Add option to use futex APIs for spinlock

During the discussion of [1] it has been suggested that the congestion could be reduced by relying on C++20 futex APIs. Additionally we incorporated the `try_lock()` design from [2]. [1]: #146 [2]: https://rigtorp.se/spinlock/ Co-authored-by: Milian Wolff <[email protected]>
efficient · Jul 7, 2022 · 6175e91 · 6175e91
1 parent 735a74b
commit 6175e91
Show file tree

Hide file tree

Showing 3 changed files with 44 additions and 5 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -5,6 +5,7 @@ project(libcuckoo
 
 # put these in the cache so they show up in ccmake
 option (BUILD_EXAMPLES "build example libcuckoo programs")
+option (LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT "Optimize the spinlock implementation with C++20 futex support" OFF)
 
 # Add the libcuckoo interface target
 add_subdirectory(libcuckoo)

diff --git a/libcuckoo/CMakeLists.txt b/libcuckoo/CMakeLists.txt
@@ -17,6 +17,10 @@ add_library(libcuckoo::libcuckoo ALIAS libcuckoo)
 # will have c++11 turned on in their compile when they use this target.
 # XXX: newer cmakes have a "cxx_std_11" feature that could be used
 target_compile_features (libcuckoo INTERFACE cxx_constexpr)
+if (LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT)
+    target_compile_features(libcuckoo INTERFACE cxx_std_20) 
+    target_compile_definitions(libcuckoo INTERFACE LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT=1)
+endif ()
 
 # Include relative to the base directory
 target_include_directories(libcuckoo INTERFACE

diff --git a/libcuckoo/cuckoohash_map.hh b/libcuckoo/cuckoohash_map.hh
@@ -819,12 +819,29 @@ private:
   LIBCUCKOO_SQUELCH_PADDING_WARNING
   class LIBCUCKOO_ALIGNAS(64) spinlock {
   public:
-    spinlock() : elem_counter_(0), is_migrated_(true) { lock_.clear(); }
+    spinlock() noexcept
+        :
+#if LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
+          lock_(),
+#endif
+          elem_counter_(0), is_migrated_(true) {
+
+#if !LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
+      lock_.clear();
+#endif
+    }
 
     spinlock(const spinlock &other) noexcept
-        : elem_counter_(other.elem_counter()),
+        :
+#if LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
+          lock_(),
+#endif
+          elem_counter_(other.elem_counter()),
           is_migrated_(other.is_migrated()) {
+
+#if !LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
       lock_.clear();
+#endif
     }
 
     spinlock &operator=(const spinlock &other) noexcept {
@@ -834,14 +851,31 @@ private:
     }
 
     void lock() noexcept {
-      while (lock_.test_and_set(std::memory_order_acq_rel))
-        ;
+      // Optimistically assume the lock is free
+      while (lock_.test_and_set(std::memory_order_acq_rel)) {
+#if LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
+        // Wait for lock to be released utilizing C++20 futex support
+        lock_.wait(true, std::memory_order::relaxed);
+#endif
+      }
     }
 
-    void unlock() noexcept { lock_.clear(std::memory_order_release); }
+    void unlock() noexcept {
+      lock_.clear(std::memory_order_release);
+#if LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
+      lock_.notify_one();
+#endif
+    }
 
     bool try_lock() noexcept {
+#if LIBCUCKOO_USE_CXX20_FUTEX_SUPPORT
+      // First do a relaxed load to check if lock is free in order to prevent
+      // unnecessary cache misses if someone does while(!try_lock())
+      return !lock_.test(std::memory_order_relaxed) &&
+             !lock_.test_and_set(std::memory_order_acq_rel);
+#else
       return !lock_.test_and_set(std::memory_order_acq_rel);
+#endif
     }
 
     counter_type &elem_counter() noexcept { return elem_counter_; }