Core: Properly configure address space when loading a binary

The code now properly configures the process image to match the loaded
binary segments (code, rodata, data) instead of just blindly allocating
a large chunk of dummy memory.

src/core/file_sys/archive_savedata.cpp

@@ -37,7 +37,7 @@ ArchiveFactory_SaveData::ArchiveFactory_SaveData(const std::string& sdmc_directo
 }
 
 ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SaveData::Open(const Path& path) {
-    std::string concrete_mount_point = GetSaveDataPath(mount_point, Kernel::g_current_process->program_id);
+    std::string concrete_mount_point = GetSaveDataPath(mount_point, Kernel::g_current_process->codeset->program_id);
     if (!FileUtil::Exists(concrete_mount_point)) {
         // When a SaveData archive is created for the first time, it is not yet formatted
         // and the save file/directory structure expected by the game has not yet been initialized.
@@ -52,7 +52,7 @@ ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SaveData::Open(const P
 }
 
 ResultCode ArchiveFactory_SaveData::Format(const Path& path) {
-    std::string concrete_mount_point = GetSaveDataPath(mount_point, Kernel::g_current_process->program_id);
+    std::string concrete_mount_point = GetSaveDataPath(mount_point, Kernel::g_current_process->codeset->program_id);
     FileUtil::DeleteDirRecursively(concrete_mount_point);
     FileUtil::CreateFullPath(concrete_mount_point);
     return RESULT_SUCCESS;

src/core/hle/kernel/kernel.h

@@ -47,6 +47,7 @@ enum class HandleType : u32 {
     Semaphore       = 10,
     Timer           = 11,
     ResourceLimit   = 12,
+    CodeSet         = 13,
 };
 
 enum {

src/core/hle/kernel/process.cpp

@@ -5,24 +5,39 @@
 #include "common/assert.h"
 #include "common/common_funcs.h"
 #include "common/logging/log.h"
+#include "common/make_unique.h"
 
 #include "core/hle/kernel/process.h"
 #include "core/hle/kernel/resource_limit.h"
 #include "core/hle/kernel/thread.h"
+#include "core/hle/kernel/vm_manager.h"
+#include "core/mem_map.h"
 #include "core/memory.h"
 
 namespace Kernel {
 
+SharedPtr<CodeSet> CodeSet::Create(std::string name, u64 program_id) {
+    SharedPtr<CodeSet> codeset(new CodeSet);
+
+    codeset->name = std::move(name);
+    codeset->program_id = program_id;
+
+    return codeset;
+}
+
+CodeSet::CodeSet() {}
+CodeSet::~CodeSet() {}
+
 u32 Process::next_process_id;
 
-SharedPtr<Process> Process::Create(std::string name, u64 program_id) {
+SharedPtr<Process> Process::Create(SharedPtr<CodeSet> code_set) {
     SharedPtr<Process> process(new Process);
 
-    process->name = std::move(name);
+    process->codeset = std::move(code_set);
-    process->program_id = program_id;
-
     process->flags.raw = 0;
     process->flags.memory_region = MemoryRegion::APPLICATION;
+    process->address_space = Common::make_unique<VMManager>();
+    Memory::InitLegacyAddressSpace(*process->address_space);
 
     return process;
 }
@@ -87,8 +102,19 @@ void Process::ParseKernelCaps(const u32* kernel_caps, size_t len) {
     }
 }
 
-void Process::Run(VAddr entry_point, s32 main_thread_priority, u32 stack_size) {
+void Process::Run(s32 main_thread_priority, u32 stack_size) {
-    Kernel::SetupMainThread(entry_point, main_thread_priority);
+    auto MapSegment = [&](CodeSet::Segment& segment, VMAPermission permissions, MemoryState memory_state) {
+        auto vma = address_space->MapMemoryBlock(segment.addr, codeset->memory,
+                segment.offset, segment.size, memory_state).Unwrap();
+        address_space->Reprotect(vma, permissions);
+    };
+
+    MapSegment(codeset->code,   VMAPermission::ReadExecute, MemoryState::Code);
+    MapSegment(codeset->rodata, VMAPermission::Read,        MemoryState::Code);
+    MapSegment(codeset->data,   VMAPermission::ReadWrite,   MemoryState::Private);
+
+    address_space->LogLayout();
+    Kernel::SetupMainThread(codeset->entrypoint, main_thread_priority);
 }
 
 Kernel::Process::Process() {}

src/core/hle/kernel/process.h

@@ -47,23 +47,51 @@ union ProcessFlags {
 };
 
 class ResourceLimit;
+class VMManager;
+
+struct CodeSet final : public Object {
+    static SharedPtr<CodeSet> Create(std::string name, u64 program_id);
+
+    std::string GetTypeName() const override { return "CodeSet"; }
+    std::string GetName() const override { return name; }
+
+    static const HandleType HANDLE_TYPE = HandleType::CodeSet;
+    HandleType GetHandleType() const override { return HANDLE_TYPE; }
+
+    /// Name of the process
+    std::string name;
+    /// Title ID corresponding to the process
+    u64 program_id;
+
+    std::shared_ptr<std::vector<u8>> memory;
+
+    struct Segment {
+        size_t offset = 0;
+        VAddr addr = 0;
+        u32 size = 0;
+    };
+
+    Segment code, rodata, data;
+    VAddr entrypoint;
+
+private:
+    CodeSet();
+    ~CodeSet() override;
+};
 
 class Process final : public Object {
 public:
-    static SharedPtr<Process> Create(std::string name, u64 program_id);
+    static SharedPtr<Process> Create(SharedPtr<CodeSet> code_set);
 
     std::string GetTypeName() const override { return "Process"; }
-    std::string GetName() const override { return name; }
+    std::string GetName() const override { return codeset->name; }
 
     static const HandleType HANDLE_TYPE = HandleType::Process;
     HandleType GetHandleType() const override { return HANDLE_TYPE; }
 
     static u32 next_process_id;
 
-    /// Name of the process
+    SharedPtr<CodeSet> codeset;
-    std::string name;
-    /// Title ID corresponding to the process
-    u64 program_id;
     /// Resource limit descriptor for this process
     SharedPtr<ResourceLimit> resource_limit;
 
@@ -81,6 +109,7 @@ public:
 
     /// Bitmask of the used TLS slots
     std::bitset<300> used_tls_slots;
+    std::unique_ptr<VMManager> address_space;
 
     /**
      * Parses a list of kernel capability descriptors (as found in the ExHeader) and applies them
@@ -91,7 +120,7 @@ public:
     /**
      * Applies address space changes and launches the process main thread.
      */
-    void Run(VAddr entry_point, s32 main_thread_priority, u32 stack_size);
+    void Run(s32 main_thread_priority, u32 stack_size);
 
 private:
     Process();

src/core/hle/kernel/vm_manager.cpp

@@ -35,6 +35,10 @@ VMManager::VMManager() {
     Reset();
 }
 
+VMManager::~VMManager() {
+    Reset();
+}
+
 void VMManager::Reset() {
     vma_map.clear();
 
@@ -130,6 +134,16 @@ void VMManager::Reprotect(VMAHandle vma_handle, VMAPermission new_perms) {
     MergeAdjacent(iter);
 }
 
+void VMManager::LogLayout() const {
+    for (const auto& p : vma_map) {
+        const VirtualMemoryArea& vma = p.second;
+        LOG_DEBUG(Kernel, "%08X - %08X  size: %8X %c%c%c", vma.base, vma.base + vma.size, vma.size,
+            (u8)vma.permissions & (u8)VMAPermission::Read    ? 'R' : '-',
+            (u8)vma.permissions & (u8)VMAPermission::Write   ? 'W' : '-',
+            (u8)vma.permissions & (u8)VMAPermission::Execute ? 'X' : '-');
+    }
+}
+
 VMManager::VMAIter VMManager::StripIterConstness(const VMAHandle & iter) {
     // This uses a neat C++ trick to convert a const_iterator to a regular iterator, given
     // non-const access to its container.

src/core/hle/kernel/vm_manager.h

@@ -101,7 +101,7 @@ struct VirtualMemoryArea {
  *  - http://duartes.org/gustavo/blog/post/how-the-kernel-manages-your-memory/
  *  - http://duartes.org/gustavo/blog/post/page-cache-the-affair-between-memory-and-files/
  */
-class VMManager {
+class VMManager final {
     // TODO(yuriks): Make page tables switchable to support multiple VMManagers
 public:
     /**
@@ -121,6 +121,7 @@ public:
     using VMAHandle = decltype(vma_map)::const_iterator;
 
     VMManager();
+    ~VMManager();
 
     /// Clears the address space map, re-initializing with a single free area.
     void Reset();
@@ -168,6 +169,9 @@ public:
     /// Changes the permissions of the given VMA.
     void Reprotect(VMAHandle vma, VMAPermission new_perms);
 
+    /// Dumps the address space layout to the log, for debugging
+    void LogLayout() const;
+
 private:
     using VMAIter = decltype(vma_map)::iterator;
 

src/core/loader/3dsx.cpp

@@ -101,7 +101,10 @@ static u32 TranslateAddr(u32 addr, const THREEloadinfo *loadinfo, u32* offsets)
     return loadinfo->seg_addrs[2] + addr - offsets[1];
 }
 
-static THREEDSX_Error Load3DSXFile(FileUtil::IOFile& file, u32 base_addr)
+using Kernel::SharedPtr;
+using Kernel::CodeSet;
+
+static THREEDSX_Error Load3DSXFile(FileUtil::IOFile& file, u32 base_addr, SharedPtr<CodeSet>* out_codeset)
 {
     if (!file.IsOpen())
         return ERROR_FILE;
@@ -201,13 +204,29 @@ static THREEDSX_Error Load3DSXFile(FileUtil::IOFile& file, u32 base_addr)
         }
     }
 
-    // Write the data
+    // Create the CodeSet
-    memcpy(Memory::GetPointer(base_addr), program_image.data(), program_image.size());
+    SharedPtr<CodeSet> code_set = CodeSet::Create("", 0);
+
+    code_set->code.offset = loadinfo.seg_ptrs[0] - program_image.data();
+    code_set->code.addr   = loadinfo.seg_addrs[0];
+    code_set->code.size   = loadinfo.seg_sizes[0];
+
+    code_set->rodata.offset = loadinfo.seg_ptrs[1] - program_image.data();
+    code_set->rodata.addr   = loadinfo.seg_addrs[1];
+    code_set->rodata.size   = loadinfo.seg_sizes[1];
+
+    code_set->data.offset = loadinfo.seg_ptrs[2] - program_image.data();
+    code_set->data.addr   = loadinfo.seg_addrs[2];
+    code_set->data.size   = loadinfo.seg_sizes[2];
+
+    code_set->entrypoint = code_set->code.addr;
+    code_set->memory = std::make_shared<std::vector<u8>>(std::move(program_image));
 
     LOG_DEBUG(Loader, "code size:   0x%X", loadinfo.seg_sizes[0]);
     LOG_DEBUG(Loader, "rodata size: 0x%X", loadinfo.seg_sizes[1]);
     LOG_DEBUG(Loader, "data size:   0x%X (including 0x%X of bss)", loadinfo.seg_sizes[2], hdr.bss_size);
 
+    *out_codeset = code_set;
     return ERROR_NONE;
 }
 
@@ -230,17 +249,19 @@ ResultStatus AppLoader_THREEDSX::Load() {
     if (!file->IsOpen())
         return ResultStatus::Error;
 
-    Kernel::g_current_process = Kernel::Process::Create(filename, 0);
+    SharedPtr<CodeSet> codeset;
+    if (Load3DSXFile(*file, Memory::PROCESS_IMAGE_VADDR, &codeset) != ERROR_NONE)
+        return ResultStatus::Error;
+    codeset->name = filename;
+
+    Kernel::g_current_process = Kernel::Process::Create(std::move(codeset));
     Kernel::g_current_process->svc_access_mask.set();
     Kernel::g_current_process->address_mappings = default_address_mappings;
 
     // Attach the default resource limit (APPLICATION) to the process
     Kernel::g_current_process->resource_limit = Kernel::ResourceLimit::GetForCategory(Kernel::ResourceLimitCategory::APPLICATION);
 
-    if (Load3DSXFile(*file, Memory::PROCESS_IMAGE_VADDR) != ERROR_NONE)
+    Kernel::g_current_process->Run(48, Kernel::DEFAULT_STACK_SIZE);
-        return ResultStatus::Error;
-
-    Kernel::g_current_process->Run(Memory::PROCESS_IMAGE_VADDR, 48, Kernel::DEFAULT_STACK_SIZE);
 
     is_loaded = true;
     return ResultStatus::Success;

src/core/loader/elf.cpp

@@ -16,6 +16,9 @@
 #include "core/loader/elf.h"
 #include "core/memory.h"
 
+using Kernel::SharedPtr;
+using Kernel::CodeSet;
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // ELF Header Constants
 
@@ -97,6 +100,12 @@ enum ElfSectionFlags
 #define PT_LOPROC  0x70000000
 #define PT_HIPROC  0x7FFFFFFF
 
+// Segment flags
+#define PF_X               0x1
+#define PF_W               0x2
+#define PF_R               0x4
+#define PF_MASKPROC 0xF0000000
+
 typedef unsigned int   Elf32_Addr;
 typedef unsigned short Elf32_Half;
 typedef unsigned int   Elf32_Off;
@@ -193,7 +202,7 @@ public:
     ElfMachine GetMachine() const { return (ElfMachine)(header->e_machine); }
     u32 GetEntryPoint() const { return entryPoint; }
     u32 GetFlags() const { return (u32)(header->e_flags); }
-    void LoadInto(u32 vaddr);
+    SharedPtr<CodeSet> LoadInto(u32 vaddr);
     bool LoadSymbols();
 
     int GetNumSegments() const { return (int)(header->e_phnum); }
@@ -249,7 +258,7 @@ const char *ElfReader::GetSectionName(int section) const {
     return nullptr;
 }
 
-void ElfReader::LoadInto(u32 vaddr) {
+SharedPtr<CodeSet> ElfReader::LoadInto(u32 vaddr) {
     LOG_DEBUG(Loader, "String section: %i", header->e_shstrndx);
 
     // Should we relocate?
@@ -267,19 +276,61 @@ void ElfReader::LoadInto(u32 vaddr) {
     u32 segment_addr[32];
     u32 base_addr = relocate ? vaddr : 0;
 
-    for (unsigned i = 0; i < header->e_phnum; i++) {
+    u32 total_image_size = 0;
-        Elf32_Phdr* p = segments + i;
+    for (unsigned int i = 0; i < header->e_phnum; ++i) {
-        LOG_DEBUG(Loader, "Type: %i Vaddr: %08x Filesz: %i Memsz: %i ", p->p_type, p->p_vaddr,
+        Elf32_Phdr* p = &segments[i];
+        if (p->p_type == PT_LOAD) {
+            total_image_size += (p->p_memsz + 0xFFF) & ~0xFFF;
+        }
+    }
+
+    std::vector<u8> program_image(total_image_size);
+    size_t current_image_position = 0;
+
+    SharedPtr<CodeSet> codeset = CodeSet::Create("", 0);
+
+    for (unsigned int i = 0; i < header->e_phnum; ++i) {
+        Elf32_Phdr* p = &segments[i];
+        LOG_DEBUG(Loader, "Type: %i Vaddr: %08X Filesz: %8X Memsz: %8X ", p->p_type, p->p_vaddr,
                   p->p_filesz, p->p_memsz);
 
         if (p->p_type == PT_LOAD) {
-            segment_addr[i] = base_addr + p->p_vaddr;
+            CodeSet::Segment* codeset_segment;
-            memcpy(Memory::GetPointer(segment_addr[i]), GetSegmentPtr(i), p->p_filesz);
+            u32 permission_flags = p->p_flags & (PF_R | PF_W | PF_X);
-            LOG_DEBUG(Loader, "Loadable Segment Copied to %08x, size %08x", segment_addr[i],
+            if (permission_flags == (PF_R | PF_X)) {
-                      p->p_memsz);
+                codeset_segment = &codeset->code;
+            } else if (permission_flags == (PF_R)) {
+                codeset_segment = &codeset->rodata;
+            } else if (permission_flags == (PF_R | PF_W)) {
+                codeset_segment = &codeset->data;
+            } else {
+                LOG_ERROR(Loader, "Unexpected ELF PT_LOAD segment id %u with flags %X", i, p->p_flags);
+                continue;
+            }
+
+            if (codeset_segment->size != 0) {
+                LOG_ERROR(Loader, "ELF has more than one segment of the same type. Skipping extra segment (id %i)", i);
+                continue;
+            }
+
+            u32 segment_addr = base_addr + p->p_vaddr;
+            u32 aligned_size = (p->p_memsz + 0xFFF) & ~0xFFF;
+
+            codeset_segment->offset = current_image_position;
+            codeset_segment->addr = segment_addr;
+            codeset_segment->size = aligned_size;
+
+            memcpy(&program_image[current_image_position], GetSegmentPtr(i), p->p_filesz);
+            current_image_position += aligned_size;
         }
     }
+
+    codeset->entrypoint = base_addr + header->e_entry;
+    codeset->memory = std::make_shared<std::vector<u8>>(std::move(program_image));
+
     LOG_DEBUG(Loader, "Done loading.");
+
+    return codeset;
 }
 
 SectionID ElfReader::GetSectionByName(const char *name, int firstSection) const {
@@ -352,18 +403,18 @@ ResultStatus AppLoader_ELF::Load() {
     if (file->ReadBytes(&buffer[0], size) != size)
         return ResultStatus::Error;
 
-    Kernel::g_current_process = Kernel::Process::Create(filename, 0);
+    ElfReader elf_reader(&buffer[0]);
+    SharedPtr<CodeSet> codeset = elf_reader.LoadInto(Memory::PROCESS_IMAGE_VADDR);
+    codeset->name = filename;
+
+    Kernel::g_current_process = Kernel::Process::Create(std::move(codeset));
     Kernel::g_current_process->svc_access_mask.set();
     Kernel::g_current_process->address_mappings = default_address_mappings;
 
     // Attach the default resource limit (APPLICATION) to the process
     Kernel::g_current_process->resource_limit = Kernel::ResourceLimit::GetForCategory(Kernel::ResourceLimitCategory::APPLICATION);
 
-    ElfReader elf_reader(&buffer[0]);
+    Kernel::g_current_process->Run(48, Kernel::DEFAULT_STACK_SIZE);
-    elf_reader.LoadInto(Memory::PROCESS_IMAGE_VADDR);
-    // TODO: Fill application title
-
-    Kernel::g_current_process->Run(elf_reader.GetEntryPoint(), 48, Kernel::DEFAULT_STACK_SIZE);
 
     is_loaded = true;
     return ResultStatus::Success;

src/core/loader/ncch.cpp

@@ -118,6 +118,9 @@ FileType AppLoader_NCCH::IdentifyType(FileUtil::IOFile& file) {
 }
 
 ResultStatus AppLoader_NCCH::LoadExec() const {
+    using Kernel::SharedPtr;
+    using Kernel::CodeSet;
+
     if (!is_loaded)
         return ResultStatus::ErrorNotLoaded;
 
@@ -126,7 +129,30 @@ ResultStatus AppLoader_NCCH::LoadExec() const {
         std::string process_name = Common::StringFromFixedZeroTerminatedBuffer(
                 (const char*)exheader_header.codeset_info.name, 8);
         u64 program_id = *reinterpret_cast<u64_le const*>(&ncch_header.program_id[0]);
-        Kernel::g_current_process = Kernel::Process::Create(process_name, program_id);
+
+        SharedPtr<CodeSet> codeset = CodeSet::Create(process_name, program_id);
+
+        codeset->code.offset = 0;
+        codeset->code.addr = exheader_header.codeset_info.text.address;
+        codeset->code.size = exheader_header.codeset_info.text.num_max_pages * Memory::PAGE_SIZE;
+
+        codeset->rodata.offset = codeset->code.offset + codeset->code.size;
+        codeset->rodata.addr = exheader_header.codeset_info.ro.address;
+        codeset->rodata.size = exheader_header.codeset_info.ro.num_max_pages * Memory::PAGE_SIZE;
+
+        // TODO(yuriks): Not sure if the bss size is added to the page-aligned .data size or just
+        //               to the regular size. Playing it safe for now.
+        u32 bss_page_size = (exheader_header.codeset_info.bss_size + 0xFFF) & ~0xFFF;
+        code.resize(code.size() + bss_page_size, 0);
+
+        codeset->data.offset = codeset->rodata.offset + codeset->rodata.size;
+        codeset->data.addr = exheader_header.codeset_info.data.address;
+        codeset->data.size = exheader_header.codeset_info.data.num_max_pages * Memory::PAGE_SIZE + bss_page_size;
+
+        codeset->entrypoint = codeset->code.addr;
+        codeset->memory = std::make_shared<std::vector<u8>>(std::move(code));
+
+        Kernel::g_current_process = Kernel::Process::Create(std::move(codeset));
 
         // Attach a resource limit to the process based on the resource limit category
         Kernel::g_current_process->resource_limit = Kernel::ResourceLimit::GetForCategory(
@@ -137,11 +163,9 @@ ResultStatus AppLoader_NCCH::LoadExec() const {
         std::copy_n(exheader_header.arm11_kernel_caps.descriptors, kernel_caps.size(), begin(kernel_caps));
         Kernel::g_current_process->ParseKernelCaps(kernel_caps.data(), kernel_caps.size());
 
-        Memory::WriteBlock(entry_point, &code[0], code.size());
-
         s32 priority = exheader_header.arm11_system_local_caps.priority;
         u32 stack_size = exheader_header.codeset_info.stack_size;
-        Kernel::g_current_process->Run(entry_point, priority, stack_size);
+        Kernel::g_current_process->Run(priority, stack_size);
         return ResultStatus::Success;
     }
     return ResultStatus::Error;

src/core/mem_map.cpp

@@ -32,7 +32,6 @@ struct MemoryArea {
 
 // We don't declare the IO regions in here since its handled by other means.
 static MemoryArea memory_areas[] = {
-    {PROCESS_IMAGE_VADDR, PROCESS_IMAGE_MAX_SIZE, "Process Image"}, // ExeFS:/.code is loaded here
     {HEAP_VADDR,          HEAP_SIZE,              "Heap"},          // Application heap (main memory)
     {SHARED_MEMORY_VADDR, SHARED_MEMORY_SIZE,     "Shared Memory"}, // Shared memory
     {LINEAR_HEAP_VADDR,   LINEAR_HEAP_SIZE,       "Linear Heap"},   // Linear heap (main memory)
@@ -132,13 +131,13 @@ VAddr PhysicalToVirtualAddress(const PAddr addr) {
     return addr | 0x80000000;
 }
 
-// TODO(yuriks): Move this into Process
-static Kernel::VMManager address_space;
-
 void Init() {
-    using namespace Kernel;
-
     InitMemoryMap();
+    LOG_DEBUG(HW_Memory, "initialized OK");
+}
+
+void InitLegacyAddressSpace(Kernel::VMManager& address_space) {
+    using namespace Kernel;
 
     for (MemoryArea& area : memory_areas) {
         auto block = std::make_shared<std::vector<u8>>(area.size);
@@ -152,14 +151,11 @@ void Init() {
     auto shared_page_vma = address_space.MapBackingMemory(SHARED_PAGE_VADDR,
             (u8*)&SharedPage::shared_page, SHARED_PAGE_SIZE, MemoryState::Shared).MoveFrom();
     address_space.Reprotect(shared_page_vma, VMAPermission::Read);
-
-    LOG_DEBUG(HW_Memory, "initialized OK");
 }
 
 void Shutdown() {
     heap_map.clear();
     heap_linear_map.clear();
-    address_space.Reset();
 
     LOG_DEBUG(HW_Memory, "shutdown OK");
 }

src/core/mem_map.h

@@ -6,9 +6,14 @@
 
 #include "common/common_types.h"
 
+namespace Kernel {
+class VMManager;
+}
+
 namespace Memory {
 
 void Init();
+void InitLegacyAddressSpace(Kernel::VMManager& address_space);
 void Shutdown();
 
 /**