2 * Copyright 2012 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #ifndef FOLLY_SMALLLOCKS_H_
18 #define FOLLY_SMALLLOCKS_H_
21 * This header defines a few very small mutex types. These are useful
22 * in highly memory-constrained environments where contention is
25 * Note: these locks are for use when you aren't likely to contend on
26 * the critical section, or when the critical section is incredibly
27 * small. Given that, both of the locks defined in this header are
28 * inherently unfair: that is, the longer a thread is waiting, the
29 * longer it waits between attempts to acquire, so newer waiters are
30 * more likely to get the mutex. For the intended use-case this is
33 * @author Keith Adams <kma@fb.com>
34 * @author Jordan DeLong <delong.j@fb.com>
39 #include <type_traits>
41 #include <boost/noncopyable.hpp>
46 #include <glog/logging.h>
49 # error "SmallLocks.h is currently x64-only."
54 //////////////////////////////////////////////////////////////////////
59 * A helper object for the condended case. Starts off with eager
60 * spinning, and falls back to sleeping for small quantums.
63 static const uint32_t kMaxActiveSpin = 4000;
68 Sleeper() : spinCount(0) {}
71 if (spinCount < kMaxActiveSpin) {
73 asm volatile("pause");
76 * Always sleep 0.5ms, assuming this will make the kernel put
77 * us down for whatever its minimum timer resolution is (in
78 * linux this varies by kernel version from 1ms to 10ms).
80 struct timespec ts = { 0, 500000 };
88 //////////////////////////////////////////////////////////////////////
91 * A really, *really* small spinlock for fine-grained locking of lots
94 * Zero initializing these is guaranteed to be as good as calling
95 * init(), since the free state is guaranteed to be all-bits zero.
97 * This class should be kept a POD, so we can used it in other packed
98 * structs (gcc does not allow __attribute__((packed)) on structs that
99 * contain non-POD data). This means avoid adding a constructor, or
100 * making some members private, etc.
102 struct MicroSpinLock {
103 enum { FREE = 0, LOCKED = 1 };
107 * Atomically move lock_ from "compare" to "newval". Return boolean
108 * success. Do not play on or around.
110 bool cas(uint8_t compare, uint8_t newVal) {
112 asm volatile("lock; cmpxchgb %2, (%3);"
115 : "a" (compare), // cmpxchgb constrains this to be in %al
116 "q" (newVal), // Needs to be byte-accessible
118 : "memory", "flags");
122 // Initialize this MSL. It is unnecessary to call this if you
123 // zero-initialize the MicroSpinLock.
129 return cas(FREE, LOCKED);
133 detail::Sleeper sleeper;
135 while (lock_ != FREE) {
136 asm volatile("" : : : "memory");
139 } while (!try_lock());
140 DCHECK(lock_ == LOCKED);
144 CHECK(lock_ == LOCKED);
145 asm volatile("" : : : "memory");
146 lock_ = FREE; // release barrier on x86
150 //////////////////////////////////////////////////////////////////////
153 * Spin lock on a single bit in an integral type. You can use this
154 * with 16, 32, or 64-bit integral types.
156 * This is useful if you want a small lock and already have an int
157 * with a bit in it that you aren't using. But note that it can't be
158 * as small as MicroSpinLock (1 byte), if you don't already have a
159 * convenient int with an unused bit lying around to put it on.
161 * To construct these, either use init() or zero initialize. We don't
162 * have a real constructor because we want this to be a POD type so we
163 * can put it into packed structs.
165 template<class IntType, int Bit = sizeof(IntType) * 8 - 1>
166 struct PicoSpinLock {
167 // Internally we deal with the unsigned version of the type.
168 typedef typename std::make_unsigned<IntType>::type UIntType;
170 static_assert(std::is_integral<IntType>::value,
171 "PicoSpinLock needs an integral type");
172 static_assert(sizeof(IntType) == 2 || sizeof(IntType) == 4 ||
173 sizeof(IntType) == 8,
174 "PicoSpinLock can't work on integers smaller than 2 bytes");
177 static const UIntType kLockBitMask_ = UIntType(1) << Bit;
181 * You must call this function before using this class, if you
182 * default constructed it. If you zero-initialized it you can
183 * assume the PicoSpinLock is in a valid unlocked state with
186 * (This doesn't use a constructor because we want to be a POD.)
188 void init(IntType initialValue = 0) {
189 CHECK(!(initialValue & kLockBitMask_));
190 lock_ = initialValue;
194 * Returns the value of the integer we using for our lock, except
195 * with the bit we are using as a lock cleared, regardless of
196 * whether the lock is held.
198 * It is 'safe' to call this without holding the lock. (As in: you
199 * get the same guarantees for simultaneous accesses to an integer
200 * as you normally get.)
202 IntType getData() const {
203 return static_cast<IntType>(lock_ & ~kLockBitMask_);
207 * Set the value of the other bits in our integer.
209 * Don't use this when you aren't holding the lock, unless it can be
210 * guaranteed that no other threads may be trying to use this.
212 void setData(IntType w) {
213 CHECK(!(w & kLockBitMask_));
214 lock_ = (lock_ & kLockBitMask_) | w;
218 * Try to get the lock without blocking: returns whether or not we
221 bool try_lock() const {
224 #define FB_DOBTS(size) \
225 asm volatile("lock; bts" #size " %1, (%2); setnc %0" \
231 switch (sizeof(IntType)) {
232 case 2: FB_DOBTS(w); break;
233 case 4: FB_DOBTS(l); break;
234 case 8: FB_DOBTS(q); break;
243 * Block until we can acquire the lock. Uses Sleeper to wait.
246 detail::Sleeper sleeper;
247 while (!try_lock()) {
253 * Release the lock, without changing the value of the rest of the
256 void unlock() const {
257 #define FB_DOBTR(size) \
258 asm volatile("lock; btr" #size " %0, (%1)" \
265 // Reads and writes can not be reordered wrt locked instructions,
266 // so we don't need a memory fence here.
267 switch (sizeof(IntType)) {
268 case 2: FB_DOBTR(w); break;
269 case 4: FB_DOBTR(l); break;
270 case 8: FB_DOBTR(q); break;
277 //////////////////////////////////////////////////////////////////////
280 * Array of spinlocks where each one is padded to prevent false sharing.
281 * Useful for shard-based locking implementations in environments where
282 * contention is unlikely.
285 // TODO: generate it from configure (`getconf LEVEL1_DCACHE_LINESIZE`)
286 #define FOLLY_CACHE_LINE_SIZE 64
288 template <class T, size_t N>
289 struct SpinLockArray {
290 T& operator[](size_t i) {
291 return data_[i].lock;
294 const T& operator[](size_t i) const {
295 return data_[i].lock;
298 constexpr size_t size() const { return N; }
301 struct PaddedSpinLock {
302 PaddedSpinLock() : lock() { }
304 char padding[FOLLY_CACHE_LINE_SIZE - sizeof(T)];
306 static_assert(sizeof(PaddedSpinLock) == FOLLY_CACHE_LINE_SIZE,
307 "Invalid size of PaddedSpinLock");
309 // Check if T can theoretically cross a cache line.
310 // NOTE: It should be alignof(std::max_align_t), but max_align_t
311 // isn't supported by gcc 4.6.2.
312 struct MaxAlign { char c; } __attribute__((aligned));
313 static_assert(alignof(MaxAlign) > 0 &&
314 FOLLY_CACHE_LINE_SIZE % alignof(MaxAlign) == 0 &&
315 sizeof(T) <= alignof(MaxAlign),
316 "T can cross cache line boundaries");
318 char padding_[FOLLY_CACHE_LINE_SIZE];
319 std::array<PaddedSpinLock, N> data_;
320 } __attribute__((aligned));
322 //////////////////////////////////////////////////////////////////////
324 typedef std::lock_guard<MicroSpinLock> MSLGuard;
326 //////////////////////////////////////////////////////////////////////