kernel/thread: Change owner_process to std::weak_ptr (#5325)

* kernel/thread: Change owner_process to std::weak_ptr

Previously this leaked almost all kernel objects. In short, Threads own Processes which own HandleTables which own maps of Objects which include Threads.

Changing this to weak_ptr at least got the camera interfaces to destruct properly. Did not really check the other objects though, and I think there are probably more leaks.

* hle/kernel: Lock certain objects while deserializing

When deserializing other kernel objects, these objects (`MemoryRegion`s and `VMManager`s) can possibly get modified. To avoid inconsistent state caused by destructor side-effects, we may as well simply lock them until loading is fully completed.

* Fix silly typo

Somehow this didn't break?!

src/citra_qt/debugger/wait_tree.cpp

@@ -243,7 +243,6 @@ std::vector<std::unique_ptr<WaitTreeItem>> WaitTreeThread::GetChildren() const {
     std::vector<std::unique_ptr<WaitTreeItem>> list(WaitTreeWaitObject::GetChildren());
 
     const auto& thread = static_cast<const Kernel::Thread&>(object);
-    const auto& process = thread.owner_process;
 
     QString processor;
     switch (thread.processor_id) {
@@ -267,10 +266,12 @@ std::vector<std::unique_ptr<WaitTreeItem>> WaitTreeThread::GetChildren() const {
     list.push_back(std::make_unique<WaitTreeText>(tr("object id = %1").arg(thread.GetObjectId())));
     list.push_back(std::make_unique<WaitTreeText>(tr("processor = %1").arg(processor)));
     list.push_back(std::make_unique<WaitTreeText>(tr("thread id = %1").arg(thread.GetThreadId())));
+    if (auto process = thread.owner_process.lock()) {
         list.push_back(
             std::make_unique<WaitTreeText>(tr("process = %1 (%2)")
                                                .arg(QString::fromStdString(process->GetName()))
                                                .arg(process->process_id)));
+    }
     list.push_back(std::make_unique<WaitTreeText>(tr("priority = %1(current) / %2(normal)")
                                                       .arg(thread.current_priority)
                                                       .arg(thread.nominal_priority)));

src/core/hle/kernel/hle_ipc.cpp

@@ -29,7 +29,9 @@ public:
             callback->WakeUp(thread, *context, reason);
         }
 
-        auto& process = thread->owner_process;
+        auto process = thread->owner_process.lock();
+        ASSERT(process);
+
         // We must copy the entire command buffer *plus* the entire static buffers area, since
         // the translation might need to read from it in order to retrieve the StaticBuffer
         // target addresses.

src/core/hle/kernel/ipc.cpp

@@ -24,8 +24,9 @@ ResultCode TranslateCommandBuffer(Kernel::KernelSystem& kernel, Memory::MemorySy
                                   VAddr dst_address,
                                   std::vector<MappedBufferContext>& mapped_buffer_context,
                                   bool reply) {
-    auto& src_process = src_thread->owner_process;
-    auto& dst_process = dst_thread->owner_process;
+    auto src_process = src_thread->owner_process.lock();
+    auto dst_process = dst_thread->owner_process.lock();
+    ASSERT(src_process && dst_process);
 
     IPC::Header header;
     // TODO(Subv): Replace by Memory::Read32 when possible.

src/core/hle/kernel/ipc_debugger/recorder.cpp

@@ -50,9 +50,10 @@ void Recorder::RegisterRequest(const std::shared_ptr<Kernel::ClientSession>& cli
                                const std::shared_ptr<Kernel::Thread>& client_thread) {
     const u32 thread_id = client_thread->GetThreadId();
 
+    if (auto owner_process = client_thread->owner_process.lock()) {
         RequestRecord record = {/* id */ ++record_count,
                                 /* status */ RequestStatus::Sent,
-                            /* client_process */ GetObjectInfo(client_thread->owner_process.get()),
+                                /* client_process */ GetObjectInfo(owner_process.get()),
                                 /* client_thread */ GetObjectInfo(client_thread.get()),
                                 /* client_session */ GetObjectInfo(client_session.get()),
                                 /* client_port */ GetObjectInfo(client_session->parent->port.get()),
@@ -64,6 +65,7 @@ void Recorder::RegisterRequest(const std::shared_ptr<Kernel::ClientSession>& cli
 
         InvokeCallbacks(record);
     }
+}
 
 void Recorder::SetRequestInfo(const std::shared_ptr<Kernel::Thread>& client_thread,
                               std::vector<u32> untranslated_cmdbuf,
@@ -82,7 +84,9 @@ void Recorder::SetRequestInfo(const std::shared_ptr<Kernel::Thread>& client_thre
     record.translated_request_cmdbuf = std::move(translated_cmdbuf);
 
     if (server_thread) {
-        record.server_process = GetObjectInfo(server_thread->owner_process.get());
+        if (auto owner_process = server_thread->owner_process.lock()) {
+            record.server_process = GetObjectInfo(owner_process.get());
+        }
         record.server_thread = GetObjectInfo(server_thread.get());
     } else {
         record.is_hle = true;

src/core/hle/kernel/kernel.cpp

@@ -177,6 +177,15 @@ void KernelSystem::serialize(Archive& ar, const unsigned int file_version) {
     ar& stored_processes;
     ar& next_thread_id;
     // Deliberately don't include debugger info to allow debugging through loads
+
+    if (Archive::is_loading::value) {
+        for (auto& memory_region : memory_regions) {
+            memory_region->Unlock();
+        }
+        for (auto& process : process_list) {
+            process->vm_manager.Unlock();
+        }
+    }
 }
 
 SERIALIZE_IMPL(KernelSystem)

src/core/hle/kernel/memory.cpp

@@ -174,6 +174,8 @@ void KernelSystem::MapSharedPages(VMManager& address_space) {
 }
 
 void MemoryRegionInfo::Reset(u32 base, u32 size) {
+    ASSERT(!is_locked);
+
     this->base = base;
     this->size = size;
     used = 0;
@@ -184,6 +186,8 @@ void MemoryRegionInfo::Reset(u32 base, u32 size) {
 }
 
 MemoryRegionInfo::IntervalSet MemoryRegionInfo::HeapAllocate(u32 size) {
+    ASSERT(!is_locked);
+
     IntervalSet result;
     u32 rest = size;
 
@@ -211,6 +215,8 @@ MemoryRegionInfo::IntervalSet MemoryRegionInfo::HeapAllocate(u32 size) {
 }
 
 bool MemoryRegionInfo::LinearAllocate(u32 offset, u32 size) {
+    ASSERT(!is_locked);
+
     Interval interval(offset, offset + size);
     if (!boost::icl::contains(free_blocks, interval)) {
         // The requested range is already allocated
@@ -222,6 +228,8 @@ bool MemoryRegionInfo::LinearAllocate(u32 offset, u32 size) {
 }
 
 std::optional<u32> MemoryRegionInfo::LinearAllocate(u32 size) {
+    ASSERT(!is_locked);
+
     // Find the first sufficient continuous block from the lower address
     for (const auto& interval : free_blocks) {
         ASSERT(interval.bounds() == boost::icl::interval_bounds::right_open());
@@ -238,10 +246,18 @@ std::optional<u32> MemoryRegionInfo::LinearAllocate(u32 size) {
 }
 
 void MemoryRegionInfo::Free(u32 offset, u32 size) {
+    if (is_locked) {
+        return;
+    }
+
     Interval interval(offset, offset + size);
     ASSERT(!boost::icl::intersects(free_blocks, interval)); // must be allocated blocks
     free_blocks += interval;
     used -= size;
 }
 
+void MemoryRegionInfo::Unlock() {
+    is_locked = false;
+}
+
 } // namespace Kernel

src/core/hle/kernel/memory.h

@@ -26,6 +26,10 @@ struct MemoryRegionInfo {
 
     IntervalSet free_blocks;
 
+    // When locked, Free calls will be ignored, while Allocate calls will hit an assert. A memory
+    // region locks itself after deserialization.
+    bool is_locked{};
+
     /**
      * Reset the allocator state
      * @param base The base offset the beginning of FCRAM.
@@ -63,6 +67,11 @@ struct MemoryRegionInfo {
      */
     void Free(u32 offset, u32 size);
 
+    /**
+     * Unlock the MemoryRegion. Used after loading is completed.
+     */
+    void Unlock();
+
 private:
     friend class boost::serialization::access;
     template <class Archive>
@@ -71,6 +80,9 @@ private:
         ar& size;
         ar& used;
         ar& free_blocks;
+        if (Archive::is_loading::value) {
+            is_locked = true;
+        }
     }
 };
 

src/core/hle/kernel/server_session.cpp

@@ -85,7 +85,8 @@ ResultCode ServerSession::HandleSyncRequest(std::shared_ptr<Thread> thread) {
     // If this ServerSession has an associated HLE handler, forward the request to it.
     if (hle_handler != nullptr) {
         std::array<u32_le, IPC::COMMAND_BUFFER_LENGTH + 2 * IPC::MAX_STATIC_BUFFERS> cmd_buf;
-        auto current_process = thread->owner_process;
+        auto current_process = thread->owner_process.lock();
+        ASSERT(current_process);
         kernel.memory.ReadBlock(*current_process, thread->GetCommandBufferAddress(), cmd_buf.data(),
                                 cmd_buf.size() * sizeof(u32));
 

src/core/hle/kernel/svc.cpp

@@ -283,7 +283,7 @@ void SVC::ExitProcess() {
     // Stop all the process threads that are currently waiting for objects.
     auto& thread_list = kernel.GetCurrentThreadManager().GetThreadList();
     for (auto& thread : thread_list) {
-        if (thread->owner_process != current_process)
+        if (thread->owner_process.lock() != current_process)
             continue;
 
         if (thread.get() == kernel.GetCurrentThreadManager().GetCurrentThread())
@@ -1041,8 +1041,7 @@ ResultCode SVC::GetProcessIdOfThread(u32* process_id, Handle thread_handle) {
     if (thread == nullptr)
         return ERR_INVALID_HANDLE;
 
-    const std::shared_ptr<Process> process = thread->owner_process;
-
+    const std::shared_ptr<Process> process = thread->owner_process.lock();
     ASSERT_MSG(process != nullptr, "Invalid parent process for thread={:#010X}", thread_handle);
 
     *process_id = process->process_id;

src/core/hle/kernel/thread.cpp

@@ -45,7 +45,16 @@ void Thread::serialize(Archive& ar, const unsigned int file_version) {
     ar& tls_address;
     ar& held_mutexes;
     ar& pending_mutexes;
-    ar& owner_process;
+
+    // Note: this is equivalent of what is done in boost/serialization/weak_ptr.hpp, but it's
+    // compatible with previous versions of savestates.
+    // TODO(SaveStates): When the savestate version is bumped, simplify this again.
+    std::shared_ptr<Process> shared_owner_process = owner_process.lock();
+    ar& shared_owner_process;
+    if (Archive::is_loading::value) {
+        owner_process = shared_owner_process;
+    }
+
     ar& wait_objects;
     ar& wait_address;
     ar& name;
@@ -99,7 +108,8 @@ void Thread::Stop() {
     u32 tls_page = (tls_address - Memory::TLS_AREA_VADDR) / Memory::PAGE_SIZE;
     u32 tls_slot =
         ((tls_address - Memory::TLS_AREA_VADDR) % Memory::PAGE_SIZE) / Memory::TLS_ENTRY_SIZE;
-    owner_process->tls_slots[tls_page].reset(tls_slot);
+    ASSERT(owner_process.lock());
+    owner_process.lock()->tls_slots[tls_page].reset(tls_slot);
 }
 
 void ThreadManager::SwitchContext(Thread* new_thread) {
@@ -110,7 +120,7 @@ void ThreadManager::SwitchContext(Thread* new_thread) {
 
     // Save context for previous thread
     if (previous_thread) {
-        previous_process = previous_thread->owner_process;
+        previous_process = previous_thread->owner_process.lock();
         previous_thread->last_running_ticks = cpu->GetTimer().GetTicks();
         cpu->SaveContext(previous_thread->context);
 
@@ -135,8 +145,9 @@ void ThreadManager::SwitchContext(Thread* new_thread) {
         ready_queue.remove(new_thread->current_priority, new_thread);
         new_thread->status = ThreadStatus::Running;
 
-        if (previous_process != current_thread->owner_process) {
-            kernel.SetCurrentProcessForCPU(current_thread->owner_process, cpu->GetID());
+        ASSERT(current_thread->owner_process.lock());
+        if (previous_process != current_thread->owner_process.lock()) {
+            kernel.SetCurrentProcessForCPU(current_thread->owner_process.lock(), cpu->GetID());
         }
 
         cpu->LoadContext(new_thread->context);

src/core/hle/kernel/thread.h

@@ -308,7 +308,7 @@ public:
     /// Mutexes that this thread is currently waiting for.
     boost::container::flat_set<std::shared_ptr<Mutex>> pending_mutexes{};
 
-    std::shared_ptr<Process> owner_process{}; ///< Process that owns this thread
+    std::weak_ptr<Process> owner_process{}; ///< Process that owns this thread
 
     /// Objects that the thread is waiting on, in the same order as they were
     // passed to WaitSynchronization1/N.

src/core/hle/kernel/vm_manager.cpp

@@ -45,6 +45,8 @@ VMManager::VMManager(Memory::MemorySystem& memory)
 VMManager::~VMManager() = default;
 
 void VMManager::Reset() {
+    ASSERT(!is_locked);
+
     vma_map.clear();
 
     // Initialize the map with a single free region covering the entire managed space.
@@ -67,6 +69,7 @@ VMManager::VMAHandle VMManager::FindVMA(VAddr target) const {
 
 ResultVal<VAddr> VMManager::MapBackingMemoryToBase(VAddr base, u32 region_size, MemoryRef memory,
                                                    u32 size, MemoryState state) {
+    ASSERT(!is_locked);
 
     // Find the first Free VMA.
     VMAHandle vma_handle = std::find_if(vma_map.begin(), vma_map.end(), [&](const auto& vma) {
@@ -96,6 +99,7 @@ ResultVal<VAddr> VMManager::MapBackingMemoryToBase(VAddr base, u32 region_size,
 
 ResultVal<VMManager::VMAHandle> VMManager::MapBackingMemory(VAddr target, MemoryRef memory,
                                                             u32 size, MemoryState state) {
+    ASSERT(!is_locked);
     ASSERT(memory.GetPtr() != nullptr);
 
     // This is the appropriately sized VMA that will turn into our allocation.
@@ -115,6 +119,8 @@ ResultVal<VMManager::VMAHandle> VMManager::MapBackingMemory(VAddr target, Memory
 ResultVal<VMManager::VMAHandle> VMManager::MapMMIO(VAddr target, PAddr paddr, u32 size,
                                                    MemoryState state,
                                                    Memory::MMIORegionPointer mmio_handler) {
+    ASSERT(!is_locked);
+
     // This is the appropriately sized VMA that will turn into our allocation.
     CASCADE_RESULT(VMAIter vma_handle, CarveVMA(target, size));
     VirtualMemoryArea& final_vma = vma_handle->second;
@@ -133,6 +139,10 @@ ResultVal<VMManager::VMAHandle> VMManager::MapMMIO(VAddr target, PAddr paddr, u3
 ResultCode VMManager::ChangeMemoryState(VAddr target, u32 size, MemoryState expected_state,
                                         VMAPermission expected_perms, MemoryState new_state,
                                         VMAPermission new_perms) {
+    if (is_locked) {
+        return RESULT_SUCCESS;
+    }
+
     VAddr target_end = target + size;
     VMAIter begin_vma = StripIterConstness(FindVMA(target));
     VMAIter i_end = vma_map.lower_bound(target_end);
@@ -167,6 +177,8 @@ ResultCode VMManager::ChangeMemoryState(VAddr target, u32 size, MemoryState expe
 }
 
 VMManager::VMAIter VMManager::Unmap(VMAIter vma_handle) {
+    ASSERT(!is_locked);
+
     VirtualMemoryArea& vma = vma_handle->second;
     vma.type = VMAType::Free;
     vma.permissions = VMAPermission::None;
@@ -181,6 +193,8 @@ VMManager::VMAIter VMManager::Unmap(VMAIter vma_handle) {
 }
 
 ResultCode VMManager::UnmapRange(VAddr target, u32 size) {
+    ASSERT(!is_locked);
+
     CASCADE_RESULT(VMAIter vma, CarveVMARange(target, size));
     const VAddr target_end = target + size;
 
@@ -196,6 +210,8 @@ ResultCode VMManager::UnmapRange(VAddr target, u32 size) {
 }
 
 VMManager::VMAHandle VMManager::Reprotect(VMAHandle vma_handle, VMAPermission new_perms) {
+    ASSERT(!is_locked);
+
     VMAIter iter = StripIterConstness(vma_handle);
 
     VirtualMemoryArea& vma = iter->second;
@@ -206,6 +222,8 @@ VMManager::VMAHandle VMManager::Reprotect(VMAHandle vma_handle, VMAPermission ne
 }
 
 ResultCode VMManager::ReprotectRange(VAddr target, u32 size, VMAPermission new_perms) {
+    ASSERT(!is_locked);
+
     CASCADE_RESULT(VMAIter vma, CarveVMARange(target, size));
     const VAddr target_end = target + size;
 
@@ -231,6 +249,10 @@ void VMManager::LogLayout(Log::Level log_level) const {
     }
 }
 
+void VMManager::Unlock() {
+    is_locked = false;
+}
+
 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

@@ -212,6 +212,11 @@ public:
     /// is scheduled.
     std::shared_ptr<Memory::PageTable> page_table;
 
+    /**
+     * Unlock the VMManager. Used after loading is completed.
+     */
+    void Unlock();
+
 private:
     using VMAIter = decltype(vma_map)::iterator;
 
@@ -250,10 +255,17 @@ private:
 
     Memory::MemorySystem& memory;
 
+    // When locked, ChangeMemoryState calls will be ignored, other modification calls will hit an
+    // assert. VMManager locks itself after deserialization.
+    bool is_locked{};
+
     template <class Archive>
     void serialize(Archive& ar, const unsigned int) {
         ar& vma_map;
         ar& page_table;
+        if (Archive::is_loading::value) {
+            is_locked = true;
+        }
     }
     friend class boost::serialization::access;
 };