X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=folly%2FMicroLock.h;h=511d06fa605860755b8f1052a4ebf391e0528346;hb=ba690cad3c4138fce4e55b319237bb9e01ca030f;hp=3ff158b8728570c968f3df9f022ec7842996feca;hpb=2f595596190c768c6c074c93c878ef8b9144fe55;p=folly.git

diff --git a/folly/MicroLock.h b/folly/MicroLock.h
index 3ff158b8..511d06fa 100644
--- a/folly/MicroLock.h
+++ b/folly/MicroLock.h
@@ -1,5 +1,5 @@
 /*
- * Copyright 2016 Facebook, Inc.
+ * Copyright 2017 Facebook, Inc.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -22,6 +22,12 @@
 #include <folly/detail/Futex.h>
 #include <folly/Portability.h>
 
+#if defined(__clang__)
+#define NO_SANITIZE_ADDRESS __attribute__((no_sanitize_address))
+#else
+#define NO_SANITIZE_ADDRESS
+#endif
+
 namespace folly {
 
 /**
@@ -43,11 +49,15 @@ namespace folly {
  * limit yourself to MicroLock slot zero, which will use the two
  * least-significant bits in the bottom byte.
  *
+ * (Note that such a union is safe only because MicroLock is based on
+ * a character type, and even under a strict interpretation of C++'s
+ * aliasing rules, character types may alias anything.)
+ *
  * MicroLock uses a dirty trick: it actually operates on the full
- * word-size, word-aligned bit of memory into which it is embedded.
+ * 32-bit, four-byte-aligned bit of memory into which it is embedded.
  * It never modifies bits outside the ones it's defined to modify, but
- * it _accesses_ all the bits in the word for purposes of
- * futex management.
+ * it _accesses_ all the bits in the 32-bit memory location for
+ * purposes of futex management.
  *
  * The MaxSpins template parameter controls the number of times we
  * spin trying to acquire the lock.  MaxYields controls the number of
@@ -76,16 +86,24 @@ namespace folly {
  *
  * (The virtual dispatch benchmark is provided for scale.)
  *
- * The contended case for MicroLock is likely to be worse compared to
- * std::mutex than the contended case is.  Make sure to benchmark your
- * particular workload.
+ * While the uncontended case for MicroLock is competitive with the
+ * glibc 2.2.0 implementation of std::mutex, std::mutex is likely to be
+ * faster in the contended case, because we need to wake up all waiters
+ * when we release.
+ *
+ * Make sure to benchmark your particular workload.
  *
  */
 
 class MicroLockCore {
  protected:
+#if defined(__SANITIZE_ADDRESS__) && !defined(__clang__) && \
+    (defined(__GNUC__) || defined(__GNUG__))
+  uint32_t lock_;
+#else
   uint8_t lock_;
-  inline detail::Futex<>* word() const;
+#endif
+  inline detail::Futex<>* word() const; // Well, halfword on 64-bit systems
   inline uint32_t baseShift(unsigned slot) const;
   inline uint32_t heldBit(unsigned slot) const;
   inline uint32_t waitBit(unsigned slot) const;
@@ -96,10 +114,10 @@ class MicroLockCore {
                            unsigned maxYields);
 
  public:
-  inline void unlock(unsigned slot);
+  inline void unlock(unsigned slot) NO_SANITIZE_ADDRESS;
   inline void unlock() { unlock(0); }
-  inline void init(unsigned slot) { lock_ &= ~(3U << (2 * slot)); }
-  inline void init() { init(0); }
+  // Initializes all the slots.
+  inline void init() { lock_ = 0; }
 };
 
 inline detail::Futex<>* MicroLockCore::word() const {
@@ -113,9 +131,8 @@ inline unsigned MicroLockCore::baseShift(unsigned slot) const {
 
   unsigned offset_bytes = (unsigned)((uintptr_t)&lock_ - (uintptr_t)word());
 
-  return kIsLittleEndian
-             ? offset_bytes * CHAR_BIT + slot * 2
-             : CHAR_BIT * (sizeof(uint32_t) - offset_bytes - 1) + slot * 2;
+  return (
+      unsigned)(kIsLittleEndian ? offset_bytes * CHAR_BIT + slot * 2 : CHAR_BIT * (sizeof(uint32_t) - offset_bytes - 1) + slot * 2);
 }
 
 inline uint32_t MicroLockCore::heldBit(unsigned slot) const {
@@ -147,9 +164,9 @@ void MicroLockCore::unlock(unsigned slot) {
 template <unsigned MaxSpins = 1000, unsigned MaxYields = 0>
 class MicroLockBase : public MicroLockCore {
  public:
-  inline void lock(unsigned slot);
+  inline void lock(unsigned slot) NO_SANITIZE_ADDRESS;
   inline void lock() { lock(0); }
-  inline bool try_lock(unsigned slot);
+  inline bool try_lock(unsigned slot) NO_SANITIZE_ADDRESS;
   inline bool try_lock() { return try_lock(0); }
 };
 
@@ -193,7 +210,7 @@ void MicroLockBase<MaxSpins, MaxYields>::lock(unsigned slot) {
                                      oldWord | heldBit(slot),
                                      std::memory_order_acquire,
                                      std::memory_order_relaxed)) {
-    // Fast uncontended case: seq_cst above is our memory barrier
+    // Fast uncontended case: memory_order_acquire above is our barrier
   } else {
     // lockSlowPath doesn't have any slot-dependent computation; it
     // just shifts the input bit.  Make sure its shifting produces the