-
-#if !FOLLY_X64
-# error "SmallLocks.h is currently x64-only."
-#endif
-
-namespace folly {
-
-//////////////////////////////////////////////////////////////////////
-
-namespace detail {
-
- /*
- * A helper object for the contended case. Starts off with eager
- * spinning, and falls back to sleeping for small quantums.
- */
- class Sleeper {
- static const uint32_t kMaxActiveSpin = 4000;
-
- uint32_t spinCount;
-
- public:
- Sleeper() : spinCount(0) {}
-
- void wait() {
- if (spinCount < kMaxActiveSpin) {
- ++spinCount;
- asm volatile("pause");
- } else {
- /*
- * Always sleep 0.5ms, assuming this will make the kernel put
- * us down for whatever its minimum timer resolution is (in
- * linux this varies by kernel version from 1ms to 10ms).
- */
- struct timespec ts = { 0, 500000 };
- nanosleep(&ts, nullptr);
- }
- }
- };
-
-}
-
-//////////////////////////////////////////////////////////////////////
-
-/*
- * A really, *really* small spinlock for fine-grained locking of lots
- * of teeny-tiny data.
- *
- * Zero initializing these is guaranteed to be as good as calling
- * init(), since the free state is guaranteed to be all-bits zero.
- *
- * This class should be kept a POD, so we can used it in other packed
- * structs (gcc does not allow __attribute__((__packed__)) on structs that
- * contain non-POD data). This means avoid adding a constructor, or
- * making some members private, etc.
- */
-struct MicroSpinLock {
- enum { FREE = 0, LOCKED = 1 };
- // lock_ can't be std::atomic<> to preserve POD-ness.
- uint8_t lock_;
-
- // Initialize this MSL. It is unnecessary to call this if you
- // zero-initialize the MicroSpinLock.
- void init() {
- payload()->store(FREE);
- }
-
- bool try_lock() {
- return cas(FREE, LOCKED);
- }
-
- void lock() {
- detail::Sleeper sleeper;
- do {
- while (payload()->load() != FREE) {
- sleeper.wait();
- }
- } while (!try_lock());
- DCHECK(payload()->load() == LOCKED);
- }
-
- void unlock() {
- CHECK(payload()->load() == LOCKED);
- payload()->store(FREE, std::memory_order_release);
- }
-
- private:
- std::atomic<uint8_t>* payload() {
- return reinterpret_cast<std::atomic<uint8_t>*>(&this->lock_);
- }
-
- bool cas(uint8_t compare, uint8_t newVal) {
- return std::atomic_compare_exchange_strong_explicit(payload(), &compare, newVal,
- std::memory_order_acquire,
- std::memory_order_relaxed);
- }
-};
-
-//////////////////////////////////////////////////////////////////////
-
-/*
- * Spin lock on a single bit in an integral type. You can use this
- * with 16, 32, or 64-bit integral types.
- *
- * This is useful if you want a small lock and already have an int
- * with a bit in it that you aren't using. But note that it can't be
- * as small as MicroSpinLock (1 byte), if you don't already have a
- * convenient int with an unused bit lying around to put it on.
- *
- * To construct these, either use init() or zero initialize. We don't
- * have a real constructor because we want this to be a POD type so we
- * can put it into packed structs.
- */
-template<class IntType, int Bit = sizeof(IntType) * 8 - 1>
-struct PicoSpinLock {
- // Internally we deal with the unsigned version of the type.
- typedef typename std::make_unsigned<IntType>::type UIntType;
-
- static_assert(std::is_integral<IntType>::value,
- "PicoSpinLock needs an integral type");
- static_assert(sizeof(IntType) == 2 || sizeof(IntType) == 4 ||
- sizeof(IntType) == 8,
- "PicoSpinLock can't work on integers smaller than 2 bytes");
-
- public:
- static const UIntType kLockBitMask_ = UIntType(1) << Bit;
- UIntType lock_;
-
- /*
- * You must call this function before using this class, if you
- * default constructed it. If you zero-initialized it you can
- * assume the PicoSpinLock is in a valid unlocked state with
- * getData() == 0.
- *
- * (This doesn't use a constructor because we want to be a POD.)
- */
- void init(IntType initialValue = 0) {
- CHECK(!(initialValue & kLockBitMask_));
- lock_ = initialValue;
- }
-
- /*
- * Returns the value of the integer we using for our lock, except
- * with the bit we are using as a lock cleared, regardless of
- * whether the lock is held.
- *
- * It is 'safe' to call this without holding the lock. (As in: you
- * get the same guarantees for simultaneous accesses to an integer
- * as you normally get.)
- */
- IntType getData() const {
- return static_cast<IntType>(lock_ & ~kLockBitMask_);
- }
-
- /*
- * Set the value of the other bits in our integer.
- *
- * Don't use this when you aren't holding the lock, unless it can be
- * guaranteed that no other threads may be trying to use this.
- */
- void setData(IntType w) {
- CHECK(!(w & kLockBitMask_));
- lock_ = (lock_ & kLockBitMask_) | w;
- }
-
- /*
- * Try to get the lock without blocking: returns whether or not we
- * got it.
- */
- bool try_lock() const {
- bool ret = false;
-
-#define FB_DOBTS(size) \
- asm volatile("lock; bts" #size " %1, (%2); setnc %0" \
- : "=r" (ret) \
- : "i" (Bit), \
- "r" (&lock_) \
- : "memory", "flags")
-
- switch (sizeof(IntType)) {
- case 2: FB_DOBTS(w); break;
- case 4: FB_DOBTS(l); break;
- case 8: FB_DOBTS(q); break;
- }
-
-#undef FB_DOBTS
-
- return ret;
- }
-
- /*
- * Block until we can acquire the lock. Uses Sleeper to wait.
- */
- void lock() const {
- detail::Sleeper sleeper;
- while (!try_lock()) {
- sleeper.wait();
- }
- }
-
- /*
- * Release the lock, without changing the value of the rest of the
- * integer.
- */
- void unlock() const {
-#define FB_DOBTR(size) \
- asm volatile("lock; btr" #size " %0, (%1)" \
- : \
- : "i" (Bit), \
- "r" (&lock_) \
- : "memory", "flags")
-
-
- // Reads and writes can not be reordered wrt locked instructions,
- // so we don't need a memory fence here.
- switch (sizeof(IntType)) {
- case 2: FB_DOBTR(w); break;
- case 4: FB_DOBTR(l); break;
- case 8: FB_DOBTR(q); break;
- }
-
-#undef FB_DOBTR
- }
-};
-
-//////////////////////////////////////////////////////////////////////
-
-/**
- * Array of spinlocks where each one is padded to prevent false sharing.
- * Useful for shard-based locking implementations in environments where
- * contention is unlikely.
- */
-
-// TODO: generate it from configure (`getconf LEVEL1_DCACHE_LINESIZE`)
-#define FOLLY_CACHE_LINE_SIZE 64
-
-template <class T, size_t N>
-struct SpinLockArray {
- T& operator[](size_t i) {
- return data_[i].lock;
- }
-
- const T& operator[](size_t i) const {
- return data_[i].lock;
- }
-
- constexpr size_t size() const { return N; }
-
- private:
- struct PaddedSpinLock {
- PaddedSpinLock() : lock() {}
- T lock;
- char padding[FOLLY_CACHE_LINE_SIZE - sizeof(T)];
- };
- static_assert(sizeof(PaddedSpinLock) == FOLLY_CACHE_LINE_SIZE,
- "Invalid size of PaddedSpinLock");
-
- // Check if T can theoretically cross a cache line.
- // NOTE: It should be alignof(std::max_align_t), but max_align_t
- // isn't supported by gcc 4.6.2.
- static_assert(alignof(MaxAlign) > 0 &&
- FOLLY_CACHE_LINE_SIZE % alignof(MaxAlign) == 0 &&
- sizeof(T) <= alignof(MaxAlign),
- "T can cross cache line boundaries");
-
- char padding_[FOLLY_CACHE_LINE_SIZE];
- std::array<PaddedSpinLock, N> data_;
-} __attribute__((__aligned__));
-
-//////////////////////////////////////////////////////////////////////
-
-typedef std::lock_guard<MicroSpinLock> MSLGuard;
-
-//////////////////////////////////////////////////////////////////////
-
-}
-