From f2fed21c1139c8d5c030bc5caee5c612dfe7979f Mon Sep 17 00:00:00 2001
From: Liam <byteslice@airmail.cc>
Date: Sun, 7 Jan 2024 13:59:48 -0500
Subject: kernel: fix page leak on process termination

---
 src/core/hle/kernel/k_page_table_base.cpp | 75 ++++++++++++++++++++++++++++++-
 src/core/hle/kernel/k_page_table_base.h   |  1 +
 src/core/hle/kernel/k_process.cpp         |  6 +++
 3 files changed, 81 insertions(+), 1 deletion(-)

(limited to 'src/core/hle/kernel')

diff --git a/src/core/hle/kernel/k_page_table_base.cpp b/src/core/hle/kernel/k_page_table_base.cpp
index 73fbda331..f01eaa164 100644
--- a/src/core/hle/kernel/k_page_table_base.cpp
+++ b/src/core/hle/kernel/k_page_table_base.cpp
@@ -431,9 +431,82 @@ Result KPageTableBase::InitializeForProcess(Svc::CreateProcessFlag as_type, bool
                                                m_memory_block_slab_manager));
 }
 
+Result KPageTableBase::FinalizeProcess() {
+    // Only process tables should be finalized.
+    ASSERT(!this->IsKernel());
+
+    // HLE processes don't have memory mapped.
+    R_SUCCEED_IF(m_impl == nullptr);
+
+    // NOTE: Here Nintendo calls an unknown OnFinalize function.
+    // this->OnFinalize();
+
+    // NOTE: Here Nintendo calls a second unknown OnFinalize function.
+    // this->OnFinalize2();
+
+    // Get implementation objects.
+    auto& impl = this->GetImpl();
+    auto& mm = m_kernel.MemoryManager();
+
+    // Traverse, freeing all pages.
+    {
+        // Get the address space size.
+        const size_t as_size = this->GetAddressSpaceSize();
+
+        // Begin the traversal.
+        TraversalContext context;
+        TraversalEntry cur_entry = {
+            .phys_addr = 0,
+            .block_size = 0,
+        };
+
+        bool cur_valid = false;
+        TraversalEntry next_entry;
+        bool next_valid;
+        size_t tot_size = 0;
+
+        next_valid = impl.BeginTraversal(std::addressof(next_entry), std::addressof(context),
+                                         this->GetAddressSpaceStart());
+
+        // Iterate over entries.
+        while (true) {
+            if ((!next_valid && !cur_valid) ||
+                (next_valid && cur_valid &&
+                 next_entry.phys_addr == cur_entry.phys_addr + cur_entry.block_size)) {
+                cur_entry.block_size += next_entry.block_size;
+            } else {
+                if (cur_valid && IsHeapPhysicalAddressForFinalize(cur_entry.phys_addr)) {
+                    mm.Close(cur_entry.phys_addr, cur_entry.block_size / PageSize);
+                }
+
+                // Update tracking variables.
+                tot_size += cur_entry.block_size;
+                cur_entry = next_entry;
+                cur_valid = next_valid;
+            }
+
+            if (cur_entry.block_size + tot_size >= as_size) {
+                break;
+            }
+
+            next_valid =
+                impl.ContinueTraversal(std::addressof(next_entry), std::addressof(context));
+        }
+
+        // Handle the last block.
+        if (cur_valid && IsHeapPhysicalAddressForFinalize(cur_entry.phys_addr)) {
+            mm.Close(cur_entry.phys_addr, cur_entry.block_size / PageSize);
+        }
+    }
+
+    R_SUCCEED();
+}
+
 void KPageTableBase::Finalize() {
+    this->FinalizeProcess();
+
     auto HostUnmapCallback = [&](KProcessAddress addr, u64 size) {
-        if (Settings::IsFastmemEnabled()) {
+        if (m_impl->fastmem_arena) {
             m_system.DeviceMemory().buffer.Unmap(GetInteger(addr), size, false);
         }
     };
diff --git a/src/core/hle/kernel/k_page_table_base.h b/src/core/hle/kernel/k_page_table_base.h
index 077cafc96..748419f86 100644
--- a/src/core/hle/kernel/k_page_table_base.h
+++ b/src/core/hle/kernel/k_page_table_base.h
@@ -241,6 +241,7 @@ public:
                                 KResourceLimit* resource_limit, Core::Memory::Memory& memory,
                                 KProcessAddress aslr_space_start);
 
+    Result FinalizeProcess();
     void Finalize();
 
     bool IsKernel() const {
diff --git a/src/core/hle/kernel/k_process.cpp b/src/core/hle/kernel/k_process.cpp
index 068e71dff..ae332a550 100644
--- a/src/core/hle/kernel/k_process.cpp
+++ b/src/core/hle/kernel/k_process.cpp
@@ -171,6 +171,12 @@ void KProcess::Finalize() {
         m_resource_limit->Close();
     }
 
+    // Clear expensive resources, as the destructor is not called for guest objects.
+    for (auto& interface : m_arm_interfaces) {
+        interface.reset();
+    }
+    m_exclusive_monitor.reset();
+
     // Perform inherited finalization.
     KSynchronizationObject::Finalize();
 }
-- 
cgit v1.2.3


From f90a022d3a20c86399f49a8154847b73bc1b8fd3 Mon Sep 17 00:00:00 2001
From: Liam <byteslice@airmail.cc>
Date: Tue, 2 Jan 2024 17:12:16 -0500
Subject: kernel: fix debugger and process list lifetime

---
 src/core/debugger/debugger.cpp          | 39 +++++++++++----
 src/core/debugger/gdbstub.cpp           | 66 +++++++++++++------------
 src/core/debugger/gdbstub.h             | 15 +++++-
 src/core/hle/kernel/kernel.cpp          | 31 ++++++++++--
 src/core/hle/kernel/kernel.h            |  6 ++-
 src/core/hle/kernel/svc/svc_process.cpp |  8 ++--
 src/core/hle/service/glue/arp.cpp       |  7 +--
 src/core/hle/service/hid/hid.cpp        | 10 ++--
 src/core/hle/service/pm/pm.cpp          | 85 ++++++++++++++-------------------
 9 files changed, 160 insertions(+), 107 deletions(-)

(limited to 'src/core/hle/kernel')

diff --git a/src/core/debugger/debugger.cpp b/src/core/debugger/debugger.cpp
index 0e270eb50..e86aae846 100644
--- a/src/core/debugger/debugger.cpp
+++ b/src/core/debugger/debugger.cpp
@@ -114,7 +114,7 @@ public:
     }
 
     Kernel::KThread* GetActiveThread() override {
-        return state->active_thread;
+        return state->active_thread.GetPointerUnsafe();
     }
 
 private:
@@ -147,11 +147,14 @@ private:
 
         std::scoped_lock lk{connection_lock};
 
+        // Find the process we are going to debug.
+        SetDebugProcess();
+
         // Ensure everything is stopped.
         PauseEmulation();
 
         // Set up the new frontend.
-        frontend = std::make_unique<GDBStub>(*this, system);
+        frontend = std::make_unique<GDBStub>(*this, system, debug_process.GetPointerUnsafe());
 
         // Set the new state. This will tear down any existing state.
         state = ConnectionState{
@@ -194,15 +197,20 @@ private:
             UpdateActiveThread();
 
             if (state->info.type == SignalType::Watchpoint) {
-                frontend->Watchpoint(state->active_thread, *state->info.watchpoint);
+                frontend->Watchpoint(std::addressof(*state->active_thread),
+                                     *state->info.watchpoint);
             } else {
-                frontend->Stopped(state->active_thread);
+                frontend->Stopped(std::addressof(*state->active_thread));
             }
 
             break;
         case SignalType::ShuttingDown:
             frontend->ShuttingDown();
 
+            // Release members.
+            state->active_thread.Reset(nullptr);
+            debug_process.Reset(nullptr);
+
             // Wait for emulation to shut down gracefully now.
             state->signal_pipe.close();
             state->client_socket.shutdown(boost::asio::socket_base::shutdown_both);
@@ -222,7 +230,7 @@ private:
                 stopped = true;
                 PauseEmulation();
                 UpdateActiveThread();
-                frontend->Stopped(state->active_thread);
+                frontend->Stopped(state->active_thread.GetPointerUnsafe());
                 break;
             }
             case DebuggerAction::Continue:
@@ -232,7 +240,7 @@ private:
                 MarkResumed([&] {
                     state->active_thread->SetStepState(Kernel::StepState::StepPending);
                     state->active_thread->Resume(Kernel::SuspendType::Debug);
-                    ResumeEmulation(state->active_thread);
+                    ResumeEmulation(state->active_thread.GetPointerUnsafe());
                 });
                 break;
             case DebuggerAction::StepThreadLocked: {
@@ -255,6 +263,7 @@ private:
     }
 
     void PauseEmulation() {
+        Kernel::KScopedLightLock ll{debug_process->GetListLock()};
         Kernel::KScopedSchedulerLock sl{system.Kernel()};
 
         // Put all threads to sleep on next scheduler round.
@@ -264,6 +273,9 @@ private:
     }
 
     void ResumeEmulation(Kernel::KThread* except = nullptr) {
+        Kernel::KScopedLightLock ll{debug_process->GetListLock()};
+        Kernel::KScopedSchedulerLock sl{system.Kernel()};
+
         // Wake up all threads.
         for (auto& thread : ThreadList()) {
             if (std::addressof(thread) == except) {
@@ -277,15 +289,16 @@ private:
 
     template <typename Callback>
     void MarkResumed(Callback&& cb) {
-        Kernel::KScopedSchedulerLock sl{system.Kernel()};
         stopped = false;
         cb();
     }
 
     void UpdateActiveThread() {
+        Kernel::KScopedLightLock ll{debug_process->GetListLock()};
+
         auto& threads{ThreadList()};
         for (auto& thread : threads) {
-            if (std::addressof(thread) == state->active_thread) {
+            if (std::addressof(thread) == state->active_thread.GetPointerUnsafe()) {
                 // Thread is still alive, no need to update.
                 return;
             }
@@ -293,12 +306,18 @@ private:
         state->active_thread = std::addressof(threads.front());
     }
 
+private:
+    void SetDebugProcess() {
+        debug_process = std::move(system.Kernel().GetProcessList().back());
+    }
+
     Kernel::KProcess::ThreadList& ThreadList() {
-        return system.ApplicationProcess()->GetThreadList();
+        return debug_process->GetThreadList();
     }
 
 private:
     System& system;
+    Kernel::KScopedAutoObject<Kernel::KProcess> debug_process;
     std::unique_ptr<DebuggerFrontend> frontend;
 
     boost::asio::io_context io_context;
@@ -310,7 +329,7 @@ private:
         boost::process::async_pipe signal_pipe;
 
         SignalInfo info;
-        Kernel::KThread* active_thread;
+        Kernel::KScopedAutoObject<Kernel::KThread> active_thread;
         std::array<u8, 4096> client_data;
         bool pipe_data;
     };
diff --git a/src/core/debugger/gdbstub.cpp b/src/core/debugger/gdbstub.cpp
index 4051ed4af..80091cc7e 100644
--- a/src/core/debugger/gdbstub.cpp
+++ b/src/core/debugger/gdbstub.cpp
@@ -108,9 +108,9 @@ static std::string EscapeXML(std::string_view data) {
     return escaped;
 }
 
-GDBStub::GDBStub(DebuggerBackend& backend_, Core::System& system_)
-    : DebuggerFrontend(backend_), system{system_} {
-    if (system.ApplicationProcess()->Is64Bit()) {
+GDBStub::GDBStub(DebuggerBackend& backend_, Core::System& system_, Kernel::KProcess* debug_process_)
+    : DebuggerFrontend(backend_), system{system_}, debug_process{debug_process_} {
+    if (GetProcess()->Is64Bit()) {
         arch = std::make_unique<GDBStubA64>();
     } else {
         arch = std::make_unique<GDBStubA32>();
@@ -276,7 +276,7 @@ void GDBStub::ExecuteCommand(std::string_view packet, std::vector<DebuggerAction
         const size_t size{static_cast<size_t>(strtoll(command.data() + sep, nullptr, 16))};
 
         std::vector<u8> mem(size);
-        if (system.ApplicationMemory().ReadBlock(addr, mem.data(), size)) {
+        if (GetMemory().ReadBlock(addr, mem.data(), size)) {
             // Restore any bytes belonging to replaced instructions.
             auto it = replaced_instructions.lower_bound(addr);
             for (; it != replaced_instructions.end() && it->first < addr + size; it++) {
@@ -310,8 +310,8 @@ void GDBStub::ExecuteCommand(std::string_view packet, std::vector<DebuggerAction
         const auto mem_substr{std::string_view(command).substr(mem_sep)};
         const auto mem{Common::HexStringToVector(mem_substr, false)};
 
-        if (system.ApplicationMemory().WriteBlock(addr, mem.data(), size)) {
-            Core::InvalidateInstructionCacheRange(system.ApplicationProcess(), addr, size);
+        if (GetMemory().WriteBlock(addr, mem.data(), size)) {
+            Core::InvalidateInstructionCacheRange(GetProcess(), addr, size);
             SendReply(GDB_STUB_REPLY_OK);
         } else {
             SendReply(GDB_STUB_REPLY_ERR);
@@ -353,7 +353,7 @@ void GDBStub::HandleBreakpointInsert(std::string_view command) {
     const size_t addr{static_cast<size_t>(strtoll(command.data() + addr_sep, nullptr, 16))};
     const size_t size{static_cast<size_t>(strtoll(command.data() + size_sep, nullptr, 16))};
 
-    if (!system.ApplicationMemory().IsValidVirtualAddressRange(addr, size)) {
+    if (!GetMemory().IsValidVirtualAddressRange(addr, size)) {
         SendReply(GDB_STUB_REPLY_ERR);
         return;
     }
@@ -362,22 +362,20 @@ void GDBStub::HandleBreakpointInsert(std::string_view command) {
 
     switch (type) {
     case BreakpointType::Software:
-        replaced_instructions[addr] = system.ApplicationMemory().Read32(addr);
-        system.ApplicationMemory().Write32(addr, arch->BreakpointInstruction());
-        Core::InvalidateInstructionCacheRange(system.ApplicationProcess(), addr, sizeof(u32));
+        replaced_instructions[addr] = GetMemory().Read32(addr);
+        GetMemory().Write32(addr, arch->BreakpointInstruction());
+        Core::InvalidateInstructionCacheRange(GetProcess(), addr, sizeof(u32));
         success = true;
         break;
     case BreakpointType::WriteWatch:
-        success = system.ApplicationProcess()->InsertWatchpoint(addr, size,
-                                                                Kernel::DebugWatchpointType::Write);
+        success = GetProcess()->InsertWatchpoint(addr, size, Kernel::DebugWatchpointType::Write);
         break;
     case BreakpointType::ReadWatch:
-        success = system.ApplicationProcess()->InsertWatchpoint(addr, size,
-                                                                Kernel::DebugWatchpointType::Read);
+        success = GetProcess()->InsertWatchpoint(addr, size, Kernel::DebugWatchpointType::Read);
         break;
     case BreakpointType::AccessWatch:
-        success = system.ApplicationProcess()->InsertWatchpoint(
-            addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
+        success =
+            GetProcess()->InsertWatchpoint(addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
         break;
     case BreakpointType::Hardware:
     default:
@@ -400,7 +398,7 @@ void GDBStub::HandleBreakpointRemove(std::string_view command) {
     const size_t addr{static_cast<size_t>(strtoll(command.data() + addr_sep, nullptr, 16))};
     const size_t size{static_cast<size_t>(strtoll(command.data() + size_sep, nullptr, 16))};
 
-    if (!system.ApplicationMemory().IsValidVirtualAddressRange(addr, size)) {
+    if (!GetMemory().IsValidVirtualAddressRange(addr, size)) {
         SendReply(GDB_STUB_REPLY_ERR);
         return;
     }
@@ -411,24 +409,22 @@ void GDBStub::HandleBreakpointRemove(std::string_view command) {
     case BreakpointType::Software: {
         const auto orig_insn{replaced_instructions.find(addr)};
         if (orig_insn != replaced_instructions.end()) {
-            system.ApplicationMemory().Write32(addr, orig_insn->second);
-            Core::InvalidateInstructionCacheRange(system.ApplicationProcess(), addr, sizeof(u32));
+            GetMemory().Write32(addr, orig_insn->second);
+            Core::InvalidateInstructionCacheRange(GetProcess(), addr, sizeof(u32));
             replaced_instructions.erase(addr);
             success = true;
         }
         break;
     }
     case BreakpointType::WriteWatch:
-        success = system.ApplicationProcess()->RemoveWatchpoint(addr, size,
-                                                                Kernel::DebugWatchpointType::Write);
+        success = GetProcess()->RemoveWatchpoint(addr, size, Kernel::DebugWatchpointType::Write);
         break;
     case BreakpointType::ReadWatch:
-        success = system.ApplicationProcess()->RemoveWatchpoint(addr, size,
-                                                                Kernel::DebugWatchpointType::Read);
+        success = GetProcess()->RemoveWatchpoint(addr, size, Kernel::DebugWatchpointType::Read);
         break;
     case BreakpointType::AccessWatch:
-        success = system.ApplicationProcess()->RemoveWatchpoint(
-            addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
+        success =
+            GetProcess()->RemoveWatchpoint(addr, size, Kernel::DebugWatchpointType::ReadOrWrite);
         break;
     case BreakpointType::Hardware:
     default:
@@ -466,10 +462,10 @@ void GDBStub::HandleQuery(std::string_view command) {
         const auto target_xml{arch->GetTargetXML()};
         SendReply(PaginateBuffer(target_xml, command.substr(30)));
     } else if (command.starts_with("Offsets")) {
-        const auto main_offset = Core::FindMainModuleEntrypoint(system.ApplicationProcess());
+        const auto main_offset = Core::FindMainModuleEntrypoint(GetProcess());
         SendReply(fmt::format("TextSeg={:x}", GetInteger(main_offset)));
     } else if (command.starts_with("Xfer:libraries:read::")) {
-        auto modules = Core::FindModules(system.ApplicationProcess());
+        auto modules = Core::FindModules(GetProcess());
 
         std::string buffer;
         buffer += R"(<?xml version="1.0"?>)";
@@ -483,7 +479,7 @@ void GDBStub::HandleQuery(std::string_view command) {
         SendReply(PaginateBuffer(buffer, command.substr(21)));
     } else if (command.starts_with("fThreadInfo")) {
         // beginning of list
-        const auto& threads = system.ApplicationProcess()->GetThreadList();
+        const auto& threads = GetProcess()->GetThreadList();
         std::vector<std::string> thread_ids;
         for (const auto& thread : threads) {
             thread_ids.push_back(fmt::format("{:x}", thread.GetThreadId()));
@@ -497,7 +493,7 @@ void GDBStub::HandleQuery(std::string_view command) {
         buffer += R"(<?xml version="1.0"?>)";
         buffer += "<threads>";
 
-        const auto& threads = system.ApplicationProcess()->GetThreadList();
+        const auto& threads = GetProcess()->GetThreadList();
         for (const auto& thread : threads) {
             auto thread_name{Core::GetThreadName(&thread)};
             if (!thread_name) {
@@ -613,7 +609,7 @@ void GDBStub::HandleRcmd(const std::vector<u8>& command) {
     std::string_view command_str{reinterpret_cast<const char*>(&command[0]), command.size()};
     std::string reply;
 
-    auto* process = system.ApplicationProcess();
+    auto* process = GetProcess();
     auto& page_table = process->GetPageTable();
 
     const char* commands = "Commands:\n"
@@ -714,7 +710,7 @@ void GDBStub::HandleRcmd(const std::vector<u8>& command) {
 }
 
 Kernel::KThread* GDBStub::GetThreadByID(u64 thread_id) {
-    auto& threads{system.ApplicationProcess()->GetThreadList()};
+    auto& threads{GetProcess()->GetThreadList()};
     for (auto& thread : threads) {
         if (thread.GetThreadId() == thread_id) {
             return std::addressof(thread);
@@ -783,4 +779,12 @@ void GDBStub::SendStatus(char status) {
     backend.WriteToClient(buf);
 }
 
+Kernel::KProcess* GDBStub::GetProcess() {
+    return debug_process;
+}
+
+Core::Memory::Memory& GDBStub::GetMemory() {
+    return GetProcess()->GetMemory();
+}
+
 } // namespace Core
diff --git a/src/core/debugger/gdbstub.h b/src/core/debugger/gdbstub.h
index 368197920..232dcf49f 100644
--- a/src/core/debugger/gdbstub.h
+++ b/src/core/debugger/gdbstub.h
@@ -12,13 +12,22 @@
 #include "core/debugger/debugger_interface.h"
 #include "core/debugger/gdbstub_arch.h"
 
+namespace Kernel {
+class KProcess;
+}
+
+namespace Core::Memory {
+class Memory;
+}
+
 namespace Core {
 
 class System;
 
 class GDBStub : public DebuggerFrontend {
 public:
-    explicit GDBStub(DebuggerBackend& backend, Core::System& system);
+    explicit GDBStub(DebuggerBackend& backend, Core::System& system,
+                     Kernel::KProcess* debug_process);
     ~GDBStub() override;
 
     void Connected() override;
@@ -42,8 +51,12 @@ private:
     void SendReply(std::string_view data);
     void SendStatus(char status);
 
+    Kernel::KProcess* GetProcess();
+    Core::Memory::Memory& GetMemory();
+
 private:
     Core::System& system;
+    Kernel::KProcess* debug_process;
     std::unique_ptr<GDBStubArch> arch;
     std::vector<char> current_command;
     std::map<VAddr, u32> replaced_instructions;
diff --git a/src/core/hle/kernel/kernel.cpp b/src/core/hle/kernel/kernel.cpp
index 1030f0c12..f3683cdcc 100644
--- a/src/core/hle/kernel/kernel.cpp
+++ b/src/core/hle/kernel/kernel.cpp
@@ -112,7 +112,14 @@ struct KernelCore::Impl {
             old_process->Close();
         }
 
-        process_list.clear();
+        {
+            std::scoped_lock lk{process_list_lock};
+            for (auto* const process : process_list) {
+                process->Terminate();
+                process->Close();
+            }
+            process_list.clear();
+        }
 
         next_object_id = 0;
         next_kernel_process_id = KProcess::InitialProcessIdMin;
@@ -770,6 +777,7 @@ struct KernelCore::Impl {
     std::atomic<u64> next_thread_id{1};
 
     // Lists all processes that exist in the current session.
+    std::mutex process_list_lock;
     std::vector<KProcess*> process_list;
     std::atomic<KProcess*> application_process{};
     std::unique_ptr<Kernel::GlobalSchedulerContext> global_scheduler_context;
@@ -869,9 +877,19 @@ KResourceLimit* KernelCore::GetSystemResourceLimit() {
 }
 
 void KernelCore::AppendNewProcess(KProcess* process) {
+    process->Open();
+
+    std::scoped_lock lk{impl->process_list_lock};
     impl->process_list.push_back(process);
 }
 
+void KernelCore::RemoveProcess(KProcess* process) {
+    std::scoped_lock lk{impl->process_list_lock};
+    if (std::erase(impl->process_list, process)) {
+        process->Close();
+    }
+}
+
 void KernelCore::MakeApplicationProcess(KProcess* process) {
     impl->MakeApplicationProcess(process);
 }
@@ -884,8 +902,15 @@ const KProcess* KernelCore::ApplicationProcess() const {
     return impl->application_process;
 }
 
-const std::vector<KProcess*>& KernelCore::GetProcessList() const {
-    return impl->process_list;
+std::list<KScopedAutoObject<KProcess>> KernelCore::GetProcessList() {
+    std::list<KScopedAutoObject<KProcess>> processes;
+    std::scoped_lock lk{impl->process_list_lock};
+
+    for (auto* const process : impl->process_list) {
+        processes.emplace_back(process);
+    }
+
+    return processes;
 }
 
 Kernel::GlobalSchedulerContext& KernelCore::GlobalSchedulerContext() {
diff --git a/src/core/hle/kernel/kernel.h b/src/core/hle/kernel/kernel.h
index 5d4102145..8ea5bed1c 100644
--- a/src/core/hle/kernel/kernel.h
+++ b/src/core/hle/kernel/kernel.h
@@ -5,6 +5,7 @@
 
 #include <array>
 #include <functional>
+#include <list>
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -116,8 +117,9 @@ public:
     /// Retrieves a shared pointer to the system resource limit instance.
     KResourceLimit* GetSystemResourceLimit();
 
-    /// Adds the given shared pointer to an internal list of active processes.
+    /// Adds/removes the given pointer to an internal list of active processes.
     void AppendNewProcess(KProcess* process);
+    void RemoveProcess(KProcess* process);
 
     /// Makes the given process the new application process.
     void MakeApplicationProcess(KProcess* process);
@@ -129,7 +131,7 @@ public:
     const KProcess* ApplicationProcess() const;
 
     /// Retrieves the list of processes.
-    const std::vector<KProcess*>& GetProcessList() const;
+    std::list<KScopedAutoObject<KProcess>> GetProcessList();
 
     /// Gets the sole instance of the global scheduler
     Kernel::GlobalSchedulerContext& GlobalSchedulerContext();
diff --git a/src/core/hle/kernel/svc/svc_process.cpp b/src/core/hle/kernel/svc/svc_process.cpp
index caa8bee9a..5c3e8829f 100644
--- a/src/core/hle/kernel/svc/svc_process.cpp
+++ b/src/core/hle/kernel/svc/svc_process.cpp
@@ -74,13 +74,15 @@ Result GetProcessList(Core::System& system, s32* out_num_processes, u64 out_proc
     }
 
     auto& memory = GetCurrentMemory(kernel);
-    const auto& process_list = kernel.GetProcessList();
+    auto process_list = kernel.GetProcessList();
+    auto it = process_list.begin();
+
     const auto num_processes = process_list.size();
     const auto copy_amount =
         std::min(static_cast<std::size_t>(out_process_ids_size), num_processes);
 
-    for (std::size_t i = 0; i < copy_amount; ++i) {
-        memory.Write64(out_process_ids, process_list[i]->GetProcessId());
+    for (std::size_t i = 0; i < copy_amount && it != process_list.end(); ++i, ++it) {
+        memory.Write64(out_process_ids, (*it)->GetProcessId());
         out_process_ids += sizeof(u64);
     }
 
diff --git a/src/core/hle/service/glue/arp.cpp b/src/core/hle/service/glue/arp.cpp
index 6f1151b03..1254b6d49 100644
--- a/src/core/hle/service/glue/arp.cpp
+++ b/src/core/hle/service/glue/arp.cpp
@@ -15,9 +15,10 @@
 namespace Service::Glue {
 
 namespace {
-std::optional<u64> GetTitleIDForProcessID(const Core::System& system, u64 process_id) {
-    const auto& list = system.Kernel().GetProcessList();
-    const auto iter = std::find_if(list.begin(), list.end(), [&process_id](const auto& process) {
+std::optional<u64> GetTitleIDForProcessID(Core::System& system, u64 process_id) {
+    auto list = system.Kernel().GetProcessList();
+
+    const auto iter = std::find_if(list.begin(), list.end(), [&process_id](auto& process) {
         return process->GetProcessId() == process_id;
     });
 
diff --git a/src/core/hle/service/hid/hid.cpp b/src/core/hle/service/hid/hid.cpp
index fc03a0a5f..4ce0a9834 100644
--- a/src/core/hle/service/hid/hid.cpp
+++ b/src/core/hle/service/hid/hid.cpp
@@ -22,12 +22,10 @@ void LoopProcess(Core::System& system) {
     std::shared_ptr<HidFirmwareSettings> firmware_settings =
         std::make_shared<HidFirmwareSettings>();
 
-    // TODO: Remove this hack until this service is emulated properly.
-    const auto process_list = system.Kernel().GetProcessList();
-    if (!process_list.empty()) {
-        resource_manager->Initialize();
-        resource_manager->RegisterAppletResourceUserId(process_list[0]->GetId(), true);
-    }
+    // TODO: Remove this hack when am is emulated properly.
+    resource_manager->Initialize();
+    resource_manager->RegisterAppletResourceUserId(system.ApplicationProcess()->GetProcessId(),
+                                                   true);
 
     server_manager->RegisterNamedService(
         "hid", std::make_shared<IHidServer>(system, resource_manager, firmware_settings));
diff --git a/src/core/hle/service/pm/pm.cpp b/src/core/hle/service/pm/pm.cpp
index d92499f05..b52468e41 100644
--- a/src/core/hle/service/pm/pm.cpp
+++ b/src/core/hle/service/pm/pm.cpp
@@ -22,27 +22,26 @@ constexpr Result ResultProcessNotFound{ErrorModule::PM, 1};
 
 constexpr u64 NO_PROCESS_FOUND_PID{0};
 
-std::optional<Kernel::KProcess*> SearchProcessList(
-    const std::vector<Kernel::KProcess*>& process_list,
-    std::function<bool(Kernel::KProcess*)> predicate) {
+using ProcessList = std::list<Kernel::KScopedAutoObject<Kernel::KProcess>>;
+
+template <typename F>
+Kernel::KScopedAutoObject<Kernel::KProcess> SearchProcessList(ProcessList& process_list,
+                                                              F&& predicate) {
     const auto iter = std::find_if(process_list.begin(), process_list.end(), predicate);
 
     if (iter == process_list.end()) {
-        return std::nullopt;
+        return nullptr;
     }
 
-    return *iter;
+    return iter->GetPointerUnsafe();
 }
 
-void GetApplicationPidGeneric(HLERequestContext& ctx,
-                              const std::vector<Kernel::KProcess*>& process_list) {
-    const auto process = SearchProcessList(process_list, [](const auto& proc) {
-        return proc->GetProcessId() == Kernel::KProcess::ProcessIdMin;
-    });
+void GetApplicationPidGeneric(HLERequestContext& ctx, ProcessList& process_list) {
+    auto process = SearchProcessList(process_list, [](auto& p) { return p->IsApplication(); });
 
     IPC::ResponseBuilder rb{ctx, 4};
     rb.Push(ResultSuccess);
-    rb.Push(process.has_value() ? (*process)->GetProcessId() : NO_PROCESS_FOUND_PID);
+    rb.Push(process.IsNull() ? NO_PROCESS_FOUND_PID : process->GetProcessId());
 }
 
 } // Anonymous namespace
@@ -80,8 +79,7 @@ private:
 
 class DebugMonitor final : public ServiceFramework<DebugMonitor> {
 public:
-    explicit DebugMonitor(Core::System& system_)
-        : ServiceFramework{system_, "pm:dmnt"}, kernel{system_.Kernel()} {
+    explicit DebugMonitor(Core::System& system_) : ServiceFramework{system_, "pm:dmnt"} {
         // clang-format off
         static const FunctionInfo functions[] = {
             {0, nullptr, "GetJitDebugProcessIdList"},
@@ -106,12 +104,11 @@ private:
 
         LOG_DEBUG(Service_PM, "called, program_id={:016X}", program_id);
 
-        const auto process =
-            SearchProcessList(kernel.GetProcessList(), [program_id](const auto& proc) {
-                return proc->GetProgramId() == program_id;
-            });
+        auto list = kernel.GetProcessList();
+        auto process = SearchProcessList(
+            list, [program_id](auto& p) { return p->GetProgramId() == program_id; });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -119,12 +116,13 @@ private:
 
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(ResultSuccess);
-        rb.Push((*process)->GetProcessId());
+        rb.Push(process->GetProcessId());
     }
 
     void GetApplicationProcessId(HLERequestContext& ctx) {
         LOG_DEBUG(Service_PM, "called");
-        GetApplicationPidGeneric(ctx, kernel.GetProcessList());
+        auto list = kernel.GetProcessList();
+        GetApplicationPidGeneric(ctx, list);
     }
 
     void AtmosphereGetProcessInfo(HLERequestContext& ctx) {
@@ -135,11 +133,10 @@ private:
 
         LOG_WARNING(Service_PM, "(Partial Implementation) called, pid={:016X}", pid);
 
-        const auto process = SearchProcessList(kernel.GetProcessList(), [pid](const auto& proc) {
-            return proc->GetProcessId() == pid;
-        });
+        auto list = kernel.GetProcessList();
+        auto process = SearchProcessList(list, [pid](auto& p) { return p->GetProcessId() == pid; });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -159,7 +156,7 @@ private:
 
         OverrideStatus override_status{};
         ProgramLocation program_location{
-            .program_id = (*process)->GetProgramId(),
+            .program_id = process->GetProgramId(),
             .storage_id = 0,
         };
 
@@ -169,14 +166,11 @@ private:
         rb.PushRaw(program_location);
         rb.PushRaw(override_status);
     }
-
-    const Kernel::KernelCore& kernel;
 };
 
 class Info final : public ServiceFramework<Info> {
 public:
-    explicit Info(Core::System& system_, const std::vector<Kernel::KProcess*>& process_list_)
-        : ServiceFramework{system_, "pm:info"}, process_list{process_list_} {
+    explicit Info(Core::System& system_) : ServiceFramework{system_, "pm:info"} {
         static const FunctionInfo functions[] = {
             {0, &Info::GetProgramId, "GetProgramId"},
             {65000, &Info::AtmosphereGetProcessId, "AtmosphereGetProcessId"},
@@ -193,11 +187,11 @@ private:
 
         LOG_DEBUG(Service_PM, "called, process_id={:016X}", process_id);
 
-        const auto process = SearchProcessList(process_list, [process_id](const auto& proc) {
-            return proc->GetProcessId() == process_id;
-        });
+        auto list = kernel.GetProcessList();
+        auto process = SearchProcessList(
+            list, [process_id](auto& p) { return p->GetProcessId() == process_id; });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -205,7 +199,7 @@ private:
 
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(ResultSuccess);
-        rb.Push((*process)->GetProgramId());
+        rb.Push(process->GetProgramId());
     }
 
     void AtmosphereGetProcessId(HLERequestContext& ctx) {
@@ -214,11 +208,11 @@ private:
 
         LOG_DEBUG(Service_PM, "called, program_id={:016X}", program_id);
 
-        const auto process = SearchProcessList(process_list, [program_id](const auto& proc) {
-            return proc->GetProgramId() == program_id;
-        });
+        auto list = system.Kernel().GetProcessList();
+        auto process = SearchProcessList(
+            list, [program_id](auto& p) { return p->GetProgramId() == program_id; });
 
-        if (!process.has_value()) {
+        if (process.IsNull()) {
             IPC::ResponseBuilder rb{ctx, 2};
             rb.Push(ResultProcessNotFound);
             return;
@@ -226,16 +220,13 @@ private:
 
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(ResultSuccess);
-        rb.Push((*process)->GetProcessId());
+        rb.Push(process->GetProcessId());
     }
-
-    const std::vector<Kernel::KProcess*>& process_list;
 };
 
 class Shell final : public ServiceFramework<Shell> {
 public:
-    explicit Shell(Core::System& system_)
-        : ServiceFramework{system_, "pm:shell"}, kernel{system_.Kernel()} {
+    explicit Shell(Core::System& system_) : ServiceFramework{system_, "pm:shell"} {
         // clang-format off
         static const FunctionInfo functions[] = {
             {0, nullptr, "LaunchProgram"},
@@ -257,10 +248,9 @@ public:
 private:
     void GetApplicationProcessIdForShell(HLERequestContext& ctx) {
         LOG_DEBUG(Service_PM, "called");
-        GetApplicationPidGeneric(ctx, kernel.GetProcessList());
+        auto list = kernel.GetProcessList();
+        GetApplicationPidGeneric(ctx, list);
     }
-
-    const Kernel::KernelCore& kernel;
 };
 
 void LoopProcess(Core::System& system) {
@@ -268,8 +258,7 @@ void LoopProcess(Core::System& system) {
 
     server_manager->RegisterNamedService("pm:bm", std::make_shared<BootMode>(system));
     server_manager->RegisterNamedService("pm:dmnt", std::make_shared<DebugMonitor>(system));
-    server_manager->RegisterNamedService(
-        "pm:info", std::make_shared<Info>(system, system.Kernel().GetProcessList()));
+    server_manager->RegisterNamedService("pm:info", std::make_shared<Info>(system));
     server_manager->RegisterNamedService("pm:shell", std::make_shared<Shell>(system));
     ServerManager::RunServer(std::move(server_manager));
 }
-- 
cgit v1.2.3


From 2f0b57ca13fb91730d7e210f6f4504357ef6cd0a Mon Sep 17 00:00:00 2001
From: Liam <byteslice@airmail.cc>
Date: Fri, 12 Jan 2024 19:19:07 -0500
Subject: kernel: optimize page free on shutdown

---
 src/core/hle/kernel/k_memory_block_manager.cpp |  4 +-
 src/core/hle/kernel/k_memory_block_manager.h   |  4 +-
 src/core/hle/kernel/k_page_table_base.cpp      | 73 +++++---------------------
 3 files changed, 18 insertions(+), 63 deletions(-)

(limited to 'src/core/hle/kernel')

diff --git a/src/core/hle/kernel/k_memory_block_manager.cpp b/src/core/hle/kernel/k_memory_block_manager.cpp
index 58a1e7216..f08a6e448 100644
--- a/src/core/hle/kernel/k_memory_block_manager.cpp
+++ b/src/core/hle/kernel/k_memory_block_manager.cpp
@@ -28,14 +28,14 @@ Result KMemoryBlockManager::Initialize(KProcessAddress st, KProcessAddress nd,
 }
 
 void KMemoryBlockManager::Finalize(KMemoryBlockSlabManager* slab_manager,
-                                   HostUnmapCallback&& host_unmap_callback) {
+                                   BlockCallback&& block_callback) {
     // Erase every block until we have none left.
     auto it = m_memory_block_tree.begin();
     while (it != m_memory_block_tree.end()) {
         KMemoryBlock* block = std::addressof(*it);
         it = m_memory_block_tree.erase(it);
+        block_callback(block->GetAddress(), block->GetSize());
         slab_manager->Free(block);
-        host_unmap_callback(block->GetAddress(), block->GetSize());
     }
 
     ASSERT(m_memory_block_tree.empty());
diff --git a/src/core/hle/kernel/k_memory_block_manager.h b/src/core/hle/kernel/k_memory_block_manager.h
index cb7b6f430..377628504 100644
--- a/src/core/hle/kernel/k_memory_block_manager.h
+++ b/src/core/hle/kernel/k_memory_block_manager.h
@@ -85,11 +85,11 @@ public:
 public:
     KMemoryBlockManager();
 
-    using HostUnmapCallback = std::function<void(Common::ProcessAddress, u64)>;
+    using BlockCallback = std::function<void(Common::ProcessAddress, u64)>;
 
     Result Initialize(KProcessAddress st, KProcessAddress nd,
                       KMemoryBlockSlabManager* slab_manager);
-    void Finalize(KMemoryBlockSlabManager* slab_manager, HostUnmapCallback&& host_unmap_callback);
+    void Finalize(KMemoryBlockSlabManager* slab_manager, BlockCallback&& block_callback);
 
     iterator end() {
         return m_memory_block_tree.end();
diff --git a/src/core/hle/kernel/k_page_table_base.cpp b/src/core/hle/kernel/k_page_table_base.cpp
index f01eaa164..3f0a39d33 100644
--- a/src/core/hle/kernel/k_page_table_base.cpp
+++ b/src/core/hle/kernel/k_page_table_base.cpp
@@ -435,69 +435,14 @@ Result KPageTableBase::FinalizeProcess() {
     // Only process tables should be finalized.
     ASSERT(!this->IsKernel());
 
-    // HLE processes don't have memory mapped.
-    R_SUCCEED_IF(m_impl == nullptr);
-
     // NOTE: Here Nintendo calls an unknown OnFinalize function.
     // this->OnFinalize();
 
     // NOTE: Here Nintendo calls a second unknown OnFinalize function.
     // this->OnFinalize2();
 
-    // Get implementation objects.
-    auto& impl = this->GetImpl();
-    auto& mm = m_kernel.MemoryManager();
-
-    // Traverse, freeing all pages.
-    {
-        // Get the address space size.
-        const size_t as_size = this->GetAddressSpaceSize();
-
-        // Begin the traversal.
-        TraversalContext context;
-        TraversalEntry cur_entry = {
-            .phys_addr = 0,
-            .block_size = 0,
-        };
-
-        bool cur_valid = false;
-        TraversalEntry next_entry;
-        bool next_valid;
-        size_t tot_size = 0;
-
-        next_valid = impl.BeginTraversal(std::addressof(next_entry), std::addressof(context),
-                                         this->GetAddressSpaceStart());
-
-        // Iterate over entries.
-        while (true) {
-            if ((!next_valid && !cur_valid) ||
-                (next_valid && cur_valid &&
-                 next_entry.phys_addr == cur_entry.phys_addr + cur_entry.block_size)) {
-                cur_entry.block_size += next_entry.block_size;
-            } else {
-                if (cur_valid && IsHeapPhysicalAddressForFinalize(cur_entry.phys_addr)) {
-                    mm.Close(cur_entry.phys_addr, cur_entry.block_size / PageSize);
-                }
-
-                // Update tracking variables.
-                tot_size += cur_entry.block_size;
-                cur_entry = next_entry;
-                cur_valid = next_valid;
-            }
-
-            if (cur_entry.block_size + tot_size >= as_size) {
-                break;
-            }
-
-            next_valid =
-                impl.ContinueTraversal(std::addressof(next_entry), std::addressof(context));
-        }
-
-        // Handle the last block.
-        if (cur_valid && IsHeapPhysicalAddressForFinalize(cur_entry.phys_addr)) {
-            mm.Close(cur_entry.phys_addr, cur_entry.block_size / PageSize);
-        }
-    }
+    // NOTE: Here Nintendo does a page table walk to discover heap pages to free.
+    // We will use the block manager finalization below to free them.
 
     R_SUCCEED();
 }
@@ -505,14 +450,24 @@ Result KPageTableBase::FinalizeProcess() {
 void KPageTableBase::Finalize() {
     this->FinalizeProcess();
 
-    auto HostUnmapCallback = [&](KProcessAddress addr, u64 size) {
+    auto BlockCallback = [&](KProcessAddress addr, u64 size) {
         if (m_impl->fastmem_arena) {
             m_system.DeviceMemory().buffer.Unmap(GetInteger(addr), size, false);
         }
+
+        // Get physical pages.
+        KPageGroup pg(m_kernel, m_block_info_manager);
+        this->MakePageGroup(pg, addr, size / PageSize);
+
+        // Free the pages.
+        pg.CloseAndReset();
     };
 
     // Finalize memory blocks.
-    m_memory_block_manager.Finalize(m_memory_block_slab_manager, std::move(HostUnmapCallback));
+    {
+        KScopedLightLock lk(m_general_lock);
+        m_memory_block_manager.Finalize(m_memory_block_slab_manager, std::move(BlockCallback));
+    }
 
     // Free any unsafe mapped memory.
     if (m_mapped_unsafe_physical_memory) {
-- 
cgit v1.2.3


From d4acdac168e4b445fc8cb2e0d5e15616c5dc7dbf Mon Sep 17 00:00:00 2001
From: GPUCode <geoster3d@gmail.com>
Date: Wed, 3 Jan 2024 23:37:41 +0200
Subject: core: Support multiple modules per patcher

---
 src/core/arm/nce/patcher.cpp                    | 83 ++++++++++++++++---------
 src/core/arm/nce/patcher.h                      | 27 +++++---
 src/core/hle/kernel/k_process.cpp               |  4 +-
 src/core/loader/deconstructed_rom_directory.cpp | 75 ++++++++++++++++------
 src/core/loader/nso.cpp                         | 41 ++++++------
 src/core/loader/nso.h                           |  3 +-
 6 files changed, 154 insertions(+), 79 deletions(-)

(limited to 'src/core/hle/kernel')

diff --git a/src/core/arm/nce/patcher.cpp b/src/core/arm/nce/patcher.cpp
index 47a7a8880..c7285e3a0 100644
--- a/src/core/arm/nce/patcher.cpp
+++ b/src/core/arm/nce/patcher.cpp
@@ -22,14 +22,10 @@ using NativeExecutionParameters = Kernel::KThread::NativeExecutionParameters;
 constexpr size_t MaxRelativeBranch = 128_MiB;
 constexpr u32 ModuleCodeIndex = 0x24 / sizeof(u32);
 
-Patcher::Patcher() : c(m_patch_instructions) {}
-
-Patcher::~Patcher() = default;
-
-void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
-                        const Kernel::CodeSet::Segment& code) {
-    // Branch to the first instruction of the module.
-    this->BranchToModule(0);
+Patcher::Patcher() : c(m_patch_instructions) {
+    // The first word of the patch section is always a branch to the first instruction of the
+    // module.
+    c.dw(0);
 
     // Write save context helper function.
     c.l(m_save_context);
@@ -38,6 +34,25 @@ void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
     // Write load context helper function.
     c.l(m_load_context);
     WriteLoadContext();
+}
+
+Patcher::~Patcher() = default;
+
+bool Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
+                        const Kernel::CodeSet::Segment& code) {
+    // If we have patched modules but cannot reach the new module, then it needs its own patcher.
+    const size_t image_size = program_image.size();
+    if (total_program_size + image_size > MaxRelativeBranch && total_program_size > 0) {
+        return false;
+    }
+
+    // Add a new module patch to our list
+    modules.emplace_back();
+    curr_patch = &modules.back();
+
+    // The first word of the patch section is always a branch to the first instruction of the
+    // module.
+    curr_patch->m_branch_to_module_relocations.push_back({0, 0});
 
     // Retrieve text segment data.
     const auto text = std::span{program_image}.subspan(code.offset, code.size);
@@ -94,16 +109,17 @@ void Patcher::PatchText(const Kernel::PhysicalMemory& program_image,
         }
 
         if (auto exclusive = Exclusive{inst}; exclusive.Verify()) {
-            m_exclusives.push_back(i);
+            curr_patch->m_exclusives.push_back(i);
         }
     }
 
     // Determine patching mode for the final relocation step
-    const size_t image_size = program_image.size();
+    total_program_size += image_size;
     this->mode = image_size > MaxRelativeBranch ? PatchMode::PreText : PatchMode::PostData;
+    return true;
 }
 
-void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
+bool Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
                               const Kernel::CodeSet::Segment& code,
                               Kernel::PhysicalMemory& program_image,
                               EntryTrampolines* out_trampolines) {
@@ -120,7 +136,7 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
         if (mode == PatchMode::PreText) {
             rc.B(rel.patch_offset - patch_size - rel.module_offset);
         } else {
-            rc.B(image_size - rel.module_offset + rel.patch_offset);
+            rc.B(total_program_size - rel.module_offset + rel.patch_offset);
         }
     };
 
@@ -129,7 +145,7 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
         if (mode == PatchMode::PreText) {
             rc.B(patch_size - rel.patch_offset + rel.module_offset);
         } else {
-            rc.B(rel.module_offset - image_size - rel.patch_offset);
+            rc.B(rel.module_offset - total_program_size - rel.patch_offset);
         }
     };
 
@@ -137,7 +153,7 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
         if (mode == PatchMode::PreText) {
             return GetInteger(load_base) + patch_offset;
         } else {
-            return GetInteger(load_base) + image_size + patch_offset;
+            return GetInteger(load_base) + total_program_size + patch_offset;
         }
     };
 
@@ -150,39 +166,50 @@ void Patcher::RelocateAndCopy(Common::ProcessAddress load_base,
     };
 
     // We are now ready to relocate!
-    for (const Relocation& rel : m_branch_to_patch_relocations) {
+    auto& patch = modules[m_relocate_module_index++];
+    for (const Relocation& rel : patch.m_branch_to_patch_relocations) {
         ApplyBranchToPatchRelocation(text_words.data() + rel.module_offset / sizeof(u32), rel);
     }
-    for (const Relocation& rel : m_branch_to_module_relocations) {
+    for (const Relocation& rel : patch.m_branch_to_module_relocations) {
         ApplyBranchToModuleRelocation(m_patch_instructions.data() + rel.patch_offset / sizeof(u32),
                                       rel);
     }
 
     // Rewrite PC constants and record post trampolines
-    for (const Relocation& rel : m_write_module_pc_relocations) {
+    for (const Relocation& rel : patch.m_write_module_pc_relocations) {
         oaknut::CodeGenerator rc{m_patch_instructions.data() + rel.patch_offset / sizeof(u32)};
         rc.dx(RebasePc(rel.module_offset));
     }
-    for (const Trampoline& rel : m_trampolines) {
+    for (const Trampoline& rel : patch.m_trampolines) {
         out_trampolines->insert({RebasePc(rel.module_offset), RebasePatch(rel.patch_offset)});
     }
 
     // Cortex-A57 seems to treat all exclusives as ordered, but newer processors do not.
     // Convert to ordered to preserve this assumption.
-    for (const ModuleTextAddress i : m_exclusives) {
+    for (const ModuleTextAddress i : patch.m_exclusives) {
         auto exclusive = Exclusive{text_words[i]};
         text_words[i] = exclusive.AsOrdered();
     }
 
-    // Copy to program image
-    if (this->mode == PatchMode::PreText) {
-        std::memcpy(program_image.data(), m_patch_instructions.data(),
-                    m_patch_instructions.size() * sizeof(u32));
-    } else {
-        program_image.resize(image_size + patch_size);
-        std::memcpy(program_image.data() + image_size, m_patch_instructions.data(),
-                    m_patch_instructions.size() * sizeof(u32));
+    // Remove the patched module size from the total. This is done so total_program_size
+    // always represents the distance from the currently patched module to the patch section.
+    total_program_size -= image_size;
+
+    // Only copy to the program image of the last module
+    if (m_relocate_module_index == modules.size()) {
+        if (this->mode == PatchMode::PreText) {
+            ASSERT(image_size == total_program_size);
+            std::memcpy(program_image.data(), m_patch_instructions.data(),
+                        m_patch_instructions.size() * sizeof(u32));
+        } else {
+            program_image.resize(image_size + patch_size);
+            std::memcpy(program_image.data() + image_size, m_patch_instructions.data(),
+                        m_patch_instructions.size() * sizeof(u32));
+        }
+        return true;
     }
+
+    return false;
 }
 
 size_t Patcher::GetSectionSize() const noexcept {
@@ -322,7 +349,7 @@ void Patcher::WriteSvcTrampoline(ModuleDestLabel module_dest, u32 svc_id) {
 
     // Write the post-SVC trampoline address, which will jump back to the guest after restoring its
     // state.
-    m_trampolines.push_back({c.offset(), module_dest});
+    curr_patch->m_trampolines.push_back({c.offset(), module_dest});
 
     // Host called this location. Save the return address so we can
     // unwind the stack properly when jumping back.
diff --git a/src/core/arm/nce/patcher.h b/src/core/arm/nce/patcher.h
index c6d1608c1..a44f385e2 100644
--- a/src/core/arm/nce/patcher.h
+++ b/src/core/arm/nce/patcher.h
@@ -31,9 +31,9 @@ public:
     explicit Patcher();
     ~Patcher();
 
-    void PatchText(const Kernel::PhysicalMemory& program_image,
+    bool PatchText(const Kernel::PhysicalMemory& program_image,
                    const Kernel::CodeSet::Segment& code);
-    void RelocateAndCopy(Common::ProcessAddress load_base, const Kernel::CodeSet::Segment& code,
+    bool RelocateAndCopy(Common::ProcessAddress load_base, const Kernel::CodeSet::Segment& code,
                          Kernel::PhysicalMemory& program_image, EntryTrampolines* out_trampolines);
     size_t GetSectionSize() const noexcept;
 
@@ -61,16 +61,16 @@ private:
 
 private:
     void BranchToPatch(uintptr_t module_dest) {
-        m_branch_to_patch_relocations.push_back({c.offset(), module_dest});
+        curr_patch->m_branch_to_patch_relocations.push_back({c.offset(), module_dest});
     }
 
     void BranchToModule(uintptr_t module_dest) {
-        m_branch_to_module_relocations.push_back({c.offset(), module_dest});
+        curr_patch->m_branch_to_module_relocations.push_back({c.offset(), module_dest});
         c.dw(0);
     }
 
     void WriteModulePc(uintptr_t module_dest) {
-        m_write_module_pc_relocations.push_back({c.offset(), module_dest});
+        curr_patch->m_write_module_pc_relocations.push_back({c.offset(), module_dest});
         c.dx(0);
     }
 
@@ -84,15 +84,22 @@ private:
         uintptr_t module_offset; ///< Offset in bytes from the start of the text section.
     };
 
+    struct ModulePatch {
+        std::vector<Trampoline> m_trampolines;
+        std::vector<Relocation> m_branch_to_patch_relocations{};
+        std::vector<Relocation> m_branch_to_module_relocations{};
+        std::vector<Relocation> m_write_module_pc_relocations{};
+        std::vector<ModuleTextAddress> m_exclusives{};
+    };
+
     oaknut::VectorCodeGenerator c;
-    std::vector<Trampoline> m_trampolines;
-    std::vector<Relocation> m_branch_to_patch_relocations{};
-    std::vector<Relocation> m_branch_to_module_relocations{};
-    std::vector<Relocation> m_write_module_pc_relocations{};
-    std::vector<ModuleTextAddress> m_exclusives{};
     oaknut::Label m_save_context{};
     oaknut::Label m_load_context{};
     PatchMode mode{PatchMode::None};
+    size_t total_program_size{};
+    size_t m_relocate_module_index{};
+    std::vector<ModulePatch> modules;
+    ModulePatch* curr_patch;
 };
 
 } // namespace Core::NCE
diff --git a/src/core/hle/kernel/k_process.cpp b/src/core/hle/kernel/k_process.cpp
index 068e71dff..8839ddbc2 100644
--- a/src/core/hle/kernel/k_process.cpp
+++ b/src/core/hle/kernel/k_process.cpp
@@ -1233,10 +1233,10 @@ void KProcess::LoadModule(CodeSet code_set, KProcessAddress base_addr) {
     ReprotectSegment(code_set.DataSegment(), Svc::MemoryPermission::ReadWrite);
 
 #ifdef HAS_NCE
-    if (this->IsApplication() && Settings::IsNceEnabled()) {
+    const auto& patch = code_set.PatchSegment();
+    if (this->IsApplication() && Settings::IsNceEnabled() && patch.size != 0) {
         auto& buffer = m_kernel.System().DeviceMemory().buffer;
         const auto& code = code_set.CodeSegment();
-        const auto& patch = code_set.PatchSegment();
         buffer.Protect(GetInteger(base_addr + code.addr), code.size,
                        Common::MemoryPermission::Read | Common::MemoryPermission::Execute);
         buffer.Protect(GetInteger(base_addr + patch.addr), patch.size,
diff --git a/src/core/loader/deconstructed_rom_directory.cpp b/src/core/loader/deconstructed_rom_directory.cpp
index c9f8707b7..b2173f697 100644
--- a/src/core/loader/deconstructed_rom_directory.cpp
+++ b/src/core/loader/deconstructed_rom_directory.cpp
@@ -19,8 +19,54 @@
 #include "core/arm/nce/patcher.h"
 #endif
 
+#ifndef HAS_NCE
+namespace Core::NCE {
+class Patcher {};
+} // namespace Core::NCE
+#endif
+
 namespace Loader {
 
+struct PatchCollection {
+    explicit PatchCollection(bool is_application_) : is_application{is_application_} {
+        module_patcher_indices.fill(-1);
+        patchers.emplace_back();
+    }
+
+    std::vector<Core::NCE::Patcher>* GetPatchers() {
+        if (is_application && Settings::IsNceEnabled()) {
+            return &patchers;
+        }
+        return nullptr;
+    }
+
+    size_t GetTotalPatchSize() const {
+        size_t total_size{};
+#ifdef HAS_NCE
+        for (auto& patcher : patchers) {
+            total_size += patcher.GetSectionSize();
+        }
+#endif
+        return total_size;
+    }
+
+    void SaveIndex(size_t module) {
+        module_patcher_indices[module] = static_cast<s32>(patchers.size() - 1);
+    }
+
+    s32 GetIndex(size_t module) const {
+        return module_patcher_indices[module];
+    }
+
+    s32 GetLastIndex() const {
+        return static_cast<s32>(patchers.size()) - 1;
+    }
+
+    bool is_application;
+    std::vector<Core::NCE::Patcher> patchers;
+    std::array<s32, 13> module_patcher_indices{};
+};
+
 AppLoader_DeconstructedRomDirectory::AppLoader_DeconstructedRomDirectory(FileSys::VirtualFile file_,
                                                                          bool override_update_)
     : AppLoader(std::move(file_)), override_update(override_update_), is_hbl(false) {
@@ -142,18 +188,7 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
     std::size_t code_size{};
 
     // Define an nce patch context for each potential module.
-#ifdef HAS_NCE
-    std::array<Core::NCE::Patcher, 13> module_patchers;
-#endif
-
-    const auto GetPatcher = [&](size_t i) -> Core::NCE::Patcher* {
-#ifdef HAS_NCE
-        if (is_application && Settings::IsNceEnabled()) {
-            return &module_patchers[i];
-        }
-#endif
-        return nullptr;
-    };
+    PatchCollection patch_ctx{is_application};
 
     // Use the NSO module loader to figure out the code layout
     for (size_t i = 0; i < static_modules.size(); i++) {
@@ -164,13 +199,14 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
         }
 
         const bool should_pass_arguments = std::strcmp(module, "rtld") == 0;
-        const auto tentative_next_load_addr =
-            AppLoader_NSO::LoadModule(process, system, *module_file, code_size,
-                                      should_pass_arguments, false, {}, GetPatcher(i));
+        const auto tentative_next_load_addr = AppLoader_NSO::LoadModule(
+            process, system, *module_file, code_size, should_pass_arguments, false, {},
+            patch_ctx.GetPatchers(), patch_ctx.GetLastIndex());
         if (!tentative_next_load_addr) {
             return {ResultStatus::ErrorLoadingNSO, {}};
         }
 
+        patch_ctx.SaveIndex(i);
         code_size = *tentative_next_load_addr;
     }
 
@@ -184,6 +220,9 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
         return 0;
     }();
 
+    // Add patch size to the total module size
+    code_size += patch_ctx.GetTotalPatchSize();
+
     // Setup the process code layout
     if (process.LoadFromMetadata(metadata, code_size, fastmem_base, is_hbl).IsError()) {
         return {ResultStatus::ErrorUnableToParseKernelMetadata, {}};
@@ -204,9 +243,9 @@ AppLoader_DeconstructedRomDirectory::LoadResult AppLoader_DeconstructedRomDirect
 
         const VAddr load_addr{next_load_addr};
         const bool should_pass_arguments = std::strcmp(module, "rtld") == 0;
-        const auto tentative_next_load_addr =
-            AppLoader_NSO::LoadModule(process, system, *module_file, load_addr,
-                                      should_pass_arguments, true, pm, GetPatcher(i));
+        const auto tentative_next_load_addr = AppLoader_NSO::LoadModule(
+            process, system, *module_file, load_addr, should_pass_arguments, true, pm,
+            patch_ctx.GetPatchers(), patch_ctx.GetIndex(i));
         if (!tentative_next_load_addr) {
             return {ResultStatus::ErrorLoadingNSO, {}};
         }
diff --git a/src/core/loader/nso.cpp b/src/core/loader/nso.cpp
index b053a0d14..583b7e927 100644
--- a/src/core/loader/nso.cpp
+++ b/src/core/loader/nso.cpp
@@ -77,7 +77,8 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
                                                const FileSys::VfsFile& nso_file, VAddr load_base,
                                                bool should_pass_arguments, bool load_into_process,
                                                std::optional<FileSys::PatchManager> pm,
-                                               Core::NCE::Patcher* patch) {
+                                               std::vector<Core::NCE::Patcher>* patches,
+                                               s32 patch_index) {
     if (nso_file.GetSize() < sizeof(NSOHeader)) {
         return std::nullopt;
     }
@@ -94,8 +95,11 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
     // Allocate some space at the beginning if we are patching in PreText mode.
     const size_t module_start = [&]() -> size_t {
 #ifdef HAS_NCE
-        if (patch && patch->GetPatchMode() == Core::NCE::PatchMode::PreText) {
-            return patch->GetSectionSize();
+        if (patches && load_into_process) {
+            auto* patch = &patches->operator[](patch_index);
+            if (patch->GetPatchMode() == Core::NCE::PatchMode::PreText) {
+                return patch->GetSectionSize();
+            }
         }
 #endif
         return 0;
@@ -160,27 +164,24 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
 #ifdef HAS_NCE
     // If we are computing the process code layout and using nce backend, patch.
     const auto& code = codeset.CodeSegment();
-    if (patch && patch->GetPatchMode() == Core::NCE::PatchMode::None) {
+    auto* patch = patches ? &patches->operator[](patch_index) : nullptr;
+    if (patch && !load_into_process) {
         // Patch SVCs and MRS calls in the guest code
-        patch->PatchText(program_image, code);
-
-        // Add patch section size to the module size.
-        image_size += static_cast<u32>(patch->GetSectionSize());
+        while (!patch->PatchText(program_image, code)) {
+            patch = &patches->emplace_back();
+        }
     } else if (patch) {
         // Relocate code patch and copy to the program_image.
-        patch->RelocateAndCopy(load_base, code, program_image, &process.GetPostHandlers());
-
-        // Update patch section.
-        auto& patch_segment = codeset.PatchSegment();
-        patch_segment.addr =
-            patch->GetPatchMode() == Core::NCE::PatchMode::PreText ? 0 : image_size;
-        patch_segment.size = static_cast<u32>(patch->GetSectionSize());
-
-        // Add patch section size to the module size. In PreText mode image_size
-        // already contains the patch segment as part of module_start.
-        if (patch->GetPatchMode() == Core::NCE::PatchMode::PostData) {
-            image_size += patch_segment.size;
+        if (patch->RelocateAndCopy(load_base, code, program_image, &process.GetPostHandlers())) {
+            // Update patch section.
+            auto& patch_segment = codeset.PatchSegment();
+            patch_segment.addr =
+                patch->GetPatchMode() == Core::NCE::PatchMode::PreText ? 0 : image_size;
+            patch_segment.size = static_cast<u32>(patch->GetSectionSize());
         }
+
+        // Refresh image_size to take account the patch section if it was added by RelocateAndCopy
+        image_size = static_cast<u32>(program_image.size());
     }
 #endif
 
diff --git a/src/core/loader/nso.h b/src/core/loader/nso.h
index 29b86ed4c..6356697e3 100644
--- a/src/core/loader/nso.h
+++ b/src/core/loader/nso.h
@@ -93,7 +93,8 @@ public:
                                            const FileSys::VfsFile& nso_file, VAddr load_base,
                                            bool should_pass_arguments, bool load_into_process,
                                            std::optional<FileSys::PatchManager> pm = {},
-                                           Core::NCE::Patcher* patch = nullptr);
+                                           std::vector<Core::NCE::Patcher>* patches = nullptr,
+                                           s32 patch_index = -1);
 
     LoadResult Load(Kernel::KProcess& process, Core::System& system) override;
 
-- 
cgit v1.2.3