Revert "[heap_tracker] Use ankerl map instead of rb tree (#249)" (#382)

This reverts commit c9a3baab5d.

this commit caused issues in ender magnolia or something, need to make
sure I didn't mess up the revert

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/382
Reviewed-by: Shinmegumi <shinmegumi@eden-emu.dev>
Reviewed-by: Lizzie <lizzie@eden-emu.dev>
Reviewed-by: MaranBr <maranbr@outlook.com>

src/common/CMakeLists.txt

@@ -274,11 +274,6 @@ endif()
 target_link_libraries(common PUBLIC fmt::fmt stb::headers Threads::Threads)
 target_link_libraries(common PRIVATE lz4::lz4 LLVM::Demangle zstd::zstd)
 
-if (TARGET unordered_dense::unordered_dense)
-  # weird quirk of system installs
-  target_link_libraries(common PUBLIC unordered_dense::unordered_dense)
-endif()
-
 if(ANDROID)
   # For ASharedMemory_create
   target_link_libraries(common PRIVATE android)

src/common/heap_tracker.cpp

@@ -1,5 +1,3 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -35,60 +33,68 @@ HeapTracker::~HeapTracker() = default;
 
 void HeapTracker::Map(size_t virtual_offset, size_t host_offset, size_t length,
                       MemoryPermission perm, bool is_separate_heap) {
-    bool rebuild_required = false;
     // When mapping other memory, map pages immediately.
     if (!is_separate_heap) {
         m_buffer.Map(virtual_offset, host_offset, length, perm, false);
         return;
     }
+
     {
-        // We are mapping part of a separate heap and insert into mappings.
+        // We are mapping part of a separate heap.
         std::scoped_lock lk{m_lock};
-        m_map_count++;
-        const auto it = m_mappings.insert_or_assign(virtual_offset, SeparateHeapMap{
+
+        auto* const map = new SeparateHeapMap{
+            .vaddr = virtual_offset,
             .paddr = host_offset,
             .size = length,
             .tick = m_tick++,
             .perm = perm,
             .is_resident = false,
-        });
-        // Update tick before possible rebuild.
-        it.first->second.tick = m_tick++;
-        // Check if we need to rebuild.
-        if (m_resident_map_count >= m_max_resident_map_count)
-            rebuild_required = true;
-        // Map the area.
-        m_buffer.Map(it.first->first, it.first->second.paddr, it.first->second.size, it.first->second.perm, false);
-        // This map is now resident.
-        it.first->second.is_resident = true;
-        m_resident_map_count++;
-        m_resident_mappings.insert(*it.first);
+        };
+
+        // Insert into mappings.
+        m_map_count++;
+        m_mappings.insert(*map);
     }
-    // A rebuild was required, so perform it now.
-    if (rebuild_required)
-        this->RebuildSeparateHeapAddressSpace();
+
+    // Finally, map.
+    this->DeferredMapSeparateHeap(virtual_offset);
 }
 
 void HeapTracker::Unmap(size_t virtual_offset, size_t size, bool is_separate_heap) {
     // If this is a separate heap...
     if (is_separate_heap) {
         std::scoped_lock lk{m_lock};
+
+        const SeparateHeapMap key{
+            .vaddr = virtual_offset,
+        };
+
         // Split at the boundaries of the region we are removing.
         this->SplitHeapMapLocked(virtual_offset);
         this->SplitHeapMapLocked(virtual_offset + size);
+
         // Erase all mappings in range.
-        auto it = m_mappings.find(virtual_offset);
-        while (it != m_mappings.end() && it->first < virtual_offset + size) {
+        auto it = m_mappings.find(key);
+        while (it != m_mappings.end() && it->vaddr < virtual_offset + size) {
+            // Get underlying item.
+            auto* const item = std::addressof(*it);
+
             // If resident, erase from resident map.
-            if (it->second.is_resident) {
+            if (item->is_resident) {
                 ASSERT(--m_resident_map_count >= 0);
-                m_resident_mappings.erase(m_resident_mappings.find(it->first));
+                m_resident_mappings.erase(m_resident_mappings.iterator_to(*item));
             }
+
             // Erase from map.
             ASSERT(--m_map_count >= 0);
             it = m_mappings.erase(it);
+
+            // Free the item.
+            delete item;
         }
     }
+
     // Unmap pages.
     m_buffer.Unmap(virtual_offset, size, false);
 }
@@ -110,51 +116,110 @@ void HeapTracker::Protect(size_t virtual_offset, size_t size, MemoryPermission p
 
         {
             std::scoped_lock lk2{m_lock};
+
+            const SeparateHeapMap key{
+                .vaddr = next,
+            };
+
             // Try to get the next mapping corresponding to this address.
-            const auto it = m_mappings.find(next);
+            const auto it = m_mappings.nfind(key);
+
             if (it == m_mappings.end()) {
                 // There are no separate heap mappings remaining.
                 next = end;
                 should_protect = true;
-            } else if (it->first == cur) {
+            } else if (it->vaddr == cur) {
                 // We are in range.
                 // Update permission bits.
-                it->second.perm = perm;
+                it->perm = perm;
 
                 // Determine next address and whether we should protect.
-                next = cur + it->second.size;
-                should_protect = it->second.is_resident;
+                next = cur + it->size;
+                should_protect = it->is_resident;
             } else /* if (it->vaddr > cur) */ {
                 // We weren't in range, but there is a block coming up that will be.
-                next = it->first;
+                next = it->vaddr;
                 should_protect = true;
             }
         }
 
         // Clamp to end.
         next = std::min(next, end);
+
         // Reprotect, if we need to.
-        if (should_protect)
+        if (should_protect) {
             m_buffer.Protect(cur, next - cur, perm);
+        }
+
         // Advance.
         cur = next;
     }
 }
 
+bool HeapTracker::DeferredMapSeparateHeap(u8* fault_address) {
+    if (m_buffer.IsInVirtualRange(fault_address)) {
+        return this->DeferredMapSeparateHeap(fault_address - m_buffer.VirtualBasePointer());
+    }
+
+    return false;
+}
+
+bool HeapTracker::DeferredMapSeparateHeap(size_t virtual_offset) {
+    bool rebuild_required = false;
+
+    {
+        std::scoped_lock lk{m_lock};
+
+        // Check to ensure this was a non-resident separate heap mapping.
+        const auto it = this->GetNearestHeapMapLocked(virtual_offset);
+        if (it == m_mappings.end() || it->is_resident) {
+            return false;
+        }
+
+        // Update tick before possible rebuild.
+        it->tick = m_tick++;
+
+        // Check if we need to rebuild.
+        if (m_resident_map_count > m_max_resident_map_count) {
+            rebuild_required = true;
+        }
+
+        // Map the area.
+        m_buffer.Map(it->vaddr, it->paddr, it->size, it->perm, false);
+
+        // This map is now resident.
+        it->is_resident = true;
+        m_resident_map_count++;
+        m_resident_mappings.insert(*it);
+    }
+
+    if (rebuild_required) {
+        // A rebuild was required, so perform it now.
+        this->RebuildSeparateHeapAddressSpace();
+    }
+
+    return true;
+}
+
 void HeapTracker::RebuildSeparateHeapAddressSpace() {
     std::scoped_lock lk{m_rebuild_lock, m_lock};
+
     ASSERT(!m_resident_mappings.empty());
+
     // Dump half of the mappings.
+    //
     // Despite being worse in theory, this has proven to be better in practice than more
     // regularly dumping a smaller amount, because it significantly reduces average case
     // lock contention.
-    std::size_t const desired_count = std::min(m_resident_map_count, m_max_resident_map_count) / 2;
-    std::size_t const evict_count = m_resident_map_count - desired_count;
+    const size_t desired_count = std::min(m_resident_map_count, m_max_resident_map_count) / 2;
+    const size_t evict_count = m_resident_map_count - desired_count;
     auto it = m_resident_mappings.begin();
-    for (std::size_t i = 0; i < evict_count && it != m_resident_mappings.end(); i++) {
+
+    for (size_t i = 0; i < evict_count && it != m_resident_mappings.end(); i++) {
         // Unmark and unmap.
-        it->second.is_resident = false;
-        m_buffer.Unmap(it->first, it->second.size, false);
+        it->is_resident = false;
+        m_buffer.Unmap(it->vaddr, it->size, false);
+
         // Advance.
         ASSERT(--m_resident_map_count >= 0);
         it = m_resident_mappings.erase(it);
@@ -163,32 +228,53 @@ void HeapTracker::RebuildSeparateHeapAddressSpace() {
 
 void HeapTracker::SplitHeapMap(VAddr offset, size_t size) {
     std::scoped_lock lk{m_lock};
+
     this->SplitHeapMapLocked(offset);
     this->SplitHeapMapLocked(offset + size);
 }
 
 void HeapTracker::SplitHeapMapLocked(VAddr offset) {
-    auto it = this->GetNearestHeapMapLocked(offset);
-    if (it != m_mappings.end() && it->first != offset) {
-        // Adjust left iterator
-        auto const orig_size = it->second.size;
-        auto const left_size = offset - it->first;
-        it->second.size = left_size;
-        // Insert the new right map.
-        auto const right = SeparateHeapMap{
-            .paddr = it->second.paddr + left_size,
+    const auto it = this->GetNearestHeapMapLocked(offset);
+    if (it == m_mappings.end() || it->vaddr == offset) {
+        // Not contained or no split required.
+        return;
+    }
+
+    // Cache the original values.
+    auto* const left = std::addressof(*it);
+    const size_t orig_size = left->size;
+
+    // Adjust the left map.
+    const size_t left_size = offset - left->vaddr;
+    left->size = left_size;
+
+    // Create the new right map.
+    auto* const right = new SeparateHeapMap{
+        .vaddr = left->vaddr + left_size,
+        .paddr = left->paddr + left_size,
         .size = orig_size - left_size,
-            .tick = it->second.tick,
-            .perm = it->second.perm,
-            .is_resident = it->second.is_resident,
+        .tick = left->tick,
+        .perm = left->perm,
+        .is_resident = left->is_resident,
     };
+
+    // Insert the new right map.
     m_map_count++;
-        auto rit = m_mappings.insert_or_assign(it->first + left_size, right);
-        if (rit.first->second.is_resident) {
+    m_mappings.insert(*right);
+
+    // If resident, also insert into resident map.
+    if (right->is_resident) {
         m_resident_map_count++;
-            m_resident_mappings.insert(*rit.first);
-        }
+        m_resident_mappings.insert(*right);
     }
 }
 
+HeapTracker::AddrTree::iterator HeapTracker::GetNearestHeapMapLocked(VAddr offset) {
+    const SeparateHeapMap key{
+        .vaddr = offset,
+    };
+
+    return m_mappings.find(key);
+}
+
 } // namespace Common

src/common/heap_tracker.h

@@ -1,55 +1,93 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #pragma once
 
+#include <atomic>
 #include <mutex>
+#include <set>
 #include <shared_mutex>
-#include <ankerl/unordered_dense.h>
+
 #include "common/host_memory.h"
+#include "common/intrusive_red_black_tree.h"
 
 namespace Common {
 
 struct SeparateHeapMap {
-    PAddr paddr{}; //8
-    std::size_t size{}; //8 (16)
-    std::size_t tick{}; //8 (24)
-    // 4 bits needed, sync with host_memory.h if needed
-    MemoryPermission perm : 4 = MemoryPermission::Read;
-    bool is_resident : 1 = false;
+    Common::IntrusiveRedBlackTreeNode addr_node{};
+    Common::IntrusiveRedBlackTreeNode tick_node{};
+    VAddr vaddr{};
+    PAddr paddr{};
+    size_t size{};
+    size_t tick{};
+    MemoryPermission perm{};
+    bool is_resident{};
+};
+
+struct SeparateHeapMapAddrComparator {
+    static constexpr int Compare(const SeparateHeapMap& lhs, const SeparateHeapMap& rhs) {
+        if (lhs.vaddr < rhs.vaddr) {
+            return -1;
+        } else if (lhs.vaddr <= (rhs.vaddr + rhs.size - 1)) {
+            return 0;
+        } else {
+            return 1;
+        }
+    }
+};
+
+struct SeparateHeapMapTickComparator {
+    static constexpr int Compare(const SeparateHeapMap& lhs, const SeparateHeapMap& rhs) {
+        if (lhs.tick < rhs.tick) {
+            return -1;
+        } else if (lhs.tick > rhs.tick) {
+            return 1;
+        } else {
+            return SeparateHeapMapAddrComparator::Compare(lhs, rhs);
+        }
+    }
 };
-static_assert(sizeof(SeparateHeapMap) == 32); //half a cache line! good for coherency
 
 class HeapTracker {
 public:
     explicit HeapTracker(Common::HostMemory& buffer);
     ~HeapTracker();
-    void Map(size_t virtual_offset, size_t host_offset, size_t length, MemoryPermission perm, bool is_separate_heap);
+
+    void Map(size_t virtual_offset, size_t host_offset, size_t length, MemoryPermission perm,
+             bool is_separate_heap);
     void Unmap(size_t virtual_offset, size_t size, bool is_separate_heap);
     void Protect(size_t virtual_offset, size_t length, MemoryPermission perm);
-    inline u8* VirtualBasePointer() noexcept {
+    u8* VirtualBasePointer() {
         return m_buffer.VirtualBasePointer();
     }
+
+    bool DeferredMapSeparateHeap(u8* fault_address);
+    bool DeferredMapSeparateHeap(size_t virtual_offset);
+
 private:
-    // TODO: You may want to "fake-map" the first 2GB of 64-bit address space
-    // and dedicate it entirely to a recursive PTE mapping :)
-    // However Ankerl is way better than using an RB tree, in all senses
-    using AddrTree = ankerl::unordered_dense::map<VAddr, SeparateHeapMap>;
-    AddrTree m_mappings;
-    using TicksTree = ankerl::unordered_dense::map<VAddr, SeparateHeapMap>;
-    TicksTree m_resident_mappings;
+    using AddrTreeTraits =
+        Common::IntrusiveRedBlackTreeMemberTraitsDeferredAssert<&SeparateHeapMap::addr_node>;
+    using AddrTree = AddrTreeTraits::TreeType<SeparateHeapMapAddrComparator>;
+
+    using TickTreeTraits =
+        Common::IntrusiveRedBlackTreeMemberTraitsDeferredAssert<&SeparateHeapMap::tick_node>;
+    using TickTree = TickTreeTraits::TreeType<SeparateHeapMapTickComparator>;
+
+    AddrTree m_mappings{};
+    TickTree m_resident_mappings{};
+
 private:
     void SplitHeapMap(VAddr offset, size_t size);
     void SplitHeapMapLocked(VAddr offset);
+
+    AddrTree::iterator GetNearestHeapMapLocked(VAddr offset);
+
     void RebuildSeparateHeapAddressSpace();
-    inline HeapTracker::AddrTree::iterator GetNearestHeapMapLocked(VAddr offset) noexcept {
-        return m_mappings.find(offset);
-    }
+
 private:
     Common::HostMemory& m_buffer;
     const s64 m_max_resident_map_count;
+
     std::shared_mutex m_rebuild_lock{};
     std::mutex m_lock{};
     s64 m_map_count{};

src/core/arm/dynarmic/arm_dynarmic.cpp

@@ -3,9 +3,47 @@
 
 #ifdef __linux__
 
-//#include "common/signal_chain.h"
+#include "common/signal_chain.h"
+
 #include "core/arm/dynarmic/arm_dynarmic.h"
-//#include "core/hle/kernel/k_process.h"
-//#include "core/memory.h"
+#include "core/hle/kernel/k_process.h"
+#include "core/memory.h"
+
+namespace Core {
+
+namespace {
+
+thread_local Core::Memory::Memory* g_current_memory{};
+std::once_flag g_registered{};
+struct sigaction g_old_segv {};
+
+void HandleSigSegv(int sig, siginfo_t* info, void* ctx) {
+    if (g_current_memory && g_current_memory->InvalidateSeparateHeap(info->si_addr)) {
+        return;
+    }
+
+    return g_old_segv.sa_sigaction(sig, info, ctx);
+}
+
+} // namespace
+
+ScopedJitExecution::ScopedJitExecution(Kernel::KProcess* process) {
+    g_current_memory = std::addressof(process->GetMemory());
+}
+
+ScopedJitExecution::~ScopedJitExecution() {
+    g_current_memory = nullptr;
+}
+
+void ScopedJitExecution::RegisterHandler() {
+    std::call_once(g_registered, [] {
+        struct sigaction sa {};
+        sa.sa_sigaction = &HandleSigSegv;
+        sa.sa_flags = SA_SIGINFO | SA_ONSTACK;
+        Common::SigAction(SIGSEGV, std::addressof(sa), std::addressof(g_old_segv));
+    });
+}
+
+} // namespace Core
 
 #endif

src/core/arm/dynarmic/arm_dynarmic.h

@@ -26,4 +26,24 @@ constexpr HaltReason TranslateHaltReason(Dynarmic::HaltReason hr) {
     return static_cast<HaltReason>(hr);
 }
 
+#ifdef __linux__
+
+class ScopedJitExecution {
+public:
+    explicit ScopedJitExecution(Kernel::KProcess* process);
+    ~ScopedJitExecution();
+    static void RegisterHandler();
+};
+
+#else
+
+class ScopedJitExecution {
+public:
+    explicit ScopedJitExecution(Kernel::KProcess* process) {}
+    ~ScopedJitExecution() {}
+    static void RegisterHandler() {}
+};
+
+#endif
+
 } // namespace Core

src/core/arm/dynarmic/arm_dynarmic_32.cpp

@@ -343,11 +343,15 @@ bool ArmDynarmic32::IsInThumbMode() const {
 }
 
 HaltReason ArmDynarmic32::RunThread(Kernel::KThread* thread) {
+    ScopedJitExecution sj(thread->GetOwnerProcess());
+
     m_jit->ClearExclusiveState();
     return TranslateHaltReason(m_jit->Run());
 }
 
 HaltReason ArmDynarmic32::StepThread(Kernel::KThread* thread) {
+    ScopedJitExecution sj(thread->GetOwnerProcess());
+
     m_jit->ClearExclusiveState();
     return TranslateHaltReason(m_jit->Step());
 }
@@ -389,6 +393,7 @@ ArmDynarmic32::ArmDynarmic32(System& system, bool uses_wall_clock, Kernel::KProc
       m_cp15(std::make_shared<DynarmicCP15>(*this)), m_core_index{core_index} {
     auto& page_table_impl = process->GetPageTable().GetBasePageTable().GetImpl();
     m_jit = MakeJit(&page_table_impl);
+    ScopedJitExecution::RegisterHandler();
 }
 
 ArmDynarmic32::~ArmDynarmic32() = default;

src/core/arm/dynarmic/arm_dynarmic_64.cpp

@@ -374,11 +374,15 @@ std::shared_ptr<Dynarmic::A64::Jit> ArmDynarmic64::MakeJit(Common::PageTable* pa
 }
 
 HaltReason ArmDynarmic64::RunThread(Kernel::KThread* thread) {
+    ScopedJitExecution sj(thread->GetOwnerProcess());
+
     m_jit->ClearExclusiveState();
     return TranslateHaltReason(m_jit->Run());
 }
 
 HaltReason ArmDynarmic64::StepThread(Kernel::KThread* thread) {
+    ScopedJitExecution sj(thread->GetOwnerProcess());
+
     m_jit->ClearExclusiveState();
     return TranslateHaltReason(m_jit->Step());
 }
@@ -418,6 +422,7 @@ ArmDynarmic64::ArmDynarmic64(System& system, bool uses_wall_clock, Kernel::KProc
     auto& page_table = process->GetPageTable().GetBasePageTable();
     auto& page_table_impl = page_table.GetImpl();
     m_jit = MakeJit(&page_table_impl, page_table.GetAddressSpaceWidth());
+    ScopedJitExecution::RegisterHandler();
 }
 
 ArmDynarmic64::~ArmDynarmic64() = default;

src/core/hle/kernel/k_process.cpp

@@ -1266,6 +1266,10 @@ void KProcess::InitializeInterfaces() {
 
 #ifdef HAS_NCE
     if (this->IsApplication() && Settings::IsNceEnabled()) {
+        // Register the scoped JIT handler before creating any NCE instances
+        // so that its signal handler will appear first in the signal chain.
+        Core::ScopedJitExecution::RegisterHandler();
+
         for (size_t i = 0; i < Core::Hardware::NUM_CPU_CORES; i++) {
             m_arm_interfaces[i] = std::make_unique<Core::ArmNce>(m_kernel.System(), true, i);
         }

src/core/memory.cpp

@@ -61,7 +61,8 @@ struct Memory::Impl {
         }
 
 #ifdef __linux__
-        buffer.emplace(system.DeviceMemory().buffer);
+        heap_tracker.emplace(system.DeviceMemory().buffer);
+        buffer = std::addressof(*heap_tracker);
 #else
         buffer = std::addressof(system.DeviceMemory().buffer);
 #endif
@@ -1023,8 +1024,9 @@ struct Memory::Impl {
     std::span<Core::GPUDirtyMemoryManager> gpu_dirty_managers;
     std::mutex sys_core_guard;
 
+    std::optional<Common::HeapTracker> heap_tracker;
 #ifdef __linux__
-    std::optional<Common::HeapTracker> buffer;
+    Common::HeapTracker* buffer{};
 #else
     Common::HostMemory* buffer{};
 #endif
@@ -1228,7 +1230,22 @@ bool Memory::InvalidateNCE(Common::ProcessAddress vaddr, size_t size) {
     if (rasterizer) {
         impl->InvalidateGPUMemory(ptr, size);
     }
+
+#ifdef __linux__
+    if (!rasterizer && mapped) {
+        impl->buffer->DeferredMapSeparateHeap(GetInteger(vaddr));
+    }
+#endif
+
     return mapped && ptr != nullptr;
 }
 
+bool Memory::InvalidateSeparateHeap(void* fault_address) {
+#ifdef __linux__
+    return impl->buffer->DeferredMapSeparateHeap(static_cast<u8*>(fault_address));
+#else
+    return false;
+#endif
+}
+
 } // namespace Core::Memory

src/core/memory.h

@@ -487,8 +487,13 @@ public:
      *              marked as debug or non-debug.
      */
     void MarkRegionDebug(Common::ProcessAddress vaddr, u64 size, bool debug);
+
     void SetGPUDirtyManagers(std::span<Core::GPUDirtyMemoryManager> managers);
+
     bool InvalidateNCE(Common::ProcessAddress vaddr, size_t size);
+
+    bool InvalidateSeparateHeap(void* fault_address);
+
 private:
     Core::System& system;
 

src/dynarmic/externals/CMakeLists.txt

@@ -60,7 +60,7 @@ AddJsonPackage(
 #     endif()
 # endif()
 
-# unordered_dense - already in root
+# unordered_dense
 
 # AddJsonPackage(
 #     NAME unordered-dense