gpu: Rewrite virtual memory manager using PageTable.

src/common/page_table.cpp

@@ -16,6 +16,7 @@ void PageTable::Resize(std::size_t address_space_width_in_bits) {
 
     pointers.resize(num_page_table_entries);
     attributes.resize(num_page_table_entries);
+    backing_addr.resize(num_page_table_entries);
 
     // The default is a 39-bit address space, which causes an initial 1GB allocation size. If the
     // vector size is subsequently decreased (via resize), the vector might not automatically
@@ -24,6 +25,7 @@ void PageTable::Resize(std::size_t address_space_width_in_bits) {
 
     pointers.shrink_to_fit();
     attributes.shrink_to_fit();
+    backing_addr.shrink_to_fit();
 }
 
 } // namespace Common

src/common/page_table.h

@@ -21,6 +21,8 @@ enum class PageType : u8 {
     RasterizerCachedMemory,
     /// Page is mapped to a I/O region. Writing and reading to this page is handled by functions.
     Special,
+    /// Page is allocated for use.
+    Allocated,
 };
 
 struct SpecialRegion {
@@ -66,7 +68,7 @@ struct PageTable {
      * Contains MMIO handlers that back memory regions whose entries in the `attribute` vector is
      * of type `Special`.
      */
-    boost::icl::interval_map<VAddr, std::set<SpecialRegion>> special_regions;
+    boost::icl::interval_map<u64, std::set<SpecialRegion>> special_regions;
 
     /**
      * Vector of fine grained page attributes. If it is set to any value other than `Memory`, then
@@ -74,6 +76,8 @@ struct PageTable {
      */
     std::vector<PageType> attributes;
 
+    std::vector<u64> backing_addr;
+
     const std::size_t page_size_in_bits{};
 };
 

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

@@ -173,16 +173,8 @@ u32 nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& ou
         return 0;
     }
 
-    auto& system_instance = Core::System::GetInstance();
+    params.offset = Core::System::GetInstance().GPU().MemoryManager().UnmapBuffer(params.offset,
-
+                                                                                  itr->second.size);
-    // Remove this memory region from the rasterizer cache.
-    auto& gpu = system_instance.GPU();
-    auto cpu_addr = gpu.MemoryManager().GpuToCpuAddress(params.offset);
-    ASSERT(cpu_addr);
-    gpu.FlushAndInvalidateRegion(ToCacheAddr(Memory::GetPointer(*cpu_addr)), itr->second.size);
-
-    params.offset = gpu.MemoryManager().UnmapBuffer(params.offset, itr->second.size);
-
     buffer_mappings.erase(itr->second.offset);
 
     std::memcpy(output.data(), &params, output.size());

src/video_core/dma_pusher.h

@@ -9,7 +9,6 @@
 
 #include "common/bit_field.h"
 #include "common/common_types.h"
-#include "video_core/memory_manager.h"
 
 namespace Tegra {
 

src/video_core/engines/kepler_memory.cpp

@@ -46,7 +46,7 @@ void KeplerMemory::ProcessData(u32 data) {
     // contain a dirty surface that will have to be written back to memory.
     const GPUVAddr address{regs.dest.Address() + state.write_offset * sizeof(u32)};
     rasterizer.InvalidateRegion(ToCacheAddr(memory_manager.GetPointer(address)), sizeof(u32));
-    memory_manager.Write32(address, data);
+    memory_manager.Write<u32>(address, data);
 
     system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
 

src/video_core/engines/maxwell_3d.cpp

@@ -307,7 +307,7 @@ void Maxwell3D::ProcessQueryGet() {
             // Write the current query sequence to the sequence address.
             // TODO(Subv): Find out what happens if you use a long query type but mark it as a short
             // query.
-            memory_manager.Write32(sequence_address, sequence);
+            memory_manager.Write<u32>(sequence_address, sequence);
         } else {
             // Write the 128-bit result structure in long mode. Note: We emulate an infinitely fast
             // GPU, this command may actually take a while to complete in real hardware due to GPU
@@ -395,7 +395,7 @@ void Maxwell3D::ProcessCBData(u32 value) {
 
     u8* ptr{memory_manager.GetPointer(address)};
     rasterizer.InvalidateRegion(ToCacheAddr(ptr), sizeof(u32));
-    memory_manager.Write32(address, value);
+    memory_manager.Write<u32>(address, value);
 
     dirty_flags.OnMemoryWrite();
 
@@ -447,7 +447,7 @@ std::vector<Texture::FullTextureInfo> Maxwell3D::GetStageTextures(Regs::ShaderSt
     for (GPUVAddr current_texture = tex_info_buffer.address + TextureInfoOffset;
          current_texture < tex_info_buffer_end; current_texture += sizeof(Texture::TextureHandle)) {
 
-        const Texture::TextureHandle tex_handle{memory_manager.Read32(current_texture)};
+        const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(current_texture)};
 
         Texture::FullTextureInfo tex_info{};
         // TODO(Subv): Use the shader to determine which textures are actually accessed.
@@ -482,7 +482,7 @@ Texture::FullTextureInfo Maxwell3D::GetStageTexture(Regs::ShaderStage stage,
 
     ASSERT(tex_info_address < tex_info_buffer.address + tex_info_buffer.size);
 
-    const Texture::TextureHandle tex_handle{memory_manager.Read32(tex_info_address)};
+    const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(tex_info_address)};
 
     Texture::FullTextureInfo tex_info{};
     tex_info.index = static_cast<u32>(offset);

src/video_core/gpu.cpp

@@ -12,6 +12,7 @@
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/gpu.h"
+#include "video_core/memory_manager.h"
 #include "video_core/renderer_base.h"
 
 namespace Tegra {
@@ -287,7 +288,7 @@ void GPU::ProcessSemaphoreTriggerMethod() {
         block.timestamp = Core::System::GetInstance().CoreTiming().GetTicks();
         memory_manager->WriteBlock(regs.smaphore_address.SmaphoreAddress(), &block, sizeof(block));
     } else {
-        const u32 word{memory_manager->Read32(regs.smaphore_address.SmaphoreAddress())};
+        const u32 word{memory_manager->Read<u32>(regs.smaphore_address.SmaphoreAddress())};
         if ((op == GpuSemaphoreOperation::AcquireEqual && word == regs.semaphore_sequence) ||
             (op == GpuSemaphoreOperation::AcquireGequal &&
              static_cast<s32>(word - regs.semaphore_sequence) > 0) ||
@@ -314,11 +315,11 @@ void GPU::ProcessSemaphoreTriggerMethod() {
 }
 
 void GPU::ProcessSemaphoreRelease() {
-    memory_manager->Write32(regs.smaphore_address.SmaphoreAddress(), regs.semaphore_release);
+    memory_manager->Write<u32>(regs.smaphore_address.SmaphoreAddress(), regs.semaphore_release);
 }
 
 void GPU::ProcessSemaphoreAcquire() {
-    const u32 word = memory_manager->Read32(regs.smaphore_address.SmaphoreAddress());
+    const u32 word = memory_manager->Read<u32>(regs.smaphore_address.SmaphoreAddress());
     const auto value = regs.semaphore_acquire;
     if (word != value) {
         regs.acquire_active = true;

src/video_core/gpu.h

@@ -9,7 +9,6 @@
 #include "common/common_types.h"
 #include "core/hle/service/nvflinger/buffer_queue.h"
 #include "video_core/dma_pusher.h"
-#include "video_core/memory_manager.h"
 
 using CacheAddr = std::uintptr_t;
 inline CacheAddr ToCacheAddr(const void* host_ptr) {
@@ -124,6 +123,8 @@ enum class EngineID {
     MAXWELL_DMA_COPY_A = 0xB0B5,
 };
 
+class MemoryManager;
+
 class GPU {
 public:
     explicit GPU(Core::System& system, VideoCore::RendererBase& renderer);
@@ -244,9 +245,8 @@ protected:
 private:
     std::unique_ptr<Tegra::MemoryManager> memory_manager;
 
-    /// Mapping of command subchannels to their bound engine ids.
+    /// Mapping of command subchannels to their bound engine ids
     std::array<EngineID, 8> bound_engines = {};
-
     /// 3D engine
     std::unique_ptr<Engines::Maxwell3D> maxwell_3d;
     /// 2D engine

src/video_core/memory_manager.cpp

@@ -5,198 +5,164 @@
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "core/core.h"
 #include "core/memory.h"
+#include "video_core/gpu.h"
 #include "video_core/memory_manager.h"
+#include "video_core/rasterizer_interface.h"
+#include "video_core/renderer_base.h"
 
 namespace Tegra {
 
 MemoryManager::MemoryManager() {
-    // Mark the first page as reserved, so that 0 is not a valid GPUVAddr. Otherwise, games might
+    std::fill(page_table.pointers.begin(), page_table.pointers.end(), nullptr);
-    // try to use 0 as a valid address, which is also used to mean nullptr. This fixes a bug with
+    std::fill(page_table.attributes.begin(), page_table.attributes.end(),
-    // Undertale using 0 for a render target.
+              Common::PageType::Unmapped);
-    PageSlot(0) = static_cast<u64>(PageStatus::Reserved);
+    page_table.Resize(address_space_width);
+
+    // Initialize the map with a single free region covering the entire managed space.
+    VirtualMemoryArea initial_vma;
+    initial_vma.size = address_space_end;
+    vma_map.emplace(initial_vma.base, initial_vma);
+
+    UpdatePageTableForVMA(initial_vma);
 }
 
 GPUVAddr MemoryManager::AllocateSpace(u64 size, u64 align) {
-    const std::optional<GPUVAddr> gpu_addr{FindFreeBlock(0, size, align, PageStatus::Unmapped)};
+    const GPUVAddr gpu_addr{
-
+        FindFreeRegion(address_space_base, size, align, VirtualMemoryArea::Type::Unmapped)};
-    ASSERT_MSG(gpu_addr, "unable to find available GPU memory");
+    AllocateMemory(gpu_addr, 0, size);
-
-    for (u64 offset{}; offset < size; offset += PAGE_SIZE) {
-        VAddr& slot{PageSlot(*gpu_addr + offset)};
-
-        ASSERT(slot == static_cast<u64>(PageStatus::Unmapped));
-
-        slot = static_cast<u64>(PageStatus::Allocated);
-    }
-
-    return *gpu_addr;
-}
-
-GPUVAddr MemoryManager::AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align) {
-    for (u64 offset{}; offset < size; offset += PAGE_SIZE) {
-        VAddr& slot{PageSlot(gpu_addr + offset)};
-
-        ASSERT(slot == static_cast<u64>(PageStatus::Unmapped));
-
-        slot = static_cast<u64>(PageStatus::Allocated);
-    }
-
     return gpu_addr;
 }
 
-GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, u64 size) {
+GPUVAddr MemoryManager::AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align) {
-    const std::optional<GPUVAddr> gpu_addr{FindFreeBlock(0, size, PAGE_SIZE, PageStatus::Unmapped)};
+    AllocateMemory(gpu_addr, 0, size);
-
+    return gpu_addr;
-    ASSERT_MSG(gpu_addr, "unable to find available GPU memory");
-
-    for (u64 offset{}; offset < size; offset += PAGE_SIZE) {
-        VAddr& slot{PageSlot(*gpu_addr + offset)};
-
-        ASSERT(slot == static_cast<u64>(PageStatus::Unmapped));
-
-        slot = cpu_addr + offset;
-    }
-
-    const MappedRegion region{cpu_addr, *gpu_addr, size};
-    mapped_regions.push_back(region);
-
-    return *gpu_addr;
 }
 
-GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size) {
+GPUVAddr MemoryManager::MapBufferEx(GPUVAddr cpu_addr, u64 size) {
-    ASSERT((gpu_addr & PAGE_MASK) == 0);
+    const GPUVAddr gpu_addr{
+        FindFreeRegion(address_space_base, size, page_size, VirtualMemoryArea::Type::Unmapped)};
+    MapBackingMemory(gpu_addr, Memory::GetPointer(cpu_addr), ((size + page_mask) & ~page_mask),
+                     cpu_addr);
+    return gpu_addr;
+}
 
-    if (PageSlot(gpu_addr) != static_cast<u64>(PageStatus::Allocated)) {
+GPUVAddr MemoryManager::MapBufferEx(GPUVAddr cpu_addr, GPUVAddr gpu_addr, u64 size) {
-        // Page has been already mapped. In this case, we must find a new area of memory to use that
+    ASSERT((gpu_addr & page_mask) == 0);
-        // is different than the specified one. Super Mario Odyssey hits this scenario when changing
-        // areas, but we do not want to overwrite the old pages.
-        // TODO(bunnei): We need to write a hardware test to confirm this behavior.
 
-        LOG_ERROR(HW_GPU, "attempting to map addr 0x{:016X}, which is not available!", gpu_addr);
+    MapBackingMemory(gpu_addr, Memory::GetPointer(cpu_addr), ((size + page_mask) & ~page_mask),
-
+                     cpu_addr);
-        const std::optional<GPUVAddr> new_gpu_addr{
-            FindFreeBlock(gpu_addr, size, PAGE_SIZE, PageStatus::Allocated)};
-
-        ASSERT_MSG(new_gpu_addr, "unable to find available GPU memory");
-
-        gpu_addr = *new_gpu_addr;
-    }
-
-    for (u64 offset{}; offset < size; offset += PAGE_SIZE) {
-        VAddr& slot{PageSlot(gpu_addr + offset)};
-
-        ASSERT(slot == static_cast<u64>(PageStatus::Allocated));
-
-        slot = cpu_addr + offset;
-    }
-
-    const MappedRegion region{cpu_addr, gpu_addr, size};
-    mapped_regions.push_back(region);
 
     return gpu_addr;
 }
 
 GPUVAddr MemoryManager::UnmapBuffer(GPUVAddr gpu_addr, u64 size) {
-    ASSERT((gpu_addr & PAGE_MASK) == 0);
+    ASSERT((gpu_addr & page_mask) == 0);
 
-    for (u64 offset{}; offset < size; offset += PAGE_SIZE) {
+    const CacheAddr cache_addr{ToCacheAddr(GetPointer(gpu_addr))};
-        VAddr& slot{PageSlot(gpu_addr + offset)};
+    Core::System::GetInstance().Renderer().Rasterizer().FlushAndInvalidateRegion(cache_addr, size);
 
-        ASSERT(slot != static_cast<u64>(PageStatus::Allocated) &&
+    UnmapRange(gpu_addr, ((size + page_mask) & ~page_mask));
-               slot != static_cast<u64>(PageStatus::Unmapped));
 
-        slot = static_cast<u64>(PageStatus::Unmapped);
-    }
-
-    // Delete the region mappings that are contained within the unmapped region
-    mapped_regions.erase(std::remove_if(mapped_regions.begin(), mapped_regions.end(),
-                                        [&](const MappedRegion& region) {
-                                            return region.gpu_addr <= gpu_addr &&
-                                                   region.gpu_addr + region.size < gpu_addr + size;
-                                        }),
-                         mapped_regions.end());
     return gpu_addr;
 }
 
-GPUVAddr MemoryManager::GetRegionEnd(GPUVAddr region_start) const {
+GPUVAddr MemoryManager::FindFreeRegion(GPUVAddr region_start, u64 size, u64 align,
-    for (const auto& region : mapped_regions) {
+                                       VirtualMemoryArea::Type vma_type) {
-        const GPUVAddr region_end{region.gpu_addr + region.size};
-        if (region_start >= region.gpu_addr && region_start < region_end) {
-            return region_end;
-        }
-    }
-    return {};
-}
 
-std::optional<GPUVAddr> MemoryManager::FindFreeBlock(GPUVAddr region_start, u64 size, u64 align,
+    align = (align + page_mask) & ~page_mask;
-                                                     PageStatus status) {
-    GPUVAddr gpu_addr{region_start};
-    u64 free_space{};
-    align = (align + PAGE_MASK) & ~PAGE_MASK;
 
-    while (gpu_addr + free_space < MAX_ADDRESS) {
+    // Find the first Free VMA.
-        if (PageSlot(gpu_addr + free_space) == static_cast<u64>(status)) {
+    const GPUVAddr base = region_start;
-            free_space += PAGE_SIZE;
+    const VMAHandle vma_handle = std::find_if(vma_map.begin(), vma_map.end(), [&](const auto& vma) {
-            if (free_space >= size) {
+        if (vma.second.type != vma_type)
-                return gpu_addr;
+            return false;
-            }
-        } else {
-            gpu_addr += free_space + PAGE_SIZE;
-            free_space = 0;
-            gpu_addr = Common::AlignUp(gpu_addr, align);
-        }
-    }
 
-    return {};
+        const VAddr vma_end = vma.second.base + vma.second.size;
-}
+        return vma_end > base && vma_end >= base + size;
+    });
 
-std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) {
+    if (vma_handle == vma_map.end()) {
-    const VAddr base_addr{PageSlot(gpu_addr)};
-
-    if (base_addr == static_cast<u64>(PageStatus::Allocated) ||
-        base_addr == static_cast<u64>(PageStatus::Unmapped) ||
-        base_addr == static_cast<u64>(PageStatus::Reserved)) {
         return {};
     }
 
-    return base_addr + (gpu_addr & PAGE_MASK);
+    return std::max(base, vma_handle->second.base);
 }
 
-u8 MemoryManager::Read8(GPUVAddr addr) {
+std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) {
-    return Memory::Read8(*GpuToCpuAddress(addr));
+    VAddr cpu_addr = page_table.backing_addr[gpu_addr >> page_bits];
+    if (cpu_addr) {
+        return cpu_addr + (gpu_addr & page_mask);
+    }
+
+    return {};
 }
 
-u16 MemoryManager::Read16(GPUVAddr addr) {
+template <typename T>
-    return Memory::Read16(*GpuToCpuAddress(addr));
+T MemoryManager::Read(GPUVAddr vaddr) {
+    const u8* page_pointer = page_table.pointers[vaddr >> page_bits];
+    if (page_pointer) {
+        // NOTE: Avoid adding any extra logic to this fast-path block
+        T value;
+        std::memcpy(&value, &page_pointer[vaddr & page_mask], sizeof(T));
+        return value;
+    }
+
+    Common::PageType type = page_table.attributes[vaddr >> page_bits];
+    switch (type) {
+    case Common::PageType::Unmapped:
+        LOG_ERROR(HW_GPU, "Unmapped Read{} @ 0x{:08X}", sizeof(T) * 8, vaddr);
+        return 0;
+    case Common::PageType::Memory:
+        ASSERT_MSG(false, "Mapped memory page without a pointer @ {:016X}", vaddr);
+        break;
+    default:
+        UNREACHABLE();
+    }
+    return {};
 }
 
-u32 MemoryManager::Read32(GPUVAddr addr) {
+template <typename T>
-    return Memory::Read32(*GpuToCpuAddress(addr));
+void MemoryManager::Write(GPUVAddr vaddr, T data) {
+    u8* page_pointer = page_table.pointers[vaddr >> page_bits];
+    if (page_pointer) {
+        // NOTE: Avoid adding any extra logic to this fast-path block
+        std::memcpy(&page_pointer[vaddr & page_mask], &data, sizeof(T));
+        return;
+    }
+
+    Common::PageType type = page_table.attributes[vaddr >> page_bits];
+    switch (type) {
+    case Common::PageType::Unmapped:
+        LOG_ERROR(HW_GPU, "Unmapped Write{} 0x{:08X} @ 0x{:016X}", sizeof(data) * 8,
+                  static_cast<u32>(data), vaddr);
+        return;
+    case Common::PageType::Memory:
+        ASSERT_MSG(false, "Mapped memory page without a pointer @ {:016X}", vaddr);
+        break;
+    default:
+        UNREACHABLE();
+    }
 }
 
-u64 MemoryManager::Read64(GPUVAddr addr) {
+template u8 MemoryManager::Read<u8>(GPUVAddr addr);
-    return Memory::Read64(*GpuToCpuAddress(addr));
+template u16 MemoryManager::Read<u16>(GPUVAddr addr);
-}
+template u32 MemoryManager::Read<u32>(GPUVAddr addr);
-
+template u64 MemoryManager::Read<u64>(GPUVAddr addr);
-void MemoryManager::Write8(GPUVAddr addr, u8 data) {
+template void MemoryManager::Write<u8>(GPUVAddr addr, u8 data);
-    Memory::Write8(*GpuToCpuAddress(addr), data);
+template void MemoryManager::Write<u16>(GPUVAddr addr, u16 data);
-}
+template void MemoryManager::Write<u32>(GPUVAddr addr, u32 data);
-
+template void MemoryManager::Write<u64>(GPUVAddr addr, u64 data);
-void MemoryManager::Write16(GPUVAddr addr, u16 data) {
-    Memory::Write16(*GpuToCpuAddress(addr), data);
-}
-
-void MemoryManager::Write32(GPUVAddr addr, u32 data) {
-    Memory::Write32(*GpuToCpuAddress(addr), data);
-}
-
-void MemoryManager::Write64(GPUVAddr addr, u64 data) {
-    Memory::Write64(*GpuToCpuAddress(addr), data);
-}
 
 u8* MemoryManager::GetPointer(GPUVAddr addr) {
-    return Memory::GetPointer(*GpuToCpuAddress(addr));
+    u8* page_pointer = page_table.pointers[addr >> page_bits];
+    if (page_pointer) {
+        return page_pointer + (addr & page_mask);
+    }
+
+    LOG_ERROR(HW_GPU, "Unknown GetPointer @ 0x{:016X}", addr);
+    return {};
 }
 
 void MemoryManager::ReadBlock(GPUVAddr src_addr, void* dest_buffer, std::size_t size) {
@@ -210,13 +176,251 @@ void MemoryManager::CopyBlock(GPUVAddr dest_addr, GPUVAddr src_addr, std::size_t
     std::memcpy(GetPointer(dest_addr), GetPointer(src_addr), size);
 }
 
-VAddr& MemoryManager::PageSlot(GPUVAddr gpu_addr) {
+void MemoryManager::MapPages(GPUVAddr base, u64 size, u8* memory, Common::PageType type,
-    auto& block{page_table[(gpu_addr >> (PAGE_BITS + PAGE_TABLE_BITS)) & PAGE_TABLE_MASK]};
+                             VAddr backing_addr) {
-    if (!block) {
+    LOG_DEBUG(HW_GPU, "Mapping {} onto {:016X}-{:016X}", fmt::ptr(memory), base * page_size,
-        block = std::make_unique<PageBlock>();
+              (base + size) * page_size);
-        block->fill(static_cast<VAddr>(PageStatus::Unmapped));
+
+    VAddr end = base + size;
+    ASSERT_MSG(end <= page_table.pointers.size(), "out of range mapping at {:016X}",
+               base + page_table.pointers.size());
+
+    std::fill(page_table.attributes.begin() + base, page_table.attributes.begin() + end, type);
+
+    if (memory == nullptr) {
+        std::fill(page_table.pointers.begin() + base, page_table.pointers.begin() + end, memory);
+        std::fill(page_table.backing_addr.begin() + base, page_table.backing_addr.begin() + end,
+                  backing_addr);
+    } else {
+        while (base != end) {
+            page_table.pointers[base] = memory;
+            page_table.backing_addr[base] = backing_addr;
+
+            base += 1;
+            memory += page_size;
+            backing_addr += page_size;
+        }
+    }
+}
+
+void MemoryManager::MapMemoryRegion(GPUVAddr base, u64 size, u8* target, VAddr backing_addr) {
+    ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: {:016X}", size);
+    ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: {:016X}", base);
+    MapPages(base / page_size, size / page_size, target, Common::PageType::Memory, backing_addr);
+}
+
+void MemoryManager::UnmapRegion(GPUVAddr base, u64 size) {
+    ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: {:016X}", size);
+    ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: {:016X}", base);
+    MapPages(base / page_size, size / page_size, nullptr, Common::PageType::Unmapped);
+}
+
+bool VirtualMemoryArea::CanBeMergedWith(const VirtualMemoryArea& next) const {
+    ASSERT(base + size == next.base);
+    if (type != next.type) {
+        return {};
+    }
+    if (type == VirtualMemoryArea::Type::Allocated && (offset + size != next.offset)) {
+        return {};
+    }
+    if (type == VirtualMemoryArea::Type::Mapped && backing_memory + size != next.backing_memory) {
+        return {};
+    }
+    return true;
+}
+
+MemoryManager::VMAHandle MemoryManager::FindVMA(GPUVAddr target) const {
+    if (target >= address_space_end) {
+        return vma_map.end();
+    } else {
+        return std::prev(vma_map.upper_bound(target));
+    }
+}
+
+MemoryManager::VMAHandle MemoryManager::AllocateMemory(GPUVAddr target, std::size_t offset,
+                                                       u64 size) {
+
+    // This is the appropriately sized VMA that will turn into our allocation.
+    VMAIter vma_handle = CarveVMA(target, size);
+    VirtualMemoryArea& final_vma = vma_handle->second;
+    ASSERT(final_vma.size == size);
+
+    final_vma.type = VirtualMemoryArea::Type::Allocated;
+    final_vma.offset = offset;
+    UpdatePageTableForVMA(final_vma);
+
+    return MergeAdjacent(vma_handle);
+}
+
+MemoryManager::VMAHandle MemoryManager::MapBackingMemory(GPUVAddr target, u8* memory, u64 size,
+                                                         VAddr backing_addr) {
+    // This is the appropriately sized VMA that will turn into our allocation.
+    VMAIter vma_handle = CarveVMA(target, size);
+    VirtualMemoryArea& final_vma = vma_handle->second;
+    ASSERT(final_vma.size == size);
+
+    final_vma.type = VirtualMemoryArea::Type::Mapped;
+    final_vma.backing_memory = memory;
+    final_vma.backing_addr = backing_addr;
+    UpdatePageTableForVMA(final_vma);
+
+    return MergeAdjacent(vma_handle);
+}
+
+MemoryManager::VMAIter MemoryManager::Unmap(VMAIter vma_handle) {
+    VirtualMemoryArea& vma = vma_handle->second;
+    vma.type = VirtualMemoryArea::Type::Allocated;
+    vma.offset = 0;
+    vma.backing_memory = nullptr;
+
+    UpdatePageTableForVMA(vma);
+
+    return MergeAdjacent(vma_handle);
+}
+
+void MemoryManager::UnmapRange(GPUVAddr target, u64 size) {
+    VMAIter vma = CarveVMARange(target, size);
+    const VAddr target_end = target + size;
+
+    const VMAIter end = vma_map.end();
+    // The comparison against the end of the range must be done using addresses since VMAs can be
+    // merged during this process, causing invalidation of the iterators.
+    while (vma != end && vma->second.base < target_end) {
+        vma = std::next(Unmap(vma));
+    }
+
+    ASSERT(FindVMA(target)->second.size >= size);
+}
+
+MemoryManager::VMAIter MemoryManager::StripIterConstness(const VMAHandle& iter) {
+    // This uses a neat C++ trick to convert a const_iterator to a regular iterator, given
+    // non-const access to its container.
+    return vma_map.erase(iter, iter); // Erases an empty range of elements
+}
+
+MemoryManager::VMAIter MemoryManager::CarveVMA(GPUVAddr base, u64 size) {
+    ASSERT_MSG((size & Tegra::MemoryManager::page_mask) == 0, "non-page aligned size: 0x{:016X}",
+               size);
+    ASSERT_MSG((base & Tegra::MemoryManager::page_mask) == 0, "non-page aligned base: 0x{:016X}",
+               base);
+
+    VMAIter vma_handle = StripIterConstness(FindVMA(base));
+    if (vma_handle == vma_map.end()) {
+        // Target address is outside the range managed by the kernel
+        return {};
+    }
+
+    const VirtualMemoryArea& vma = vma_handle->second;
+    if (vma.type == VirtualMemoryArea::Type::Mapped) {
+        // Region is already allocated
+        return {};
+    }
+
+    const VAddr start_in_vma = base - vma.base;
+    const VAddr end_in_vma = start_in_vma + size;
+
+    if (end_in_vma < vma.size) {
+        // Split VMA at the end of the allocated region
+        SplitVMA(vma_handle, end_in_vma);
+    }
+    if (start_in_vma != 0) {
+        // Split VMA at the start of the allocated region
+        vma_handle = SplitVMA(vma_handle, start_in_vma);
+    }
+
+    return vma_handle;
+}
+
+MemoryManager::VMAIter MemoryManager::CarveVMARange(GPUVAddr target, u64 size) {
+    ASSERT_MSG((size & Tegra::MemoryManager::page_mask) == 0, "non-page aligned size: 0x{:016X}",
+               size);
+    ASSERT_MSG((target & Tegra::MemoryManager::page_mask) == 0, "non-page aligned base: 0x{:016X}",
+               target);
+
+    const VAddr target_end = target + size;
+    ASSERT(target_end >= target);
+    ASSERT(size > 0);
+
+    VMAIter begin_vma = StripIterConstness(FindVMA(target));
+    const VMAIter i_end = vma_map.lower_bound(target_end);
+    if (std::any_of(begin_vma, i_end, [](const auto& entry) {
+            return entry.second.type == VirtualMemoryArea::Type::Unmapped;
+        })) {
+        return {};
+    }
+
+    if (target != begin_vma->second.base) {
+        begin_vma = SplitVMA(begin_vma, target - begin_vma->second.base);
+    }
+
+    VMAIter end_vma = StripIterConstness(FindVMA(target_end));
+    if (end_vma != vma_map.end() && target_end != end_vma->second.base) {
+        end_vma = SplitVMA(end_vma, target_end - end_vma->second.base);
+    }
+
+    return begin_vma;
+}
+
+MemoryManager::VMAIter MemoryManager::SplitVMA(VMAIter vma_handle, u64 offset_in_vma) {
+    VirtualMemoryArea& old_vma = vma_handle->second;
+    VirtualMemoryArea new_vma = old_vma; // Make a copy of the VMA
+
+    // For now, don't allow no-op VMA splits (trying to split at a boundary) because it's probably
+    // a bug. This restriction might be removed later.
+    ASSERT(offset_in_vma < old_vma.size);
+    ASSERT(offset_in_vma > 0);
+
+    old_vma.size = offset_in_vma;
+    new_vma.base += offset_in_vma;
+    new_vma.size -= offset_in_vma;
+
+    switch (new_vma.type) {
+    case VirtualMemoryArea::Type::Unmapped:
+        break;
+    case VirtualMemoryArea::Type::Allocated:
+        new_vma.offset += offset_in_vma;
+        break;
+    case VirtualMemoryArea::Type::Mapped:
+        new_vma.backing_memory += offset_in_vma;
+        break;
+    }
+
+    ASSERT(old_vma.CanBeMergedWith(new_vma));
+
+    return vma_map.emplace_hint(std::next(vma_handle), new_vma.base, new_vma);
+}
+
+MemoryManager::VMAIter MemoryManager::MergeAdjacent(VMAIter iter) {
+    const VMAIter next_vma = std::next(iter);
+    if (next_vma != vma_map.end() && iter->second.CanBeMergedWith(next_vma->second)) {
+        iter->second.size += next_vma->second.size;
+        vma_map.erase(next_vma);
+    }
+
+    if (iter != vma_map.begin()) {
+        VMAIter prev_vma = std::prev(iter);
+        if (prev_vma->second.CanBeMergedWith(iter->second)) {
+            prev_vma->second.size += iter->second.size;
+            vma_map.erase(iter);
+            iter = prev_vma;
+        }
+    }
+
+    return iter;
+}
+
+void MemoryManager::UpdatePageTableForVMA(const VirtualMemoryArea& vma) {
+    switch (vma.type) {
+    case VirtualMemoryArea::Type::Unmapped:
+        UnmapRegion(vma.base, vma.size);
+        break;
+    case VirtualMemoryArea::Type::Allocated:
+        MapMemoryRegion(vma.base, vma.size, nullptr, vma.backing_addr);
+        break;
+    case VirtualMemoryArea::Type::Mapped:
+        MapMemoryRegion(vma.base, vma.size, vma.backing_memory, vma.backing_addr);
+        break;
     }
-    return (*block)[(gpu_addr >> PAGE_BITS) & PAGE_BLOCK_MASK];
 }
 
 } // namespace Tegra

src/video_core/memory_manager.h

@@ -1,79 +1,147 @@
-// Copyright 2018 yuzu emulator team
+// Copyright 2018 yuzu emulator team
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
 #pragma once
 
-#include <array>
+#include <map>
-#include <memory>
 #include <optional>
-#include <vector>
 
 #include "common/common_types.h"
+#include "common/page_table.h"
 
 namespace Tegra {
 
+/**
+ * Represents a VMA in an address space. A VMA is a contiguous region of virtual addressing space
+ * with homogeneous attributes across its extents. In this particular implementation each VMA is
+ * also backed by a single host memory allocation.
+ */
+struct VirtualMemoryArea {
+    enum class Type : u8 {
+        Unmapped,
+        Allocated,
+        Mapped,
+    };
+
+    /// Virtual base address of the region.
+    GPUVAddr base{};
+    /// Size of the region.
+    u64 size{};
+    /// Memory area mapping type.
+    Type type{Type::Unmapped};
+    /// CPU memory mapped address corresponding to this memory area.
+    VAddr backing_addr{};
+    /// Offset into the backing_memory the mapping starts from.
+    std::size_t offset{};
+    /// Pointer backing this VMA.
+    u8* backing_memory{};
+
+    /// Tests if this area can be merged to the right with `next`.
+    bool CanBeMergedWith(const VirtualMemoryArea& next) const;
+};
+
 class MemoryManager final {
 public:
     MemoryManager();
 
     GPUVAddr AllocateSpace(u64 size, u64 align);
     GPUVAddr AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align);
-    GPUVAddr MapBufferEx(VAddr cpu_addr, u64 size);
+    GPUVAddr MapBufferEx(GPUVAddr cpu_addr, u64 size);
-    GPUVAddr MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size);
+    GPUVAddr MapBufferEx(GPUVAddr cpu_addr, GPUVAddr gpu_addr, u64 size);
     GPUVAddr UnmapBuffer(GPUVAddr gpu_addr, u64 size);
-    GPUVAddr GetRegionEnd(GPUVAddr region_start) const;
     std::optional<VAddr> GpuToCpuAddress(GPUVAddr gpu_addr);
 
-    static constexpr u64 PAGE_BITS = 16;
+    template <typename T>
-    static constexpr u64 PAGE_SIZE = 1 << PAGE_BITS;
+    T Read(GPUVAddr vaddr);
-    static constexpr u64 PAGE_MASK = PAGE_SIZE - 1;
 
-    u8 Read8(GPUVAddr addr);
+    template <typename T>
-    u16 Read16(GPUVAddr addr);
+    void Write(GPUVAddr vaddr, T data);
-    u32 Read32(GPUVAddr addr);
-    u64 Read64(GPUVAddr addr);
-
-    void Write8(GPUVAddr addr, u8 data);
-    void Write16(GPUVAddr addr, u16 data);
-    void Write32(GPUVAddr addr, u32 data);
-    void Write64(GPUVAddr addr, u64 data);
 
     u8* GetPointer(GPUVAddr vaddr);
 
     void ReadBlock(GPUVAddr src_addr, void* dest_buffer, std::size_t size);
     void WriteBlock(GPUVAddr dest_addr, const void* src_buffer, std::size_t size);
-    void CopyBlock(VAddr dest_addr, VAddr src_addr, std::size_t size);
+    void CopyBlock(GPUVAddr dest_addr, GPUVAddr src_addr, std::size_t size);
 
 private:
-    enum class PageStatus : u64 {
+    using VMAMap = std::map<GPUVAddr, VirtualMemoryArea>;
-        Unmapped = 0xFFFFFFFFFFFFFFFFULL,
+    using VMAHandle = VMAMap::const_iterator;
-        Allocated = 0xFFFFFFFFFFFFFFFEULL,
+    using VMAIter = VMAMap::iterator;
-        Reserved = 0xFFFFFFFFFFFFFFFDULL,
-    };
 
-    std::optional<GPUVAddr> FindFreeBlock(GPUVAddr region_start, u64 size, u64 align,
+    void MapPages(GPUVAddr base, u64 size, u8* memory, Common::PageType type,
-                                          PageStatus status);
+                  VAddr backing_addr = 0);
-    VAddr& PageSlot(GPUVAddr gpu_addr);
+    void MapMemoryRegion(GPUVAddr base, u64 size, u8* target, VAddr backing_addr);
+    void UnmapRegion(GPUVAddr base, u64 size);
 
-    static constexpr u64 MAX_ADDRESS{0x10000000000ULL};
+    /// Finds the VMA in which the given address is included in, or `vma_map.end()`.
-    static constexpr u64 PAGE_TABLE_BITS{10};
+    VMAHandle FindVMA(GPUVAddr target) const;
-    static constexpr u64 PAGE_TABLE_SIZE{1 << PAGE_TABLE_BITS};
-    static constexpr u64 PAGE_TABLE_MASK{PAGE_TABLE_SIZE - 1};
-    static constexpr u64 PAGE_BLOCK_BITS{14};
-    static constexpr u64 PAGE_BLOCK_SIZE{1 << PAGE_BLOCK_BITS};
-    static constexpr u64 PAGE_BLOCK_MASK{PAGE_BLOCK_SIZE - 1};
 
-    using PageBlock = std::array<VAddr, PAGE_BLOCK_SIZE>;
+    VMAHandle AllocateMemory(GPUVAddr target, std::size_t offset, u64 size);
-    std::array<std::unique_ptr<PageBlock>, PAGE_TABLE_SIZE> page_table{};
 
-    struct MappedRegion {
+    /**
-        VAddr cpu_addr;
+     * Maps an unmanaged host memory pointer at a given address.
-        GPUVAddr gpu_addr;
+     *
-        u64 size;
+     * @param target The guest address to start the mapping at.
-    };
+     * @param memory The memory to be mapped.
+     * @param size Size of the mapping.
+     * @param state MemoryState tag to attach to the VMA.
+     */
+    VMAHandle MapBackingMemory(GPUVAddr target, u8* memory, u64 size, VAddr backing_addr);
 
-    std::vector<MappedRegion> mapped_regions;
+    /// Unmaps a range of addresses, splitting VMAs as necessary.
+    void UnmapRange(GPUVAddr target, u64 size);
+
+    /// Converts a VMAHandle to a mutable VMAIter.
+    VMAIter StripIterConstness(const VMAHandle& iter);
+
+    /// Unmaps the given VMA.
+    VMAIter Unmap(VMAIter vma);
+
+    /**
+     * Carves a VMA of a specific size at the specified address by splitting Free VMAs while doing
+     * the appropriate error checking.
+     */
+    VMAIter CarveVMA(GPUVAddr base, u64 size);
+
+    /**
+     * Splits the edges of the given range of non-Free VMAs so that there is a VMA split at each
+     * end of the range.
+     */
+    VMAIter CarveVMARange(GPUVAddr base, u64 size);
+
+    /**
+     * Splits a VMA in two, at the specified offset.
+     * @returns the right side of the split, with the original iterator becoming the left side.
+     */
+    VMAIter SplitVMA(VMAIter vma, u64 offset_in_vma);
+
+    /**
+     * Checks for and merges the specified VMA with adjacent ones if possible.
+     * @returns the merged VMA or the original if no merging was possible.
+     */
+    VMAIter MergeAdjacent(VMAIter vma);
+
+    /// Updates the pages corresponding to this VMA so they match the VMA's attributes.
+    void UpdatePageTableForVMA(const VirtualMemoryArea& vma);
+
+    GPUVAddr FindFreeRegion(GPUVAddr region_start, u64 size, u64 align,
+                            VirtualMemoryArea::Type vma_type);
+
+private:
+    static constexpr u64 page_bits{16};
+    static constexpr u64 page_size{1 << page_bits};
+    static constexpr u64 page_mask{page_size - 1};
+
+    /// Address space in bits, this is fairly arbitrary but sufficiently large.
+    static constexpr u32 address_space_width = 39;
+    /// Start address for mapping, this is fairly arbitrary but must be non-zero.
+    static constexpr GPUVAddr address_space_base = 0x100000;
+    /// End of address space, based on address space in bits.
+    static constexpr GPUVAddr address_space_end = 1ULL << address_space_width;
+
+    Common::PageTable page_table{page_bits};
+    VMAMap vma_map;
 };
 
 } // namespace Tegra

src/video_core/rasterizer_interface.h

@@ -9,7 +9,6 @@
 #include "common/common_types.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/gpu.h"
-#include "video_core/memory_manager.h"
 
 namespace VideoCore {
 

src/video_core/renderer_opengl/gl_global_cache.cpp

@@ -76,8 +76,8 @@ GlobalRegion GlobalRegionCacheOpenGL::GetGlobalRegion(
     const auto cbufs{gpu.Maxwell3D().state.shader_stages[static_cast<u64>(stage)]};
     const auto addr{cbufs.const_buffers[global_region.GetCbufIndex()].address +
                     global_region.GetCbufOffset()};
-    const auto actual_addr{memory_manager.Read64(addr)};
+    const auto actual_addr{memory_manager.Read<u64>(addr)};
-    const auto size{memory_manager.Read32(addr + 8)};
+    const auto size{memory_manager.Read<u32>(addr + 8)};
 
     // Look up global region in the cache based on address
     const auto& host_ptr{memory_manager.GetPointer(actual_addr)};

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -610,11 +610,11 @@ CachedSurface::CachedSurface(const SurfaceParams& params)
     // check is necessary to prevent flushing from overwriting unmapped memory.
 
     auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
-    const u64 max_size{memory_manager.GetRegionEnd(params.gpu_addr) - params.gpu_addr};
+    // const u64 max_size{memory_manager.GetRegionEnd(params.gpu_addr) - params.gpu_addr};
-    if (cached_size_in_bytes > max_size) {
+    // if (cached_size_in_bytes > max_size) {
-        LOG_ERROR(HW_GPU, "Surface size {} exceeds region size {}", params.size_in_bytes, max_size);
+    //    LOG_ERROR(HW_GPU, "Surface size {} exceeds region size {}", params.size_in_bytes,
-        cached_size_in_bytes = max_size;
+    //    max_size); cached_size_in_bytes = max_size;
-    }
+    //}
 
     cpu_addr = *memory_manager.GpuToCpuAddress(params.gpu_addr);
 }