Skip to content
This repository
tree: 148cba11ab
Fetching contributors…

Octocat-spinner-32-eaf2f5

Cannot retrieve contributors at this time

file 751 lines (640 sloc) 23.394 kb
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750
/*
* Copyright 2012 Facebook, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/

/*
* Two Read-Write spin lock implementations.
*
* Ref: http://locklessinc.com/articles/locks
*
* Both locks here are faster than pthread_rwlock and have very low
* overhead (usually 20-30ns). They don't use any system mutexes and
* are very compact (4/8 bytes), so are suitable for per-instance
* based locking, particularly when contention is not expected.
*
* In most cases, RWSpinLock is a reasonable choice. It has minimal
* overhead, and comparable contention performance when the number of
* competing threads is less than or equal to the number of logical
* CPUs. Even as the number of threads gets larger, RWSpinLock can
* still be very competitive in READ, although it is slower on WRITE,
* and also inherently unfair to writers.
*
* RWTicketSpinLock shows more balanced READ/WRITE performance. If
* your application really needs a lot more threads, and a
* higher-priority writer, prefer one of the RWTicketSpinLock locks.
*
* Caveats:
*
* RWTicketSpinLock locks can only be used with GCC on x86/x86-64
* based systems.
*
* RWTicketSpinLock<32> only allows up to 2^8 - 1 concurrent
* readers and writers.
*
* RWTicketSpinLock<64> only allows up to 2^16 - 1 concurrent
* readers and writers.
*
* RWSpinLock handles 2^30 - 1 concurrent readers.
*
* @author Xin Liu <xliux@fb.com>
*/

#ifndef FOLLY_RWSPINLOCK_H_
#define FOLLY_RWSPINLOCK_H_

/*
========================================================================
Benchmark on (Intel(R) Xeon(R) CPU L5630 @ 2.13GHz) 8 cores(16 HTs)
========================================================================

------------------------------------------------------------------------------
1. Single thread benchmark (read/write lock + unlock overhead)
Benchmark Iters Total t t/iter iter/sec
-------------------------------------------------------------------------------
* BM_RWSpinLockRead 100000 1.786 ms 17.86 ns 53.4M
+30.5% BM_RWSpinLockWrite 100000 2.331 ms 23.31 ns 40.91M
+85.7% BM_RWTicketSpinLock32Read 100000 3.317 ms 33.17 ns 28.75M
+96.0% BM_RWTicketSpinLock32Write 100000 3.5 ms 35 ns 27.25M
+85.6% BM_RWTicketSpinLock64Read 100000 3.315 ms 33.15 ns 28.77M
+96.0% BM_RWTicketSpinLock64Write 100000 3.5 ms 35 ns 27.25M
+85.7% BM_RWTicketSpinLock32FavorWriterRead 100000 3.317 ms 33.17 ns 28.75M
+29.7% BM_RWTicketSpinLock32FavorWriterWrite 100000 2.316 ms 23.16 ns 41.18M
+85.3% BM_RWTicketSpinLock64FavorWriterRead 100000 3.309 ms 33.09 ns 28.82M
+30.2% BM_RWTicketSpinLock64FavorWriterWrite 100000 2.325 ms 23.25 ns 41.02M
+ 175% BM_PThreadRWMutexRead 100000 4.917 ms 49.17 ns 19.4M
+ 166% BM_PThreadRWMutexWrite 100000 4.757 ms 47.57 ns 20.05M

------------------------------------------------------------------------------
2. Contention Benchmark 90% read 10% write
Benchmark hits average min max sigma
------------------------------------------------------------------------------
---------- 8 threads ------------
RWSpinLock Write 142666 220ns 78ns 40.8us 269ns
RWSpinLock Read 1282297 222ns 80ns 37.7us 248ns
RWTicketSpinLock Write 85692 209ns 71ns 17.9us 252ns
RWTicketSpinLock Read 769571 215ns 78ns 33.4us 251ns
pthread_rwlock_t Write 84248 2.48us 99ns 269us 8.19us
pthread_rwlock_t Read 761646 933ns 101ns 374us 3.25us

---------- 16 threads ------------
RWSpinLock Write 124236 237ns 78ns 261us 801ns
RWSpinLock Read 1115807 236ns 78ns 2.27ms 2.17us
RWTicketSpinLock Write 81781 231ns 71ns 31.4us 351ns
RWTicketSpinLock Read 734518 238ns 78ns 73.6us 379ns
pthread_rwlock_t Write 83363 7.12us 99ns 785us 28.1us
pthread_rwlock_t Read 754978 2.18us 101ns 1.02ms 14.3us

---------- 50 threads ------------
RWSpinLock Write 131142 1.37us 82ns 7.53ms 68.2us
RWSpinLock Read 1181240 262ns 78ns 6.62ms 12.7us
RWTicketSpinLock Write 83045 397ns 73ns 7.01ms 31.5us
RWTicketSpinLock Read 744133 386ns 78ns 11ms 31.4us
pthread_rwlock_t Write 80849 112us 103ns 4.52ms 263us
pthread_rwlock_t Read 728698 24us 101ns 7.28ms 194us

*/

#if defined(__GNUC__) && (defined(__i386) || defined(__x86_64__) || \
defined(ARCH_K8))
#define RW_SPINLOCK_USE_X86_INTRINSIC_
#include <x86intrin.h>
#else
#undef RW_SPINLOCK_USE_X86_INTRINSIC_
#endif

#include <atomic>
#include <string>
#include <algorithm>
#include <boost/noncopyable.hpp>

#include <sched.h>
#include <glog/logging.h>

#include "folly/Likely.h"

namespace folly {

/*
* A simple, small (4-bytes), but unfair rwlock. Use it when you want
* a nice writer and don't expect a lot of write/read contention, or
* when you need small rwlocks since you are creating a large number
* of them.
*
* Note that the unfairness here is extreme: if the lock is
* continually accessed for read, writers will never get a chance. If
* the lock can be that highly contended this class is probably not an
* ideal choice anyway.
*
* It currently implements most of the Lockable, SharedLockable and
* UpgradeLockable concepts except the TimedLockable related locking/unlocking
* interfaces.
*/
class RWSpinLock : boost::noncopyable {
  enum : int32_t { READER = 4, UPGRADED = 2, WRITER = 1 };
 public:
  RWSpinLock() : bits_(0) {}

  // Lockable Concept
  void lock() {
    int count = 0;
    while (!LIKELY(try_lock())) {
      if (++count > 1000) sched_yield();
    }
  }

  // Writer is responsible for clearing up both the UPGRADED and WRITER bits.
  void unlock() {
    static_assert(READER > WRITER + UPGRADED, "wrong bits!");
    bits_.fetch_and(~(WRITER | UPGRADED), std::memory_order_release);
  }

  // SharedLockable Concept
  void lock_shared() {
    int count = 0;
    while (!LIKELY(try_lock_shared())) {
      if (++count > 1000) sched_yield();
    }
  }

  void unlock_shared() {
    bits_.fetch_add(-READER, std::memory_order_release);
  }

  // Downgrade the lock from writer status to reader status.
  void unlock_and_lock_shared() {
    bits_.fetch_add(READER, std::memory_order_acquire);
    unlock();
  }

  // UpgradeLockable Concept
  void lock_upgrade() {
    int count = 0;
    while (!try_lock_upgrade()) {
      if (++count > 1000) sched_yield();
    }
  }

  void unlock_upgrade() {
    bits_.fetch_add(-UPGRADED, std::memory_order_acq_rel);
  }

  // unlock upgrade and try to acquire write lock
  void unlock_upgrade_and_lock() {
    int64_t count = 0;
    while (!try_unlock_upgrade_and_lock()) {
      if (++count > 1000) sched_yield();
    }
  }

  // unlock upgrade and read lock atomically
  void unlock_upgrade_and_lock_shared() {
    bits_.fetch_add(READER - UPGRADED, std::memory_order_acq_rel);
  }

  void unlock_shared_and_lock_upgrade() {
    lock_upgrade();
    unlock_shared();
  }

  // write unlock and upgrade lock atomically
  void unlock_and_lock_upgrade() {
    // need to do it in two steps here -- as the UPGRADED bit might be OR-ed at
    // the same time when other threads are trying do try_lock_upgrade().
    bits_.fetch_or(UPGRADED, std::memory_order_acquire);
    bits_.fetch_add(-WRITER, std::memory_order_release);
  }


  // Attempt to acquire writer permission. Return false if we didn't get it.
  bool try_lock() {
    int32_t expect = 0;
    return bits_.compare_exchange_strong(expect, WRITER,
      std::memory_order_acq_rel);
  }

  // Try to get reader permission on the lock. This can fail if we
  // find out someone is a writer.
  bool try_lock_shared() {
    // fetch_add is considerably (100%) faster than compare_exchange,
    // so here we are optimizing for the common (lock success) case.
    int32_t value = bits_.fetch_add(READER, std::memory_order_acquire);
    if (UNLIKELY(value & WRITER)) {
      bits_.fetch_add(-READER, std::memory_order_release);
      return false;
    }
    return true;
  }

  // try to unlock upgrade and write lock atomically
  bool try_unlock_upgrade_and_lock() {
    int32_t expect = UPGRADED;
    return bits_.compare_exchange_strong(expect, WRITER,
        std::memory_order_acq_rel);
  }

  // try to acquire an upgradable lock.
  bool try_lock_upgrade() {
    int32_t value = bits_.fetch_or(UPGRADED, std::memory_order_acquire);

    // Note: when failed, we cannot flip the UPGRADED bit back,
    // as in this case there is either another upgrade lock or a write lock.
    // If it's a write lock, the bit will get cleared up when that lock's done
    // with unlock().
    return ((value & (UPGRADED | WRITER)) == 0);
  }

  // mainly for debugging purposes.
  int32_t bits() const { return bits_.load(std::memory_order_acquire); }

  class ReadHolder;
  class UpgradedHolder;
  class WriteHolder;

  class ReadHolder {
   public:
    explicit ReadHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
      if (lock_) lock_->lock_shared();
    }

    explicit ReadHolder(RWSpinLock& lock) : lock_(&lock) {
      lock_->lock_shared();
    }

    ReadHolder(ReadHolder&& other) : lock_(other.lock_) {
      other.lock_ = nullptr;
    }

    // down-grade
    explicit ReadHolder(UpgradedHolder&& upgraded) : lock_(upgraded.lock_) {
      upgraded.lock_ = nullptr;
      if (lock_) lock_->unlock_upgrade_and_lock_shared();
    }

    explicit ReadHolder(WriteHolder&& writer) : lock_(writer.lock_) {
      writer.lock_ = nullptr;
      if (lock_) lock_->unlock_and_lock_shared();
    }

    ReadHolder& operator=(ReadHolder&& other) {
      using std::swap;
      swap(lock_, other.lock_);
      return *this;
    }

    ReadHolder(const ReadHolder& other) = delete;
    ReadHolder& operator=(const ReadHolder& other) = delete;

    ~ReadHolder() { if (lock_) lock_->unlock_shared(); }

    void reset(RWSpinLock* lock = nullptr) {
      if (lock == lock_) return;
      if (lock_) lock_->unlock_shared();
      lock_ = lock;
      if (lock_) lock_->lock_shared();
    }

    void swap(ReadHolder* other) {
      std::swap(lock_, other->lock_);
    }

   private:
    friend class UpgradedHolder;
    friend class WriteHolder;
    RWSpinLock* lock_;
  };

  class UpgradedHolder {
   public:
    explicit UpgradedHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
      if (lock_) lock_->lock_upgrade();
    }

    explicit UpgradedHolder(RWSpinLock& lock) : lock_(&lock) {
      lock_->lock_upgrade();
    }

    explicit UpgradedHolder(ReadHolder&& reader) {
      lock_ = reader.lock_;
      reader.lock_ = nullptr;
      if (lock_) lock_->unlock_shared_and_lock_upgrade();
    }

    explicit UpgradedHolder(WriteHolder&& writer) {
      lock_ = writer.lock_;
      writer.lock_ = nullptr;
      if (lock_) lock_->unlock_and_lock_upgrade();
    }

    UpgradedHolder(UpgradedHolder&& other) : lock_(other.lock_) {
      other.lock_ = nullptr;
    }

    UpgradedHolder& operator =(UpgradedHolder&& other) {
      using std::swap;
      swap(lock_, other.lock_);
      return *this;
    }

    UpgradedHolder(const UpgradedHolder& other) = delete;
    UpgradedHolder& operator =(const UpgradedHolder& other) = delete;

    ~UpgradedHolder() { if (lock_) lock_->unlock_upgrade(); }

    void reset(RWSpinLock* lock = nullptr) {
      if (lock == lock_) return;
      if (lock_) lock_->unlock_upgrade();
      lock_ = lock;
      if (lock_) lock_->lock_upgrade();
    }

    void swap(UpgradedHolder* other) {
      using std::swap;
      swap(lock_, other->lock_);
    }

   private:
    friend class WriteHolder;
    friend class ReadHolder;
    RWSpinLock* lock_;
  };

  class WriteHolder {
   public:
    explicit WriteHolder(RWSpinLock* lock = nullptr) : lock_(lock) {
      if (lock_) lock_->lock();
    }

    explicit WriteHolder(RWSpinLock& lock) : lock_(&lock) {
      lock_->lock();
    }

    // promoted from an upgrade lock holder
    explicit WriteHolder(UpgradedHolder&& upgraded) {
      lock_ = upgraded.lock_;
      upgraded.lock_ = nullptr;
      if (lock_) lock_->unlock_upgrade_and_lock();
    }

    WriteHolder(WriteHolder&& other) : lock_(other.lock_) {
      other.lock_ = nullptr;
    }

    WriteHolder& operator =(WriteHolder&& other) {
      using std::swap;
      swap(lock_, other.lock_);
      return *this;
    }

    WriteHolder(const WriteHolder& other) = delete;
    WriteHolder& operator =(const WriteHolder& other) = delete;

    ~WriteHolder () { if (lock_) lock_->unlock(); }

    void reset(RWSpinLock* lock = nullptr) {
      if (lock == lock_) return;
      if (lock_) lock_->unlock();
      lock_ = lock;
      if (lock_) lock_->lock();
    }

    void swap(WriteHolder* other) {
      using std::swap;
      swap(lock_, other->lock_);
    }

   private:
    friend class ReadHolder;
    friend class UpgradedHolder;
    RWSpinLock* lock_;
  };

  // Synchronized<> adaptors
  friend void acquireRead(RWSpinLock& l) { return l.lock_shared(); }
  friend void acquireReadWrite(RWSpinLock& l) { return l.lock(); }
  friend void releaseRead(RWSpinLock& l) { return l.unlock_shared(); }
  friend void releaseReadWrite(RWSpinLock& l) { return l.unlock(); }

 private:
  std::atomic<int32_t> bits_;
};


#ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
// A more balanced Read-Write spin lock implemented based on GCC intrinsics.

namespace detail {
template <size_t kBitWidth> struct RWTicketIntTrait {
  static_assert(kBitWidth == 32 || kBitWidth == 64,
      "bit width has to be either 32 or 64 ");
};

template <>
struct RWTicketIntTrait<64> {
  typedef uint64_t FullInt;
  typedef uint32_t HalfInt;
  typedef uint16_t QuarterInt;

#ifdef __SSE2__
  static __m128i make128(const uint16_t v[4]) {
    return _mm_set_epi16(0, 0, 0, 0, v[3], v[2], v[1], v[0]);
  }
  static inline __m128i fromInteger(uint64_t from) {
    return _mm_cvtsi64_si128(from);
  }
  static inline uint64_t toInteger(__m128i in) {
    return _mm_cvtsi128_si64(in);
  }
  static inline uint64_t addParallel(__m128i in, __m128i kDelta) {
    return toInteger(_mm_add_epi16(in, kDelta));
  }
#endif
};

template <>
struct RWTicketIntTrait<32> {
  typedef uint32_t FullInt;
  typedef uint16_t HalfInt;
  typedef uint8_t QuarterInt;

#ifdef __SSE2__
  static __m128i make128(const uint8_t v[4]) {
    return _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, v[3], v[2], v[1], v[0]);
  }
  static inline __m128i fromInteger(uint32_t from) {
    return _mm_cvtsi32_si128(from);
  }
  static inline uint32_t toInteger(__m128i in) {
    return _mm_cvtsi128_si32(in);
  }
  static inline uint32_t addParallel(__m128i in, __m128i kDelta) {
    return toInteger(_mm_add_epi8(in, kDelta));
  }
#endif
};
} // detail


template<size_t kBitWidth, bool kFavorWriter=false>
class RWTicketSpinLockT : boost::noncopyable {
  typedef detail::RWTicketIntTrait<kBitWidth> IntTraitType;
  typedef typename detail::RWTicketIntTrait<kBitWidth>::FullInt FullInt;
  typedef typename detail::RWTicketIntTrait<kBitWidth>::HalfInt HalfInt;
  typedef typename detail::RWTicketIntTrait<kBitWidth>::QuarterInt
    QuarterInt;

  union RWTicket {
    FullInt whole;
    HalfInt readWrite;
    __extension__ struct {
      QuarterInt write;
      QuarterInt read;
      QuarterInt users;
    };
  } ticket;

 private: // Some x64-specific utilities for atomic access to ticket.
  template<class T> static T load_acquire(T* addr) {
    T t = *addr; // acquire barrier
    asm volatile("" : : : "memory");
    return t;
  }

  template<class T>
  static void store_release(T* addr, T v) {
    asm volatile("" : : : "memory");
    *addr = v; // release barrier
  }

 public:

  RWTicketSpinLockT() {
    store_release(&ticket.whole, FullInt(0));
  }

  void lock() {
    if (kFavorWriter) {
      writeLockAggressive();
    } else {
      writeLockNice();
    }
  }

  /*
* Both try_lock and try_lock_shared diverge in our implementation from the
* lock algorithm described in the link above.
*
* In the read case, it is undesirable that the readers could wait
* for another reader (before increasing ticket.read in the other
* implementation). Our approach gives up on
* first-come-first-serve, but our benchmarks showed improve
* performance for both readers and writers under heavily contended
* cases, particularly when the number of threads exceeds the number
* of logical CPUs.
*
* We have writeLockAggressive() using the original implementation
* for a writer, which gives some advantage to the writer over the
* readers---for that path it is guaranteed that the writer will
* acquire the lock after all the existing readers exit.
*/
  bool try_lock() {
    RWTicket t;
    FullInt old = t.whole = load_acquire(&ticket.whole);
    if (t.users != t.write) return false;
    ++t.users;
    return __sync_bool_compare_and_swap(&ticket.whole, old, t.whole);
  }

  /*
* Call this if you want to prioritize writer to avoid starvation.
* Unlike writeLockNice, immediately acquires the write lock when
* the existing readers (arriving before the writer) finish their
* turns.
*/
  void writeLockAggressive() {
    // sched_yield() is needed here to avoid a pathology if the number
    // of threads attempting concurrent writes is >= the number of real
    // cores allocated to this process. This is less likely than the
    // corresponding situation in lock_shared(), but we still want to
    // avoid it
    int count = 0;
    QuarterInt val = __sync_fetch_and_add(&ticket.users, 1);
    while (val != load_acquire(&ticket.write)) {
      asm volatile("pause");
      if (UNLIKELY(++count > 1000)) sched_yield();
    }
  }

  // Call this when the writer should be nicer to the readers.
  void writeLockNice() {
    // Here it doesn't cpu-relax the writer.
    //
    // This is because usually we have many more readers than the
    // writers, so the writer has less chance to get the lock when
    // there are a lot of competing readers. The aggressive spinning
    // can help to avoid starving writers.
    //
    // We don't worry about sched_yield() here because the caller
    // has already explicitly abandoned fairness.
    while (!try_lock()) {}
  }

  // Atomically unlock the write-lock from writer and acquire the read-lock.
  void unlock_and_lock_shared() {
    QuarterInt val = __sync_fetch_and_add(&ticket.read, 1);
  }

  // Release writer permission on the lock.
  void unlock() {
    RWTicket t;
    t.whole = load_acquire(&ticket.whole);
    FullInt old = t.whole;

#ifdef __SSE2__
    // SSE2 can reduce the lock and unlock overhead by 10%
    static const QuarterInt kDeltaBuf[4] = { 1, 1, 0, 0 }; // write/read/user
    static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
    __m128i m = IntTraitType::fromInteger(old);
    t.whole = IntTraitType::addParallel(m, kDelta);
#else
    ++t.read;
    ++t.write;
#endif
    store_release(&ticket.readWrite, t.readWrite);
  }

  void lock_shared() {
    // sched_yield() is important here because we can't grab the
    // shared lock if there is a pending writeLockAggressive, so we
    // need to let threads that already have a shared lock complete
    int count = 0;
    while (!LIKELY(try_lock_shared())) {
      asm volatile("pause");
      if (UNLIKELY((++count & 1023) == 0)) sched_yield();
    }
  }

  bool try_lock_shared() {
    RWTicket t, old;
    old.whole = t.whole = load_acquire(&ticket.whole);
    old.users = old.read;
#ifdef __SSE2__
    // SSE2 may reduce the total lock and unlock overhead by 10%
    static const QuarterInt kDeltaBuf[4] = { 0, 1, 1, 0 }; // write/read/user
    static const __m128i kDelta = IntTraitType::make128(kDeltaBuf);
    __m128i m = IntTraitType::fromInteger(old.whole);
    t.whole = IntTraitType::addParallel(m, kDelta);
#else
    ++t.read;
    ++t.users;
#endif
    return __sync_bool_compare_and_swap(&ticket.whole, old.whole, t.whole);
  }

  void unlock_shared() {
    QuarterInt val = __sync_fetch_and_add(&ticket.write, 1);
  }

  class WriteHolder;

  typedef RWTicketSpinLockT<kBitWidth, kFavorWriter> RWSpinLock;
  class ReadHolder : boost::noncopyable {
   public:
    explicit ReadHolder(RWSpinLock *lock = nullptr) :
      lock_(lock) {
      if (lock_) lock_->lock_shared();
    }

    explicit ReadHolder(RWSpinLock &lock) : lock_ (&lock) {
      if (lock_) lock_->lock_shared();
    }

    // atomically unlock the write-lock from writer and acquire the read-lock
    explicit ReadHolder(WriteHolder *writer) : lock_(nullptr) {
      std::swap(this->lock_, writer->lock_);
      if (lock_) {
        lock_->unlock_and_lock_shared();
      }
    }

    ~ReadHolder() {
      if (lock_) lock_->unlock_shared();
    }

    void reset(RWSpinLock *lock = nullptr) {
      if (lock_) lock_->unlock_shared();
      lock_ = lock;
      if (lock_) lock_->lock_shared();
    }

    void swap(ReadHolder *other) {
      std::swap(this->lock_, other->lock_);
    }

   private:
    RWSpinLock *lock_;
  };

  class WriteHolder : boost::noncopyable {
   public:
    explicit WriteHolder(RWSpinLock *lock = nullptr) : lock_(lock) {
      if (lock_) lock_->lock();
    }
    explicit WriteHolder(RWSpinLock &lock) : lock_ (&lock) {
      if (lock_) lock_->lock();
    }

    ~WriteHolder() {
      if (lock_) lock_->unlock();
    }

    void reset(RWSpinLock *lock = nullptr) {
      if (lock == lock_) return;
      if (lock_) lock_->unlock();
      lock_ = lock;
      if (lock_) lock_->lock();
    }

    void swap(WriteHolder *other) {
      std::swap(this->lock_, other->lock_);
    }

   private:
    friend class ReadHolder;
    RWSpinLock *lock_;
  };

  // Synchronized<> adaptors.
  friend void acquireRead(RWTicketSpinLockT& mutex) {
    mutex.lock_shared();
  }
  friend void acquireReadWrite(RWTicketSpinLockT& mutex) {
    mutex.lock();
  }
  friend bool acquireReadWrite(RWTicketSpinLockT& mutex,
                               unsigned int milliseconds) {
    mutex.lock();
    return true;
  }
  friend void releaseRead(RWTicketSpinLockT& mutex) {
    mutex.unlock_shared();
  }
  friend void releaseReadWrite(RWTicketSpinLockT& mutex) {
    mutex.unlock();
  }
};

typedef RWTicketSpinLockT<32> RWTicketSpinLock32;
typedef RWTicketSpinLockT<64> RWTicketSpinLock64;

#endif // RW_SPINLOCK_USE_X86_INTRINSIC_

} // namespace folly

#ifdef RW_SPINLOCK_USE_X86_INTRINSIC_
#undef RW_SPINLOCK_USE_X86_INTRINSIC_
#endif

#endif // FOLLY_RWSPINLOCK_H_
Something went wrong with that request. Please try again.