// PE image loader. This file owns image mapping, relocation, import resolution, and export lookup.

#include <stdint.h>
#include <string.h>
#include <kernel/ldr_image.h>
#include <kernel/ldr_section.h>
#include <ntlite/boot_info.h>
#include <kernel/ke_process.h>
#include <kernel/io_file.h>
#include <kernel/mm_virt.h>
#include <kernel/ke_debug.h>
#include <ntlite/syscall.h>

#ifndef LDR_DEBUG_VERBOSE
#define LDR_DEBUG_VERBOSE 0
#endif

#if LDR_DEBUG_VERBOSE
#define LdrDbg(...) DbgPrint(__VA_ARGS__)
#else
#define LdrDbg(...) ((void)0)
#endif

int _stricmp(const char* a, const char* b);

// PE32+ structures (just what we need).
#define PE_SIGNATURE        0x00004550UL   // 'PE\0\0'
#define IMAGE_FILE_MACHINE_AMD64  0x8664
#define IMAGE_FILE_MACHINE_I386   0x014c
#define IMAGE_NT_OPTIONAL_HDR32_MAGIC 0x10b
#define IMAGE_NT_OPTIONAL_HDR64_MAGIC 0x20b

#pragma pack(push, 1)
typedef struct {
    uint16_t Magic;
    uint8_t  MajorLinkerVersion;
    uint8_t  MinorLinkerVersion;
    uint32_t SizeOfCode;
    uint32_t SizeOfInitializedData;
    uint32_t SizeOfUninitializedData;
    uint32_t AddressOfEntryPoint;
    uint32_t BaseOfCode;
    uint64_t ImageBase;
    uint32_t SectionAlignment;
    uint32_t FileAlignment;
    uint16_t MajorOSVersion;
    uint16_t MinorOSVersion;
    uint16_t MajorImageVersion;
    uint16_t MinorImageVersion;
    uint16_t MajorSubsystemVersion;
    uint16_t MinorSubsystemVersion;
    uint32_t Win32VersionValue;
    uint32_t SizeOfImage;
    uint32_t SizeOfHeaders;
    uint32_t CheckSum;
    uint16_t Subsystem;
    uint16_t DllCharacteristics;
    uint64_t SizeOfStackReserve;
    uint64_t SizeOfStackCommit;
    uint64_t SizeOfHeapReserve;
    uint64_t SizeOfHeapCommit;
    uint32_t LoaderFlags;
    uint32_t NumberOfRvaAndSizes;
    struct { uint32_t VirtualAddress; uint32_t Size; } DataDirectory[16];
} OPT_HEADER64;

typedef struct {
    uint32_t Signature;
    uint16_t Machine;
    uint16_t NumberOfSections;
    uint32_t TimeDateStamp;
    uint32_t PointerToSymbolTable;
    uint32_t NumberOfSymbols;
    uint16_t SizeOfOptionalHeader;
    uint16_t Characteristics;
    OPT_HEADER64 OptionalHeader;
} PE_HEADER;

typedef struct {
    char     Name[8];
    uint32_t VirtualSize;
    uint32_t VirtualAddress;
    uint32_t SizeOfRawData;
    uint32_t PointerToRawData;
    uint32_t PointerToRelocations;
    uint32_t PointerToLinenumbers;
    uint16_t NumberOfRelocations;
    uint16_t NumberOfLinenumbers;
    uint32_t Characteristics;
} SECTION_HEADER;

// Implements the align Pe Size operation.
static size_t KiAlignPeSize(size_t value) {
    return (value + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);
}

// Implements the compute Required Image Size operation.
static KSTATUS KiComputeRequiredImageSize(const PE_HEADER* pe, size_t fileSize, size_t* outSize) {
    if (!pe || !outSize) return NTLSTATUS_INVALID_PARAMETER;

    const OPT_HEADER64* opt = &pe->OptionalHeader;
    size_t required = KiAlignPeSize(opt->SizeOfImage);
    size_t headers = KiAlignPeSize(opt->SizeOfHeaders);
    if (headers > required) required = headers;

    const SECTION_HEADER* sec = (const SECTION_HEADER*)((const uint8_t*)&pe->OptionalHeader + pe->SizeOfOptionalHeader);
    for (int i = 0; i < pe->NumberOfSections; i++) {
        size_t va = (size_t)sec[i].VirtualAddress;
        size_t virt = (size_t)sec[i].VirtualSize;
        size_t raw = (size_t)sec[i].SizeOfRawData;
        size_t span = virt ? virt : raw;
        if (raw > span) span = raw;
        if (!span) continue;

        if (va > (size_t)-1 - span) return NTLSTATUS_NO_MEMORY;
        size_t end = KiAlignPeSize(va + span);
        if (end > required) required = end;

        if (raw) {
            size_t rawBase = (size_t)sec[i].PointerToRawData;
            if (rawBase > fileSize || raw > fileSize - rawBase) return NTLSTATUS_FAILURE;
        }
    }

    if (!required) return NTLSTATUS_FAILURE;
    *outSize = required;
    return NTLSTATUS_SUCCESS;
}

typedef struct {
    uint32_t Characteristics;
    uint32_t TimeDateStamp;
    uint16_t MajorVersion;
    uint16_t MinorVersion;
    uint32_t Name;
    uint32_t Base;
    uint32_t NumberOfFunctions;
    uint32_t NumberOfNames;
    uint32_t AddressOfFunctions;
    uint32_t AddressOfNames;
    uint32_t AddressOfNameOrdinals;
} EXPORT_DIR;

typedef struct {
    uint32_t OriginalFirstThunk;
    uint32_t TimeDateStamp;
    uint32_t ForwarderChain;
    uint32_t Name;
    uint32_t FirstThunk;
} IMPORT_DESC;

typedef struct {
    uint32_t VirtualAddress;
    uint32_t SizeOfBlock;
} BASE_RELOC;
#pragma pack(pop)

#define IMAGE_REL_BASED_DIR64 10
#define IMAGE_SCN_CNT_CODE               0x00000020u
#define IMAGE_SCN_CNT_INITIALIZED_DATA   0x00000040u
#define IMAGE_SCN_CNT_UNINITIALIZED_DATA 0x00000080u
#define IMAGE_SCN_MEM_EXECUTE            0x20000000u
#define IMAGE_SCN_MEM_READ               0x40000000u
#define IMAGE_SCN_MEM_WRITE              0x80000000u

// Global kernel-module registry. Entries are allocated as modules arrive so boot-driver count is not an ABI limit.
#define LDR_KERNEL_MODULE_NAME_MAX 64u

typedef struct MODULE_ENTRY {
    char Name[LDR_KERNEL_MODULE_NAME_MAX];
    uintptr_t Base;
    size_t Size;
    bool ExportsAvailable;
    struct MODULE_ENTRY* Next;
} MODULE_ENTRY;

static MODULE_ENTRY* gModules;

// Ensures a user image page is privately writable before the kernel patches it.
static KSTATUS KiEnsureUserImageWrite(PROCESS* process, bool user, uintptr_t address, size_t size) {
    if (!user) return NTLSTATUS_SUCCESS;
    return MmVadMakeImageRangePrivate(process, (void*)address, size);
}

// Allocates and inserts one global kernel-module registry entry.
static KSTATUS KiRegisterGlobalModule(const char* name, uintptr_t base, size_t size,
                                      bool exportsAvailable) {
    if (!name || !*name || !base || !size)
        return NTLSTATUS_INVALID_PARAMETER;

    size_t nameLength = 0;
    while (nameLength < LDR_KERNEL_MODULE_NAME_MAX && name[nameLength]) nameLength++;
    if (nameLength == 0 || nameLength >= LDR_KERNEL_MODULE_NAME_MAX ||
        !SysSvIsUtf8Text(SysSvFromParts(name, nameLength)))
        return NTLSTATUS_INVALID_PARAMETER;

    MODULE_ENTRY* entry = (MODULE_ENTRY*)MmAllocatePool(sizeof(MODULE_ENTRY));
    if (!entry)
        return NTLSTATUS_NO_MEMORY;

    memset(entry, 0, sizeof(*entry));
    memcpy(entry->Name, name, nameLength);
    entry->Name[nameLength] = '\0';
    entry->Base = base;
    entry->Size = size;
    entry->ExportsAvailable = exportsAvailable;
    entry->Next = gModules;
    gModules = entry;
    return NTLSTATUS_SUCCESS;
}

// Registers the two fixed bootstrap images described explicitly by BOOT_INFO.
KSTATUS LdrRegisterModulesFromBootInfo(const BOOT_INFO* bootInfo) {
    if (!bootInfo)
        return NTLSTATUS_INVALID_PARAMETER;

    KSTATUS status = KiRegisterGlobalModule("kernel.exe",
                                            bootInfo->KernelImage.VirtualBase,
                                            bootInfo->KernelImage.ImageSize, true);
    if (!KSUCCESS(status))
        return status;

    return KiRegisterGlobalModule("hal.dll",
                                  bootInfo->HalImage.VirtualBase,
                                  bootInfo->HalImage.ImageSize, true);
}

// Registers one loaded kernel module without imposing a fixed registry capacity.
KSTATUS LdrRegisterModule(const char* name, uintptr_t base) {
    size_t size = base ? LdrGetImageSize(base) : 0;
    return KiRegisterGlobalModule(name, base, size, false);
}

// Makes an initialized module's exports visible to future import resolution.
// Failed images remain mapped for diagnostics but must not provide exports.
KSTATUS LdrSetModuleExportsAvailable(const char* name, bool available) {
    if (!name) return NTLSTATUS_INVALID_PARAMETER;
    for (MODULE_ENTRY* entry = gModules; entry; entry = entry->Next) {
        if (_stricmp(entry->Name, name) == 0) {
            entry->ExportsAvailable = available;
            return NTLSTATUS_SUCCESS;
        }
    }
    return NTLSTATUS_NOT_FOUND;
}

// Finds the loaded kernel module that owns an address.
const char* LdrFindKernelModuleByAddress(uintptr_t address, uintptr_t* baseOut, size_t* sizeOut) {
    if (baseOut) *baseOut = 0;
    if (sizeOut) *sizeOut = 0;

    for (MODULE_ENTRY* module = gModules; module; module = module->Next) {
        uintptr_t base = module->Base;
        size_t size = module->Size;
        if (!base || !size || base > UINTPTR_MAX - size)
            continue;
        if (address >= base && address < base + size) {
            if (baseOut) *baseOut = base;
            if (sizeOut) *sizeOut = size;
            return module->Name;
        }
    }
    return "unknown";
}

// Returns the loaded global kernel module base for one exact module name.
static uintptr_t KiFindGlobalModuleBase(const char* name) {
    if (!name) return 0;
    for (MODULE_ENTRY* module = gModules; module; module = module->Next) {
        if (module->ExportsAvailable && _stricmp(module->Name, name) == 0)
            return module->Base;
    }
    return 0;
}

// Returns export At Base.
static void* KiGetExportAtBase(uintptr_t base, const char* functionName) {
    if (!base) return (void*)0;

    uint32_t peOffset = *(uint32_t*)(base + 0x3C);
    PE_HEADER* pe = (PE_HEADER*)(base + peOffset);
    if (pe->Signature != PE_SIGNATURE) return (void*)0;

    OPT_HEADER64* opt = &pe->OptionalHeader;
    if (opt->NumberOfRvaAndSizes < 1) return (void*)0;

    uint32_t expRva = opt->DataDirectory[0].VirtualAddress;
    if (!expRva) return (void*)0;

    EXPORT_DIR* exp = (EXPORT_DIR*)(base + expRva);
    uint32_t* names     = (uint32_t*)(base + exp->AddressOfNames);
    uint16_t* ordinals  = (uint16_t*)(base + exp->AddressOfNameOrdinals);
    uint32_t* functions = (uint32_t*)(base + exp->AddressOfFunctions);

    for (uint32_t i = 0; i < exp->NumberOfNames; i++) {
        const char* name = (const char*)(base + names[i]);
        if (strcmp(name, functionName) == 0) {
            uint16_t ord = ordinals[i];
            return (void*)(base + functions[ord]);
        }
    }
    return (void*)0;
}

// Resolves a named export without depending on the image's registration name.
void* LdrGetImageExport(uintptr_t imageBase, const char* functionName) {
    return functionName ? KiGetExportAtBase(imageBase, functionName) : (void*)0;
}

// Export lookup.
void* LdrGetExport(const char* moduleName, const char* functionName) {
    uintptr_t base = KiFindGlobalModuleBase(moduleName);
    return KiGetExportAtBase(base, functionName);
}

// Implements the ldr Get Image Size operation.
size_t LdrGetImageSize(uintptr_t base) {
    if (!base) return 0;
    uint32_t peOffset = *(uint32_t*)(base + 0x3C);
    PE_HEADER* pe = (PE_HEADER*)(base + peOffset);
    if (pe->Signature != PE_SIGNATURE) return 0;
    return (size_t)pe->OptionalHeader.SizeOfImage;
}

// Implements the ldr Get Mapped Image Size operation.
size_t LdrGetMappedImageSize(PROCESS* process, uintptr_t base) {
    if (!base) return 0;
    if (process) {
        size_t regionSize = MmVadGetSize(process, (const void*)base);
        if (regionSize) return regionSize;
    }
    size_t globalRegionSize = MmVadGetSizeByBase((const void*)base);
    if (globalRegionSize) return globalRegionSize;
    return LdrGetImageSize(base);
}

// Releases loaded Image For Process.
static void KiFreeLoadedImageForProcess(PROCESS* process, uintptr_t imageBase, bool user) {
    if (!imageBase) return;
    size_t imageSize = user ? LdrGetMappedImageSize(process, imageBase) : LdrGetImageSize(imageBase);
    if (user) MmFreeUserImageForProcess(process, (void*)imageBase, imageSize);
    else MmFreePool((void*)imageBase);
}

// Returns the final page flags for a PE section.
static uint64_t KiSectionPageProtection(const SECTION_HEADER* section, bool user) {
    uint64_t prot = PAGE_PRESENT | (user ? PAGE_USER : 0);

    // Some old toolchains mark BSS-like sections as uninitialized data without
    // setting MEM_WRITE. Treating those pages as read-only breaks programs that
    // legitimately write their zero-fill storage.
    if ((section->Characteristics & IMAGE_SCN_MEM_WRITE) ||
        (section->Characteristics & IMAGE_SCN_CNT_UNINITIALIZED_DATA)) {
        prot |= PAGE_WRITE;
    }

    if (!((section->Characteristics & IMAGE_SCN_MEM_EXECUTE) ||
          (section->Characteristics & IMAGE_SCN_CNT_CODE))) {
        prot |= PAGE_NX;
    }
    return prot;
}

// Applies final page protections to a mapped PE image.
KSTATUS LdrProtectImageForProcess(PROCESS* process, uintptr_t imageBase, bool user) {
    if (!imageBase) return NTLSTATUS_INVALID_PARAMETER;
    if (user && !process) return NTLSTATUS_INVALID_PARAMETER;

    size_t mappedSize = user ? LdrGetMappedImageSize(process, imageBase) : LdrGetImageSize(imageBase);
    if (!mappedSize) return NTLSTATUS_FAILURE;

    uint32_t peOffset = *(uint32_t*)(imageBase + 0x3C);
    if ((size_t)peOffset > mappedSize || mappedSize - (size_t)peOffset < sizeof(uint32_t) + 20)
        return NTLSTATUS_FAILURE;

    PE_HEADER* pe = (PE_HEADER*)(imageBase + peOffset);
    if (pe->Signature != PE_SIGNATURE) return NTLSTATUS_FAILURE;

    size_t sectionTableOffset = (size_t)peOffset + 24u + (size_t)pe->SizeOfOptionalHeader;
    size_t sectionTableSize = (size_t)pe->NumberOfSections * sizeof(SECTION_HEADER);
    if (sectionTableOffset > mappedSize || sectionTableSize > mappedSize - sectionTableOffset)
        return NTLSTATUS_FAILURE;

    size_t pageCount = (mappedSize + PAGE_SIZE - 1) / PAGE_SIZE;
    uint64_t* pageProt = (uint64_t*)MmAllocatePoolTag(pageCount * sizeof(uint64_t), MM_POOL_TAG('i', 'p', 'r', 't'));
    if (!pageProt) return NTLSTATUS_NO_MEMORY;

    uint64_t headerProt = PAGE_PRESENT | (user ? PAGE_USER : 0) | PAGE_NX;
    for (size_t i = 0; i < pageCount; ++i)
        pageProt[i] = headerProt;

    SECTION_HEADER* sec = (SECTION_HEADER*)(imageBase + sectionTableOffset);
    for (int i = 0; i < pe->NumberOfSections; i++) {
        size_t protectSize = sec[i].VirtualSize ? sec[i].VirtualSize : sec[i].SizeOfRawData;
        if (!protectSize) continue;

        uintptr_t secBase = imageBase + (uintptr_t)sec[i].VirtualAddress;
        uintptr_t secEnd = secBase + protectSize;
        uintptr_t mapEnd = imageBase + mappedSize;

        if (secBase < imageBase || secBase >= mapEnd) {
            LdrDbg("LDR: protect skipping out-of-range section %d base=%016llx size=%llx image=%016llx mapped=%llx\n",
                   i, (unsigned long long)secBase, (unsigned long long)protectSize,
                   (unsigned long long)imageBase, (unsigned long long)mappedSize);
            continue;
        }
        if (secEnd < secBase) {
            MmFreePool(pageProt);
            return NTLSTATUS_NO_MEMORY;
        }
        if (secEnd > mapEnd) {
            LdrDbg("LDR: protect clamping section %d end=%016llx to mapEnd=%016llx\n",
                   i, (unsigned long long)secEnd, (unsigned long long)mapEnd);
            secEnd = mapEnd;
        }

        size_t firstPage = (size_t)((secBase - imageBase) / PAGE_SIZE);
        size_t lastPage = (size_t)((secEnd - 1 - imageBase) / PAGE_SIZE);
        uint64_t prot = KiSectionPageProtection(&sec[i], user);
        for (size_t page = firstPage; page <= lastPage && page < pageCount; ++page) {
            // Protection is page-granular. If sections ever share a page, the
            // hardware page must allow the union of their access requirements.
            if (prot & PAGE_WRITE)
                pageProt[page] |= PAGE_WRITE;
            if ((prot & PAGE_NX) == 0)
                pageProt[page] &= ~PAGE_NX;
        }
    }

    size_t runStart = 0;
    KSTATUS s = NTLSTATUS_SUCCESS;
    while (runStart < pageCount) {
        uint64_t prot = pageProt[runStart];
        size_t runEnd = runStart + 1;
        while (runEnd < pageCount && pageProt[runEnd] == prot)
            runEnd++;

        uintptr_t va = imageBase + runStart * PAGE_SIZE;
        size_t bytes = (runEnd - runStart) * PAGE_SIZE;
        if (user)
            s = MmVadProtect(process, (void*)va, bytes, prot);
        else
            s = MmPteProtectRange(va, bytes, prot);

        if (!KSUCCESS(s)) {
            LdrDbg("LDR: protect failed va=%016llx size=%llx prot=%016llx st=%d image=%016llx mapped=%llx\n",
                   (unsigned long long)va, (unsigned long long)bytes,
                   (unsigned long long)prot, s,
                   (unsigned long long)imageBase, (unsigned long long)mappedSize);
            break;
        }
        runStart = runEnd;
    }

    MmFreePool(pageProt);
    return s;
}

// Entry point lookup.
uintptr_t LdrGetEntryPoint(uintptr_t base) {
    uint32_t pe_offset = *(uint32_t*)(base + 0x3C);
    uint32_t entry_rva = *(uint32_t*)(base + pe_offset + 0x28);
    return base + entry_rva;
}

// Base Relocation.
static KSTATUS LdrApplyRelocationsForProcess(PROCESS* process, uintptr_t base, uintptr_t delta, bool user) {
    uint32_t peOffset = *(uint32_t*)(base + 0x3C);
    PE_HEADER* pe = (PE_HEADER*)(base + peOffset);
    OPT_HEADER64* opt = &pe->OptionalHeader;

    if (opt->NumberOfRvaAndSizes < 6) return NTLSTATUS_SUCCESS;
    uint32_t relocRva = opt->DataDirectory[5].VirtualAddress;
    uint32_t relocSize = opt->DataDirectory[5].Size;
    if (!relocRva || !relocSize) return NTLSTATUS_SUCCESS;

    BASE_RELOC* block = (BASE_RELOC*)(base + relocRva);
    uintptr_t end = (uintptr_t)block + relocSize;

    while ((uintptr_t)block < end && block->SizeOfBlock > 0) {
        uint32_t count = (block->SizeOfBlock - sizeof(BASE_RELOC)) / sizeof(uint16_t);
        uint16_t* entries = (uint16_t*)(block + 1);

        for (uint32_t i = 0; i < count; i++) {
            int type = entries[i] >> 12;
            int offset = entries[i] & 0xFFF;

            // For 64-bit PEs, we mostly care about DIR64 (Type 10).
            if (type == IMAGE_REL_BASED_DIR64) {
                uint64_t* ptr = (uint64_t*)(base + block->VirtualAddress + offset);
                KSTATUS ws = KiEnsureUserImageWrite(process, user, (uintptr_t)ptr, sizeof(*ptr));
                if (!KSUCCESS(ws)) return ws;
                *ptr += delta;
            }
        }
        block = (BASE_RELOC*)((uint8_t*)block + block->SizeOfBlock);
    }
    return NTLSTATUS_SUCCESS;
}

// Resolves imports for kernel images against the kernel module registry.
KSTATUS LdrResolveImports(uintptr_t imageBase) {
    uint32_t peOffset = *(uint32_t*)(imageBase + 0x3C);
    PE_HEADER* pe = (PE_HEADER*)(imageBase + peOffset);
    OPT_HEADER64* opt = &pe->OptionalHeader;

    if (opt->NumberOfRvaAndSizes < 2)
        return NTLSTATUS_SUCCESS;

    uint32_t impRva = opt->DataDirectory[1].VirtualAddress;
    if (!impRva)
        return NTLSTATUS_SUCCESS;

    IMPORT_DESC* imp = (IMPORT_DESC*)(imageBase + impRva);
    while (imp->Name) {
        const char* dllName = (const char*)(imageBase + imp->Name);
        LdrDbg("LDR: resolve kernel import dll=%s image=%016llx\n",
               dllName, (unsigned long long)imageBase);

        uint64_t* thunk = (uint64_t*)(imageBase + imp->FirstThunk);
        uint64_t* origThunk = (uint64_t*)(imageBase +
            (imp->OriginalFirstThunk ? imp->OriginalFirstThunk : imp->FirstThunk));

        while (*origThunk) {
            if (*origThunk & (1ULL << 63)) {
                DbgPrint("LdrResolveImports: Ordinal imports not supported! (%s)\n", dllName);
                return NTLSTATUS_NOT_SUPPORTED;
            }

            const char* funcName =
                (const char*)(imageBase + (*origThunk & 0xFFFFFFFF) + 2);
            void* addr = LdrGetExport(dllName, funcName);
            if (!addr) {
                DbgPrint("LdrResolveImports: Function not found! (%s!%s)\n", dllName, funcName);
                return NTLSTATUS_NOT_FOUND;
            }

            *thunk = (uint64_t)addr;
            thunk++;
            origThunk++;
        }
        imp++;
    }
    return NTLSTATUS_SUCCESS;
}

// Requires an already-mapped bootstrap image to have no import directory.
KSTATUS LdrRequireNoImports(uintptr_t imageBase) {
    if (!imageBase) return NTLSTATUS_INVALID_PARAMETER;
    uint32_t peOffset = *(uint32_t*)(imageBase + 0x3C);
    PE_HEADER* pe = (PE_HEADER*)(imageBase + peOffset);
    if (pe->Signature != PE_SIGNATURE ||
        pe->Machine != IMAGE_FILE_MACHINE_AMD64 ||
        pe->OptionalHeader.Magic != IMAGE_NT_OPTIONAL_HDR64_MAGIC)
        return NTLSTATUS_FAILURE;
    if (pe->OptionalHeader.NumberOfRvaAndSizes >= 2 &&
        (pe->OptionalHeader.DataDirectory[1].VirtualAddress != 0 ||
         pe->OptionalHeader.DataDirectory[1].Size != 0))
        return NTLSTATUS_NOT_SUPPORTED;
    if (pe->OptionalHeader.NumberOfRvaAndSizes >= 14 &&
        (pe->OptionalHeader.DataDirectory[13].VirtualAddress != 0 ||
         pe->OptionalHeader.DataDirectory[13].Size != 0))
        return NTLSTATUS_NOT_SUPPORTED;
    return NTLSTATUS_SUCCESS;
}

// Loads a user image from an open file through the image-section object model.
KSTATUS LdrLoadImageFileForProcess(PROCESS* process, FILE_OBJECT* file, uintptr_t* baseOut) {
    if (!process || !file || !baseOut)
        return NTLSTATUS_INVALID_PARAMETER;
    *baseOut = 0;

    IMAGE_SECTION_OBJECT* section = (IMAGE_SECTION_OBJECT*)0;
    KSTATUS s = LdrCreateImageSectionFromFile(file, &section);
    if (!KSUCCESS(s))
        return s;

    void* view = (void*)0;
    s = LdrMapImageSection(process, section, &view);
    if (!KSUCCESS(s)) {
        LdrDereferenceImageSection(section);
        return s;
    }

    const IMAGE_SECTION_METADATA* metadata = LdrImageSectionMetadata(section);
    uintptr_t imageBase = (uintptr_t)view;
    uintptr_t delta = imageBase - metadata->PreferredBase;
    if (delta != 0) {
        s = LdrApplyRelocationsForProcess(process, imageBase, delta, true);
        if (!KSUCCESS(s)) {
            KiFreeLoadedImageForProcess(process, imageBase, true);
            LdrDereferenceImageSection(section);
            return s;
        }
    }

    LdrDereferenceImageSection(section);
    *baseOut = imageBase;
    return NTLSTATUS_SUCCESS;
}

// Loads a boot/kernel PE image from a raw memory buffer into kernel space.
static KSTATUS KiLdrLoadKernelImageFromBuffer(const void* buffer, size_t size, uintptr_t* baseOut) {
    if (!buffer || !size || !baseOut) return NTLSTATUS_INVALID_PARAMETER;
    if (size < 0x40) return NTLSTATUS_FAILURE;

    const uint8_t* bytes = (const uint8_t*)buffer;
    uint32_t peOffset = *(const uint32_t*)(bytes + 0x3C);
    if ((size_t)peOffset > size || size - (size_t)peOffset < 24) return NTLSTATUS_FAILURE;

    uint32_t signature = *(const uint32_t*)(bytes + peOffset);
    if (signature != PE_SIGNATURE) return NTLSTATUS_FAILURE;

    uint16_t machine = *(const uint16_t*)(bytes + peOffset + 4);
    uint16_t optionalSize = *(const uint16_t*)(bytes + peOffset + 20);
    if (optionalSize < sizeof(OPT_HEADER64) ||
        (size_t)peOffset + 24 + optionalSize > size) {
        return NTLSTATUS_FAILURE;
    }

    uint16_t optionalMagic = *(const uint16_t*)(bytes + peOffset + 24);
    if (machine == IMAGE_FILE_MACHINE_I386 || optionalMagic == IMAGE_NT_OPTIONAL_HDR32_MAGIC) {
        DbgPrint("LdrLoadKernelImage: hard error: 32-bit PE images are not supported (machine=0x%X magic=0x%X)\n",
                 machine, optionalMagic);
        return NTLSTATUS_NOT_SUPPORTED;
    }

    if (machine != IMAGE_FILE_MACHINE_AMD64 || optionalMagic != IMAGE_NT_OPTIONAL_HDR64_MAGIC) {
        DbgPrint("LdrLoadKernelImage: unsupported PE image (machine=0x%X magic=0x%X)\n",
                 machine, optionalMagic);
        return NTLSTATUS_NOT_SUPPORTED;
    }

    PE_HEADER* pe = (PE_HEADER*)(bytes + peOffset);
    OPT_HEADER64* opt = &pe->OptionalHeader;
    size_t imageSize = 0;
    KSTATUS sizeStatus = KiComputeRequiredImageSize(pe, size, &imageSize);
    if (!KSUCCESS(sizeStatus)) return sizeStatus;

    LdrDbg("LDR: kernel image header sizeOfImage=%x required=%llx preferred=%016llx sections=%u\n",
           opt->SizeOfImage,
           (unsigned long long)imageSize,
           (unsigned long long)opt->ImageBase,
           pe->NumberOfSections);

    void* mem = MmAllocatePoolTag(imageSize, MM_POOL_TAG('i', 'm', 'g', ' '));
    if (!mem) {
        DbgPrint("LDR: kernel image allocation failed required=%llx\n",
                 (unsigned long long)imageSize);
        return NTLSTATUS_NO_MEMORY;
    }
    memset(mem, 0, imageSize);

    memcpy(mem, buffer, opt->SizeOfHeaders);
    SECTION_HEADER* sec = (SECTION_HEADER*)((uint8_t*)&pe->OptionalHeader + pe->SizeOfOptionalHeader);
    for (int i = 0; i < pe->NumberOfSections; i++) {
        void* dst = (uint8_t*)mem + sec[i].VirtualAddress;
        const void* src = (const uint8_t*)buffer + sec[i].PointerToRawData;
        if (sec[i].SizeOfRawData) {
            size_t copySize = sec[i].SizeOfRawData;
            if (sec[i].VirtualSize && sec[i].VirtualSize < copySize) copySize = sec[i].VirtualSize;
            memcpy(dst, src, copySize);
        }

        MmPteProtectRange((uintptr_t)mem + sec[i].VirtualAddress,
                         sec[i].VirtualSize ? sec[i].VirtualSize : sec[i].SizeOfRawData,
                         PAGE_PRESENT | PAGE_WRITE);
    }

    uintptr_t delta = (uintptr_t)mem - opt->ImageBase;
    if (delta != 0) {
        KSTATUS rs = LdrApplyRelocationsForProcess((PROCESS*)0, (uintptr_t)mem, delta, false);
        if (!KSUCCESS(rs)) {
            MmFreePool(mem);
            return rs;
        }
    }

    // Kernel drivers are still pool-backed images. Section protections cannot
    // safely make pages read-only yet because imports and driver-owned writable
    // image data are not separated from callback code by a section object. What
    // must be true is simpler: every page in the loaded image must be executable
    // enough for driver callbacks, including section-alignment slack near the
    // end of the image.
    KSTATUS xs = MmPteProtectRange((uintptr_t)mem, imageSize, PAGE_PRESENT | PAGE_WRITE);
    if (!KSUCCESS(xs)) {
        MmFreePool(mem);
        return xs;
    }

    *baseOut = (uintptr_t)mem;
    return NTLSTATUS_SUCCESS;
}

// Loads a kernel image from a bootloader or driver raw buffer.
KSTATUS LdrLoadKernelImage(const void* buffer, size_t size, uintptr_t* baseOut) {
    return KiLdrLoadKernelImageFromBuffer(buffer, size, baseOut);
}