Merge pull request #12466 from liamwhite/sh2

 core: track separate heap allocation for linux 

5 files changed, 395 insertions, 7 deletions

diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index b58a7073f..8c57d47c6 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -64,6 +64,8 @@ add_library(common STATIC
     fs/path_util.cpp
     fs/path_util.h
     hash.h
+    heap_tracker.cpp
+    heap_tracker.h
     hex_util.cpp
     hex_util.h
     host_memory.cpp
diff --git a/src/common/heap_tracker.cpp b/src/common/heap_tracker.cpp
new file mode 100644
index 000000000..683208795
--- /dev/null
+++ b/src/common/heap_tracker.cpp
@@ -0,0 +1,281 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <fstream>
+#include <vector>
+
+#include "common/heap_tracker.h"
+#include "common/logging/log.h"
+
+namespace Common {
+
+namespace {
+
+s64 GetMaxPermissibleResidentMapCount() {
+    // Default value.
+    s64 value = 65530;
+
+    // Try to read how many mappings we can make.
+    std::ifstream s("/proc/sys/vm/max_map_count");
+    s >> value;
+
+    // Print, for debug.
+    LOG_INFO(HW_Memory, "Current maximum map count: {}", value);
+
+    // Allow 20000 maps for other code and to account for split inaccuracy.
+    return std::max<s64>(value - 20000, 0);
+}
+
+} // namespace
+
+HeapTracker::HeapTracker(Common::HostMemory& buffer)
+    : m_buffer(buffer), m_max_resident_map_count(GetMaxPermissibleResidentMapCount()) {}
+HeapTracker::~HeapTracker() = default;
+
+void HeapTracker::Map(size_t virtual_offset, size_t host_offset, size_t length,
+                      MemoryPermission perm, bool is_separate_heap) {
+    // 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.
+        std::scoped_lock lk{m_lock};
+
+        auto* const map = new SeparateHeapMap{
+            .vaddr = virtual_offset,
+            .paddr = host_offset,
+            .size = length,
+            .tick = m_tick++,
+            .perm = perm,
+            .is_resident = false,
+        };
+
+        // Insert into mappings.
+        m_map_count++;
+        m_mappings.insert(*map);
+    }
+
+    // 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(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 (item->is_resident) {
+                ASSERT(--m_resident_map_count >= 0);
+                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);
+}
+
+void HeapTracker::Protect(size_t virtual_offset, size_t size, MemoryPermission perm) {
+    // Ensure no rebuild occurs while reprotecting.
+    std::shared_lock lk{m_rebuild_lock};
+
+    // Split at the boundaries of the region we are reprotecting.
+    this->SplitHeapMap(virtual_offset, size);
+
+    // Declare tracking variables.
+    const VAddr end = virtual_offset + size;
+    VAddr cur = virtual_offset;
+
+    while (cur < end) {
+        VAddr next = cur;
+        bool should_protect = false;
+
+        {
+            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.nfind(key);
+
+            if (it == m_mappings.end()) {
+                // There are no separate heap mappings remaining.
+                next = end;
+                should_protect = true;
+            } else if (it->vaddr == cur) {
+                // We are in range.
+                // Update permission bits.
+                it->perm = perm;
+
+                // Determine next address and whether we should protect.
+                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->vaddr;
+                should_protect = true;
+            }
+        }
+
+        // Clamp to end.
+        next = std::min(next, end);
+
+        // Reprotect, if we need to.
+        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.
+    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 (size_t i = 0; i < evict_count && it != m_resident_mappings.end(); i++) {
+        // Unmark and unmap.
+        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);
+    }
+}
+
+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) {
+    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 = left->tick,
+        .perm = left->perm,
+        .is_resident = left->is_resident,
+    };
+
+    // Insert the new right map.
+    m_map_count++;
+    m_mappings.insert(*right);
+
+    // If resident, also insert into resident map.
+    if (right->is_resident) {
+        m_resident_map_count++;
+        m_resident_mappings.insert(*right);
+    }
+}
+
+HeapTracker::AddrTree::iterator HeapTracker::GetNearestHeapMapLocked(VAddr offset) {
+    const SeparateHeapMap key{
+        .vaddr = offset,
+    };
+
+    return m_mappings.find(key);
+}
+
+} // namespace Common
diff --git a/src/common/heap_tracker.h b/src/common/heap_tracker.h
new file mode 100644
index 000000000..ee5b0bf43
--- /dev/null
+++ b/src/common/heap_tracker.h
@@ -0,0 +1,98 @@
+// 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 "common/host_memory.h"
+#include "common/intrusive_red_black_tree.h"
+
+namespace Common {
+
+struct SeparateHeapMap {
+    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);
+        }
+    }
+};
+
+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 Unmap(size_t virtual_offset, size_t size, bool is_separate_heap);
+    void Protect(size_t virtual_offset, size_t length, MemoryPermission perm);
+    u8* VirtualBasePointer() {
+        return m_buffer.VirtualBasePointer();
+    }
+
+    bool DeferredMapSeparateHeap(u8* fault_address);
+    bool DeferredMapSeparateHeap(size_t virtual_offset);
+
+private:
+    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();
+
+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{};
+    s64 m_resident_map_count{};
+    size_t m_tick{};
+};
+
+} // namespace Common
diff --git a/src/common/host_memory.cpp b/src/common/host_memory.cpp
index e540375b8..860c39e6a 100644
--- a/src/common/host_memory.cpp
+++ b/src/common/host_memory.cpp
@@ -679,7 +679,7 @@ HostMemory::HostMemory(HostMemory&&) noexcept = default;
 HostMemory& HostMemory::operator=(HostMemory&&) noexcept = default;
 
 void HostMemory::Map(size_t virtual_offset, size_t host_offset, size_t length,
-                     MemoryPermission perms) {
+                     MemoryPermission perms, bool separate_heap) {
     ASSERT(virtual_offset % PageAlignment == 0);
     ASSERT(host_offset % PageAlignment == 0);
     ASSERT(length % PageAlignment == 0);
@@ -691,7 +691,7 @@ void HostMemory::Map(size_t virtual_offset, size_t host_offset, size_t length,
     impl->Map(virtual_offset + virtual_base_offset, host_offset, length, perms);
 }
 
-void HostMemory::Unmap(size_t virtual_offset, size_t length) {
+void HostMemory::Unmap(size_t virtual_offset, size_t length, bool separate_heap) {
     ASSERT(virtual_offset % PageAlignment == 0);
     ASSERT(length % PageAlignment == 0);
     ASSERT(virtual_offset + length <= virtual_size);
@@ -701,14 +701,16 @@ void HostMemory::Unmap(size_t virtual_offset, size_t length) {
     impl->Unmap(virtual_offset + virtual_base_offset, length);
 }
 
-void HostMemory::Protect(size_t virtual_offset, size_t length, bool read, bool write,
-                         bool execute) {
+void HostMemory::Protect(size_t virtual_offset, size_t length, MemoryPermission perm) {
     ASSERT(virtual_offset % PageAlignment == 0);
     ASSERT(length % PageAlignment == 0);
     ASSERT(virtual_offset + length <= virtual_size);
     if (length == 0 || !virtual_base || !impl) {
         return;
     }
+    const bool read = True(perm & MemoryPermission::Read);
+    const bool write = True(perm & MemoryPermission::Write);
+    const bool execute = True(perm & MemoryPermission::Execute);
     impl->Protect(virtual_offset + virtual_base_offset, length, read, write, execute);
 }
 
diff --git a/src/common/host_memory.h b/src/common/host_memory.h
index 747c5850c..72fbb05af 100644
--- a/src/common/host_memory.h
+++ b/src/common/host_memory.h
@@ -40,11 +40,12 @@ public:
     HostMemory(HostMemory&& other) noexcept;
     HostMemory& operator=(HostMemory&& other) noexcept;
 
-    void Map(size_t virtual_offset, size_t host_offset, size_t length, MemoryPermission perms);
+    void Map(size_t virtual_offset, size_t host_offset, size_t length, MemoryPermission perms,
+             bool separate_heap);
 
-    void Unmap(size_t virtual_offset, size_t length);
+    void Unmap(size_t virtual_offset, size_t length, bool separate_heap);
 
-    void Protect(size_t virtual_offset, size_t length, bool read, bool write, bool execute = false);
+    void Protect(size_t virtual_offset, size_t length, MemoryPermission perms);
 
     void EnableDirectMappedAddress();
 
@@ -64,6 +65,10 @@ public:
         return virtual_base;
     }
 
+    bool IsInVirtualRange(void* address) const noexcept {
+        return address >= virtual_base && address < virtual_base + virtual_size;
+    }
+
 private:
     size_t backing_size{};
     size_t virtual_size{};