Kernel: Properly implement ControlMemory FREE and COMMIT

src/core/hle/kernel/process.cpp

@@ -36,8 +36,7 @@ SharedPtr<Process> Process::Create(SharedPtr<CodeSet> code_set) {
     process->codeset = std::move(code_set);
     process->flags.raw = 0;
     process->flags.memory_region = MemoryRegion::APPLICATION;
-    process->address_space = Common::make_unique<VMManager>();
+    Memory::InitLegacyAddressSpace(process->vm_manager);
-    Memory::InitLegacyAddressSpace(*process->address_space);
 
     return process;
 }
@@ -104,19 +103,130 @@ void Process::ParseKernelCaps(const u32* kernel_caps, size_t len) {
 
 void Process::Run(s32 main_thread_priority, u32 stack_size) {
     auto MapSegment = [&](CodeSet::Segment& segment, VMAPermission permissions, MemoryState memory_state) {
-        auto vma = address_space->MapMemoryBlock(segment.addr, codeset->memory,
+        auto vma = vm_manager.MapMemoryBlock(segment.addr, codeset->memory,
                 segment.offset, segment.size, memory_state).Unwrap();
-        address_space->Reprotect(vma, permissions);
+        vm_manager.Reprotect(vma, permissions);
     };
 
+    // Map CodeSet segments
     MapSegment(codeset->code,   VMAPermission::ReadExecute, MemoryState::Code);
     MapSegment(codeset->rodata, VMAPermission::Read,        MemoryState::Code);
     MapSegment(codeset->data,   VMAPermission::ReadWrite,   MemoryState::Private);
 
-    address_space->LogLayout(Log::Level::Debug);
+    // Allocate and map stack
+    vm_manager.MapMemoryBlock(Memory::HEAP_VADDR_END - stack_size,
+            std::make_shared<std::vector<u8>>(stack_size, 0), 0, stack_size, MemoryState::Locked
+            ).Unwrap();
+
+    vm_manager.LogLayout(Log::Level::Debug);
     Kernel::SetupMainThread(codeset->entrypoint, main_thread_priority);
 }
 
+ResultVal<VAddr> Process::HeapAllocate(VAddr target, u32 size, VMAPermission perms) {
+    if (target < Memory::HEAP_VADDR || target + size > Memory::HEAP_VADDR_END || target + size < target) {
+        return ERR_INVALID_ADDRESS;
+    }
+
+    if (heap_memory == nullptr) {
+        // Initialize heap
+        heap_memory = std::make_shared<std::vector<u8>>();
+        heap_start = heap_end = target;
+    }
+
+    // If necessary, expand backing vector to cover new heap extents.
+    if (target < heap_start) {
+        heap_memory->insert(begin(*heap_memory), heap_start - target, 0);
+        heap_start = target;
+        vm_manager.RefreshMemoryBlockMappings(heap_memory.get());
+    }
+    if (target + size > heap_end) {
+        heap_memory->insert(end(*heap_memory), (target + size) - heap_end, 0);
+        heap_end = target + size;
+        vm_manager.RefreshMemoryBlockMappings(heap_memory.get());
+    }
+    ASSERT(heap_end - heap_start == heap_memory->size());
+
+    CASCADE_RESULT(auto vma, vm_manager.MapMemoryBlock(target, heap_memory, target - heap_start, size, MemoryState::Private));
+    vm_manager.Reprotect(vma, perms);
+
+    return MakeResult<VAddr>(heap_end - size);
+}
+
+ResultCode Process::HeapFree(VAddr target, u32 size) {
+    if (target < Memory::HEAP_VADDR || target + size > Memory::HEAP_VADDR_END || target + size < target) {
+        return ERR_INVALID_ADDRESS;
+    }
+
+    ResultCode result = vm_manager.UnmapRange(target, size);
+    if (result.IsError()) return result;
+
+    return RESULT_SUCCESS;
+}
+
+ResultVal<VAddr> Process::LinearAllocate(VAddr target, u32 size, VMAPermission perms) {
+    if (linear_heap_memory == nullptr) {
+        // Initialize heap
+        linear_heap_memory = std::make_shared<std::vector<u8>>();
+    }
+
+    VAddr heap_end = Memory::LINEAR_HEAP_VADDR + (u32)linear_heap_memory->size();
+    // Games and homebrew only ever seem to pass 0 here (which lets the kernel decide the address),
+    // but explicit addresses are also accepted and respected.
+    if (target == 0) {
+        target = heap_end;
+    }
+
+    if (target < Memory::LINEAR_HEAP_VADDR || target + size > Memory::LINEAR_HEAP_VADDR_END ||
+        target > heap_end || target + size < target) {
+
+        return ERR_INVALID_ADDRESS;
+    }
+
+    // Expansion of the linear heap is only allowed if you do an allocation immediatelly at its
+    // end. It's possible to free gaps in the middle of the heap and then reallocate them later,
+    // but expansions are only allowed at the end.
+    if (target == heap_end) {
+        linear_heap_memory->insert(linear_heap_memory->end(), size, 0);
+        vm_manager.RefreshMemoryBlockMappings(linear_heap_memory.get());
+    }
+
+    size_t offset = target - Memory::LINEAR_HEAP_VADDR;
+    CASCADE_RESULT(auto vma, vm_manager.MapMemoryBlock(target, linear_heap_memory, offset, size, MemoryState::Continuous));
+    vm_manager.Reprotect(vma, perms);
+
+    return MakeResult<VAddr>(target);
+}
+
+ResultCode Process::LinearFree(VAddr target, u32 size) {
+    if (linear_heap_memory == nullptr || target < Memory::LINEAR_HEAP_VADDR ||
+        target + size > Memory::LINEAR_HEAP_VADDR_END || target + size < target) {
+
+        return ERR_INVALID_ADDRESS;
+    }
+
+    VAddr heap_end = Memory::LINEAR_HEAP_VADDR + (u32)linear_heap_memory->size();
+    if (target + size > heap_end) {
+        return ERR_INVALID_ADDRESS_STATE;
+    }
+
+    ResultCode result = vm_manager.UnmapRange(target, size);
+    if (result.IsError()) return result;
+
+    if (target + size == heap_end) {
+        // End of linear heap has been freed, so check what's the last allocated block in it and
+        // reduce the size.
+        auto vma = vm_manager.FindVMA(target);
+        ASSERT(vma != vm_manager.vma_map.end());
+        ASSERT(vma->second.type == VMAType::Free);
+        VAddr new_end = vma->second.base;
+        if (new_end >= Memory::LINEAR_HEAP_VADDR) {
+            linear_heap_memory->resize(new_end - Memory::LINEAR_HEAP_VADDR);
+        }
+    }
+
+    return RESULT_SUCCESS;
+}
+
 Kernel::Process::Process() {}
 Kernel::Process::~Process() {}
 

src/core/hle/kernel/process.h

@@ -15,6 +15,7 @@
 #include "common/common_types.h"
 
 #include "core/hle/kernel/kernel.h"
+#include "core/hle/kernel/vm_manager.h"
 
 namespace Kernel {
 
@@ -48,7 +49,6 @@ union ProcessFlags {
 };
 
 class ResourceLimit;
-class VMManager;
 
 struct CodeSet final : public Object {
     static SharedPtr<CodeSet> Create(std::string name, u64 program_id);
@@ -108,10 +108,6 @@ public:
     /// The id of this process
     u32 process_id = next_process_id++;
 
-    /// 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
      * to this process.
@@ -123,6 +119,31 @@ public:
      */
     void Run(s32 main_thread_priority, u32 stack_size);
 
+
+    ///////////////////////////////////////////////////////////////////////////////////////////////
+    // Memory Management
+
+    VMManager vm_manager;
+
+    // Memory used to back the allocations in the regular heap. A single vector is used to cover
+    // the entire virtual address space extents that bound the allocations, including any holes.
+    // This makes deallocation and reallocation of holes fast and keeps process memory contiguous
+    // in the emulator address space, allowing Memory::GetPointer to be reasonably safe.
+    std::shared_ptr<std::vector<u8>> heap_memory;
+    // The left/right bounds of the address space covered by heap_memory.
+    VAddr heap_start = 0, heap_end = 0;
+
+    std::shared_ptr<std::vector<u8>> linear_heap_memory;
+
+    /// Bitmask of the used TLS slots
+    std::bitset<300> used_tls_slots;
+
+    ResultVal<VAddr> HeapAllocate(VAddr target, u32 size, VMAPermission perms);
+    ResultCode HeapFree(VAddr target, u32 size);
+
+    ResultVal<VAddr> LinearAllocate(VAddr target, u32 size, VMAPermission perms);
+    ResultCode LinearFree(VAddr target, u32 size);
+
 private:
     Process();
     ~Process() override;

src/core/hle/kernel/vm_manager.cpp

@@ -60,8 +60,12 @@ void VMManager::Reset() {
 }
 
 VMManager::VMAHandle VMManager::FindVMA(VAddr target) const {
+    if (target >= MAX_ADDRESS) {
+        return vma_map.end();
+    } else {
         return std::prev(vma_map.upper_bound(target));
     }
+}
 
 ResultVal<VMManager::VMAHandle> VMManager::MapMemoryBlock(VAddr target,
         std::shared_ptr<std::vector<u8>> block, size_t offset, u32 size, MemoryState state) {
@@ -115,10 +119,8 @@ ResultVal<VMManager::VMAHandle> VMManager::MapMMIO(VAddr target, PAddr paddr, u3
     return MakeResult<VMAHandle>(MergeAdjacent(vma_handle));
 }
 
-void VMManager::Unmap(VMAHandle vma_handle) {
+VMManager::VMAIter VMManager::Unmap(VMAIter vma_handle) {
-    VMAIter iter = StripIterConstness(vma_handle);
+    VirtualMemoryArea& vma = vma_handle->second;
-
-    VirtualMemoryArea& vma = iter->second;
     vma.type = VMAType::Free;
     vma.permissions = VMAPermission::None;
     vma.meminfo_state = MemoryState::Free;
@@ -130,17 +132,57 @@ void VMManager::Unmap(VMAHandle vma_handle) {
 
     UpdatePageTableForVMA(vma);
 
-    MergeAdjacent(iter);
+    return MergeAdjacent(vma_handle);
 }
 
-void VMManager::Reprotect(VMAHandle vma_handle, VMAPermission new_perms) {
+ResultCode VMManager::UnmapRange(VAddr target, u32 size) {
+    CASCADE_RESULT(VMAIter vma, CarveVMARange(target, size));
+    VAddr target_end = target + size;
+
+    VMAIter end = vma_map.end();
+    // The comparison against the end of the range must be done using addresses since VMAs can be
+    // merged during this process, causing invalidation of the iterators.
+    while (vma != end && vma->second.base < target_end) {
+        vma = std::next(Unmap(vma));
+    }
+
+    ASSERT(FindVMA(target)->second.size >= size);
+    return RESULT_SUCCESS;
+}
+
+VMManager::VMAHandle VMManager::Reprotect(VMAHandle vma_handle, VMAPermission new_perms) {
     VMAIter iter = StripIterConstness(vma_handle);
 
     VirtualMemoryArea& vma = iter->second;
     vma.permissions = new_perms;
     UpdatePageTableForVMA(vma);
 
-    MergeAdjacent(iter);
+    return MergeAdjacent(iter);
+}
+
+ResultCode VMManager::ReprotectRange(VAddr target, u32 size, VMAPermission new_perms) {
+    CASCADE_RESULT(VMAIter vma, CarveVMARange(target, size));
+    VAddr target_end = target + size;
+
+    VMAIter end = vma_map.end();
+    // The comparison against the end of the range must be done using addresses since VMAs can be
+    // merged during this process, causing invalidation of the iterators.
+    while (vma != end && vma->second.base < target_end) {
+        vma = std::next(StripIterConstness(Reprotect(vma, new_perms)));
+    }
+
+    return RESULT_SUCCESS;
+}
+
+void VMManager::RefreshMemoryBlockMappings(const std::vector<u8>* block) {
+    // If this ever proves to have a noticeable performance impact, allow users of the function to
+    // specify a specific range of addresses to limit the scan to.
+    for (const auto& p : vma_map) {
+        const VirtualMemoryArea& vma = p.second;
+        if (block == vma.backing_block.get()) {
+            UpdatePageTableForVMA(vma);
+        }
+    }
 }
 
 void VMManager::LogLayout(Log::Level log_level) const {
@@ -161,8 +203,8 @@ VMManager::VMAIter VMManager::StripIterConstness(const VMAHandle & iter) {
 }
 
 ResultVal<VMManager::VMAIter> VMManager::CarveVMA(VAddr base, u32 size) {
-    ASSERT_MSG((size & Memory::PAGE_MASK) == 0, "non-page aligned size: %8X", size);
+    ASSERT_MSG((size & Memory::PAGE_MASK) == 0, "non-page aligned size: 0x%8X", size);
-    ASSERT_MSG((base & Memory::PAGE_MASK) == 0, "non-page aligned base: %08X", base);
+    ASSERT_MSG((base & Memory::PAGE_MASK) == 0, "non-page aligned base: 0x%08X", base);
 
     VMAIter vma_handle = StripIterConstness(FindVMA(base));
     if (vma_handle == vma_map.end()) {
@@ -196,6 +238,35 @@ ResultVal<VMManager::VMAIter> VMManager::CarveVMA(VAddr base, u32 size) {
     return MakeResult<VMAIter>(vma_handle);
 }
 
+ResultVal<VMManager::VMAIter> VMManager::CarveVMARange(VAddr target, u32 size) {
+    ASSERT_MSG((size & Memory::PAGE_MASK) == 0, "non-page aligned size: 0x%8X", size);
+    ASSERT_MSG((target & Memory::PAGE_MASK) == 0, "non-page aligned base: 0x%08X", target);
+
+    VAddr target_end = target + size;
+    ASSERT(target_end >= target);
+    ASSERT(target_end <= MAX_ADDRESS);
+    ASSERT(size > 0);
+
+    VMAIter begin_vma = StripIterConstness(FindVMA(target));
+    VMAIter i_end = vma_map.lower_bound(target_end);
+    for (auto i = begin_vma; i != i_end; ++i) {
+        if (i->second.type == VMAType::Free) {
+            return ERR_INVALID_ADDRESS_STATE;
+        }
+    }
+
+    if (target != begin_vma->second.base) {
+        begin_vma = SplitVMA(begin_vma, target - begin_vma->second.base);
+    }
+
+    VMAIter end_vma = StripIterConstness(FindVMA(target_end));
+    if (end_vma != vma_map.end() && target_end != end_vma->second.base) {
+        end_vma = SplitVMA(end_vma, target_end - end_vma->second.base);
+    }
+
+    return MakeResult<VMAIter>(begin_vma);
+}
+
 VMManager::VMAIter VMManager::SplitVMA(VMAIter vma_handle, u32 offset_in_vma) {
     VirtualMemoryArea& old_vma = vma_handle->second;
     VirtualMemoryArea new_vma = old_vma; // Make a copy of the VMA

src/core/hle/kernel/vm_manager.h

@@ -171,11 +171,20 @@ public:
      */
     ResultVal<VMAHandle> MapMMIO(VAddr target, PAddr paddr, u32 size, MemoryState state);
 
-    /// Unmaps the given VMA.
+    /// Unmaps a range of addresses, splitting VMAs as necessary.
-    void Unmap(VMAHandle vma);
+    ResultCode UnmapRange(VAddr target, u32 size);
 
     /// Changes the permissions of the given VMA.
-    void Reprotect(VMAHandle vma, VMAPermission new_perms);
+    VMAHandle Reprotect(VMAHandle vma, VMAPermission new_perms);
+
+    /// Changes the permissions of a range of addresses, splitting VMAs as necessary.
+    ResultCode ReprotectRange(VAddr target, u32 size, VMAPermission new_perms);
+
+    /**
+     * Scans all VMAs and updates the page table range of any that use the given vector as backing
+     * memory. This should be called after any operation that causes reallocation of the vector.
+     */
+    void RefreshMemoryBlockMappings(const std::vector<u8>* block);
 
     /// Dumps the address space layout to the log, for debugging
     void LogLayout(Log::Level log_level) const;
@@ -186,12 +195,21 @@ private:
     /// Converts a VMAHandle to a mutable VMAIter.
     VMAIter StripIterConstness(const VMAHandle& iter);
 
+    /// Unmaps the given VMA.
+    VMAIter Unmap(VMAIter vma);
+
     /**
      * Carves a VMA of a specific size at the specified address by splitting Free VMAs while doing
      * the appropriate error checking.
      */
     ResultVal<VMAIter> CarveVMA(VAddr base, u32 size);
 
+    /**
+     * Splits the edges of the given range of non-Free VMAs so that there is a VMA split at each
+     * end of the range.
+     */
+    ResultVal<VMAIter> CarveVMARange(VAddr base, u32 size);
+
     /**
      * Splits a VMA in two, at the specified offset.
      * @returns the right side of the split, with the original iterator becoming the left side.

src/core/hle/svc.cpp

@@ -41,32 +41,114 @@ const ResultCode ERR_NOT_FOUND(ErrorDescription::NotFound, ErrorModule::Kernel,
 const ResultCode ERR_PORT_NAME_TOO_LONG(ErrorDescription(30), ErrorModule::OS,
         ErrorSummary::InvalidArgument, ErrorLevel::Usage); // 0xE0E0181E
 
+const ResultCode ERR_MISALIGNED_ADDRESS{ // 0xE0E01BF1
+        ErrorDescription::MisalignedAddress, ErrorModule::OS,
+        ErrorSummary::InvalidArgument, ErrorLevel::Usage};
+const ResultCode ERR_MISALIGNED_SIZE{ // 0xE0E01BF2
+        ErrorDescription::MisalignedSize, ErrorModule::OS,
+        ErrorSummary::InvalidArgument, ErrorLevel::Usage};
+const ResultCode ERR_INVALID_COMBINATION{ // 0xE0E01BEE
+        ErrorDescription::InvalidCombination, ErrorModule::OS,
+        ErrorSummary::InvalidArgument, ErrorLevel::Usage};
+
 enum ControlMemoryOperation {
-    MEMORY_OPERATION_HEAP       = 0x00000003,
+    MEMOP_FREE    = 1,
-    MEMORY_OPERATION_GSP_HEAP   = 0x00010003,
+    MEMOP_RESERVE = 2, // This operation seems to be unsupported in the kernel
+    MEMOP_COMMIT  = 3,
+    MEMOP_MAP     = 4,
+    MEMOP_UNMAP   = 5,
+    MEMOP_PROTECT = 6,
+    MEMOP_OPERATION_MASK = 0xFF,
+
+    MEMOP_REGION_APP    = 0x100,
+    MEMOP_REGION_SYSTEM = 0x200,
+    MEMOP_REGION_BASE   = 0x300,
+    MEMOP_REGION_MASK   = 0xF00,
+
+    MEMOP_LINEAR = 0x10000,
 };
 
 /// Map application or GSP heap memory
 static ResultCode ControlMemory(u32* out_addr, u32 operation, u32 addr0, u32 addr1, u32 size, u32 permissions) {
-    LOG_TRACE(Kernel_SVC,"called operation=0x%08X, addr0=0x%08X, addr1=0x%08X, size=%08X, permissions=0x%08X",
+    using namespace Kernel;
+
+    LOG_DEBUG(Kernel_SVC,"called operation=0x%08X, addr0=0x%08X, addr1=0x%08X, size=0x%X, permissions=0x%08X",
         operation, addr0, addr1, size, permissions);
 
-    switch (operation) {
+    if ((addr0 & Memory::PAGE_MASK) != 0 || (addr1 & Memory::PAGE_MASK) != 0) {
+        return ERR_MISALIGNED_ADDRESS;
+    }
+    if ((size & Memory::PAGE_MASK) != 0) {
+        return ERR_MISALIGNED_SIZE;
+    }
 
-    // Map normal heap memory
+    u32 region = operation & MEMOP_REGION_MASK;
-    case MEMORY_OPERATION_HEAP:
+    operation &= ~MEMOP_REGION_MASK;
-        *out_addr = Memory::MapBlock_Heap(size, operation, permissions);
+
+    if (region != 0) {
+        LOG_WARNING(Kernel_SVC, "ControlMemory with specified region not supported, region=%X", region);
+    }
+
+    if ((permissions & (u32)MemoryPermission::ReadWrite) != permissions) {
+        return ERR_INVALID_COMBINATION;
+    }
+    VMAPermission vma_permissions = (VMAPermission)permissions;
+
+    auto& process = *g_current_process;
+
+    switch (operation & MEMOP_OPERATION_MASK) {
+    case MEMOP_FREE:
+    {
+        if (addr0 >= Memory::HEAP_VADDR && addr0 < Memory::HEAP_VADDR_END) {
+            ResultCode result = process.HeapFree(addr0, size);
+            if (result.IsError()) return result;
+        } else if (addr0 >= Memory::LINEAR_HEAP_VADDR && addr0 < Memory::LINEAR_HEAP_VADDR_END) {
+            ResultCode result = process.LinearFree(addr0, size);
+            if (result.IsError()) return result;
+        } else {
+            return ERR_INVALID_ADDRESS;
+        }
+        *out_addr = addr0;
         break;
+    }
 
-    // Map GSP heap memory
+    case MEMOP_COMMIT:
-    case MEMORY_OPERATION_GSP_HEAP:
+    {
-        *out_addr = Memory::MapBlock_HeapLinear(size, operation, permissions);
+        if (operation & MEMOP_LINEAR) {
+            CASCADE_RESULT(*out_addr, process.LinearAllocate(addr0, size, vma_permissions));
+        } else {
+            CASCADE_RESULT(*out_addr, process.HeapAllocate(addr0, size, vma_permissions));
+        }
         break;
+    }
+
+    case MEMOP_MAP: // TODO: This is just a hack to avoid regressions until memory aliasing is implemented
+    {
+        CASCADE_RESULT(*out_addr, process.HeapAllocate(addr0, size, vma_permissions));
+        break;
+    }
+
+    case MEMOP_UNMAP: // TODO: This is just a hack to avoid regressions until memory aliasing is implemented
+    {
+        ResultCode result = process.HeapFree(addr0, size);
+        if (result.IsError()) return result;
+        break;
+    }
+
+    case MEMOP_PROTECT:
+    {
+        ResultCode result = process.vm_manager.ReprotectRange(addr0, size, vma_permissions);
+        if (result.IsError()) return result;
+        break;
+    }
 
-    // Unknown ControlMemory operation
     default:
         LOG_ERROR(Kernel_SVC, "unknown operation=0x%08X", operation);
+        return ERR_INVALID_COMBINATION;
     }
+
+    process.vm_manager.LogLayout(Log::Level::Trace);
+
     return RESULT_SUCCESS;
 }
 
@@ -537,9 +619,9 @@ static ResultCode QueryProcessMemory(MemoryInfo* memory_info, PageInfo* page_inf
     if (process == nullptr)
         return ERR_INVALID_HANDLE;
 
-    auto vma = process->address_space->FindVMA(addr);
+    auto vma = process->vm_manager.FindVMA(addr);
 
-    if (vma == process->address_space->vma_map.end())
+    if (vma == Kernel::g_current_process->vm_manager.vma_map.end())
         return ResultCode(ErrorDescription::InvalidAddress, ErrorModule::OS, ErrorSummary::InvalidArgument, ErrorLevel::Usage);
 
     memory_info->base_address = vma->second.base;

src/core/mem_map.cpp

@@ -32,9 +32,7 @@ struct MemoryArea {
 
 // We don't declare the IO regions in here since its handled by other means.
 static MemoryArea memory_areas[] = {
-    {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)
     {VRAM_VADDR,          VRAM_SIZE,              "VRAM"},          // Video memory (VRAM)
     {DSP_RAM_VADDR,       DSP_RAM_SIZE,           "DSP RAM"},       // DSP memory
     {TLS_AREA_VADDR,      TLS_AREA_SIZE,          "TLS Area"},      // TLS memory