69 @@
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_base.h"
 
+#pragma optimize("", off)
+
 namespace Tegra {
 
 std::atomic<size_t> MemoryManager::unique_identifier_generator{};
 
-MemoryManager::MemoryManager(Core::System& system_, u64 address_space_bits_, u64 page_bits_)
-    : system{system_}, address_space_bits{address_space_bits_}, page_bits{page_bits_}, entries{},
-      page_table{address_space_bits, address_space_bits + page_bits - 38, page_bits},
+MemoryManager::MemoryManager(Core::System& system_, u64 address_space_bits_, u64 big_page_bits_,
+                             u64 page_bits_)
+    : system{system_}, memory{system.Memory()}, device_memory{system.DeviceMemory()},
+      address_space_bits{address_space_bits_}, page_bits{page_bits_}, big_page_bits{big_page_bits_},
+      entries{}, big_entries{}, page_table{address_space_bits, address_space_bits + page_bits - 38,
+                                           page_bits != big_page_bits ? page_bits : 0},
       unique_identifier{unique_identifier_generator.fetch_add(1, std::memory_order_acq_rel)} {
     address_space_size = 1ULL << address_space_bits;
-    allocate_start = address_space_bits > 32 ? 1ULL << 32 : 0;
     page_size = 1ULL << page_bits;
     page_mask = page_size - 1ULL;
-    const u64 page_table_bits = address_space_bits - cpu_page_bits;
+    big_page_size = 1ULL << big_page_bits;
+    big_page_mask = big_page_size - 1ULL;
+    const u64 page_table_bits = address_space_bits - page_bits;
+    const u64 big_page_table_bits = address_space_bits - big_page_bits;
     const u64 page_table_size = 1ULL << page_table_bits;
+    const u64 big_page_table_size = 1ULL << big_page_table_bits;
     page_table_mask = page_table_size - 1;
+    big_page_table_mask = big_page_table_size - 1;
 
+    big_entries.resize(big_page_table_size / 32, 0);
+    big_page_table_cpu.resize(big_page_table_size);
+    big_page_table_physical.resize(big_page_table_size);
     entries.resize(page_table_size / 32, 0);
 }
 
 MemoryManager::~MemoryManager() = default;
 
+template <bool is_big_page>
 MemoryManager::EntryType MemoryManager::GetEntry(size_t position) const {
-    position = position >> page_bits;
-    const u64 entry_mask = entries[position / 32];
-    const size_t sub_index = position % 32;
-    return static_cast<EntryType>((entry_mask >> (2 * sub_index)) & 0x03ULL);
+    if constexpr (is_big_page) {
+        position = position >> big_page_bits;
+        const u64 entry_mask = big_entries[position / 32];
+        const size_t sub_index = position % 32;
+        return static_cast<EntryType>((entry_mask >> (2 * sub_index)) & 0x03ULL);
+    } else {
+        position = position >> page_bits;
+        const u64 entry_mask = entries[position / 32];
+        const size_t sub_index = position % 32;
+        return static_cast<EntryType>((entry_mask >> (2 * sub_index)) & 0x03ULL);
+    }
 }
 
+template <bool is_big_page>
 void MemoryManager::SetEntry(size_t position, MemoryManager::EntryType entry) {
-    position = position >> page_bits;
-    const u64 entry_mask = entries[position / 32];
-    const size_t sub_index = position % 32;
-    entries[position / 32] =
-        (~(3ULL << sub_index * 2) & entry_mask) | (static_cast<u64>(entry) << sub_index * 2);
+    if constexpr (is_big_page) {
+        position = position >> big_page_bits;
+        const u64 entry_mask = big_entries[position / 32];
+        const size_t sub_index = position % 32;
+        big_entries[position / 32] =
+            (~(3ULL << sub_index * 2) & entry_mask) | (static_cast<u64>(entry) << sub_index * 2);
+    } else {
+        position = position >> page_bits;
+        const u64 entry_mask = entries[position / 32];
+        const size_t sub_index = position % 32;
+        entries[position / 32] =
+            (~(3ULL << sub_index * 2) & entry_mask) | (static_cast<u64>(entry) << sub_index * 2);
+    }
 }
 
 template <MemoryManager::EntryType entry_type>
@@ -59,48 +89,66 @@ GPUVAddr MemoryManager::PageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] VAddr cp
     }
     for (u64 offset{}; offset < size; offset += page_size) {
         const GPUVAddr current_gpu_addr = gpu_addr + offset;
-        [[maybe_unused]] const auto current_entry_type = GetEntry(current_gpu_addr);
-        SetEntry(current_gpu_addr, entry_type);
+        [[maybe_unused]] const auto current_entry_type = GetEntry<false>(current_gpu_addr);
+        SetEntry<false>(current_gpu_addr, entry_type);
         if (current_entry_type != entry_type) {
             rasterizer->ModifyGPUMemory(unique_identifier, gpu_addr, page_size);
         }
         if constexpr (entry_type == EntryType::Mapped) {
             const VAddr current_cpu_addr = cpu_addr + offset;
-            const auto index = PageEntryIndex(current_gpu_addr);
-            const u32 sub_value = static_cast<u32>(current_cpu_addr >> 12ULL);
-            if (current_entry_type == entry_type && sub_value != page_table[index]) {
-                rasterizer->InvalidateRegion(static_cast<VAddr>(page_table[index]) << 12ULL,
-                                             page_size);
-            }
-            page_table[index] = static_cast<u32>(current_cpu_addr >> 12ULL);
+            const auto index = PageEntryIndex<false>(current_gpu_addr);
+            const u32 sub_value = static_cast<u32>(current_cpu_addr >> cpu_page_bits);
+            page_table[index] = sub_value;
         }
         remaining_size -= page_size;
     }
     return gpu_addr;
 }
 
+template <MemoryManager::EntryType entry_type>
+GPUVAddr MemoryManager::BigPageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] VAddr cpu_addr,
+                                       size_t size) {
+    u64 remaining_size{size};
+    for (u64 offset{}; offset < size; offset += big_page_size) {
+        const GPUVAddr current_gpu_addr = gpu_addr + offset;
+        [[maybe_unused]] const auto current_entry_type = GetEntry<true>(current_gpu_addr);
+        SetEntry<true>(current_gpu_addr, entry_type);
+        if (current_entry_type != entry_type) {
+            rasterizer->ModifyGPUMemory(unique_identifier, gpu_addr, big_page_size);
+        }
+        if constexpr (entry_type == EntryType::Mapped) {
+            const VAddr current_cpu_addr = cpu_addr + offset;
+            const auto index = PageEntryIndex<true>(current_gpu_addr);
+            const u32 sub_value = static_cast<u32>(current_cpu_addr >> cpu_page_bits);
+            big_page_table_cpu[index] = sub_value;
+            const PAddr phys_address =
+                device_memory.GetPhysicalAddr(memory.GetPointer(current_cpu_addr));
+            big_page_table_physical[index] = static_cast<u32>(phys_address);
+        }
+        remaining_size -= big_page_size;
+    }
+    return gpu_addr;
+}
+
 void MemoryManager::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
     rasterizer = rasterizer_;
 }
 
-GPUVAddr MemoryManager::Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size) {
+GPUVAddr MemoryManager::Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size,
+                            bool is_big_pages) {
+    if (is_big_pages) [[likely]] {
+        return BigPageTableOp<EntryType::Mapped>(gpu_addr, cpu_addr, size);
+    }
     return PageTableOp<EntryType::Mapped>(gpu_addr, cpu_addr, size);
 }
 
-GPUVAddr MemoryManager::MapSparse(GPUVAddr gpu_addr, std::size_t size) {
+GPUVAddr MemoryManager::MapSparse(GPUVAddr gpu_addr, std::size_t size, bool is_big_pages) {
+    if (is_big_pages) [[likely]] {
+        return BigPageTableOp<EntryType::Reserved>(gpu_addr, 0, size);
+    }
     return PageTableOp<EntryType::Reserved>(gpu_addr, 0, size);
 }
 
-GPUVAddr MemoryManager::MapAllocate(VAddr cpu_addr, std::size_t size, std::size_t align) {
-    return Map(*FindFreeRange(size, align), cpu_addr, size);
-}
-
-GPUVAddr MemoryManager::MapAllocate32(VAddr cpu_addr, std::size_t size) {
-    const std::optional<GPUVAddr> gpu_addr = FindFreeRange(size, 1, true);
-    ASSERT(gpu_addr);
-    return Map(*gpu_addr, cpu_addr, size);
-}
-
 void MemoryManager::Unmap(GPUVAddr gpu_addr, std::size_t size) {
     if (size == 0) {
         return;
@@ -115,61 +163,24 @@ void MemoryManager::Unmap(GPUVAddr gpu_addr, std::size_t size) {
         rasterizer->UnmapMemory(*cpu_addr, map_size);
     }
 
+    BigPageTableOp<EntryType::Free>(gpu_addr, 0, size);
     PageTableOp<EntryType::Free>(gpu_addr, 0, size);
 }
 
-std::optional<GPUVAddr> MemoryManager::AllocateFixed(GPUVAddr gpu_addr, std::size_t size) {
-    for (u64 offset{}; offset < size; offset += page_size) {
-        if (GetEntry(gpu_addr + offset) != EntryType::Free) {
+std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) const {
+    if (GetEntry<true>(gpu_addr) != EntryType::Mapped) [[unlikely]] {
+        if (GetEntry<false>(gpu_addr) != EntryType::Mapped) {
             return std::nullopt;
         }
-    }
 
-    return PageTableOp<EntryType::Reserved>(gpu_addr, 0, size);
-}
-
-GPUVAddr MemoryManager::Allocate(std::size_t size, std::size_t align) {
-    return *AllocateFixed(*FindFreeRange(size, align), size);
-}
-
-std::optional<GPUVAddr> MemoryManager::FindFreeRange(std::size_t size, std::size_t align,
-                                                     bool start_32bit_address) const {
-    if (!align) {
-        align = page_size;
-    } else {
-        align = Common::AlignUp(align, page_size);
-    }
-
-    u64 available_size{};
-    GPUVAddr gpu_addr{start_32bit_address ? 0 : allocate_start};
-    while (gpu_addr + available_size < address_space_size) {
-        if (GetEntry(gpu_addr + available_size) == EntryType::Free) {
-            available_size += page_size;
-
-            if (available_size >= size) {
-                return gpu_addr;
-            }
-        } else {
-            gpu_addr += available_size + page_size;
-            available_size = 0;
-
-            const auto remainder{gpu_addr % align};
-            if (remainder) {
-                gpu_addr = (gpu_addr - remainder) + align;
-            }
-        }
-    }
-
-    return std::nullopt;
-}
-
-std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) const {
-    if (GetEntry(gpu_addr) != EntryType::Mapped) {
-        return std::nullopt;
+        const VAddr cpu_addr_base = static_cast<VAddr>(page_table[PageEntryIndex<false>(gpu_addr)])
+                                    << cpu_page_bits;
+        return cpu_addr_base + (gpu_addr & page_mask);
     }
 
-    const VAddr cpu_addr_base = static_cast<VAddr>(page_table[PageEntryIndex(gpu_addr)]) << 12ULL;
-    return cpu_addr_base + (gpu_addr & page_mask);
+    const VAddr cpu_addr_base =
+        static_cast<VAddr>(big_page_table_cpu[PageEntryIndex<true>(gpu_addr)]) << cpu_page_bits;
+    return cpu_addr_base + (gpu_addr & big_page_mask);
 }
 
 std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr addr, std::size_t size) const {
@@ -225,7 +236,7 @@ u8* MemoryManager::GetPointer(GPUVAddr gpu_addr) {
         return {};
     }
 
-    return system.Memory().GetPointer(*address);
+    return memory.GetPointer(*address);
 }
 
 const u8* MemoryManager::GetPointer(GPUVAddr gpu_addr) const {
@@ -234,98 +245,161 @@ const u8* MemoryManager::GetPointer(GPUVAddr gpu_addr) const {
         return {};
     }
 
-    return system.Memory().GetPointer(*address);
+    return memory.GetPointer(*address);
 }
 
-void MemoryManager::ReadBlockImpl(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size,
-                                  bool is_safe) const {
+#pragma inline_recursion(on)
+
+template <bool is_big_pages, typename FuncMapped, typename FuncReserved, typename FuncUnmapped>
+inline void MemoryManager::MemoryOperation(GPUVAddr gpu_src_addr, std::size_t size,
+                                           FuncMapped&& func_mapped, FuncReserved&& func_reserved,
+                                           FuncUnmapped&& func_unmapped) const {
+    u64 used_page_size;
+    u64 used_page_mask;
+    u64 used_page_bits;
+    if constexpr (is_big_pages) {
+        used_page_size = big_page_size;
+        used_page_mask = big_page_mask;
+        used_page_bits = big_page_bits;
+    } else {
+        used_page_size = page_size;
+        used_page_mask = page_mask;
+        used_page_bits = page_bits;
+    }
     std::size_t remaining_size{size};
-    std::size_t page_index{gpu_src_addr >> page_bits};
-    std::size_t page_offset{gpu_src_addr & page_mask};
+    std::size_t page_index{gpu_src_addr >> used_page_bits};
+    std::size_t page_offset{gpu_src_addr & used_page_mask};
+    GPUVAddr current_address = gpu_src_addr;
 
     while (remaining_size > 0) {
         const std::size_t copy_amount{
-            std::min(static_cast<std::size_t>(page_size) - page_offset, remaining_size)};
-        const auto page_addr{GpuToCpuAddress(page_index << page_bits)};
-        if (page_addr) {
-            const auto src_addr{*page_addr + page_offset};
-            if (is_safe) {
-                // Flush must happen on the rasterizer interface, such that memory is always
-                // synchronous when it is read (even when in asynchronous GPU mode).
-                // Fixes Dead Cells title menu.
-                rasterizer->FlushRegion(src_addr, copy_amount);
-            }
-            system.Memory().ReadBlockUnsafe(src_addr, dest_buffer, copy_amount);
-        } else {
-            std::memset(dest_buffer, 0, copy_amount);
+            std::min(static_cast<std::size_t>(used_page_size) - page_offset, remaining_size)};
+        auto entry = GetEntry<is_big_pages>(current_address);
+        if (entry == EntryType::Mapped) [[likely]] {
+            func_mapped(page_index, page_offset, copy_amount);
+        } else if (entry == EntryType::Reserved) {
+            func_reserved(page_index, page_offset, copy_amount);
+        } else [[unlikely]] {
+            func_unmapped(page_index, page_offset, copy_amount);
         }
-
         page_index++;
         page_offset = 0;
-        dest_buffer = static_cast<u8*>(dest_buffer) + copy_amount;
         remaining_size -= copy_amount;
+        current_address += copy_amount;
     }
 }
 
+template <bool is_safe>
+void MemoryManager::ReadBlockImpl(GPUVAddr gpu_src_addr, void* dest_buffer,
+                                  std::size_t size) const {
+    auto set_to_zero = [&]([[maybe_unused]] std::size_t page_index,
+                           [[maybe_unused]] std::size_t offset, std::size_t copy_amount) {
+        std::memset(dest_buffer, 0, copy_amount);
+        dest_buffer = static_cast<u8*>(dest_buffer) + copy_amount;
+    };
+    auto mapped_normal = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(page_table[page_index]) << cpu_page_bits) + offset;
+        if constexpr (is_safe) {
+            rasterizer->FlushRegion(cpu_addr_base, copy_amount);
+        }
+        memory.ReadBlockUnsafe(cpu_addr_base, dest_buffer, copy_amount);
+        dest_buffer = static_cast<u8*>(dest_buffer) + copy_amount;
+    };
+    auto mapped_big = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(big_page_table_cpu[page_index]) << cpu_page_bits) + offset;
+        if constexpr (is_safe) {
+            rasterizer->FlushRegion(cpu_addr_base, copy_amount);
+        }
+        memory.ReadBlockUnsafe(cpu_addr_base, dest_buffer, copy_amount);
+        // u8* physical = device_memory.GetPointer(big_page_table_physical[page_index] + offset);
+        // std::memcpy(dest_buffer, physical, copy_amount);
+        dest_buffer = static_cast<u8*>(dest_buffer) + copy_amount;
+    };
+    auto read_short_pages = [&](std::size_t page_index, std::size_t offset,
+                                std::size_t copy_amount) {
+        GPUVAddr base = (page_index << big_page_bits) + offset;
+        MemoryOperation<false>(base, copy_amount, mapped_normal, set_to_zero, set_to_zero);
+    };
+    MemoryOperation<true>(gpu_src_addr, size, mapped_big, set_to_zero, read_short_pages);
+}
+
 void MemoryManager::ReadBlock(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size) const {
-    ReadBlockImpl(gpu_src_addr, dest_buffer, size, true);
+    ReadBlockImpl<true>(gpu_src_addr, dest_buffer, size);
 }
 
 void MemoryManager::ReadBlockUnsafe(GPUVAddr gpu_src_addr, void* dest_buffer,
                                     const std::size_t size) const {
-    ReadBlockImpl(gpu_src_addr, dest_buffer, size, false);
+    ReadBlockImpl<false>(gpu_src_addr, dest_buffer, size);
 }
 
-void MemoryManager::WriteBlockImpl(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size,
-                                   bool is_safe) {
-    std::size_t remaining_size{size};
-    std::size_t page_index{gpu_dest_addr >> page_bits};
-    std::size_t page_offset{gpu_dest_addr & page_mask};
-
-    while (remaining_size > 0) {
-        const std::size_t copy_amount{
-            std::min(static_cast<std::size_t>(page_size) - page_offset, remaining_size)};
-        const auto page_addr{GpuToCpuAddress(page_index << page_bits)};
-        if (page_addr) {
-            const auto dest_addr{*page_addr + page_offset};
-
-            if (is_safe) {
-                // Invalidate must happen on the rasterizer interface, such that memory is always
-                // synchronous when it is written (even when in asynchronous GPU mode).
-                rasterizer->InvalidateRegion(dest_addr, copy_amount);
-            }
-            system.Memory().WriteBlockUnsafe(dest_addr, src_buffer, copy_amount);
+template <bool is_safe>
+void MemoryManager::WriteBlockImpl(GPUVAddr gpu_dest_addr, const void* src_buffer,
+                                   std::size_t size) {
+    auto just_advance = [&]([[maybe_unused]] std::size_t page_index,
+                            [[maybe_unused]] std::size_t offset, std::size_t copy_amount) {
+        src_buffer = static_cast<const u8*>(src_buffer) + copy_amount;
+    };
+    auto mapped_normal = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(page_table[page_index]) << cpu_page_bits) + offset;
+        if constexpr (is_safe) {
+            rasterizer->InvalidateRegion(cpu_addr_base, copy_amount);
         }
-
-        page_index++;
-        page_offset = 0;
+        memory.WriteBlockUnsafe(cpu_addr_base, src_buffer, copy_amount);
         src_buffer = static_cast<const u8*>(src_buffer) + copy_amount;
-        remaining_size -= copy_amount;
-    }
+    };
+    auto mapped_big = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(big_page_table_cpu[page_index]) << cpu_page_bits) + offset;
+        if constexpr (is_safe) {
+            rasterizer->InvalidateRegion(cpu_addr_base, copy_amount);
+        }
+        memory.WriteBlockUnsafe(cpu_addr_base, src_buffer, copy_amount);
+        /*u8* physical =
+            device_memory.GetPointer(big_page_table_physical[page_index] << cpu_page_bits) + offset;
+        std::memcpy(physical, src_buffer, copy_amount);*/
+        src_buffer = static_cast<const u8*>(src_buffer) + copy_amount;
+    };
+    auto write_short_pages = [&](std::size_t page_index, std::size_t offset,
+                                 std::size_t copy_amount) {
+        GPUVAddr base = (page_index << big_page_bits) + offset;
+        MemoryOperation<false>(base, copy_amount, mapped_normal, just_advance, just_advance);
+    };
+    MemoryOperation<true>(gpu_dest_addr, size, mapped_big, just_advance, write_short_pages);
 }
 
 void MemoryManager::WriteBlock(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size) {
-    WriteBlockImpl(gpu_dest_addr, src_buffer, size, true);
+    WriteBlockImpl<true>(gpu_dest_addr, src_buffer, size);
 }
 
 void MemoryManager::WriteBlockUnsafe(GPUVAddr gpu_dest_addr, const void* src_buffer,
                                      std::size_t size) {
-    WriteBlockImpl(gpu_dest_addr, src_buffer, size, false);
+    WriteBlockImpl<false>(gpu_dest_addr, src_buffer, size);
 }
 
 void MemoryManager::FlushRegion(GPUVAddr gpu_addr, size_t size) const {
-    size_t remaining_size{size};
-    size_t page_index{gpu_addr >> page_bits};
-    size_t page_offset{gpu_addr & page_mask};
-    while (remaining_size > 0) {
-        const size_t num_bytes{std::min(page_size - page_offset, remaining_size)};
-        if (const auto page_addr{GpuToCpuAddress(page_index << page_bits)}; page_addr) {
-            rasterizer->FlushRegion(*page_addr + page_offset, num_bytes);
-        }
-        ++page_index;
-        page_offset = 0;
-        remaining_size -= num_bytes;
-    }
+    auto do_nothing = [&]([[maybe_unused]] std::size_t page_index,
+                          [[maybe_unused]] std::size_t offset,
+                          [[maybe_unused]] std::size_t copy_amount) {};
+
+    auto mapped_normal = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(page_table[page_index]) << cpu_page_bits) + offset;
+        rasterizer->FlushRegion(cpu_addr_base, copy_amount);
+    };
+    auto mapped_big = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(big_page_table_cpu[page_index]) << cpu_page_bits) + offset;
+        rasterizer->FlushRegion(cpu_addr_base, copy_amount);
+    };
+    auto flush_short_pages = [&](std::size_t page_index, std::size_t offset,
+                                 std::size_t copy_amount) {
+        GPUVAddr base = (page_index << big_page_bits) + offset;
+        MemoryOperation<false>(base, copy_amount, mapped_normal, do_nothing, do_nothing);
+    };
+    MemoryOperation<true>(gpu_addr, size, mapped_big, do_nothing, flush_short_pages);
 }
 
 void MemoryManager::CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size) {
@@ -348,7 +422,7 @@ bool MemoryManager::IsGranularRange(GPUVAddr gpu_addr, std::size_t size) const {
 }
 
 bool MemoryManager::IsContinousRange(GPUVAddr gpu_addr, std::size_t size) const {
-    size_t page_index{gpu_addr >> page_bits};
+    size_t page_index{gpu_addr >> big_page_bits};
     const size_t page_last{(gpu_addr + size + page_size - 1) >> page_bits};
     std::optional<VAddr> old_page_addr{};
     while (page_index != page_last) {
@@ -371,7 +445,7 @@ bool MemoryManager::IsFullyMappedRange(GPUVAddr gpu_addr, std::size_t size) cons
     size_t page_index{gpu_addr >> page_bits};
     const size_t page_last{(gpu_addr + size + page_size - 1) >> page_bits};
     while (page_index < page_last) {
-        if (GetEntry(page_index << page_bits) == EntryType::Free) {
+        if (GetEntry<false>(page_index << page_bits) == EntryType::Free) {
             return false;
         }
         ++page_index;
@@ -379,47 +453,63 @@ bool MemoryManager::IsFullyMappedRange(GPUVAddr gpu_addr, std::size_t size) cons
     return true;
 }
 
+#pragma inline_recursion(on)
+
 std::vector<std::pair<GPUVAddr, std::size_t>> MemoryManager::GetSubmappedRange(
     GPUVAddr gpu_addr, std::size_t size) const {
     std::vector<std::pair<GPUVAddr, std::size_t>> result{};
-    size_t page_index{gpu_addr >> page_bits};
-    size_t remaining_size{size};
-    size_t page_offset{gpu_addr & page_mask};
     std::optional<std::pair<GPUVAddr, std::size_t>> last_segment{};
     std::optional<VAddr> old_page_addr{};
-    const auto extend_size = [this, &last_segment, &page_index, &page_offset](std::size_t bytes) {
-        if (!last_segment) {
-            const GPUVAddr new_base_addr = (page_index << page_bits) + page_offset;
-            last_segment = {new_base_addr, bytes};
-        } else {
-            last_segment->second += bytes;
-        }
-    };
-    const auto split = [&last_segment, &result] {
+    const auto split = [&last_segment, &result]([[maybe_unused]] std::size_t page_index,
+                                                [[maybe_unused]] std::size_t offset,
+                                                [[maybe_unused]] std::size_t copy_amount) {
         if (last_segment) {
             result.push_back(*last_segment);
             last_segment = std::nullopt;
         }
     };
-    while (remaining_size > 0) {
-        const size_t num_bytes{std::min(page_size - page_offset, remaining_size)};
-        const auto page_addr{GpuToCpuAddress(page_index << page_bits)};
-        if (!page_addr || *page_addr == 0) {
-            split();
-        } else if (old_page_addr) {
-            if (*old_page_addr + page_size != *page_addr) {
-                split();
+    const auto extend_size_big = [this, &split, &old_page_addr,
+                                  &last_segment](std::size_t page_index, std::size_t offset,
+                                                 std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(big_page_table_cpu[page_index]) << cpu_page_bits) + offset;
+        if (old_page_addr) {
+            if (*old_page_addr != cpu_addr_base) {
+                split(0, 0, 0);
             }
-            extend_size(num_bytes);
+        }
+        old_page_addr = {cpu_addr_base + copy_amount};
+        if (!last_segment) {
+            const GPUVAddr new_base_addr = (page_index << big_page_bits) + offset;
+            last_segment = {new_base_addr, copy_amount};
         } else {
-            extend_size(num_bytes);
+            last_segment->second += copy_amount;
         }
-        ++page_index;
-        page_offset = 0;
-        remaining_size -= num_bytes;
-        old_page_addr = page_addr;
-    }
-    split();
+    };
+    const auto extend_size_short = [this, &split, &old_page_addr,
+                                    &last_segment](std::size_t page_index, std::size_t offset,
+                                                   std::size_t copy_amount) {
+        const VAddr cpu_addr_base =
+            (static_cast<VAddr>(page_table[page_index]) << cpu_page_bits) + offset;
+        if (old_page_addr) {
+            if (*old_page_addr != cpu_addr_base) {
+                split(0, 0, 0);
+            }
+        }
+        old_page_addr = {cpu_addr_base + copy_amount};
+        if (!last_segment) {
+            const GPUVAddr new_base_addr = (page_index << page_bits) + offset;
+            last_segment = {new_base_addr, copy_amount};
+        } else {
+            last_segment->second += copy_amount;
+        }
+    };
+    auto do_short_pages = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
+        GPUVAddr base = (page_index << big_page_bits) + offset;
+        MemoryOperation<false>(base, copy_amount, extend_size_short, split, split);
+    };
+    MemoryOperation<true>(gpu_addr, size, extend_size_big, split, do_short_pages);
+    split(0, 0, 0);
     return result;
 }
 

src/video_core/memory_manager.h

@@ -10,21 +10,26 @@
 
 #include "common/common_types.h"
 #include "common/multi_level_page_table.h"
+#include "common/virtual_buffer.h"
 
 namespace VideoCore {
 class RasterizerInterface;
 }
 
 namespace Core {
+class DeviceMemory;
+namespace Memory {
+class Memory;
+} // namespace Memory
 class System;
-}
+} // namespace Core
 
 namespace Tegra {
 
 class MemoryManager final {
 public:
     explicit MemoryManager(Core::System& system_, u64 address_space_bits_ = 40,
-                           u64 page_bits_ = 16);
+                           u64 big_page_bits_ = 16, u64 page_bits_ = 12);
     ~MemoryManager();
 
     size_t GetID() const {
@@ -93,12 +98,8 @@ public:
     std::vector<std::pair<GPUVAddr, std::size_t>> GetSubmappedRange(GPUVAddr gpu_addr,
                                                                     std::size_t size) const;
 
-    GPUVAddr Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size);
-    GPUVAddr MapSparse(GPUVAddr gpu_addr, std::size_t size);
-    [[nodiscard]] GPUVAddr MapAllocate(VAddr cpu_addr, std::size_t size, std::size_t align);
-    [[nodiscard]] GPUVAddr MapAllocate32(VAddr cpu_addr, std::size_t size);
-    [[nodiscard]] std::optional<GPUVAddr> AllocateFixed(GPUVAddr gpu_addr, std::size_t size);
-    [[nodiscard]] GPUVAddr Allocate(std::size_t size, std::size_t align);
+    GPUVAddr Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size, bool is_big_pages = true);
+    GPUVAddr MapSparse(GPUVAddr gpu_addr, std::size_t size, bool is_big_pages = true);
     void Unmap(GPUVAddr gpu_addr, std::size_t size);
 
     void FlushRegion(GPUVAddr gpu_addr, size_t size) const;
@@ -107,26 +108,42 @@ private:
     [[nodiscard]] std::optional<GPUVAddr> FindFreeRange(std::size_t size, std::size_t align,
                                                         bool start_32bit_address = false) const;
 
-    void ReadBlockImpl(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size,
-                       bool is_safe) const;
-    void WriteBlockImpl(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size,
-                        bool is_safe);
+    template <bool is_big_pages, typename FuncMapped, typename FuncReserved, typename FuncUnmapped>
+    inline void MemoryOperation(GPUVAddr gpu_src_addr, std::size_t size, FuncMapped&& func_mapped,
+                                FuncReserved&& func_reserved, FuncUnmapped&& func_unmapped) const;
+
+    template <bool is_safe>
+    void ReadBlockImpl(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size) const;
+
+    template <bool is_safe>
+    void WriteBlockImpl(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size);
 
+    template <bool is_big_page>
     [[nodiscard]] inline std::size_t PageEntryIndex(GPUVAddr gpu_addr) const {
-        return (gpu_addr >> page_bits) & page_table_mask;
+        if constexpr (is_big_page) {
+            return (gpu_addr >> big_page_bits) & big_page_table_mask;
+        } else {
+            return (gpu_addr >> page_bits) & page_table_mask;
+        }
     }
 
     Core::System& system;
+    Core::Memory::Memory& memory;
+    Core::DeviceMemory& device_memory;
 
     const u64 address_space_bits;
     const u64 page_bits;
     u64 address_space_size;
-    u64 allocate_start;
     u64 page_size;
     u64 page_mask;
     u64 page_table_mask;
     static constexpr u64 cpu_page_bits{12};
 
+    const u64 big_page_bits;
+    u64 big_page_size;
+    u64 big_page_mask;
+    u64 big_page_table_mask;
+
     VideoCore::RasterizerInterface* rasterizer = nullptr;
 
     enum class EntryType : u64 {
@@ -136,15 +153,23 @@ private:
     };
 
     std::vector<u64> entries;
+    std::vector<u64> big_entries;
 
     template <EntryType entry_type>
     GPUVAddr PageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] VAddr cpu_addr, size_t size);
 
-    EntryType GetEntry(size_t position) const;
+    template <EntryType entry_type>
+    GPUVAddr BigPageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] VAddr cpu_addr, size_t size);
+
+    template <bool is_big_page>
+    inline EntryType GetEntry(size_t position) const;
 
-    void SetEntry(size_t position, EntryType entry);
+    template <bool is_big_page>
+    inline void SetEntry(size_t position, EntryType entry);
 
     Common::MultiLevelPageTable<u32> page_table;
+    Common::VirtualBuffer<u32> big_page_table_cpu;
+    Common::VirtualBuffer<u32> big_page_table_physical;
 
     const size_t unique_identifier;