smmu: use new range mutex construction for protecting counters

4 files changed, 97 insertions, 11 deletions

diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index e30fea268..85926fc8f 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -106,6 +106,7 @@ add_library(common STATIC
     precompiled_headers.h
     quaternion.h
     range_map.h
+    range_mutex.h
     reader_writer_queue.h
     ring_buffer.h
     ${CMAKE_CURRENT_BINARY_DIR}/scm_rev.cpp
diff --git a/src/common/range_mutex.h b/src/common/range_mutex.h
new file mode 100644
index 000000000..d6c949811
--- /dev/null
+++ b/src/common/range_mutex.h
@@ -0,0 +1,93 @@
+// SPDX-FileCopyrightText: 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <condition_variable>
+#include <mutex>
+
+#include "common/intrusive_list.h"
+
+namespace Common {
+
+class ScopedRangeLock;
+
+class RangeMutex {
+public:
+    explicit RangeMutex() = default;
+    ~RangeMutex() = default;
+
+private:
+    friend class ScopedRangeLock;
+
+    void Lock(ScopedRangeLock& l);
+    void Unlock(ScopedRangeLock& l);
+    bool HasIntersectionLocked(ScopedRangeLock& l);
+
+private:
+    std::mutex m_mutex;
+    std::condition_variable m_cv;
+
+    using LockList = Common::IntrusiveListBaseTraits<ScopedRangeLock>::ListType;
+    LockList m_list;
+};
+
+class ScopedRangeLock : public Common::IntrusiveListBaseNode<ScopedRangeLock> {
+public:
+    explicit ScopedRangeLock(RangeMutex& mutex, u64 address, u64 size)
+        : m_mutex(mutex), m_address(address), m_size(size) {
+        if (m_size > 0) {
+            m_mutex.Lock(*this);
+        }
+    }
+    ~ScopedRangeLock() {
+        if (m_size > 0) {
+            m_mutex.Unlock(*this);
+        }
+    }
+
+    u64 GetAddress() const {
+        return m_address;
+    }
+
+    u64 GetSize() const {
+        return m_size;
+    }
+
+private:
+    RangeMutex& m_mutex;
+    const u64 m_address{};
+    const u64 m_size{};
+};
+
+inline void RangeMutex::Lock(ScopedRangeLock& l) {
+    std::unique_lock lk{m_mutex};
+    m_cv.wait(lk, [&] { return !HasIntersectionLocked(l); });
+    m_list.push_back(l);
+}
+
+inline void RangeMutex::Unlock(ScopedRangeLock& l) {
+    {
+        std::scoped_lock lk{m_mutex};
+        m_list.erase(m_list.iterator_to(l));
+    }
+    m_cv.notify_all();
+}
+
+inline bool RangeMutex::HasIntersectionLocked(ScopedRangeLock& l) {
+    const auto cur_begin = l.GetAddress();
+    const auto cur_last = l.GetAddress() + l.GetSize() - 1;
+
+    for (const auto& other : m_list) {
+        const auto other_begin = other.GetAddress();
+        const auto other_last = other.GetAddress() + other.GetSize() - 1;
+
+        if (cur_begin <= other_last && other_begin <= cur_last) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+} // namespace Common
diff --git a/src/core/device_memory_manager.h b/src/core/device_memory_manager.h
index ffeed46cc..63823602c 100644
--- a/src/core/device_memory_manager.h
+++ b/src/core/device_memory_manager.h
@@ -10,6 +10,7 @@
 #include <mutex>
 
 #include "common/common_types.h"
+#include "common/range_mutex.h"
 #include "common/scratch_buffer.h"
 #include "common/virtual_buffer.h"
 
@@ -204,7 +205,7 @@ private:
         (1ULL << (device_virtual_bits - page_bits)) / subentries;
     using CachedPages = std::array<CounterEntry, num_counter_entries>;
     std::unique_ptr<CachedPages> cached_pages;
-    std::mutex counter_guard;
+    Common::RangeMutex counter_guard;
     std::mutex mapping_guard;
 };
 
diff --git a/src/core/device_memory_manager.inc b/src/core/device_memory_manager.inc
index eab8a2731..0a59000aa 100644
--- a/src/core/device_memory_manager.inc
+++ b/src/core/device_memory_manager.inc
@@ -508,12 +508,7 @@ void DeviceMemoryManager<Traits>::UnregisterProcess(Asid asid) {
 
 template <typename Traits>
 void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
-    std::unique_lock<std::mutex> lk(counter_guard, std::defer_lock);
-    const auto Lock = [&] {
-        if (!lk) {
-            lk.lock();
-        }
-    };
+    Common::ScopedRangeLock lk(counter_guard, addr, size);
     u64 uncache_begin = 0;
     u64 cache_begin = 0;
     u64 uncache_bytes = 0;
@@ -548,7 +543,6 @@ void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size
             }
             uncache_bytes += Memory::YUZU_PAGESIZE;
         } else if (uncache_bytes > 0) {
-            Lock();
             MarkRegionCaching(memory_device_inter, uncache_begin << Memory::YUZU_PAGEBITS,
                               uncache_bytes, false);
             uncache_bytes = 0;
@@ -559,7 +553,6 @@ void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size
             }
             cache_bytes += Memory::YUZU_PAGESIZE;
         } else if (cache_bytes > 0) {
-            Lock();
             MarkRegionCaching(memory_device_inter, cache_begin << Memory::YUZU_PAGEBITS, cache_bytes,
                               true);
             cache_bytes = 0;
@@ -567,12 +560,10 @@ void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size
         vpage++;
     }
     if (uncache_bytes > 0) {
-        Lock();
         MarkRegionCaching(memory_device_inter, uncache_begin << Memory::YUZU_PAGEBITS, uncache_bytes,
                           false);
     }
     if (cache_bytes > 0) {
-        Lock();
         MarkRegionCaching(memory_device_inter, cache_begin << Memory::YUZU_PAGEBITS, cache_bytes,
                           true);
     }