NVDRV: Remake ASGPU

src/common/CMakeLists.txt

@@ -17,6 +17,8 @@ endif ()
 include(GenerateSCMRev)
 
 add_library(common STATIC
+    address_space.cpp
+    address_space.h
     algorithm.h
     alignment.h
     announce_multiplayer_room.h

src/common/address_space.cpp

@@ -0,0 +1,11 @@
+// Copyright © 2021 Skyline Team and Contributors (https://github.com/skyline-emu/)
+// Licensed under GPLv3 or any later version
+// Refer to the license.txt file included.
+
+#include "common/address_space.inc"
+
+namespace Common {
+
+template class Common::FlatAllocator<u32, 0, 32>;
+
+}

src/common/address_space.h

@@ -0,0 +1,134 @@
+// Copyright © 2021 Skyline Team and Contributors (https://github.com/skyline-emu/)
+// Licensed under GPLv3 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <concepts>
+#include <functional>
+#include <mutex>
+#include <vector>
+
+#include "common/common_types.h"
+
+namespace Common {
+template <typename VaType, size_t AddressSpaceBits>
+concept AddressSpaceValid = std::is_unsigned_v<VaType> && sizeof(VaType) * 8 >= AddressSpaceBits;
+
+struct EmptyStruct {};
+
+/**
+ * @brief FlatAddressSpaceMap provides a generic VA->PA mapping implementation using a sorted vector
+ */
+template <typename VaType, VaType UnmappedVa, typename PaType, PaType UnmappedPa,
+          bool PaContigSplit, size_t AddressSpaceBits, typename ExtraBlockInfo = EmptyStruct>
+requires AddressSpaceValid<VaType, AddressSpaceBits> class FlatAddressSpaceMap {
+private:
+    std::function<void(VaType, VaType)>
+        unmapCallback{}; //!< Callback called when the mappings in an region have changed
+
+protected:
+    /**
+     * @brief Represents a block of memory in the AS, the physical mapping is contiguous until
+     * another block with a different phys address is hit
+     */
+    struct Block {
+        VaType virt{UnmappedVa}; //!< VA of the block
+        PaType phys{UnmappedPa}; //!< PA of the block, will increase 1-1 with VA until a new block
+                                 //!< is encountered
+        [[no_unique_address]] ExtraBlockInfo extraInfo;
+
+        Block() = default;
+
+        Block(VaType virt, PaType phys, ExtraBlockInfo extraInfo)
+            : virt(virt), phys(phys), extraInfo(extraInfo) {}
+
+        constexpr bool Valid() {
+            return virt != UnmappedVa;
+        }
+
+        constexpr bool Mapped() {
+            return phys != UnmappedPa;
+        }
+
+        constexpr bool Unmapped() {
+            return phys == UnmappedPa;
+        }
+
+        bool operator<(const VaType& pVirt) const {
+            return virt < pVirt;
+        }
+    };
+
+    std::mutex blockMutex;
+    std::vector<Block> blocks{Block{}};
+
+    /**
+     * @brief Maps a PA range into the given AS region
+     * @note blockMutex MUST be locked when calling this
+     */
+    void MapLocked(VaType virt, PaType phys, VaType size, ExtraBlockInfo extraInfo);
+
+    /**
+     * @brief Unmaps the given range and merges it with other unmapped regions
+     * @note blockMutex MUST be locked when calling this
+     */
+    void UnmapLocked(VaType virt, VaType size);
+
+public:
+    static constexpr VaType VaMaximum{(1ULL << (AddressSpaceBits - 1)) +
+                                      ((1ULL << (AddressSpaceBits - 1)) -
+                                       1)}; //!< The maximum VA that this AS can technically reach
+
+    VaType vaLimit{VaMaximum}; //!< A soft limit on the maximum VA of the AS
+
+    FlatAddressSpaceMap(VaType vaLimit, std::function<void(VaType, VaType)> unmapCallback = {});
+
+    FlatAddressSpaceMap() = default;
+
+    void Map(VaType virt, PaType phys, VaType size, ExtraBlockInfo extraInfo = {}) {
+        std::scoped_lock lock(blockMutex);
+        MapLocked(virt, phys, size, extraInfo);
+    }
+
+    void Unmap(VaType virt, VaType size) {
+        std::scoped_lock lock(blockMutex);
+        UnmapLocked(virt, size);
+    }
+};
+
+/**
+ * @brief FlatMemoryManager specialises FlatAddressSpaceMap to work as an allocator, with an
+ * initial, fast linear pass and a subsequent slower pass that iterates until it finds a free block
+ */
+template <typename VaType, VaType UnmappedVa, size_t AddressSpaceBits>
+requires AddressSpaceValid<VaType, AddressSpaceBits> class FlatAllocator
+    : public FlatAddressSpaceMap<VaType, UnmappedVa, bool, false, false, AddressSpaceBits> {
+private:
+    using Base = FlatAddressSpaceMap<VaType, UnmappedVa, bool, false, false, AddressSpaceBits>;
+
+    VaType currentLinearAllocEnd; //!< The end address for the initial linear allocation pass, once
+                                  //!< this reaches the AS limit the slower allocation path will be
+                                  //!< used
+
+public:
+    VaType vaStart; //!< The base VA of the allocator, no allocations will be below this
+
+    FlatAllocator(VaType vaStart, VaType vaLimit = Base::VaMaximum);
+
+    /**
+     * @brief Allocates a region in the AS of the given size and returns its address
+     */
+    VaType Allocate(VaType size);
+
+    /**
+     * @brief Marks the given region in the AS as allocated
+     */
+    void AllocateFixed(VaType virt, VaType size);
+
+    /**
+     * @brief Frees an AS region so it can be used again
+     */
+    void Free(VaType virt, VaType size);
+};
+} // namespace Common

src/common/address_space.inc

@@ -0,0 +1,338 @@
+// SPDX-License-Identifier: GPLv3 or later
+// Copyright © 2021 Skyline Team and Contributors (https://github.com/skyline-emu/)
+
+#include "common/address_space.h"
+#include "common/assert.h"
+
+#define MAP_MEMBER(returnType)                                                                     \
+    template <typename VaType, VaType UnmappedVa, typename PaType, PaType UnmappedPa,              \
+              bool PaContigSplit, size_t AddressSpaceBits, typename ExtraBlockInfo>                \
+    requires AddressSpaceValid<VaType, AddressSpaceBits> returnType FlatAddressSpaceMap<           \
+        VaType, UnmappedVa, PaType, UnmappedPa, PaContigSplit, AddressSpaceBits, ExtraBlockInfo>
+#define MAP_MEMBER_CONST()                                                                         \
+    template <typename VaType, VaType UnmappedVa, typename PaType, PaType UnmappedPa,              \
+              bool PaContigSplit, size_t AddressSpaceBits, typename ExtraBlockInfo>                \
+    requires AddressSpaceValid<VaType, AddressSpaceBits> FlatAddressSpaceMap<                      \
+        VaType, UnmappedVa, PaType, UnmappedPa, PaContigSplit, AddressSpaceBits, ExtraBlockInfo>
+
+#define MM_MEMBER(returnType)                                                                      \
+    template <typename VaType, VaType UnmappedVa, size_t AddressSpaceBits>                         \
+    requires AddressSpaceValid<VaType, AddressSpaceBits> returnType                                \
+        FlatMemoryManager<VaType, UnmappedVa, AddressSpaceBits>
+
+#define ALLOC_MEMBER(returnType)                                                                   \
+    template <typename VaType, VaType UnmappedVa, size_t AddressSpaceBits>                         \
+    requires AddressSpaceValid<VaType, AddressSpaceBits> returnType                                \
+        FlatAllocator<VaType, UnmappedVa, AddressSpaceBits>
+#define ALLOC_MEMBER_CONST()                                                                       \
+    template <typename VaType, VaType UnmappedVa, size_t AddressSpaceBits>                         \
+    requires AddressSpaceValid<VaType, AddressSpaceBits>                                           \
+        FlatAllocator<VaType, UnmappedVa, AddressSpaceBits>
+
+namespace Common {
+MAP_MEMBER_CONST()::FlatAddressSpaceMap(VaType vaLimit,
+                                        std::function<void(VaType, VaType)> unmapCallback)
+    : unmapCallback(std::move(unmapCallback)), vaLimit(vaLimit) {
+    if (vaLimit > VaMaximum)
+        UNREACHABLE_MSG("Invalid VA limit!");
+}
+
+MAP_MEMBER(void)::MapLocked(VaType virt, PaType phys, VaType size, ExtraBlockInfo extraInfo) {
+    VaType virtEnd{virt + size};
+
+    if (virtEnd > vaLimit)
+        UNREACHABLE_MSG("Trying to map a block past the VA limit: virtEnd: 0x{:X}, vaLimit: 0x{:X}",
+                        virtEnd, vaLimit);
+
+    auto blockEndSuccessor{std::lower_bound(blocks.begin(), blocks.end(), virtEnd)};
+    if (blockEndSuccessor == blocks.begin())
+        UNREACHABLE_MSG("Trying to map a block before the VA start: virtEnd: 0x{:X}", virtEnd);
+
+    auto blockEndPredecessor{std::prev(blockEndSuccessor)};
+
+    if (blockEndSuccessor != blocks.end()) {
+        // We have blocks in front of us, if one is directly in front then we don't have to add a
+        // tail
+        if (blockEndSuccessor->virt != virtEnd) {
+            PaType tailPhys{[&]() -> PaType {
+                if constexpr (!PaContigSplit) {
+                    return blockEndPredecessor
+                        ->phys; // Always propagate unmapped regions rather than calculating offset
+                } else {
+                    if (blockEndPredecessor->Unmapped())
+                        return blockEndPredecessor->phys; // Always propagate unmapped regions
+                                                          // rather than calculating offset
+                    else
+                        return blockEndPredecessor->phys + virtEnd - blockEndPredecessor->virt;
+                }
+            }()};
+
+            if (blockEndPredecessor->virt >= virt) {
+                // If this block's start would be overlapped by the map then reuse it as a tail
+                // block
+                blockEndPredecessor->virt = virtEnd;
+                blockEndPredecessor->phys = tailPhys;
+                blockEndPredecessor->extraInfo = blockEndPredecessor->extraInfo;
+
+                // No longer predecessor anymore
+                blockEndSuccessor = blockEndPredecessor--;
+            } else {
+                // Else insert a new one and we're done
+                blocks.insert(blockEndSuccessor,
+                              {Block(virt, phys, extraInfo),
+                               Block(virtEnd, tailPhys, blockEndPredecessor->extraInfo)});
+                if (unmapCallback)
+                    unmapCallback(virt, size);
+
+                return;
+            }
+        }
+    } else {
+        // blockEndPredecessor will always be unmapped as blocks has to be terminated by an unmapped
+        // chunk
+        if (blockEndPredecessor != blocks.begin() && blockEndPredecessor->virt >= virt) {
+            // Move the unmapped block start backwards
+            blockEndPredecessor->virt = virtEnd;
+
+            // No longer predecessor anymore
+            blockEndSuccessor = blockEndPredecessor--;
+        } else {
+            // Else insert a new one and we're done
+            blocks.insert(blockEndSuccessor,
+                          {Block(virt, phys, extraInfo), Block(virtEnd, UnmappedPa, {})});
+            if (unmapCallback)
+                unmapCallback(virt, size);
+
+            return;
+        }
+    }
+
+    auto blockStartSuccessor{blockEndSuccessor};
+
+    // Walk the block vector to find the start successor as this is more efficient than another
+    // binary search in most scenarios
+    while (std::prev(blockStartSuccessor)->virt >= virt)
+        blockStartSuccessor--;
+
+    // Check that the start successor is either the end block or something in between
+    if (blockStartSuccessor->virt > virtEnd) {
+        UNREACHABLE_MSG("Unsorted block in AS map: virt: 0x{:X}", blockStartSuccessor->virt);
+    } else if (blockStartSuccessor->virt == virtEnd) {
+        // We need to create a new block as there are none spare that we would overwrite
+        blocks.insert(blockStartSuccessor, Block(virt, phys, extraInfo));
+    } else {
+        // Erase overwritten blocks
+        if (auto eraseStart{std::next(blockStartSuccessor)}; eraseStart != blockEndSuccessor)
+            blocks.erase(eraseStart, blockEndSuccessor);
+
+        // Reuse a block that would otherwise be overwritten as a start block
+        blockStartSuccessor->virt = virt;
+        blockStartSuccessor->phys = phys;
+        blockStartSuccessor->extraInfo = extraInfo;
+    }
+
+    if (unmapCallback)
+        unmapCallback(virt, size);
+}
+
+MAP_MEMBER(void)::UnmapLocked(VaType virt, VaType size) {
+    VaType virtEnd{virt + size};
+
+    if (virtEnd > vaLimit)
+        UNREACHABLE_MSG("Trying to map a block past the VA limit: virtEnd: 0x{:X}, vaLimit: 0x{:X}",
+                        virtEnd, vaLimit);
+
+    auto blockEndSuccessor{std::lower_bound(blocks.begin(), blocks.end(), virtEnd)};
+    if (blockEndSuccessor == blocks.begin())
+        UNREACHABLE_MSG("Trying to unmap a block before the VA start: virtEnd: 0x{:X}", virtEnd);
+
+    auto blockEndPredecessor{std::prev(blockEndSuccessor)};
+
+    auto walkBackToPredecessor{[&](auto iter) {
+        while (iter->virt >= virt)
+            iter--;
+
+        return iter;
+    }};
+
+    auto eraseBlocksWithEndUnmapped{[&](auto unmappedEnd) {
+        auto blockStartPredecessor{walkBackToPredecessor(unmappedEnd)};
+        auto blockStartSuccessor{std::next(blockStartPredecessor)};
+
+        auto eraseEnd{[&]() {
+            if (blockStartPredecessor->Unmapped()) {
+                // If the start predecessor is unmapped then we can erase everything in our region
+                // and be done
+                return std::next(unmappedEnd);
+            } else {
+                // Else reuse the end predecessor as the start of our unmapped region then erase all
+                // up to it
+                unmappedEnd->virt = virt;
+                return unmappedEnd;
+            }
+        }()};
+
+        // We can't have two unmapped regions after each other
+        if (eraseEnd != blocks.end() &&
+            (eraseEnd == blockStartSuccessor ||
+             (blockStartPredecessor->Unmapped() && eraseEnd->Unmapped())))
+            UNREACHABLE_MSG("Multiple contiguous unmapped regions are unsupported!");
+
+        blocks.erase(blockStartSuccessor, eraseEnd);
+    }};
+
+    // We can avoid any splitting logic if these are the case
+    if (blockEndPredecessor->Unmapped()) {
+        if (blockEndPredecessor->virt > virt)
+            eraseBlocksWithEndUnmapped(blockEndPredecessor);
+
+        if (unmapCallback)
+            unmapCallback(virt, size);
+
+        return; // The region is unmapped, bail out early
+    } else if (blockEndSuccessor->virt == virtEnd && blockEndSuccessor->Unmapped()) {
+        eraseBlocksWithEndUnmapped(blockEndSuccessor);
+
+        if (unmapCallback)
+            unmapCallback(virt, size);
+
+        return; // The region is unmapped here and doesn't need splitting, bail out early
+    } else if (blockEndSuccessor == blocks.end()) {
+        // This should never happen as the end should always follow an unmapped block
+        UNREACHABLE_MSG("Unexpected Memory Manager state!");
+    } else if (blockEndSuccessor->virt != virtEnd) {
+        // If one block is directly in front then we don't have to add a tail
+
+        // The previous block is mapped so we will need to add a tail with an offset
+        PaType tailPhys{[&]() {
+            if constexpr (PaContigSplit)
+                return blockEndPredecessor->phys + virtEnd - blockEndPredecessor->virt;
+            else
+                return blockEndPredecessor->phys;
+        }()};
+
+        if (blockEndPredecessor->virt >= virt) {
+            // If this block's start would be overlapped by the unmap then reuse it as a tail block
+            blockEndPredecessor->virt = virtEnd;
+            blockEndPredecessor->phys = tailPhys;
+
+            // No longer predecessor anymore
+            blockEndSuccessor = blockEndPredecessor--;
+        } else {
+            blocks.insert(blockEndSuccessor,
+                          {Block(virt, UnmappedPa, {}),
+                           Block(virtEnd, tailPhys, blockEndPredecessor->extraInfo)});
+            if (unmapCallback)
+                unmapCallback(virt, size);
+
+            return; // The previous block is mapped and ends before
+        }
+    }
+
+    // Walk the block vector to find the start predecessor as this is more efficient than another
+    // binary search in most scenarios
+    auto blockStartPredecessor{walkBackToPredecessor(blockEndSuccessor)};
+    auto blockStartSuccessor{std::next(blockStartPredecessor)};
+
+    if (blockStartSuccessor->virt > virtEnd) {
+        UNREACHABLE_MSG("Unsorted block in AS map: virt: 0x{:X}", blockStartSuccessor->virt);
+    } else if (blockStartSuccessor->virt == virtEnd) {
+        // There are no blocks between the start and the end that would let us skip inserting a new
+        // one for head
+
+        // The previous block is may be unmapped, if so we don't need to insert any unmaps after it
+        if (blockStartPredecessor->Mapped())
+            blocks.insert(blockStartSuccessor, Block(virt, UnmappedPa, {}));
+    } else if (blockStartPredecessor->Unmapped()) {
+        // If the previous block is unmapped
+        blocks.erase(blockStartSuccessor, blockEndPredecessor);
+    } else {
+        // Erase overwritten blocks, skipping the first one as we have written the unmapped start
+        // block there
+        if (auto eraseStart{std::next(blockStartSuccessor)}; eraseStart != blockEndSuccessor)
+            blocks.erase(eraseStart, blockEndSuccessor);
+
+        // Add in the unmapped block header
+        blockStartSuccessor->virt = virt;
+        blockStartSuccessor->phys = UnmappedPa;
+    }
+
+    if (unmapCallback)
+        unmapCallback(virt, size);
+}
+
+ALLOC_MEMBER_CONST()::FlatAllocator(VaType vaStart, VaType vaLimit)
+    : Base(vaLimit), currentLinearAllocEnd(vaStart), vaStart(vaStart) {}
+
+ALLOC_MEMBER(VaType)::Allocate(VaType size) {
+    std::scoped_lock lock(this->blockMutex);
+
+    VaType allocStart{UnmappedVa};
+    VaType allocEnd{currentLinearAllocEnd + size};
+
+    // Avoid searching backwards in the address space if possible
+    if (allocEnd >= currentLinearAllocEnd && allocEnd <= this->vaLimit) {
+        auto allocEndSuccessor{
+            std::lower_bound(this->blocks.begin(), this->blocks.end(), allocEnd)};
+        if (allocEndSuccessor == this->blocks.begin())
+            UNREACHABLE_MSG("First block in AS map is invalid!");
+
+        auto allocEndPredecessor{std::prev(allocEndSuccessor)};
+        if (allocEndPredecessor->virt <= currentLinearAllocEnd) {
+            allocStart = currentLinearAllocEnd;
+        } else {
+            // Skip over fixed any mappings in front of us
+            while (allocEndSuccessor != this->blocks.end()) {
+                if (allocEndSuccessor->virt - allocEndPredecessor->virt < size ||
+                    allocEndPredecessor->Mapped()) {
+                    allocStart = allocEndPredecessor->virt;
+                    break;
+                }
+
+                allocEndPredecessor = allocEndSuccessor++;
+
+                // Use the VA limit to calculate if we can fit in the final block since it has no
+                // successor
+                if (allocEndSuccessor == this->blocks.end()) {
+                    allocEnd = allocEndPredecessor->virt + size;
+
+                    if (allocEnd >= allocEndPredecessor->virt && allocEnd <= this->vaLimit)
+                        allocStart = allocEndPredecessor->virt;
+                }
+            }
+        }
+    }
+
+    if (allocStart != UnmappedVa) {
+        currentLinearAllocEnd = allocStart + size;
+    } else { // If linear allocation overflows the AS then find a gap
+        if (this->blocks.size() <= 2)
+            UNREACHABLE_MSG("Unexpected allocator state!");
+
+        auto searchPredecessor{this->blocks.begin()};
+        auto searchSuccessor{std::next(searchPredecessor)};
+
+        while (searchSuccessor != this->blocks.end() &&
+               (searchSuccessor->virt - searchPredecessor->virt < size ||
+                searchPredecessor->Mapped())) {
+            searchPredecessor = searchSuccessor++;
+        }
+
+        if (searchSuccessor != this->blocks.end())
+            allocStart = searchPredecessor->virt;
+        else
+            return {}; // AS is full
+    }
+
+    this->MapLocked(allocStart, true, size, {});
+    return allocStart;
+}
+
+ALLOC_MEMBER(void)::AllocateFixed(VaType virt, VaType size) {
+    this->Map(virt, true, size);
+}
+
+ALLOC_MEMBER(void)::Free(VaType virt, VaType size) {
+    this->Unmap(virt, size);
+}
+} // namespace Common

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

@@ -6,6 +6,7 @@
 #include <cstring>
 #include <utility>
 
+#include "common/alignment.h"
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "core/core.h"
@@ -21,8 +22,8 @@
 namespace Service::Nvidia::Devices {
 
 nvhost_as_gpu::nvhost_as_gpu(Core::System& system_, Module& module_, NvCore::Container& core)
-    : nvdevice{system_}, module{module_}, container{core}, nvmap{core.GetNvMapFile()},
+    : nvdevice{system_}, module{module_}, container{core}, nvmap{core.GetNvMapFile()}, vm{},
-      gmmu{std::make_shared<Tegra::MemoryManager>(system)} {}
+      gmmu{} {}
 nvhost_as_gpu::~nvhost_as_gpu() = default;
 
 NvResult nvhost_as_gpu::Ioctl1(DeviceFD fd, Ioctl command, const std::vector<u8>& input,
@@ -89,12 +90,49 @@ NvResult nvhost_as_gpu::AllocAsEx(const std::vector<u8>& input, std::vector<u8>&
     IoctlAllocAsEx params{};
     std::memcpy(&params, input.data(), input.size());
 
-    LOG_WARNING(Service_NVDRV, "(STUBBED) called, big_page_size=0x{:X}", params.big_page_size);
+    LOG_DEBUG(Service_NVDRV, "called, big_page_size=0x{:X}", params.big_page_size);
-    if (params.big_page_size == 0) {
+
-        params.big_page_size = DEFAULT_BIG_PAGE_SIZE;
+    std::scoped_lock lock(mutex);
+
+    if (vm.initialised) {
+        UNREACHABLE_MSG("Cannot initialise an address space twice!");
+        return NvResult::InvalidState;
     }
 
-    big_page_size = params.big_page_size;
+    if (params.big_page_size) {
+        if (!std::has_single_bit(params.big_page_size)) {
+            LOG_ERROR(Service_NVDRV, "Non power-of-2 big page size: 0x{:X}!", params.big_page_size);
+            return NvResult::BadValue;
+        }
+
+        if (!(params.big_page_size & VM::SUPPORTED_BIG_PAGE_SIZES)) {
+            LOG_ERROR(Service_NVDRV, "Unsupported big page size: 0x{:X}!", params.big_page_size);
+            return NvResult::BadValue;
+        }
+
+        vm.big_page_size = params.big_page_size;
+        vm.big_page_size_bits = static_cast<u32>(std::countr_zero(params.big_page_size));
+
+        vm.va_range_start = params.big_page_size << VM::VA_START_SHIFT;
+    }
+
+    // If this is unspecified then default values should be used
+    if (params.va_range_start) {
+        vm.va_range_start = params.va_range_start;
+        vm.va_range_split = params.va_range_split;
+        vm.va_range_end = params.va_range_end;
+    }
+
+    const u64 start_pages{vm.va_range_start >> VM::PAGE_SIZE_BITS};
+    const u64 end_pages{vm.va_range_split >> VM::PAGE_SIZE_BITS};
+    vm.small_page_allocator = std::make_shared<VM::Allocator>(start_pages, end_pages);
+
+    const u64 start_big_pages{vm.va_range_split >> vm.big_page_size_bits};
+    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);
+    vm.initialised = true;
 
     return NvResult::Success;
 }
@@ -106,21 +144,73 @@ NvResult nvhost_as_gpu::AllocateSpace(const std::vector<u8>& input, std::vector<
     LOG_DEBUG(Service_NVDRV, "called, pages={:X}, page_size={:X}, flags={:X}", params.pages,
               params.page_size, params.flags);
 
-    const auto size{static_cast<u64>(params.pages) * static_cast<u64>(params.page_size)};
+    std::scoped_lock lock(mutex);
-    if ((params.flags & AddressSpaceFlags::FixedOffset) != AddressSpaceFlags::None) {
+
-        params.offset = *(gmmu->AllocateFixed(params.offset, size));
+    if (!vm.initialised) {
-    } else {
+        return NvResult::BadValue;
-        params.offset = gmmu->Allocate(size, params.align);
     }
 
-    auto result = NvResult::Success;
+    if (params.page_size != VM::YUZU_PAGESIZE && params.page_size != vm.big_page_size) {
-    if (!params.offset) {
+        return NvResult::BadValue;
-        LOG_CRITICAL(Service_NVDRV, "allocation failed for size {}", size);
-        result = NvResult::InsufficientMemory;
     }
 
+    if (params.page_size != vm.big_page_size &&
+        ((params.flags & MappingFlags::Sparse) != MappingFlags::None)) {
+        UNIMPLEMENTED_MSG("Sparse small pages are not implemented!");
+        return NvResult::NotImplemented;
+    }
+
+    const u32 page_size_bits{params.page_size == VM::YUZU_PAGESIZE ? VM::PAGE_SIZE_BITS
+                                                                   : vm.big_page_size_bits};
+
+    auto& allocator{params.page_size == VM::YUZU_PAGESIZE ? *vm.small_page_allocator
+                                                          : *vm.big_page_allocator};
+
+    if ((params.flags & MappingFlags::Fixed) != MappingFlags::None) {
+        allocator.AllocateFixed(static_cast<u32>(params.offset >> page_size_bits), params.pages);
+    } else {
+        params.offset = static_cast<u64>(allocator.Allocate(params.pages)) << page_size_bits;
+        if (!params.offset) {
+            UNREACHABLE_MSG("Failed to allocate free space in the GPU AS!");
+            return NvResult::InsufficientMemory;
+        }
+    }
+
+    u64 size{static_cast<u64>(params.pages) * params.page_size};
+
+    if ((params.flags & MappingFlags::Sparse) != MappingFlags::None) {
+        gmmu->MapSparse(params.offset, size);
+    }
+
+    allocation_map[params.offset] = {
+        .size = size,
+        .page_size = params.page_size,
+        .sparse = (params.flags & MappingFlags::Sparse) != MappingFlags::None,
+    };
+
     std::memcpy(output.data(), &params, output.size());
-    return result;
+    return NvResult::Success;
+}
+
+void nvhost_as_gpu::FreeMappingLocked(u64 offset) {
+    auto mapping{mapping_map.at(offset)};
+
+    if (!mapping->fixed) {
+        auto& allocator{mapping->big_page ? *vm.big_page_allocator : *vm.small_page_allocator};
+        u32 page_size_bits{mapping->big_page ? vm.big_page_size_bits : VM::PAGE_SIZE_BITS};
+
+        allocator.Free(static_cast<u32>(mapping->offset >> page_size_bits),
+                       static_cast<u32>(mapping->size >> page_size_bits));
+    }
+
+    // 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);
+    else
+        gmmu->Unmap(offset, mapping->size);
+
+    mapping_map.erase(offset);
 }
 
 NvResult nvhost_as_gpu::FreeSpace(const std::vector<u8>& input, std::vector<u8>& output) {
@@ -130,7 +220,40 @@ NvResult nvhost_as_gpu::FreeSpace(const std::vector<u8>& input, std::vector<u8>&
     LOG_DEBUG(Service_NVDRV, "called, offset={:X}, pages={:X}, page_size={:X}", params.offset,
               params.pages, params.page_size);
 
-    gmmu->Unmap(params.offset, static_cast<std::size_t>(params.pages) * params.page_size);
+    std::scoped_lock lock(mutex);
+
+    if (!vm.initialised) {
+        return NvResult::BadValue;
+    }
+
+    try {
+        auto allocation{allocation_map[params.offset]};
+
+        if (allocation.page_size != params.page_size ||
+            allocation.size != (static_cast<u64>(params.pages) * params.page_size)) {
+            return NvResult::BadValue;
+        }
+
+        for (const auto& mapping : allocation.mappings) {
+            FreeMappingLocked(mapping->offset);
+        }
+
+        // Unset sparse flag if required
+        if (allocation.sparse) {
+            gmmu->Unmap(params.offset, allocation.size);
+        }
+
+        auto& allocator{params.page_size == VM::YUZU_PAGESIZE ? *vm.small_page_allocator
+                                                              : *vm.big_page_allocator};
+        u32 page_size_bits{params.page_size == VM::YUZU_PAGESIZE ? VM::PAGE_SIZE_BITS
+                                                                 : vm.big_page_size_bits};
+
+        allocator.Free(static_cast<u32>(params.offset >> page_size_bits),
+                       static_cast<u32>(allocation.size >> page_size_bits));
+        allocation_map.erase(params.offset);
+    } catch ([[maybe_unused]] const std::out_of_range& e) {
+        return NvResult::BadValue;
+    }
 
     std::memcpy(output.data(), &params, output.size());
     return NvResult::Success;
@@ -141,43 +264,51 @@ NvResult nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& out
 
     LOG_DEBUG(Service_NVDRV, "called, num_entries=0x{:X}", num_entries);
 
-    auto result = NvResult::Success;
     std::vector<IoctlRemapEntry> entries(num_entries);
     std::memcpy(entries.data(), input.data(), input.size());
 
+    std::scoped_lock lock(mutex);
+
+    if (!vm.initialised) {
+        return NvResult::BadValue;
+    }
+
     for (const auto& entry : entries) {
-        LOG_DEBUG(Service_NVDRV, "remap entry, offset=0x{:X} handle=0x{:X} pages=0x{:X}",
+        GPUVAddr virtual_address{static_cast<u64>(entry.as_offset_big_pages)
-                  entry.offset, entry.nvmap_handle, entry.pages);
+                                 << vm.big_page_size_bits};
+        u64 size{static_cast<u64>(entry.big_pages) << vm.big_page_size_bits};
 
-        if (entry.nvmap_handle == 0) {
+        auto alloc{allocation_map.upper_bound(virtual_address)};
-            // If nvmap handle is null, we should unmap instead.
+
-            const auto offset{static_cast<GPUVAddr>(entry.offset) << 0x10};
+        if (alloc-- == allocation_map.begin() ||
-            const auto size{static_cast<u64>(entry.pages) << 0x10};
+            (virtual_address - alloc->first) + size > alloc->second.size) {
-            gmmu->Unmap(offset, size);
+            LOG_WARNING(Service_NVDRV, "Cannot remap into an unallocated region!");
-            continue;
+            return NvResult::BadValue;
         }
 
-        const auto object{nvmap.GetHandle(entry.nvmap_handle)};
+        if (!alloc->second.sparse) {
-        if (!object) {
+            LOG_WARNING(Service_NVDRV, "Cannot remap a non-sparse mapping!");
-            LOG_CRITICAL(Service_NVDRV, "invalid nvmap_handle={:X}", entry.nvmap_handle);
+            return NvResult::BadValue;
-            result = NvResult::InvalidState;
-            break;
         }
 
-        const auto offset{static_cast<GPUVAddr>(entry.offset) << 0x10};
+        if (!entry.handle) {
-        const auto size{static_cast<u64>(entry.pages) << 0x10};
+            gmmu->MapSparse(virtual_address, size);
-        const auto map_offset{static_cast<u64>(entry.map_offset) << 0x10};
+        } else {
-        const auto addr{gmmu->Map(object->address + map_offset, offset, size)};
+            auto handle{nvmap.GetHandle(entry.handle)};
+            if (!handle) {
+                return NvResult::BadValue;
+            }
 
-        if (!addr) {
+            VAddr cpu_address{static_cast<VAddr>(
-            LOG_CRITICAL(Service_NVDRV, "map returned an invalid address!");
+                handle->address +
-            result = NvResult::InvalidState;
+                (static_cast<u64>(entry.handle_offset_big_pages) << vm.big_page_size_bits))};
-            break;
+
+            gmmu->Map(virtual_address, cpu_address, size);
         }
     }
 
     std::memcpy(output.data(), entries.data(), output.size());
-    return result;
+    return NvResult::Success;
 }
 
 NvResult nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8>& output) {
@@ -187,75 +318,96 @@ NvResult nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8
     LOG_DEBUG(Service_NVDRV,
               "called, flags={:X}, nvmap_handle={:X}, buffer_offset={}, mapping_size={}"
               ", offset={}",
-              params.flags, params.nvmap_handle, params.buffer_offset, params.mapping_size,
+              params.flags, params.handle, params.buffer_offset, params.mapping_size,
               params.offset);
 
-    if ((params.flags & AddressSpaceFlags::Remap) != AddressSpaceFlags::None) {
+    std::scoped_lock lock(mutex);
-        if (const auto buffer_map{FindBufferMap(params.offset)}; buffer_map) {
-            const auto cpu_addr{static_cast<VAddr>(buffer_map->CpuAddr() + params.buffer_offset)};
-            const auto gpu_addr{static_cast<GPUVAddr>(params.offset + params.buffer_offset)};
 
-            if (!gmmu->Map(cpu_addr, gpu_addr, params.mapping_size)) {
+    if (!vm.initialised) {
-                LOG_CRITICAL(Service_NVDRV,
+        return NvResult::BadValue;
-                             "remap failed, flags={:X}, nvmap_handle={:X}, buffer_offset={}, "
+    }
-                             "mapping_size = {}, offset={}",
-                             params.flags, params.nvmap_handle, params.buffer_offset,
-                             params.mapping_size, params.offset);
 
-                std::memcpy(output.data(), &params, output.size());
+    // Remaps a subregion of an existing mapping to a different PA
-                return NvResult::InvalidState;
+    if ((params.flags & MappingFlags::Remap) != MappingFlags::None) {
+        try {
+            auto mapping{mapping_map.at(params.offset)};
+
+            if (mapping->size < params.mapping_size) {
+                LOG_WARNING(Service_NVDRV,
+                            "Cannot remap a partially mapped GPU address space region: 0x{:X}",
+                            params.offset);
+                return NvResult::BadValue;
+            }
+
+            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);
+
+            return NvResult::Success;
+        } catch ([[maybe_unused]] const std::out_of_range& e) {
+            LOG_WARNING(Service_NVDRV, "Cannot remap an unmapped GPU address space region: 0x{:X}",
+                        params.offset);
+            return NvResult::BadValue;
+        }
+    }
+
+    auto handle{nvmap.GetHandle(params.handle)};
+    if (!handle) {
+        return NvResult::BadValue;
+    }
+
+    VAddr cpu_address{static_cast<VAddr>(handle->address + params.buffer_offset)};
+    u64 size{params.mapping_size ? params.mapping_size : handle->orig_size};
+
+    if ((params.flags & MappingFlags::Fixed) != MappingFlags::None) {
+        auto alloc{allocation_map.upper_bound(params.offset)};
+
+        if (alloc-- == allocation_map.begin() ||
+            (params.offset - alloc->first) + size > alloc->second.size) {
+            UNREACHABLE_MSG("Cannot perform a fixed mapping into an unallocated region!");
+            return NvResult::BadValue;
+        }
+
+        gmmu->Map(params.offset, cpu_address, size);
+
+        auto mapping{std::make_shared<Mapping>(cpu_address, params.offset, size, true, false,
+                                               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};
+        u32 page_size_bits{big_page ? vm.big_page_size_bits : VM::PAGE_SIZE_BITS};
+
+        params.offset = static_cast<u64>(allocator.Allocate(
+                            static_cast<u32>(Common::AlignUp(size, page_size) >> page_size_bits)))
+                        << page_size_bits;
+        if (!params.offset) {
+            UNREACHABLE_MSG("Failed to allocate free space in the GPU AS!");
+            return NvResult::InsufficientMemory;
+        }
+
+        gmmu->Map(params.offset, cpu_address, size);
+
+        auto mapping{
+            std::make_shared<Mapping>(cpu_address, params.offset, size, false, big_page, false)};
+        mapping_map[params.offset] = mapping;
     }
 
     std::memcpy(output.data(), &params, output.size());
     return NvResult::Success;
-        } else {
-            LOG_CRITICAL(Service_NVDRV, "address not mapped offset={}", params.offset);
-
-            std::memcpy(output.data(), &params, output.size());
-            return NvResult::InvalidState;
-        }
-    }
-
-    const auto object{nvmap.GetHandle(params.nvmap_handle)};
-    if (!object) {
-        LOG_CRITICAL(Service_NVDRV, "invalid nvmap_handle={:X}", params.nvmap_handle);
-        std::memcpy(output.data(), &params, output.size());
-        return NvResult::InvalidState;
-    }
-
-    // The real nvservices doesn't make a distinction between handles and ids, and
-    // object can only have one handle and it will be the same as its id. Assert that this is the
-    // case to prevent unexpected behavior.
-    ASSERT(object->id == params.nvmap_handle);
-
-    u64 page_size{params.page_size};
-    if (!page_size) {
-        page_size = object->align;
-    }
-
-    const auto physical_address{object->address + params.buffer_offset};
-    u64 size{params.mapping_size};
-    if (!size) {
-        size = object->size;
-    }
-
-    const bool is_alloc{(params.flags & AddressSpaceFlags::FixedOffset) == AddressSpaceFlags::None};
-    if (is_alloc) {
-        params.offset = gmmu->MapAllocate(physical_address, size, page_size);
-    } else {
-        params.offset = gmmu->Map(physical_address, params.offset, size);
-    }
-
-    auto result = NvResult::Success;
-    if (!params.offset) {
-        LOG_CRITICAL(Service_NVDRV, "failed to map size={}", size);
-        result = NvResult::InvalidState;
-    } else {
-        AddBufferMap(params.offset, size, physical_address, is_alloc);
-    }
-
-    std::memcpy(output.data(), &params, output.size());
-    return result;
 }
 
 NvResult nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& output) {
@@ -264,13 +416,36 @@ NvResult nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8
 
     LOG_DEBUG(Service_NVDRV, "called, offset=0x{:X}", params.offset);
 
-    if (const auto size{RemoveBufferMap(params.offset)}; size) {
+    std::scoped_lock lock(mutex);
-        gmmu->Unmap(params.offset, *size);
+
-    } else {
+    if (!vm.initialised) {
-        LOG_ERROR(Service_NVDRV, "invalid offset=0x{:X}", params.offset);
+        return NvResult::BadValue;
+    }
+
+    try {
+        auto mapping{mapping_map.at(params.offset)};
+
+        if (!mapping->fixed) {
+            auto& allocator{mapping->big_page ? *vm.big_page_allocator : *vm.small_page_allocator};
+            u32 page_size_bits{mapping->big_page ? vm.big_page_size_bits : VM::PAGE_SIZE_BITS};
+
+            allocator.Free(static_cast<u32>(mapping->offset >> page_size_bits),
+                           static_cast<u32>(mapping->size >> page_size_bits));
+        }
+
+        // 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);
+        } else {
+            gmmu->Unmap(params.offset, mapping->size);
+        }
+
+        mapping_map.erase(params.offset);
+    } catch ([[maybe_unused]] const std::out_of_range& e) {
+        LOG_WARNING(Service_NVDRV, "Couldn't find region to unmap at 0x{:X}", params.offset);
     }
 
-    std::memcpy(output.data(), &params, output.size());
     return NvResult::Success;
 }
 
@@ -284,28 +459,37 @@ NvResult nvhost_as_gpu::BindChannel(const std::vector<u8>& input, std::vector<u8
     return NvResult::Success;
 }
 
+void nvhost_as_gpu::GetVARegionsImpl(IoctlGetVaRegions& params) {
+    params.buf_size = 2 * sizeof(VaRegion);
+
+    params.regions = std::array<VaRegion, 2>{
+        VaRegion{
+            .offset = vm.small_page_allocator->vaStart << VM::PAGE_SIZE_BITS,
+            .page_size = VM::YUZU_PAGESIZE,
+            .pages = vm.small_page_allocator->vaLimit - vm.small_page_allocator->vaStart,
+        },
+        VaRegion{
+            .offset = vm.big_page_allocator->vaStart << vm.big_page_size_bits,
+            .page_size = vm.big_page_size,
+            .pages = vm.big_page_allocator->vaLimit - vm.big_page_allocator->vaStart,
+        },
+    };
+}
+
 NvResult nvhost_as_gpu::GetVARegions(const std::vector<u8>& input, std::vector<u8>& output) {
     IoctlGetVaRegions params{};
     std::memcpy(&params, input.data(), input.size());
 
-    LOG_WARNING(Service_NVDRV, "(STUBBED) called, buf_addr={:X}, buf_size={:X}", params.buf_addr,
+    LOG_DEBUG(Service_NVDRV, "called, buf_addr={:X}, buf_size={:X}", params.buf_addr,
               params.buf_size);
 
-    params.buf_size = 0x30;
+    std::scoped_lock lock(mutex);
 
-    params.small = IoctlVaRegion{
+    if (!vm.initialised) {
-        .offset = 0x04000000,
+        return NvResult::BadValue;
-        .page_size = DEFAULT_SMALL_PAGE_SIZE,
+    }
-        .pages = 0x3fbfff,
-    };
 
-    params.big = IoctlVaRegion{
+    GetVARegionsImpl(params);
-        .offset = 0x04000000,
-        .page_size = big_page_size,
-        .pages = 0x1bffff,
-    };
-
-    // TODO(ogniK): This probably can stay stubbed but should add support way way later
 
     std::memcpy(output.data(), &params, output.size());
     return NvResult::Success;
@@ -316,64 +500,24 @@ NvResult nvhost_as_gpu::GetVARegions(const std::vector<u8>& input, std::vector<u
     IoctlGetVaRegions params{};
     std::memcpy(&params, input.data(), input.size());
 
-    LOG_WARNING(Service_NVDRV, "(STUBBED) called, buf_addr={:X}, buf_size={:X}", params.buf_addr,
+    LOG_DEBUG(Service_NVDRV, "called, buf_addr={:X}, buf_size={:X}", params.buf_addr,
               params.buf_size);
 
-    params.buf_size = 0x30;
+    std::scoped_lock lock(mutex);
 
-    params.small = IoctlVaRegion{
+    if (!vm.initialised) {
-        .offset = 0x04000000,
+        return NvResult::BadValue;
-        .page_size = 0x1000,
+    }
-        .pages = 0x3fbfff,
-    };
 
-    params.big = IoctlVaRegion{
+    GetVARegionsImpl(params);
-        .offset = 0x04000000,
-        .page_size = big_page_size,
-        .pages = 0x1bffff,
-    };
-
-    // TODO(ogniK): This probably can stay stubbed but should add support way way later
 
     std::memcpy(output.data(), &params, output.size());
-    std::memcpy(inline_output.data(), &params.small, sizeof(IoctlVaRegion));
+    std::memcpy(inline_output.data(), &params.regions[0], sizeof(VaRegion));
-    std::memcpy(inline_output.data() + sizeof(IoctlVaRegion), &params.big, sizeof(IoctlVaRegion));
+    std::memcpy(inline_output.data() + sizeof(VaRegion), &params.regions[1], sizeof(VaRegion));
 
     return NvResult::Success;
 }
 
-std::optional<nvhost_as_gpu::BufferMap> nvhost_as_gpu::FindBufferMap(GPUVAddr gpu_addr) const {
-    const auto end{buffer_mappings.upper_bound(gpu_addr)};
-    for (auto iter{buffer_mappings.begin()}; iter != end; ++iter) {
-        if (gpu_addr >= iter->second.StartAddr() && gpu_addr < iter->second.EndAddr()) {
-            return iter->second;
-        }
-    }
-
-    return std::nullopt;
-}
-
-void nvhost_as_gpu::AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr,
-                                 bool is_allocated) {
-    buffer_mappings[gpu_addr] = {gpu_addr, size, cpu_addr, is_allocated};
-}
-
-std::optional<std::size_t> nvhost_as_gpu::RemoveBufferMap(GPUVAddr gpu_addr) {
-    if (const auto iter{buffer_mappings.find(gpu_addr)}; iter != buffer_mappings.end()) {
-        std::size_t size{};
-
-        if (iter->second.IsAllocated()) {
-            size = iter->second.Size();
-        }
-
-        buffer_mappings.erase(iter);
-
-        return size;
-    }
-
-    return std::nullopt;
-}
-
 Kernel::KEvent* nvhost_as_gpu::QueryEvent(u32 event_id) {
     LOG_CRITICAL(Service_NVDRV, "Unknown AS GPU Event {}", event_id);
     return nullptr;

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

@@ -5,14 +5,19 @@
 
 #pragma once
 
+#include <bit>
+#include <list>
 #include <map>
 #include <memory>
+#include <mutex>
 #include <optional>
 #include <vector>
 
+#include "common/address_space.h"
 #include "common/common_funcs.h"
 #include "common/common_types.h"
 #include "common/swap.h"
+#include "core/hle/service/nvdrv/core/nvmap.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
 
 namespace Tegra {
@@ -30,17 +35,13 @@ class NvMap;
 
 namespace Service::Nvidia::Devices {
 
-constexpr u32 DEFAULT_BIG_PAGE_SIZE = 1 << 16;
+enum class MappingFlags : u32 {
-constexpr u32 DEFAULT_SMALL_PAGE_SIZE = 1 << 12;
+    None = 0,
-
+    Fixed = 1 << 0,
-class nvmap;
+    Sparse = 1 << 1,
-
+    Remap = 1 << 8,
-enum class AddressSpaceFlags : u32 {
-    None = 0x0,
-    FixedOffset = 0x1,
-    Remap = 0x100,
 };
-DECLARE_ENUM_FLAG_OPERATORS(AddressSpaceFlags);
+DECLARE_ENUM_FLAG_OPERATORS(MappingFlags);
 
 class nvhost_as_gpu final : public nvdevice {
 public:
@@ -59,46 +60,15 @@ public:
 
     Kernel::KEvent* QueryEvent(u32 event_id) override;
 
-private:
+    struct VaRegion {
-    class BufferMap final {
+        u64 offset;
-    public:
+        u32 page_size;
-        constexpr BufferMap() = default;
+        u32 _pad0_;
-
+        u64 pages;
-        constexpr BufferMap(GPUVAddr start_addr_, std::size_t size_)
-            : start_addr{start_addr_}, end_addr{start_addr_ + size_} {}
-
-        constexpr BufferMap(GPUVAddr start_addr_, std::size_t size_, VAddr cpu_addr_,
-                            bool is_allocated_)
-            : start_addr{start_addr_}, end_addr{start_addr_ + size_}, cpu_addr{cpu_addr_},
-              is_allocated{is_allocated_} {}
-
-        constexpr VAddr StartAddr() const {
-            return start_addr;
-        }
-
-        constexpr VAddr EndAddr() const {
-            return end_addr;
-        }
-
-        constexpr std::size_t Size() const {
-            return end_addr - start_addr;
-        }
-
-        constexpr VAddr CpuAddr() const {
-            return cpu_addr;
-        }
-
-        constexpr bool IsAllocated() const {
-            return is_allocated;
-        }
-
-    private:
-        GPUVAddr start_addr{};
-        GPUVAddr end_addr{};
-        VAddr cpu_addr{};
-        bool is_allocated{};
     };
+    static_assert(sizeof(VaRegion) == 0x18);
 
+private:
     struct IoctlAllocAsEx {
         u32_le flags{}; // usually passes 1
         s32_le as_fd{}; // ignored; passes 0
@@ -113,7 +83,7 @@ private:
     struct IoctlAllocSpace {
         u32_le pages{};
         u32_le page_size{};
-        AddressSpaceFlags flags{};
+        MappingFlags flags{};
         INSERT_PADDING_WORDS(1);
         union {
             u64_le offset;
@@ -130,19 +100,19 @@ private:
     static_assert(sizeof(IoctlFreeSpace) == 16, "IoctlFreeSpace is incorrect size");
 
     struct IoctlRemapEntry {
-        u16_le flags{};
+        u16 flags;
-        u16_le kind{};
+        u16 kind;
-        u32_le nvmap_handle{};
+        NvCore::NvMap::Handle::Id handle;
-        u32_le map_offset{};
+        u32 handle_offset_big_pages;
-        u32_le offset{};
+        u32 as_offset_big_pages;
-        u32_le pages{};
+        u32 big_pages;
     };
     static_assert(sizeof(IoctlRemapEntry) == 20, "IoctlRemapEntry is incorrect size");
 
     struct IoctlMapBufferEx {
-        AddressSpaceFlags flags{}; // bit0: fixed_offset, bit2: cacheable
+        MappingFlags flags{}; // bit0: fixed_offset, bit2: cacheable
         u32_le kind{};        // -1 is default
-        u32_le nvmap_handle{};
+        NvCore::NvMap::Handle::Id handle;
         u32_le page_size{}; // 0 means don't care
         s64_le buffer_offset{};
         u64_le mapping_size{};
@@ -160,27 +130,15 @@ private:
     };
     static_assert(sizeof(IoctlBindChannel) == 4, "IoctlBindChannel is incorrect size");
 
-    struct IoctlVaRegion {
-        u64_le offset{};
-        u32_le page_size{};
-        INSERT_PADDING_WORDS(1);
-        u64_le pages{};
-    };
-    static_assert(sizeof(IoctlVaRegion) == 24, "IoctlVaRegion is incorrect size");
-
     struct IoctlGetVaRegions {
         u64_le buf_addr{}; // (contained output user ptr on linux, ignored)
         u32_le buf_size{}; // forced to 2*sizeof(struct va_region)
         u32_le reserved{};
-        IoctlVaRegion small{};
+        std::array<VaRegion, 2> regions{};
-        IoctlVaRegion big{};
     };
-    static_assert(sizeof(IoctlGetVaRegions) == 16 + sizeof(IoctlVaRegion) * 2,
+    static_assert(sizeof(IoctlGetVaRegions) == 16 + sizeof(VaRegion) * 2,
                   "IoctlGetVaRegions is incorrect size");
 
-    s32 channel{};
-    u32 big_page_size{DEFAULT_BIG_PAGE_SIZE};
-
     NvResult AllocAsEx(const std::vector<u8>& input, std::vector<u8>& output);
     NvResult AllocateSpace(const std::vector<u8>& input, std::vector<u8>& output);
     NvResult Remap(const std::vector<u8>& input, std::vector<u8>& output);
@@ -189,23 +147,74 @@ private:
     NvResult FreeSpace(const std::vector<u8>& input, std::vector<u8>& output);
     NvResult BindChannel(const std::vector<u8>& input, std::vector<u8>& output);
 
+    void GetVARegionsImpl(IoctlGetVaRegions& params);
     NvResult GetVARegions(const std::vector<u8>& input, std::vector<u8>& output);
     NvResult GetVARegions(const std::vector<u8>& input, std::vector<u8>& output,
                           std::vector<u8>& inline_output);
 
-    std::optional<BufferMap> FindBufferMap(GPUVAddr gpu_addr) const;
+    void FreeMappingLocked(u64 offset);
-    void AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr, bool is_allocated);
-    std::optional<std::size_t> RemoveBufferMap(GPUVAddr gpu_addr);
 
     Module& module;
 
     NvCore::Container& container;
     NvCore::NvMap& nvmap;
 
+    struct Mapping {
+        VAddr ptr;
+        u64 offset;
+        u64 size;
+        bool fixed;
+        bool big_page; // Only valid if fixed == false
+        bool sparse_alloc;
+
+        Mapping(VAddr ptr_, u64 offset_, u64 size_, bool fixed_, bool big_page_, bool sparse_alloc_)
+            : ptr(ptr_), offset(offset_), size(size_), fixed(fixed_), big_page(big_page_),
+              sparse_alloc(sparse_alloc_) {}
+    };
+
+    struct Allocation {
+        u64 size;
+        std::list<std::shared_ptr<Mapping>> mappings;
+        u32 page_size;
+        bool sparse;
+    };
+
+    std::map<u64, std::shared_ptr<Mapping>>
+        mapping_map; //!< This maps the base addresses of mapped buffers to their total sizes and
+                     //!< mapping type, this is needed as what was originally a single buffer may
+                     //!< have been split into multiple GPU side buffers with the remap flag.
+    std::map<u64, Allocation> allocation_map; //!< Holds allocations created by AllocSpace from
+                                              //!< which fixed buffers can be mapped into
+    std::mutex mutex;                         //!< Locks all AS operations
+
+    struct VM {
+        static constexpr u32 YUZU_PAGESIZE{0x1000};
+        static constexpr u32 PAGE_SIZE_BITS{std::countr_zero(YUZU_PAGESIZE)};
+
+        static constexpr u32 SUPPORTED_BIG_PAGE_SIZES{0x30000};
+        static constexpr u32 DEFAULT_BIG_PAGE_SIZE{0x20000};
+        u32 big_page_size{DEFAULT_BIG_PAGE_SIZE};
+        u32 big_page_size_bits{std::countr_zero(DEFAULT_BIG_PAGE_SIZE)};
+
+        static constexpr u32 VA_START_SHIFT{10};
+        static constexpr u64 DEFAULT_VA_SPLIT{1ULL << 34};
+        static constexpr u64 DEFAULT_VA_RANGE{1ULL << 37};
+        u64 va_range_start{DEFAULT_BIG_PAGE_SIZE << VA_START_SHIFT};
+        u64 va_range_split{DEFAULT_VA_SPLIT};
+        u64 va_range_end{DEFAULT_VA_RANGE};
+
+        using Allocator = Common::FlatAllocator<u32, 0, 32>;
+
+        std::unique_ptr<Allocator> big_page_allocator;
+        std::shared_ptr<Allocator>
+            small_page_allocator; //! Shared as this is also used by nvhost::GpuChannel
+
+        bool initialised{};
+    } vm;
     std::shared_ptr<Tegra::MemoryManager> gmmu;
 
-    // This is expected to be ordered, therefore we must use a map, not unordered_map
+    // s32 channel{};
-    std::map<GPUVAddr, BufferMap> buffer_mappings;
+    // u32 big_page_size{VM::DEFAULT_BIG_PAGE_SIZE};
 };
 
 } // namespace Service::Nvidia::Devices

src/video_core/memory_manager.cpp

@@ -71,18 +71,22 @@ void MemoryManager::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_)
     rasterizer = rasterizer_;
 }
 
-GPUVAddr MemoryManager::Map(VAddr cpu_addr, GPUVAddr gpu_addr, std::size_t size) {
+GPUVAddr MemoryManager::Map(GPUVAddr gpu_addr, VAddr cpu_addr, std::size_t size) {
     return PageTableOp<EntryType::Mapped>(gpu_addr, cpu_addr, size);
 }
 
+GPUVAddr MemoryManager::MapSparse(GPUVAddr gpu_addr, std::size_t 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(cpu_addr, *FindFreeRange(size, align), size);
+    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(cpu_addr, *gpu_addr, size);
+    return Map(*gpu_addr, cpu_addr, size);
 }
 
 void MemoryManager::Unmap(GPUVAddr gpu_addr, std::size_t size) {

src/video_core/memory_manager.h

@@ -88,7 +88,8 @@ public:
     std::vector<std::pair<GPUVAddr, std::size_t>> GetSubmappedRange(GPUVAddr gpu_addr,
                                                                     std::size_t size) const;
 
-    [[nodiscard]] GPUVAddr Map(VAddr cpu_addr, GPUVAddr gpu_addr, std::size_t size);
+    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);