[Linux] Implement fast mutex similar to Windows version

src/xenia/base/mutex.cc

@@ -10,6 +10,10 @@
 #include "xenia/base/mutex.h"
 #if XE_PLATFORM_WIN32 == 1
 #include "xenia/base/platform_win.h"
+#elif XE_PLATFORM_LINUX == 1
+#include <linux/futex.h>
+#include <sys/syscall.h>
+#include <unistd.h>
 #endif
 
 namespace xe {
@@ -70,6 +74,179 @@ bool xe_fast_mutex::try_lock() {
   }
   return false;
 }
+#elif XE_PLATFORM_LINUX == 1 && XE_ENABLE_FAST_LINUX_MUTEX == 1
+
+namespace {
+
+inline int futex_wait(std::atomic<uint32_t>* addr, uint32_t expected) {
+  return syscall(SYS_futex, addr, FUTEX_WAIT_PRIVATE, expected, nullptr,
+                 nullptr, 0);
+}
+
+inline int futex_wake(std::atomic<uint32_t>* addr, int count) {
+  return syscall(SYS_futex, addr, FUTEX_WAKE_PRIVATE, count, nullptr, nullptr,
+                 0);
+}
+
+inline pid_t gettid() { return static_cast<pid_t>(syscall(SYS_gettid)); }
+
+}  // namespace
+
+// xe_global_mutex implementation (recursive)
+void xe_global_mutex::lock() {
+  pid_t self = gettid();
+
+  // Fast path: check if we already own it (recursive lock)
+  if (owner_.load(std::memory_order_relaxed) == self) {
+    ++recursion_count_;
+    return;
+  }
+
+  // Try to acquire with a simple CAS first (uncontended case)
+  uint32_t expected = 0;
+  if (XE_LIKELY(state_.compare_exchange_strong(
+          expected, 1, std::memory_order_acquire, std::memory_order_relaxed))) {
+    owner_.store(self, std::memory_order_relaxed);
+    recursion_count_ = 1;
+    return;
+  }
+
+  lock_slow();
+}
+
+void xe_global_mutex::lock_slow() {
+  pid_t self = gettid();
+
+  // Spin phase
+  for (int i = 0; i < XE_LINUX_MUTEX_SPINCOUNT; ++i) {
+#if XE_ARCH_AMD64 == 1
+    _mm_pause();
+#endif
+    uint32_t expected = 0;
+    if (state_.compare_exchange_strong(expected, 1, std::memory_order_acquire,
+                                       std::memory_order_relaxed)) {
+      owner_.store(self, std::memory_order_relaxed);
+      recursion_count_ = 1;
+      return;
+    }
+  }
+
+  // Slow path: use futex
+  while (true) {
+    // Mark as contended (state = 2) and wait
+    uint32_t state = state_.exchange(2, std::memory_order_acquire);
+    if (state == 0) {
+      // We got the lock while marking contended
+      owner_.store(self, std::memory_order_relaxed);
+      recursion_count_ = 1;
+      return;
+    }
+
+    // Wait on futex
+    futex_wait(&state_, 2);
+
+    // Try to acquire after wakeup
+    uint32_t expected = 0;
+    if (state_.compare_exchange_strong(expected, 2, std::memory_order_acquire,
+                                       std::memory_order_relaxed)) {
+      owner_.store(self, std::memory_order_relaxed);
+      recursion_count_ = 1;
+      return;
+    }
+  }
+}
+
+void xe_global_mutex::unlock() {
+  if (--recursion_count_ > 0) {
+    return;  // Still have recursive locks
+  }
+
+  owner_.store(0, std::memory_order_relaxed);
+
+  // If state was 2 (contended), we need to wake a waiter
+  if (state_.exchange(0, std::memory_order_release) == 2) {
+    futex_wake(&state_, 1);
+  }
+}
+
+bool xe_global_mutex::try_lock() {
+  pid_t self = gettid();
+
+  // Check for recursive lock
+  if (owner_.load(std::memory_order_relaxed) == self) {
+    ++recursion_count_;
+    return true;
+  }
+
+  uint32_t expected = 0;
+  if (state_.compare_exchange_strong(expected, 1, std::memory_order_acquire,
+                                     std::memory_order_relaxed)) {
+    owner_.store(self, std::memory_order_relaxed);
+    recursion_count_ = 1;
+    return true;
+  }
+  return false;
+}
+
+// xe_fast_mutex implementation (non-recursive)
+void xe_fast_mutex::lock() {
+  // Fast path: uncontended
+  uint32_t expected = 0;
+  if (XE_LIKELY(state_.compare_exchange_strong(
+          expected, 1, std::memory_order_acquire, std::memory_order_relaxed))) {
+    return;
+  }
+
+  lock_slow();
+}
+
+void xe_fast_mutex::lock_slow() {
+  // Spin phase
+  for (int i = 0; i < XE_LINUX_MUTEX_SPINCOUNT; ++i) {
+#if XE_ARCH_AMD64 == 1
+    _mm_pause();
+#endif
+    uint32_t expected = 0;
+    if (state_.compare_exchange_strong(expected, 1, std::memory_order_acquire,
+                                       std::memory_order_relaxed)) {
+      return;
+    }
+  }
+
+  // Slow path: use futex
+  while (true) {
+    // Mark as contended (state = 2) and wait
+    uint32_t state = state_.exchange(2, std::memory_order_acquire);
+    if (state == 0) {
+      // We got the lock while marking contended
+      return;
+    }
+
+    // Wait on futex
+    futex_wait(&state_, 2);
+
+    // Try to acquire after wakeup
+    uint32_t expected = 0;
+    if (state_.compare_exchange_strong(expected, 2, std::memory_order_acquire,
+                                       std::memory_order_relaxed)) {
+      return;
+    }
+  }
+}
+
+void xe_fast_mutex::unlock() {
+  // If state was 2 (contended), we need to wake a waiter
+  if (state_.exchange(0, std::memory_order_release) == 2) {
+    futex_wake(&state_, 1);
+  }
+}
+
+bool xe_fast_mutex::try_lock() {
+  uint32_t expected = 0;
+  return state_.compare_exchange_strong(expected, 1, std::memory_order_acquire,
+                                        std::memory_order_relaxed);
+}
+
 #endif
 global_mutex_type& global_critical_region::mutex() {
   static global_mutex_type global_mutex;

src/xenia/base/mutex.h

@@ -9,6 +9,7 @@
 
 #ifndef XENIA_BASE_MUTEX_H_
 #define XENIA_BASE_MUTEX_H_
+#include <atomic>
 #include <mutex>
 #include <thread>
 #include "platform.h"
@@ -19,6 +20,7 @@
 #endif
 #include "memory.h"
 #define XE_ENABLE_FAST_WIN32_MUTEX 1
+#define XE_ENABLE_FAST_LINUX_MUTEX 1
 namespace xe {
 
 #if XE_PLATFORM_WIN32 == 1 && XE_ENABLE_FAST_WIN32_MUTEX == 1
@@ -80,6 +82,78 @@ class xe_unlikely_mutex {
   void unlock() { mut.exchange(0); }
   bool try_lock() { return _tryget(); }
 };
+using xe_mutex = xe_fast_mutex;
+#elif XE_PLATFORM_LINUX == 1 && XE_ENABLE_FAST_LINUX_MUTEX == 1
+
+#define XE_LINUX_MUTEX_SPINCOUNT 128
+
+// Fast recursive mutex for Linux using futex
+// Mimics Windows CRITICAL_SECTION behavior: spin before blocking
+class alignas(4096) xe_global_mutex {
+  std::atomic<uint32_t> state_{0};  // 0 = unlocked, 1 = locked, 2 = contended
+  std::atomic<pid_t> owner_{0};
+  uint32_t recursion_count_{0};
+
+  void lock_slow();
+
+ public:
+  xe_global_mutex() = default;
+  ~xe_global_mutex() = default;
+
+  void lock();
+  void unlock();
+  bool try_lock();
+};
+using global_mutex_type = xe_global_mutex;
+
+// Fast non-recursive mutex for Linux using futex
+class alignas(64) xe_fast_mutex {
+  std::atomic<uint32_t> state_{0};  // 0 = unlocked, 1 = locked, 2 = contended
+
+  void lock_slow();
+
+ public:
+  xe_fast_mutex() = default;
+  ~xe_fast_mutex() = default;
+
+  void lock();
+  void unlock();
+  bool try_lock();
+};
+
+// xe_unlikely_mutex remains a simple spinlock for Linux too
+class xe_unlikely_mutex {
+  std::atomic<uint32_t> mut{0};
+  bool _tryget() {
+    uint32_t lock_expected = 0;
+    return mut.compare_exchange_strong(
+        lock_expected, 1, std::memory_order_acquire, std::memory_order_relaxed);
+  }
+
+ public:
+  xe_unlikely_mutex() = default;
+  ~xe_unlikely_mutex() = default;
+
+  void lock() {
+    if (XE_LIKELY(_tryget())) {
+      return;
+    }
+    // Spin a bit before yielding
+    for (int i = 0; i < XE_LINUX_MUTEX_SPINCOUNT; ++i) {
+#if XE_ARCH_AMD64 == 1
+      _mm_pause();
+#endif
+      if (_tryget()) return;
+    }
+    // Fall back to yielding
+    while (!_tryget()) {
+      std::this_thread::yield();
+    }
+  }
+  void unlock() { mut.store(0, std::memory_order_release); }
+  bool try_lock() { return _tryget(); }
+};
+
 using xe_mutex = xe_fast_mutex;
 #else
 using global_mutex_type = std::recursive_mutex;
@@ -150,11 +224,7 @@ class global_critical_region {
     return global_unique_lock_type(mutex());
   }
 
-  static inline void PrepareToAcquire() {
+  static inline void PrepareToAcquire() { swcache::PrefetchW(&mutex()); }
-#if XE_PLATFORM_WIN32 == 1
-    swcache::PrefetchW(&mutex());
-#endif
-  }
 
   // Acquires a deferred lock on the global critical section.
   static inline global_unique_lock_type AcquireDeferred() {