LRUCacheShard cache line size alignment

Summary:
combining #2568 and #2612.
Closes https://github.com/facebook/rocksdb/pull/2620

Differential Revision: D5464394

Pulled By: IslamAbdelRahman

fbshipit-source-id: 9f71d3058dd6adaf02ce3b2de3a81a1228009778

cache/lru_cache.cc

@@ -22,7 +22,7 @@
 
 namespace rocksdb {
 
-LRUHandleTable::LRUHandleTable() : length_(0), elems_(0), list_(nullptr) {
+LRUHandleTable::LRUHandleTable() : list_(nullptr), length_(0), elems_(0) {
   Resize();
 }
 
@@ -100,7 +100,7 @@ void LRUHandleTable::Resize() {
 }
 
 LRUCacheShard::LRUCacheShard()
-    : usage_(0), lru_usage_(0), high_pri_pool_usage_(0) {
+    : high_pri_pool_usage_(0), usage_(0), lru_usage_(0) {
   // Make empty circular linked list
   lru_.next = &lru_;
   lru_.prev = &lru_;
@@ -233,6 +233,14 @@ void LRUCacheShard::EvictFromLRU(size_t charge,
   }
 }
 
+void* LRUCacheShard::operator new(size_t size) {
+  return rocksdb::port::cacheline_aligned_alloc(size);
+}
+
+void LRUCacheShard::operator delete(void *memblock) {
+  rocksdb::port::cacheline_aligned_free(memblock);
+}
+
 void LRUCacheShard::SetCapacity(size_t capacity) {
   autovector<LRUHandle*> last_reference_list;
   {
@@ -449,7 +457,14 @@ LRUCache::LRUCache(size_t capacity, int num_shard_bits,
                    bool strict_capacity_limit, double high_pri_pool_ratio)
     : ShardedCache(capacity, num_shard_bits, strict_capacity_limit) {
   num_shards_ = 1 << num_shard_bits;
+#if defined(_MSC_VER)
+#pragma warning(push)
+#pragma warning(disable: 4316) // We've validated the alignment with the new operators
+#endif
   shards_ = new LRUCacheShard[num_shards_];
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
   SetCapacity(capacity);
   SetStrictCapacityLimit(strict_capacity_limit);
   for (int i = 0; i < num_shards_; i++) {

cache/lru_cache.h

@@ -148,13 +148,13 @@ class LRUHandleTable {
 
   // The table consists of an array of buckets where each bucket is
   // a linked list of cache entries that hash into the bucket.
+  LRUHandle** list_;
   uint32_t length_;
   uint32_t elems_;
-  LRUHandle** list_;
 };
 
 // A single shard of sharded cache.
-class LRUCacheShard : public CacheShard {
+class ALIGN_AS(CACHE_LINE_SIZE) LRUCacheShard : public CacheShard {
  public:
   LRUCacheShard();
   virtual ~LRUCacheShard();
@@ -202,6 +202,11 @@ class LRUCacheShard : public CacheShard {
   //  not threadsafe
   size_t TEST_GetLRUSize();
 
+  // Overloading to aligned it to cache line size
+  void* operator new(size_t);
+
+  void operator delete(void *);
+
  private:
   void LRU_Remove(LRUHandle* e);
   void LRU_Insert(LRUHandle* e);
@@ -223,12 +228,6 @@ class LRUCacheShard : public CacheShard {
   // Initialized before use.
   size_t capacity_;
 
-  // Memory size for entries residing in the cache
-  size_t usage_;
-
-  // Memory size for entries residing only in the LRU list
-  size_t lru_usage_;
-
   // Memory size for entries in high-pri pool.
   size_t high_pri_pool_usage_;
 
@@ -242,11 +241,6 @@ class LRUCacheShard : public CacheShard {
   // Remember the value to avoid recomputing each time.
   double high_pri_pool_capacity_;
 
-  // mutex_ protects the following state.
-  // We don't count mutex_ as the cache's internal state so semantically we
-  // don't mind mutex_ invoking the non-const actions.
-  mutable port::Mutex mutex_;
-
   // Dummy head of LRU list.
   // lru.prev is newest entry, lru.next is oldest entry.
   // LRU contains items which can be evicted, ie reference only by cache
@@ -255,7 +249,29 @@ class LRUCacheShard : public CacheShard {
   // Pointer to head of low-pri pool in LRU list.
   LRUHandle* lru_low_pri_;
 
+  // ------------^^^^^^^^^^^^^-----------
+  // Not frequently modified data members
+  // ------------------------------------
+  //
+  // We separate data members that are updated frequently from the ones that
+  // are not frequently updated so that they don't share the same cache line
+  // which will lead into false cache sharing
+  //
+  // ------------------------------------
+  // Frequently modified data members
+  // ------------vvvvvvvvvvvvv-----------
   LRUHandleTable table_;
+
+  // Memory size for entries residing in the cache
+  size_t usage_;
+
+  // Memory size for entries residing only in the LRU list
+  size_t lru_usage_;
+
+  // mutex_ protects the following state.
+  // We don't count mutex_ as the cache's internal state so semantically we
+  // don't mind mutex_ invoking the non-const actions.
+  mutable port::Mutex mutex_;
 };
 
 class LRUCache : public ShardedCache {

cache/lru_cache_test.cc

@@ -17,7 +17,16 @@ class LRUCacheTest : public testing::Test {
   ~LRUCacheTest() {}
 
   void NewCache(size_t capacity, double high_pri_pool_ratio = 0.0) {
-    cache_.reset(new LRUCacheShard());
+    cache_.reset(
+#if defined(_MSC_VER)
+#pragma warning(push)
+#pragma warning(disable: 4316) // We've validated the alignment with the new operators
+#endif
+      new LRUCacheShard()
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
+    );
     cache_->SetCapacity(capacity);
     cache_->SetStrictCapacityLimit(false);
     cache_->SetHighPriorityPoolRatio(high_pri_pool_ratio);

port/port_posix.cc

@@ -184,5 +184,22 @@ int GetMaxOpenFiles() {
   return -1;
 }
 
+void *cacheline_aligned_alloc(size_t size) {
+#if defined (_ISOC11_SOURCE)
+  return aligned_alloc(CACHE_LINE_SIZE, size);
+#elif ( _POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600 || defined(__APPLE__))
+  void *m;
+  errno = posix_memalign(&m, CACHE_LINE_SIZE, size);
+  return errno ? NULL : m;
+#else
+  return malloc(size);
+#endif
+}
+
+void cacheline_aligned_free(void *memblock) {
+  free(memblock);
+}
+
+
 }  // namespace port
 }  // namespace rocksdb

port/port_posix.h

@@ -193,6 +193,13 @@ extern void InitOnce(OnceType* once, void (*initializer)());
   #endif
 #endif
 
+
+extern void *cacheline_aligned_alloc(size_t size);
+
+extern void cacheline_aligned_free(void *memblock);
+
+#define ALIGN_AS(n) alignas(n)
+
 #define PREFETCH(addr, rw, locality) __builtin_prefetch(addr, rw, locality)
 
 extern void Crash(const std::string& srcfile, int srcline);

port/win/port_win.h

@@ -27,6 +27,7 @@
 #include <mutex>
 #include <limits>
 #include <condition_variable>
+#include <malloc.h>
 
 #include <stdint.h>
 
@@ -239,6 +240,23 @@ extern void InitOnce(OnceType* once, void (*initializer)());
 #define CACHE_LINE_SIZE 64U
 #endif
 
+
+inline void *cacheline_aligned_alloc(size_t size) {
+  return _aligned_malloc(CACHE_LINE_SIZE, size);
+}
+
+inline void cacheline_aligned_free(void *memblock) {
+  _aligned_free(memblock);
+}
+
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=52991 for MINGW32
+// could not be worked around with by -mno-ms-bitfields
+#ifndef __MINGW32__
+#define ALIGN_AS(n) __declspec(align(n))
+#else
+#define ALIGN_AS(n)
+#endif
+
 static inline void AsmVolatilePause() {
 #if defined(_M_IX86) || defined(_M_X64)
   YieldProcessor();