Merge pull request #181 from shadps4-emu/new_memory

core: Rework memory manager

src/common/error.cpp

@@ -20,7 +20,7 @@ std::string NativeErrorToString(int e) {
 
     DWORD res = FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_ALLOCATE_BUFFER |
                                    FORMAT_MESSAGE_IGNORE_INSERTS,
-                               nullptr, e, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
+                               nullptr, e, MAKELANGID(LANG_ENGLISH, SUBLANG_ENGLISH_US),
                                reinterpret_cast<LPSTR>(&err_str), 1, nullptr);
     if (!res) {
         return "(FormatMessageA failed to format error)";

src/core/address_space.cpp

@@ -6,7 +6,6 @@
 #include "common/error.h"
 #include "core/address_space.h"
 #include "core/libraries/kernel/memory_management.h"
-#include "core/virtual_memory.h"
 
 #ifdef _WIN32
 #include <windows.h>
@@ -17,11 +16,40 @@
 namespace Core {
 
 static constexpr size_t BackingSize = SCE_KERNEL_MAIN_DMEM_SIZE;
-static constexpr size_t VirtualSize = USER_MAX - USER_MIN + 1;
 
 #ifdef _WIN32
 struct AddressSpace::Impl {
     Impl() : process{GetCurrentProcess()} {
+        // Allocate virtual address placeholder for our address space.
+        MEM_ADDRESS_REQUIREMENTS req{};
+        MEM_EXTENDED_PARAMETER param{};
+        req.LowestStartingAddress = reinterpret_cast<PVOID>(SYSTEM_MANAGED_MIN);
+        // The ending address must align to page boundary - 1
+        // https://stackoverflow.com/questions/54223343/virtualalloc2-with-memextendedparameteraddressrequirements-always-produces-error
+        req.HighestEndingAddress = reinterpret_cast<PVOID>(USER_MIN + UserSize - 1);
+        req.Alignment = 0;
+        param.Type = MemExtendedParameterAddressRequirements;
+        param.Pointer = &req;
+
+        // Typically, lower parts of system managed area is already reserved in windows.
+        // If reservation fails attempt again by reducing the area size a little bit.
+        // System managed is about 31GB in size so also cap the number of times we can reduce it
+        // to a reasonable amount.
+        static constexpr size_t ReductionOnFail = 1_GB;
+        static constexpr size_t MaxReductions = 10;
+        virtual_size = SystemSize + UserSize + ReductionOnFail;
+        for (u32 i = 0; i < MaxReductions && !virtual_base; i++) {
+            virtual_size -= ReductionOnFail;
+            virtual_base = static_cast<u8*>(VirtualAlloc2(process, NULL, virtual_size,
+                                                          MEM_RESERVE | MEM_RESERVE_PLACEHOLDER,
+                                                          PAGE_NOACCESS, &param, 1));
+        }
+        ASSERT_MSG(virtual_base, "Unable to reserve virtual address space!");
+
+        // Initializer placeholder tracker
+        const uintptr_t virtual_addr = reinterpret_cast<uintptr_t>(virtual_base);
+        placeholders.insert({virtual_addr, virtual_addr + virtual_size});
+
         // Allocate backing file that represents the total physical memory.
         backing_handle =
             CreateFileMapping2(INVALID_HANDLE_VALUE, nullptr, FILE_MAP_WRITE | FILE_MAP_READ,
@@ -35,21 +63,6 @@ struct AddressSpace::Impl {
         void* const ret = MapViewOfFile3(backing_handle, process, backing_base, 0, BackingSize,
                                          MEM_REPLACE_PLACEHOLDER, PAGE_READWRITE, nullptr, 0);
         ASSERT(ret == backing_base);
-        // Allocate virtual address placeholder for our address space.
-        MEM_ADDRESS_REQUIREMENTS req{};
-        MEM_EXTENDED_PARAMETER param{};
-        req.LowestStartingAddress = reinterpret_cast<PVOID>(USER_MIN);
-        req.HighestEndingAddress = reinterpret_cast<PVOID>(USER_MAX);
-        req.Alignment = 0;
-        param.Type = MemExtendedParameterAddressRequirements;
-        param.Pointer = &req;
-        virtual_base = static_cast<u8*>(VirtualAlloc2(process, nullptr, VirtualSize,
-                                                      MEM_RESERVE | MEM_RESERVE_PLACEHOLDER,
-                                                      PAGE_NOACCESS, &param, 1));
-        ASSERT(virtual_base);
-
-        const uintptr_t virtual_addr = reinterpret_cast<uintptr_t>(virtual_base);
-        placeholders.insert({virtual_addr, virtual_addr + VirtualSize});
     }
 
     ~Impl() {
@@ -71,7 +84,7 @@ struct AddressSpace::Impl {
         }
     }
 
-    void* MapUser(VAddr virtual_addr, PAddr phys_addr, size_t size, ULONG prot) {
+    void* Map(VAddr virtual_addr, PAddr phys_addr, size_t size, ULONG prot) {
         const auto it = placeholders.find(virtual_addr);
         ASSERT_MSG(it != placeholders.end(), "Cannot map already mapped region");
         ASSERT_MSG(virtual_addr >= it->lower() && virtual_addr + size <= it->upper(),
@@ -106,54 +119,25 @@ struct AddressSpace::Impl {
             ptr = MapViewOfFile3(backing_handle, process, reinterpret_cast<PVOID>(virtual_addr),
                                  phys_addr, size, MEM_REPLACE_PLACEHOLDER, prot, nullptr, 0);
         } else {
-            ptr = VirtualAlloc2(process, reinterpret_cast<PVOID>(virtual_addr), size,
+            ptr =
-                                MEM_REPLACE_PLACEHOLDER, prot, nullptr, 0);
+                VirtualAlloc2(process, reinterpret_cast<PVOID>(virtual_addr), size,
+                              MEM_RESERVE | MEM_COMMIT | MEM_REPLACE_PLACEHOLDER, prot, nullptr, 0);
         }
         ASSERT_MSG(ptr, "{}", Common::GetLastErrorMsg());
         return ptr;
     }
 
-    void* MapPrivate(VAddr virtual_addr, size_t size, u64 alignment, ULONG prot,
+    void Unmap(VAddr virtual_addr, PAddr phys_addr, size_t size) {
-                     bool no_commit = false) {
+        bool ret;
-        // Map a private allocation
+        if (phys_addr != -1) {
-        PVOID addr = reinterpret_cast<PVOID>(virtual_addr);
+            ret = UnmapViewOfFile2(process, reinterpret_cast<PVOID>(virtual_addr),
-        MEM_ADDRESS_REQUIREMENTS req{};
+                                   MEM_PRESERVE_PLACEHOLDER);
-        MEM_EXTENDED_PARAMETER param{};
-        // req.LowestStartingAddress =
-        //     (virtual_addr == 0 ? reinterpret_cast<PVOID>(SYSTEM_MANAGED_MIN)
-        //                        : reinterpret_cast<PVOID>(virtual_addr));
-        req.HighestEndingAddress = reinterpret_cast<PVOID>(SYSTEM_MANAGED_MAX);
-        req.Alignment = alignment < 64_KB ? 0 : alignment;
-        param.Type = MemExtendedParameterAddressRequirements;
-        param.Pointer = &req;
-        ULONG alloc_type = MEM_RESERVE | (alignment > 2_MB ? MEM_LARGE_PAGES : 0);
-        if (!no_commit) {
-            alloc_type |= MEM_COMMIT;
-        }
-        // Check if the area has been reserved beforehand (typically for tesselation buffer)
-        // and in that case don't reserve it again as Windows complains.
-        if (virtual_addr) {
-            MEMORY_BASIC_INFORMATION info;
-            VirtualQuery(addr, &info, sizeof(info));
-            if (info.State == MEM_RESERVE) {
-                alloc_type &= ~MEM_RESERVE;
-            }
-        }
-        void* ptr{};
-        if (virtual_addr) {
-            ptr = VirtualAlloc2(process, addr, size, alloc_type, prot, NULL, 0);
-            ASSERT_MSG(ptr && VAddr(ptr) == virtual_addr, "{}", Common::GetLastErrorMsg());
         } else {
-            ptr = VirtualAlloc2(process, nullptr, size, alloc_type, prot, &param, 1);
+            ret = VirtualFreeEx(process, reinterpret_cast<PVOID>(virtual_addr), size,
-            ASSERT_MSG(ptr, "{}", Common::GetLastErrorMsg());
+                                MEM_RELEASE | MEM_PRESERVE_PLACEHOLDER);
         }
-        return ptr;
+        ASSERT_MSG(ret, "Unmap operation on virtual_addr={:#X} failed: {}", virtual_addr,
-    }
+                   Common::GetLastErrorMsg());
-
-    void UnmapUser(VAddr virtual_addr, size_t size) {
-        const bool ret = UnmapViewOfFile2(process, reinterpret_cast<PVOID>(virtual_addr),
-                                          MEM_PRESERVE_PLACEHOLDER);
-        ASSERT_MSG(ret, "Unmap operation on virtual_addr={:#X} failed", virtual_addr);
 
         // The unmap call will create a new placeholder region. We need to see if we can coalesce it
         // with neighbors.
@@ -186,12 +170,6 @@ struct AddressSpace::Impl {
         placeholders.insert({placeholder_start, placeholder_end});
     }
 
-    void UnmapPrivate(VAddr virtual_addr, size_t size) {
-        const bool ret =
-            VirtualFreeEx(process, reinterpret_cast<LPVOID>(virtual_addr), 0, MEM_RELEASE);
-        ASSERT_MSG(ret, "{}", Common::GetLastErrorMsg());
-    }
-
     void Protect(VAddr virtual_addr, size_t size, bool read, bool write, bool execute) {
         DWORD new_flags{};
         if (read && write) {
@@ -221,6 +199,7 @@ struct AddressSpace::Impl {
     HANDLE backing_handle{};
     u8* backing_base{};
     u8* virtual_base{};
+    size_t virtual_size{};
     boost::icl::separate_interval_set<uintptr_t> placeholders;
 };
 #else
@@ -239,22 +218,12 @@ struct AddressSpace::Impl {
         UNREACHABLE();
     }
 
-    void* MapUser(VAddr virtual_addr, PAddr phys_addr, size_t size, PosixPageProtection prot) {
+    void* Map(VAddr virtual_addr, PAddr phys_addr, size_t size, PosixPageProtection prot) {
         UNREACHABLE();
         return nullptr;
     }
 
-    void* MapPrivate(VAddr virtual_addr, size_t size, u64 alignment, PosixPageProtection prot,
+    void Unmap(VAddr virtual_addr, PAddr phys_addr, size_t size) {
-                     bool no_commit = false) {
-        UNREACHABLE();
-        return nullptr;
-    }
-
-    void UnmapUser(VAddr virtual_addr, size_t size) {
-        UNREACHABLE();
-    }
-
-    void UnmapPrivate(VAddr virtual_addr, size_t size) {
         UNREACHABLE();
     }
 
@@ -264,36 +233,30 @@ struct AddressSpace::Impl {
 
     u8* backing_base{};
     u8* virtual_base{};
+    size_t virtual_size{};
 };
 #endif
 
 AddressSpace::AddressSpace() : impl{std::make_unique<Impl>()} {
     virtual_base = impl->virtual_base;
     backing_base = impl->backing_base;
+    virtual_size = impl->virtual_size;
 }
 
 AddressSpace::~AddressSpace() = default;
 
-void* AddressSpace::Map(VAddr virtual_addr, size_t size, u64 alignment, PAddr phys_addr) {
+void* AddressSpace::Map(VAddr virtual_addr, size_t size, u64 alignment, PAddr phys_addr,
-    if (virtual_addr >= USER_MIN) {
+                        bool is_exec) {
-        return impl->MapUser(virtual_addr, phys_addr, size, PAGE_READWRITE);
+    return impl->Map(virtual_addr, phys_addr, size,
-    }
+                     is_exec ? PAGE_EXECUTE_READWRITE : PAGE_READWRITE);
-    return impl->MapPrivate(virtual_addr, size, alignment, PAGE_READWRITE);
 }
 
-void AddressSpace::Unmap(VAddr virtual_addr, size_t size) {
+void AddressSpace::Unmap(VAddr virtual_addr, size_t size, PAddr phys_addr) {
-    if (virtual_addr >= USER_MIN) {
+    return impl->Unmap(virtual_addr, phys_addr, size);
-        return impl->UnmapUser(virtual_addr, size);
-    }
-    return impl->UnmapPrivate(virtual_addr, size);
 }
 
 void AddressSpace::Protect(VAddr virtual_addr, size_t size, MemoryPermission perms) {
     return impl->Protect(virtual_addr, size, true, true, true);
 }
 
-void* AddressSpace::Reserve(size_t size, u64 alignment) {
-    return impl->MapPrivate(0, size, alignment, PAGE_READWRITE, true);
-}
-
 } // namespace Core

src/core/address_space.h

@@ -14,9 +14,23 @@ enum class MemoryPermission : u32 {
     Write = 1 << 1,
     ReadWrite = Read | Write,
     Execute = 1 << 2,
+    ReadWriteExecute = Read | Write | Execute,
 };
 DECLARE_ENUM_FLAG_OPERATORS(MemoryPermission)
 
+constexpr VAddr SYSTEM_RESERVED = 0x800000000ULL;
+constexpr VAddr CODE_BASE_OFFSET = 0x100000000ULL;
+constexpr VAddr SYSTEM_MANAGED_MIN = 0x0000040000ULL;
+constexpr VAddr SYSTEM_MANAGED_MAX = 0x07FFFFBFFFULL;
+constexpr VAddr USER_MIN = 0x1000000000ULL;
+constexpr VAddr USER_MAX = 0xFBFFFFFFFFULL;
+
+// User area size is normally larger than this. However games are unlikely to map to high
+// regions of that area, so by default we allocate a smaller virtual address space (about 1/4th).
+// to save space on page tables.
+static constexpr size_t UserSize = 1ULL << 38;
+static constexpr size_t SystemSize = USER_MIN - SYSTEM_MANAGED_MIN;
+
 /**
  * Represents the user virtual address space backed by a dmem memory block
  */
@@ -25,12 +39,15 @@ public:
     explicit AddressSpace();
     ~AddressSpace();
 
-    [[nodiscard]] u8* VirtualBase() noexcept {
+    [[nodiscard]] VAddr VirtualBase() noexcept {
-        return virtual_base;
+        return reinterpret_cast<VAddr>(virtual_base);
     }
     [[nodiscard]] const u8* VirtualBase() const noexcept {
         return virtual_base;
     }
+    [[nodiscard]] size_t VirtualSize() const noexcept {
+        return virtual_size;
+    }
 
     /**
      * @brief Maps memory to the specified virtual address.
@@ -42,20 +59,20 @@ public:
      *                  If zero is provided the mapping is considered as private.
      * @return A pointer to the mapped memory.
      */
-    void* Map(VAddr virtual_addr, size_t size, u64 alignment = 0, PAddr phys_addr = -1);
+    void* Map(VAddr virtual_addr, size_t size, u64 alignment = 0, PAddr phys_addr = -1,
+              bool exec = false);
 
     /// Unmaps specified virtual memory area.
-    void Unmap(VAddr virtual_addr, size_t size);
+    void Unmap(VAddr virtual_addr, size_t size, PAddr phys_addr);
 
     void Protect(VAddr virtual_addr, size_t size, MemoryPermission perms);
 
-    void* Reserve(size_t size, u64 alignment);
-
 private:
     struct Impl;
     std::unique_ptr<Impl> impl;
     u8* backing_base{};
     u8* virtual_base{};
+    size_t virtual_size{};
 };
 
 } // namespace Core

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -10,7 +10,6 @@
 #include "core/libraries/gnmdriver/gnmdriver.h"
 #include "core/libraries/libs.h"
 #include "core/libraries/videoout/video_out.h"
-#include "core/memory.h"
 #include "core/platform.h"
 #include "video_core/amdgpu/liverpool.h"
 #include "video_core/amdgpu/pm4_cmds.h"
@@ -40,10 +39,7 @@ struct AscQueueInfo {
     u32 ring_size_dw;
 };
 static VideoCore::SlotVector<AscQueueInfo> asc_queues{};
-
+static constexpr VAddr tessellation_factors_ring_addr = 0xFF0000000ULL;
-static constexpr u32 TessellationFactorRingSize = 128_KB;
-static constexpr u32 TessellationFactorRingAlignment = 64_KB; // toolkit is using this alignment
-VAddr tessellation_factors_ring_addr{0};
 
 static void DumpCommandList(std::span<const u32> cmd_list, const std::string& postfix) {
     using namespace Common::FS;
@@ -624,11 +620,6 @@ int PS4_SYSV_ABI sceGnmGetShaderStatus() {
 VAddr PS4_SYSV_ABI sceGnmGetTheTessellationFactorRingBufferBaseAddress() {
     LOG_TRACE(Lib_GnmDriver, "called");
     // Actual virtual buffer address is hardcoded in the driver to 0xff00'000
-    if (tessellation_factors_ring_addr == 0) {
-        auto* memory = Core::Memory::Instance();
-        tessellation_factors_ring_addr =
-            memory->Reserve(TessellationFactorRingSize, TessellationFactorRingAlignment);
-    }
     return tessellation_factors_ring_addr;
 }
 

src/core/libraries/kernel/memory_management.cpp

@@ -55,6 +55,20 @@ s32 PS4_SYSV_ABI sceKernelAllocateMainDirectMemory(size_t len, size_t alignment,
                                          physAddrOut);
 }
 
+s32 PS4_SYSV_ABI sceKernelCheckedReleaseDirectMemory(u64 start, size_t len) {
+    LOG_INFO(Kernel_Vmm, "called start = {:#x}, len = {:#x}", start, len);
+    auto* memory = Core::Memory::Instance();
+    memory->Free(start, len);
+    return ORBIS_OK;
+}
+
+s32 PS4_SYSV_ABI sceKernelVirtualQuery(const void* addr, int flags, OrbisVirtualQueryInfo* info,
+                                       size_t infoSize) {
+    LOG_INFO(Kernel_Vmm, "called addr = {}, flags = {:#x}", fmt::ptr(addr), flags);
+    auto* memory = Core::Memory::Instance();
+    return memory->VirtualQuery(std::bit_cast<VAddr>(addr), flags, info);
+}
+
 int PS4_SYSV_ABI sceKernelMapNamedDirectMemory(void** addr, u64 len, int prot, int flags,
                                                s64 directMemoryStart, u64 alignment,
                                                const char* name) {
@@ -83,7 +97,7 @@ int PS4_SYSV_ABI sceKernelMapNamedDirectMemory(void** addr, u64 len, int prot, i
     const auto map_flags = static_cast<Core::MemoryMapFlags>(flags);
     auto* memory = Core::Memory::Instance();
     return memory->MapMemory(addr, in_addr, len, mem_prot, map_flags, Core::VMAType::Direct, "",
-                             directMemoryStart, alignment);
+                             false, directMemoryStart, alignment);
 }
 
 int PS4_SYSV_ABI sceKernelMapDirectMemory(void** addr, u64 len, int prot, int flags,

src/core/libraries/kernel/memory_management.h

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include "common/bit_field.h"
 #include "common/types.h"
 
 constexpr u64 SCE_KERNEL_MAIN_DMEM_SIZE = 5376_MB; // ~ 6GB
@@ -36,6 +37,22 @@ struct OrbisQueryInfo {
     int memoryType;
 };
 
+struct OrbisVirtualQueryInfo {
+    uintptr_t start;
+    uintptr_t end;
+    size_t offset;
+    s32 protection;
+    s32 memory_type;
+    union {
+        BitField<0, 1, u32> is_flexible;
+        BitField<1, 1, u32> is_direct;
+        BitField<2, 1, u32> is_stack;
+        BitField<3, 1, u32> is_pooled;
+        BitField<4, 1, u32> is_commited;
+    };
+    std::array<char, 32> name;
+};
+
 u64 PS4_SYSV_ABI sceKernelGetDirectMemorySize();
 int PS4_SYSV_ABI sceKernelAllocateDirectMemory(s64 searchStart, s64 searchEnd, u64 len,
                                                u64 alignment, int memoryType, s64* physAddrOut);
@@ -46,6 +63,9 @@ int PS4_SYSV_ABI sceKernelMapDirectMemory(void** addr, u64 len, int prot, int fl
                                           s64 directMemoryStart, u64 alignment);
 s32 PS4_SYSV_ABI sceKernelAllocateMainDirectMemory(size_t len, size_t alignment, int memoryType,
                                                    s64* physAddrOut);
+s32 PS4_SYSV_ABI sceKernelCheckedReleaseDirectMemory(u64 start, size_t len);
+s32 PS4_SYSV_ABI sceKernelVirtualQuery(const void* addr, int flags, OrbisVirtualQueryInfo* info,
+                                       size_t infoSize);
 s32 PS4_SYSV_ABI sceKernelMapNamedFlexibleMemory(void** addrInOut, std::size_t len, int prot,
                                                  int flags, const char* name);
 s32 PS4_SYSV_ABI sceKernelMapFlexibleMemory(void** addr_in_out, std::size_t len, int prot,

src/core/memory.cpp

@@ -1,7 +1,6 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <algorithm>
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/scope_exit.h"
@@ -13,49 +12,82 @@
 namespace Core {
 
 MemoryManager::MemoryManager() {
-    // Insert a virtual memory area that covers the user area.
+    // Insert an area that covers direct memory physical block.
-    const size_t user_size = USER_MAX - USER_MIN - 1;
+    dmem_map.emplace(0, DirectMemoryArea{0, SCE_KERNEL_MAIN_DMEM_SIZE});
-    vma_map.emplace(USER_MIN, VirtualMemoryArea{USER_MIN, user_size});
 
-    // Insert a virtual memory area that covers the system managed area.
+    // Insert a virtual memory area that covers the entire area we manage.
-    const size_t sys_size = SYSTEM_MANAGED_MAX - SYSTEM_MANAGED_MIN - 1;
+    const VAddr virtual_base = impl.VirtualBase();
-    vma_map.emplace(SYSTEM_MANAGED_MIN, VirtualMemoryArea{SYSTEM_MANAGED_MIN, sys_size});
+    const size_t virtual_size = impl.VirtualSize();
+    vma_map.emplace(virtual_base, VirtualMemoryArea{virtual_base, virtual_size});
 }
 
 MemoryManager::~MemoryManager() = default;
 
 PAddr MemoryManager::Allocate(PAddr search_start, PAddr search_end, size_t size, u64 alignment,
                               int memory_type) {
-    PAddr free_addr = search_start;
+    std::scoped_lock lk{mutex};
 
-    // Iterate through allocated blocked and find the next free position
+    auto dmem_area = FindDmemArea(search_start);
-    for (const auto& block : allocations) {
+
-        const PAddr end = block.base + block.size;
+    const auto is_suitable = [&] {
-        free_addr = std::max(end, free_addr);
+        return dmem_area->second.is_free && dmem_area->second.size >= size;
+    };
+    while (!is_suitable() && dmem_area->second.GetEnd() <= search_end) {
+        dmem_area++;
     }
+    ASSERT_MSG(is_suitable(), "Unable to find free direct memory area");
 
     // Align free position
+    PAddr free_addr = dmem_area->second.base;
     free_addr = alignment > 0 ? Common::AlignUp(free_addr, alignment) : free_addr;
-    ASSERT(free_addr >= search_start && free_addr + size <= search_end);
 
     // Add the allocated region to the list and commit its pages.
-    allocations.emplace_back(free_addr, size, memory_type);
+    auto& area = AddDmemAllocation(free_addr, size);
+    area.memory_type = memory_type;
+    area.is_free = false;
     return free_addr;
 }
 
 void MemoryManager::Free(PAddr phys_addr, size_t size) {
-    const auto it = std::ranges::find_if(allocations, [&](const auto& alloc) {
+    std::scoped_lock lk{mutex};
-        return alloc.base == phys_addr && alloc.size == size;
-    });
-    ASSERT(it != allocations.end());
 
-    // Free the ranges.
+    const auto dmem_area = FindDmemArea(phys_addr);
-    allocations.erase(it);
+    ASSERT(dmem_area != dmem_map.end() && dmem_area->second.base == phys_addr &&
+           dmem_area->second.size == size);
+
+    // Release any dmem mappings that reference this physical block.
+    std::vector<std::pair<VAddr, u64>> remove_list;
+    for (const auto& [addr, mapping] : vma_map) {
+        if (mapping.type != VMAType::Direct) {
+            continue;
+        }
+        if (mapping.phys_base <= phys_addr && phys_addr < mapping.phys_base + mapping.size) {
+            LOG_INFO(Kernel_Vmm, "Unmaping direct mapping {:#x} with size {:#x}", addr,
+                     mapping.size);
+            // Unmaping might erase from vma_map. We can't do it here.
+            remove_list.emplace_back(addr, mapping.size);
+        }
+    }
+    for (const auto& [addr, size] : remove_list) {
+        UnmapMemory(addr, size);
+    }
+
+    // Mark region as free and attempt to coalesce it with neighbours.
+    auto& area = dmem_area->second;
+    area.is_free = true;
+    area.memory_type = 0;
+    MergeAdjacent(dmem_map, dmem_area);
 }
 
 int MemoryManager::MapMemory(void** out_addr, VAddr virtual_addr, size_t size, MemoryProt prot,
                              MemoryMapFlags flags, VMAType type, std::string_view name,
-                             PAddr phys_addr, u64 alignment) {
+                             bool is_exec, PAddr phys_addr, u64 alignment) {
+    std::scoped_lock lk{mutex};
+
+    // When virtual addr is zero, force it to virtual_base. The guest cannot pass Fixed
+    // flag so we will take the branch that searches for free (or reserved) mappings.
+    virtual_addr = (virtual_addr == 0) ? impl.VirtualBase() : virtual_addr;
+
     VAddr mapped_addr = alignment > 0 ? Common::AlignUp(virtual_addr, alignment) : virtual_addr;
     SCOPE_EXIT {
         auto& new_vma = AddMapping(mapped_addr, size);
@@ -65,18 +97,11 @@ int MemoryManager::MapMemory(void** out_addr, VAddr virtual_addr, size_t size, M
         new_vma.type = type;
 
         if (type == VMAType::Direct) {
+            new_vma.phys_base = phys_addr;
             MapVulkanMemory(mapped_addr, size);
         }
     };
 
-    // When virtual addr is zero let the address space manager pick the address.
-    // Alignment matters here as we let the OS pick the address.
-    if (virtual_addr == 0) {
-        *out_addr = impl.Map(virtual_addr, size, alignment);
-        mapped_addr = std::bit_cast<VAddr>(*out_addr);
-        return ORBIS_OK;
-    }
-
     // Fixed mapping means the virtual address must exactly match the provided one.
     if (True(flags & MemoryMapFlags::Fixed) && True(flags & MemoryMapFlags::NoOverwrite)) {
         // This should return SCE_KERNEL_ERROR_ENOMEM but shouldn't normally happen.
@@ -92,21 +117,28 @@ int MemoryManager::MapMemory(void** out_addr, VAddr virtual_addr, size_t size, M
             it++;
         }
         ASSERT(it != vma_map.end());
-        mapped_addr = alignment > 0 ? Common::AlignUp(it->second.base, alignment) : it->second.base;
+        const VAddr base = it->second.base;
+        mapped_addr = alignment > 0 ? Common::AlignUp(base, alignment) : base;
     }
 
     // Perform the mapping.
-    *out_addr = impl.Map(mapped_addr, size, alignment, phys_addr);
+    *out_addr = impl.Map(mapped_addr, size, alignment, phys_addr, is_exec);
     return ORBIS_OK;
 }
 
 void MemoryManager::UnmapMemory(VAddr virtual_addr, size_t size) {
+    std::scoped_lock lk{mutex};
+
     // TODO: Partial unmaps are technically supported by the guest.
     const auto it = vma_map.find(virtual_addr);
     ASSERT_MSG(it != vma_map.end() && it->first == virtual_addr,
                "Attempting to unmap partially mapped range");
 
-    if (it->second.type == VMAType::Direct) {
+    const auto type = it->second.type;
+    fmt::print("{}\n", u32(type));
+    std::fflush(stdout);
+    const PAddr phys_addr = type == VMAType::Direct ? it->second.phys_base : -1;
+    if (type == VMAType::Direct) {
         UnmapVulkanMemory(virtual_addr, size);
     }
 
@@ -115,13 +147,15 @@ void MemoryManager::UnmapMemory(VAddr virtual_addr, size_t size) {
     vma.type = VMAType::Free;
     vma.prot = MemoryProt::NoAccess;
     vma.phys_base = 0;
-    MergeAdjacent(it);
+    MergeAdjacent(vma_map, it);
 
     // Unmap the memory region.
-    impl.Unmap(virtual_addr, size);
+    impl.Unmap(virtual_addr, size, phys_addr);
 }
 
 int MemoryManager::QueryProtection(VAddr addr, void** start, void** end, u32* prot) {
+    std::scoped_lock lk{mutex};
+
     const auto it = FindVMA(addr);
     const auto& vma = it->second;
     ASSERT_MSG(vma.type != VMAType::Free, "Provided address is not mapped");
@@ -132,18 +166,70 @@ int MemoryManager::QueryProtection(VAddr addr, void** start, void** end, u32* pr
     return ORBIS_OK;
 }
 
-int MemoryManager::DirectMemoryQuery(PAddr addr, bool find_next,
+int MemoryManager::VirtualQuery(VAddr addr, int flags,
-                                     Libraries::Kernel::OrbisQueryInfo* out_info) {
+                                Libraries::Kernel::OrbisVirtualQueryInfo* info) {
-    const auto it = std::ranges::find_if(allocations, [&](const DirectMemoryArea& alloc) {
+    auto it = FindVMA(addr);
-        return alloc.base <= addr && addr < alloc.base + alloc.size;
+    if (it->second.type == VMAType::Free && flags == 1) {
-    });
+        it++;
-    if (it == allocations.end()) {
+    }
-        return SCE_KERNEL_ERROR_EACCES;
+    if (it->second.type == VMAType::Free) {
+        LOG_WARNING(Kernel_Vmm, "VirtualQuery on free memory region");
+        return ORBIS_KERNEL_ERROR_EACCES;
     }
 
-    out_info->start = it->base;
+    const auto& vma = it->second;
-    out_info->end = it->base + it->size;
+    info->start = vma.base;
-    out_info->memoryType = it->memory_type;
+    info->end = vma.base + vma.size;
+    info->is_flexible.Assign(vma.type == VMAType::Flexible);
+    info->is_direct.Assign(vma.type == VMAType::Direct);
+    info->is_commited.Assign(vma.type != VMAType::Free);
+    if (vma.type == VMAType::Direct) {
+        const auto dmem_it = FindDmemArea(vma.phys_base);
+        ASSERT(dmem_it != dmem_map.end());
+        info->memory_type = dmem_it->second.memory_type;
+    }
+
+    return ORBIS_OK;
+}
+
+int MemoryManager::DirectMemoryQuery(PAddr addr, bool find_next,
+                                     Libraries::Kernel::OrbisQueryInfo* out_info) {
+    std::scoped_lock lk{mutex};
+
+    auto dmem_area = FindDmemArea(addr);
+    if (dmem_area->second.is_free && find_next) {
+        dmem_area++;
+    }
+
+    if (dmem_area == dmem_map.end() || dmem_area->second.is_free) {
+        LOG_ERROR(Core, "Unable to find allocated direct memory region to query!");
+        return ORBIS_KERNEL_ERROR_EACCES;
+    }
+
+    const auto& area = dmem_area->second;
+    out_info->start = area.base;
+    out_info->end = area.GetEnd();
+    out_info->memoryType = area.memory_type;
+    return ORBIS_OK;
+}
+
+int MemoryManager::DirectQueryAvailable(PAddr search_start, PAddr search_end, size_t alignment,
+                                        PAddr* phys_addr_out, size_t* size_out) {
+    std::scoped_lock lk{mutex};
+
+    auto dmem_area = FindDmemArea(search_start);
+    PAddr paddr{};
+    size_t max_size{};
+    while (dmem_area != dmem_map.end() && dmem_area->second.GetEnd() <= search_end) {
+        if (dmem_area->second.size > max_size) {
+            paddr = dmem_area->second.base;
+            max_size = dmem_area->second.size;
+        }
+        dmem_area++;
+    }
+
+    *phys_addr_out = alignment > 0 ? Common::AlignUp(paddr, alignment) : paddr;
+    *size_out = max_size;
     return ORBIS_OK;
 }
 
@@ -178,6 +264,30 @@ VirtualMemoryArea& MemoryManager::AddMapping(VAddr virtual_addr, size_t size) {
     return vma_handle->second;
 }
 
+DirectMemoryArea& MemoryManager::AddDmemAllocation(PAddr addr, size_t size) {
+    auto dmem_handle = FindDmemArea(addr);
+    ASSERT_MSG(dmem_handle != dmem_map.end(), "Physical address not in dmem_map");
+
+    const DirectMemoryArea& area = dmem_handle->second;
+    ASSERT_MSG(area.is_free && area.base <= addr,
+               "Adding an allocation to already allocated region");
+
+    const PAddr start_in_area = addr - area.base;
+    const PAddr end_in_vma = start_in_area + size;
+    ASSERT_MSG(end_in_vma <= area.size, "Mapping cannot fit inside free region");
+
+    if (end_in_vma != area.size) {
+        // Split VMA at the end of the allocated region
+        Split(dmem_handle, end_in_vma);
+    }
+    if (start_in_area != 0) {
+        // Split VMA at the start of the allocated region
+        dmem_handle = Split(dmem_handle, start_in_area);
+    }
+
+    return dmem_handle->second;
+}
+
 MemoryManager::VMAHandle MemoryManager::Split(VMAHandle vma_handle, size_t offset_in_vma) {
     auto& old_vma = vma_handle->second;
     ASSERT(offset_in_vma < old_vma.size && offset_in_vma > 0);
@@ -193,24 +303,17 @@ MemoryManager::VMAHandle MemoryManager::Split(VMAHandle vma_handle, size_t offse
     return vma_map.emplace_hint(std::next(vma_handle), new_vma.base, new_vma);
 }
 
-MemoryManager::VMAHandle MemoryManager::MergeAdjacent(VMAHandle iter) {
+MemoryManager::DMemHandle MemoryManager::Split(DMemHandle dmem_handle, size_t offset_in_area) {
-    const auto next_vma = std::next(iter);
+    auto& old_area = dmem_handle->second;
-    if (next_vma != vma_map.end() && iter->second.CanMergeWith(next_vma->second)) {
+    ASSERT(offset_in_area < old_area.size && offset_in_area > 0);
-        iter->second.size += next_vma->second.size;
-        vma_map.erase(next_vma);
-    }
 
-    if (iter != vma_map.begin()) {
+    auto new_area = old_area;
-        auto prev_vma = std::prev(iter);
+    old_area.size = offset_in_area;
-        if (prev_vma->second.CanMergeWith(iter->second)) {
+    new_area.base += offset_in_area;
-            prev_vma->second.size += iter->second.size;
+    new_area.size -= offset_in_area;
-            vma_map.erase(iter);
-            iter = prev_vma;
-        }
-    }
 
-    return iter;
+    return dmem_map.emplace_hint(std::next(dmem_handle), new_area.base, new_area);
-}
+};
 
 void MemoryManager::MapVulkanMemory(VAddr addr, size_t size) {
     return;

src/core/memory.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include <functional>
+#include <mutex>
 #include <string_view>
 #include <vector>
 #include <boost/icl/split_interval_map.hpp>
@@ -11,6 +12,7 @@
 #include "common/singleton.h"
 #include "common/types.h"
 #include "core/address_space.h"
+#include "core/libraries/kernel/memory_management.h"
 #include "video_core/renderer_vulkan/vk_common.h"
 
 namespace Vulkan {
@@ -47,12 +49,28 @@ enum class VMAType : u32 {
     Flexible = 3,
     Pooled = 4,
     Stack = 5,
+    Code = 6,
 };
 
 struct DirectMemoryArea {
     PAddr base = 0;
     size_t size = 0;
     int memory_type = 0;
+    bool is_free = true;
+
+    PAddr GetEnd() const {
+        return base + size;
+    }
+
+    bool CanMergeWith(const DirectMemoryArea& next) const {
+        if (base + size != next.base) {
+            return false;
+        }
+        if (is_free != next.is_free) {
+            return false;
+        }
+        return true;
+    }
 };
 
 struct VirtualMemoryArea {
@@ -81,14 +99,10 @@ struct VirtualMemoryArea {
     }
 };
 
-constexpr VAddr SYSTEM_RESERVED = 0x800000000ULL;
-constexpr VAddr CODE_BASE_OFFSET = 0x100000000ULL;
-constexpr VAddr SYSTEM_MANAGED_MIN = 0x0000040000ULL;
-constexpr VAddr SYSTEM_MANAGED_MAX = 0x07FFFFBFFFULL;
-constexpr VAddr USER_MIN = 0x1000000000ULL;
-constexpr VAddr USER_MAX = 0xFBFFFFFFFFULL;
-
 class MemoryManager {
+    using DMemMap = std::map<PAddr, DirectMemoryArea>;
+    using DMemHandle = DMemMap::iterator;
+
     using VMAMap = std::map<VAddr, VirtualMemoryArea>;
     using VMAHandle = VMAMap::iterator;
 
@@ -107,35 +121,57 @@ public:
 
     int MapMemory(void** out_addr, VAddr virtual_addr, size_t size, MemoryProt prot,
                   MemoryMapFlags flags, VMAType type, std::string_view name = "",
-                  PAddr phys_addr = -1, u64 alignment = 0);
+                  bool is_exec = false, PAddr phys_addr = -1, u64 alignment = 0);
 
     void UnmapMemory(VAddr virtual_addr, size_t size);
 
     int QueryProtection(VAddr addr, void** start, void** end, u32* prot);
 
+    int VirtualQuery(VAddr addr, int flags, Libraries::Kernel::OrbisVirtualQueryInfo* info);
+
     int DirectMemoryQuery(PAddr addr, bool find_next, Libraries::Kernel::OrbisQueryInfo* out_info);
 
-    VAddr Reserve(size_t size, u64 alignment) {
+    int DirectQueryAvailable(PAddr search_start, PAddr search_end, size_t alignment,
-        return reinterpret_cast<VAddr>(impl.Reserve(size, alignment));
+                             PAddr* phys_addr_out, size_t* size_out);
-    }
 
     std::pair<vk::Buffer, size_t> GetVulkanBuffer(VAddr addr);
 
 private:
     VMAHandle FindVMA(VAddr target) {
-        // Return first the VMA with base >= target.
+        return std::prev(vma_map.upper_bound(target));
-        const auto it = vma_map.lower_bound(target);
+    }
-        if (it != vma_map.end() && it->first == target) {
+
-            return it;
+    DMemHandle FindDmemArea(PAddr target) {
+        return std::prev(dmem_map.upper_bound(target));
+    }
+
+    template <typename Handle>
+    Handle MergeAdjacent(auto& handle_map, Handle iter) {
+        const auto next_vma = std::next(iter);
+        if (next_vma != handle_map.end() && iter->second.CanMergeWith(next_vma->second)) {
+            iter->second.size += next_vma->second.size;
+            handle_map.erase(next_vma);
         }
-        return std::prev(it);
+
+        if (iter != handle_map.begin()) {
+            auto prev_vma = std::prev(iter);
+            if (prev_vma->second.CanMergeWith(iter->second)) {
+                prev_vma->second.size += iter->second.size;
+                handle_map.erase(iter);
+                iter = prev_vma;
+            }
+        }
+
+        return iter;
     }
 
     VirtualMemoryArea& AddMapping(VAddr virtual_addr, size_t size);
 
+    DirectMemoryArea& AddDmemAllocation(PAddr addr, size_t size);
+
     VMAHandle Split(VMAHandle vma_handle, size_t offset_in_vma);
 
-    VMAHandle MergeAdjacent(VMAHandle iter);
+    DMemHandle Split(DMemHandle dmem_handle, size_t offset_in_area);
 
     void MapVulkanMemory(VAddr addr, size_t size);
 
@@ -143,8 +179,9 @@ private:
 
 private:
     AddressSpace impl;
-    std::vector<DirectMemoryArea> allocations;
+    DMemMap dmem_map;
     VMAMap vma_map;
+    std::recursive_mutex mutex;
 
     struct MappedMemory {
         vk::UniqueBuffer buffer;

src/core/module.cpp

@@ -7,6 +7,7 @@
 #include "common/logging/log.h"
 #include "common/string_util.h"
 #include "core/aerolib/aerolib.h"
+#include "core/memory.h"
 #include "core/module.h"
 #include "core/tls.h"
 #include "core/virtual_memory.h"
@@ -81,8 +82,11 @@ void Module::LoadModuleToMemory() {
     aligned_base_size = Common::AlignUp(base_size, BlockAlign);
 
     // Map module segments (and possible TLS trampolines)
-    base_virtual_addr = VirtualMemory::memory_alloc(LoadAddress, aligned_base_size + TrampolineSize,
+    auto* memory = Core::Memory::Instance();
-                                                    VirtualMemory::MemoryMode::ExecuteReadWrite);
+    void** out_addr = reinterpret_cast<void**>(&base_virtual_addr);
+    const auto name = file.filename().string();
+    memory->MapMemory(out_addr, LoadAddress, aligned_base_size + TrampolineSize,
+                      MemoryProt::CpuReadWrite, MemoryMapFlags::Fixed, VMAType::Code, name, true);
     LoadAddress += CODE_BASE_INCR * (1 + aligned_base_size / CODE_BASE_INCR);
 
     // Initialize trampoline generator.

src/main.cpp

@@ -20,6 +20,7 @@
 #include "core/libraries/libs.h"
 #include "core/libraries/videoout/video_out.h"
 #include "core/linker.h"
+#include "core/memory.h"
 #include "input/controller.h"
 #include "sdl_window.h"
 
@@ -30,6 +31,8 @@ int main(int argc, char* argv[]) {
         fmt::print("Usage: {} <elf or eboot.bin path>\n", argv[0]);
         return -1;
     }
+    // Initialize memory system as early as possible to reserve mappings.
+    [[maybe_unused]] const auto* memory = Core::Memory::Instance();
     const auto config_dir = Common::FS::GetUserPath(Common::FS::PathType::UserDir);
     Config::load(config_dir / "config.toml");
     Common::Log::Initialize();