LifoSem.h

Go to the documentation of this file.

/*
 * Copyright 2014-present 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.
 */
#pragma once

#include <algorithm>
#include <atomic>
#include <cstdint>
#include <cstring>
#include <memory>
#include <system_error>

#include <folly/CPortability.h>
#include <folly/CachelinePadded.h>
#include <folly/IndexedMemPool.h>
#include <folly/Likely.h>
#include <folly/lang/SafeAssert.h>
#include <folly/synchronization/AtomicStruct.h>
#include <folly/synchronization/SaturatingSemaphore.h>

namespace folly {

template <
    template <typename> class Atom = std::atomic,
    class BatonType = SaturatingSemaphore<true, Atom>>
struct LifoSemImpl;

typedef LifoSemImpl<> LifoSem;

struct FOLLY_EXPORT ShutdownSemError : public std::runtime_error {
  explicit ShutdownSemError(const std::string& msg);
  ~ShutdownSemError() noexcept override;
};

namespace detail {

// Internally, a LifoSem is either a value or a linked list of wait nodes.
// This union is captured in the LifoSemHead type, which holds either a
// value or an indexed pointer to the list.  LifoSemHead itself is a value
// type, the head is a mutable atomic box containing a LifoSemHead value.
// Each wait node corresponds to exactly one waiter.  Values can flow
// through the semaphore either by going into and out of the head's value,
// or by direct communication from a poster to a waiter.  The former path
// is taken when there are no pending waiters, the latter otherwise.  The
// general flow of a post is to try to increment the value or pop-and-post
// a wait node.  Either of those have the effect of conveying one semaphore
// unit.  Waiting is the opposite, either a decrement of the value or
// push-and-wait of a wait node.  The generic LifoSemBase abstracts the
// actual mechanism by which a wait node's post->wait communication is
// performed, which is why we have LifoSemRawNode and LifoSemNode.

template <template <typename> class Atom>
struct LifoSemRawNode {
  std::aligned_storage<sizeof(void*), alignof(void*)>::type raw;

  uint32_t next;

  bool isShutdownNotice() const {
    return next == uint32_t(-1);
  }
  void clearShutdownNotice() {
    next = 0;
  }
  void setShutdownNotice() {
    next = uint32_t(-1);
  }

  typedef folly::IndexedMemPool<LifoSemRawNode<Atom>, 32, 200, Atom> Pool;

  static Pool& pool();
};

#define LIFOSEM_DECLARE_POOL(Atom, capacity)                 \
  namespace folly {                                          \
  namespace detail {                                         \
  template <>                                                \
  LifoSemRawNode<Atom>::Pool& LifoSemRawNode<Atom>::pool() { \
    static Pool* instance = new Pool((capacity));            \
    return *instance;                                        \
  }                                                          \
  }                                                          \
  }

template <typename Handoff, template <typename> class Atom>
struct LifoSemNode : public LifoSemRawNode<Atom> {
  static_assert(
      sizeof(Handoff) <= sizeof(LifoSemRawNode<Atom>::raw),
      "Handoff too big for small-object optimization, use indirection");
  static_assert(
      alignof(Handoff) <= alignof(decltype(LifoSemRawNode<Atom>::raw)),
      "Handoff alignment constraint not satisfied");

  template <typename... Args>
  void init(Args&&... args) {
    new (&this->raw) Handoff(std::forward<Args>(args)...);
  }

  void destroy() {
    handoff().~Handoff();
#ifndef NDEBUG
    memset(&this->raw, 'F', sizeof(this->raw));
#endif
  }

  Handoff& handoff() {
    return *static_cast<Handoff*>(static_cast<void*>(&this->raw));
  }

  const Handoff& handoff() const {
    return *static_cast<const Handoff*>(static_cast<const void*>(&this->raw));
  }
};

template <typename Handoff, template <typename> class Atom>
struct LifoSemNodeRecycler {
  void operator()(LifoSemNode<Handoff, Atom>* elem) const {
    elem->destroy();
    auto idx = LifoSemRawNode<Atom>::pool().locateElem(elem);
    LifoSemRawNode<Atom>::pool().recycleIndex(idx);
  }
};

class LifoSemHead {
  // What we really want are bitfields:
  //  uint64_t data : 32; uint64_t isNodeIdx : 1; uint64_t seq : 31;
  // Unfortunately g++ generates pretty bad code for this sometimes (I saw
  // -O3 code from gcc 4.7.1 copying the bitfields one at a time instead of
  // in bulk, for example).  We can generate better code anyway by assuming
  // that setters won't be given values that cause under/overflow, and
  // putting the sequence at the end where its planned overflow doesn't
  // need any masking.
  //
  // data == 0 (empty list) with isNodeIdx is conceptually the same
  // as data == 0 (no unclaimed increments) with !isNodeIdx, we always
  // convert the former into the latter to make the logic simpler.
  enum {
    IsNodeIdxShift = 32,
    IsShutdownShift = 33,
    IsLockedShift = 34,
    SeqShift = 35,
  };
  enum : uint64_t {
    IsNodeIdxMask = uint64_t(1) << IsNodeIdxShift,
    IsShutdownMask = uint64_t(1) << IsShutdownShift,
    IsLockedMask = uint64_t(1) << IsLockedShift,
    SeqIncr = uint64_t(1) << SeqShift,
    SeqMask = ~(SeqIncr - 1),
  };

 public:
  uint64_t bits;


  inline uint32_t idx() const {
    assert(isNodeIdx());
    assert(uint32_t(bits) != 0);
    return uint32_t(bits);
  }
  inline uint32_t value() const {
    assert(!isNodeIdx());
    return uint32_t(bits);
  }
  inline constexpr bool isNodeIdx() const {
    return (bits & IsNodeIdxMask) != 0;
  }
  inline constexpr bool isShutdown() const {
    return (bits & IsShutdownMask) != 0;
  }
  inline constexpr bool isLocked() const {
    return (bits & IsLockedMask) != 0;
  }
  inline constexpr uint32_t seq() const {
    return uint32_t(bits >> SeqShift);
  }


  static inline constexpr LifoSemHead fresh(uint32_t value) {
    return LifoSemHead{value};
  }

  inline LifoSemHead withPop(uint32_t idxNext) const {
    assert(!isLocked());
    assert(isNodeIdx());
    if (idxNext == 0) {
      // no isNodeIdx bit or data bits.  Wraparound of seq bits is okay
      return LifoSemHead{(bits & (SeqMask | IsShutdownMask)) + SeqIncr};
    } else {
      // preserve sequence bits (incremented with wraparound okay) and
      // isNodeIdx bit, replace all data bits
      return LifoSemHead{(bits & (SeqMask | IsShutdownMask | IsNodeIdxMask)) +
                         SeqIncr + idxNext};
    }
  }

  inline LifoSemHead withPush(uint32_t _idx) const {
    assert(!isLocked());
    assert(isNodeIdx() || value() == 0);
    assert(!isShutdown());
    assert(_idx != 0);
    return LifoSemHead{(bits & SeqMask) | IsNodeIdxMask | _idx};
  }

  inline LifoSemHead withValueIncr(uint32_t delta) const {
    assert(!isLocked());
    assert(!isNodeIdx());
    auto rv = LifoSemHead{bits + SeqIncr + delta};
    if (UNLIKELY(rv.isNodeIdx())) {
      // value has overflowed into the isNodeIdx bit
      rv = LifoSemHead{(rv.bits & ~IsNodeIdxMask) | (IsNodeIdxMask - 1)};
    }
    return rv;
  }

  inline LifoSemHead withValueDecr(uint32_t delta) const {
    assert(!isLocked());
    assert(delta > 0 && delta <= value());
    return LifoSemHead{bits + SeqIncr - delta};
  }

  inline LifoSemHead withShutdown() const {
    return LifoSemHead{bits | IsShutdownMask};
  }

  // Returns LifoSemHead with lock bit set, but rest of bits unchanged.
  inline LifoSemHead withLock() const {
    assert(!isLocked());
    return LifoSemHead{bits | IsLockedMask};
  }

  // Returns LifoSemHead with lock bit unset, and updated seqno based
  // on idx.
  inline LifoSemHead withoutLock(uint32_t idxNext) const {
    assert(isLocked());
    // We need to treat this as a pop, as we may change the list head.
    return LifoSemHead{bits & ~IsLockedMask}.withPop(idxNext);
  }

  inline constexpr bool operator==(const LifoSemHead& rhs) const {
    return bits == rhs.bits;
  }
  inline constexpr bool operator!=(const LifoSemHead& rhs) const {
    return !(*this == rhs);
  }
};

template <typename Handoff, template <typename> class Atom = std::atomic>
struct LifoSemBase {
  constexpr explicit LifoSemBase(uint32_t initialValue = 0)
      : head_(LifoSemHead::fresh(initialValue)) {}

  LifoSemBase(LifoSemBase const&) = delete;
  LifoSemBase& operator=(LifoSemBase const&) = delete;

  bool post() {
    auto idx = incrOrPop(1);
    if (idx != 0) {
      idxToNode(idx).handoff().post();
      return true;
    }
    return false;
  }

  void post(uint32_t n) {
    uint32_t idx;
    while (n > 0 && (idx = incrOrPop(n)) != 0) {
      // pop accounts for only 1
      idxToNode(idx).handoff().post();
      --n;
    }
  }

  bool isShutdown() const {
    return UNLIKELY(head_->load(std::memory_order_acquire).isShutdown());
  }

  void shutdown() {
    // first set the shutdown bit
    auto h = head_->load(std::memory_order_acquire);
    while (!h.isShutdown()) {
      if (head_->compare_exchange_strong(h, h.withShutdown())) {
        // success
        h = h.withShutdown();
        break;
      }
      // compare_exchange_strong rereads h, retry
    }

    // now wake up any waiters
    while (h.isNodeIdx()) {
      if (h.isLocked()) {
        std::this_thread::yield();
        h = head_->load(std::memory_order_acquire);
        continue;
      }
      auto& node = idxToNode(h.idx());
      auto repl = h.withPop(node.next);
      if (head_->compare_exchange_strong(h, repl)) {
        // successful pop, wake up the waiter and move on.  The next
        // field is used to convey that this wakeup didn't consume a value
        node.setShutdownNotice();
        node.handoff().post();
        h = repl;
      }
    }
  }

  bool tryWait() {
    uint32_t n = 1;
    auto rv = decrOrPush(n, 0);
    assert(
        (rv == WaitResult::DECR && n == 0) ||
        (rv != WaitResult::DECR && n == 1));
    // SHUTDOWN is okay here, since we don't actually wait
    return rv == WaitResult::DECR;
  }

  uint32_t tryWait(uint32_t n) {
    auto const orig = n;
    while (n > 0) {
#ifndef NDEBUG
      auto prev = n;
#endif
      auto rv = decrOrPush(n, 0);
      assert(
          (rv == WaitResult::DECR && n < prev) ||
          (rv != WaitResult::DECR && n == prev));
      if (rv != WaitResult::DECR) {
        break;
      }
    }
    return orig - n;
  }

  void wait() {
    auto const deadline = std::chrono::steady_clock::time_point::max();
    auto res = try_wait_until(deadline);
    FOLLY_SAFE_DCHECK(res, "infinity time has passed");
  }

  template <typename Rep, typename Period>
  bool try_wait_for(const std::chrono::duration<Rep, Period>& timeout) {
    return try_wait_until(timeout + std::chrono::steady_clock::now());
  }

  template <typename Clock, typename Duration>
  bool try_wait_until(
      const std::chrono::time_point<Clock, Duration>& deadline) {
    // early check isn't required for correctness, but is an important
    // perf win if we can avoid allocating and deallocating a node
    if (tryWait()) {
      return true;
    }

    // allocateNode() won't compile unless Handoff has a default
    // constructor
    UniquePtr node = allocateNode();

    auto rv = tryWaitOrPush(*node);
    if (UNLIKELY(rv == WaitResult::SHUTDOWN)) {
      assert(isShutdown());
      throw ShutdownSemError("wait() would block but semaphore is shut down");
    }

    if (rv == WaitResult::PUSH) {
      if (!node->handoff().try_wait_until(deadline)) {
        if (tryRemoveNode(*node)) {
          return false;
        } else {
          // We could not remove our node. Return to waiting.
          //
          // This only happens if we lose a removal race with post(),
          // so we are not likely to wait long.  This is only
          // necessary to ensure we don't return node's memory back to
          // IndexedMemPool before post() has had a chance to post to
          // handoff().  In a stronger memory reclamation scheme, such
          // as hazptr or rcu, this would not be necessary.
          node->handoff().wait();
        }
      }
      if (UNLIKELY(node->isShutdownNotice())) {
        // this wait() didn't consume a value, it was triggered by shutdown
        throw ShutdownSemError(
            "blocking wait() interrupted by semaphore shutdown");
      }

      // node->handoff().wait() can't return until after the node has
      // been popped and post()ed, so it is okay for the UniquePtr to
      // recycle the node now
    }
    // else node wasn't pushed, so it is safe to recycle
    return true;
  }

  uint32_t valueGuess() const {
    // this is actually linearizable, but we don't promise that because
    // we may want to add striping in the future to help under heavy
    // contention
    auto h = head_->load(std::memory_order_acquire);
    return h.isNodeIdx() ? 0 : h.value();
  }

 protected:
  enum class WaitResult {
    PUSH,
    DECR,
    SHUTDOWN,
  };

  typedef std::
      unique_ptr<LifoSemNode<Handoff, Atom>, LifoSemNodeRecycler<Handoff, Atom>>
          UniquePtr;

  template <typename... Args>
  UniquePtr allocateNode(Args&&... args) {
    auto idx = LifoSemRawNode<Atom>::pool().allocIndex();
    if (idx != 0) {
      auto& node = idxToNode(idx);
      node.clearShutdownNotice();
      try {
        node.init(std::forward<Args>(args)...);
      } catch (...) {
        LifoSemRawNode<Atom>::pool().recycleIndex(idx);
        throw;
      }
      return UniquePtr(&node);
    } else {
      return UniquePtr();
    }
  }

  WaitResult tryWaitOrPush(LifoSemNode<Handoff, Atom>& waiterNode) {
    uint32_t n = 1;
    return decrOrPush(n, nodeToIdx(waiterNode));
  }

 private:
  CachelinePadded<folly::AtomicStruct<LifoSemHead, Atom>> head_;

  static LifoSemNode<Handoff, Atom>& idxToNode(uint32_t idx) {
    auto raw = &LifoSemRawNode<Atom>::pool()[idx];
    return *static_cast<LifoSemNode<Handoff, Atom>*>(raw);
  }

  static uint32_t nodeToIdx(const LifoSemNode<Handoff, Atom>& node) {
    return LifoSemRawNode<Atom>::pool().locateElem(&node);
  }

  // Locks the list head (blocking concurrent pushes and pops)
  // and attempts to remove this node.  Returns true if node was
  // found and removed, false if not found.
  bool tryRemoveNode(const LifoSemNode<Handoff, Atom>& removenode) {
    auto removeidx = nodeToIdx(removenode);
    auto head = head_->load(std::memory_order_acquire);
    // Try to lock the head.
    while (true) {
      if (head.isLocked()) {
        std::this_thread::yield();
        head = head_->load(std::memory_order_acquire);
        continue;
      }
      if (!head.isNodeIdx()) {
        return false;
      }
      if (head_->compare_exchange_weak(
              head,
              head.withLock(),
              std::memory_order_acquire,
              std::memory_order_relaxed)) {
        break;
      }
    }
    // Update local var to what head_ is, for better assert() checking.
    head = head.withLock();
    bool result = false;
    auto idx = head.idx();
    if (idx == removeidx) {
      // pop from head.  Head seqno is updated.
      head_->store(
          head.withoutLock(removenode.next), std::memory_order_release);
      return true;
    }
    auto node = &idxToNode(idx);
    idx = node->next;
    while (idx) {
      if (idx == removeidx) {
        // Pop from mid-list.
        node->next = removenode.next;
        result = true;
        break;
      }
      node = &idxToNode(idx);
      idx = node->next;
    }
    // Unlock and return result
    head_->store(head.withoutLock(head.idx()), std::memory_order_release);
    return result;
  }

  uint32_t incrOrPop(uint32_t n) {
    while (true) {
      assert(n > 0);

      auto head = head_->load(std::memory_order_acquire);
      if (head.isLocked()) {
        std::this_thread::yield();
        continue;
      }
      if (head.isNodeIdx()) {
        auto& node = idxToNode(head.idx());
        if (head_->compare_exchange_strong(head, head.withPop(node.next))) {
          // successful pop
          return head.idx();
        }
      } else {
        auto after = head.withValueIncr(n);
        if (head_->compare_exchange_strong(head, after)) {
          // successful incr
          return 0;
        }
      }
      // retry
    }
  }

  WaitResult decrOrPush(uint32_t& n, uint32_t idx) {
    assert(n > 0);

    while (true) {
      auto head = head_->load(std::memory_order_acquire);

      if (head.isLocked()) {
        std::this_thread::yield();
        continue;
      }

      if (!head.isNodeIdx() && head.value() > 0) {
        // decr
        auto delta = std::min(n, head.value());
        if (head_->compare_exchange_strong(head, head.withValueDecr(delta))) {
          n -= delta;
          return WaitResult::DECR;
        }
      } else {
        // push
        if (idx == 0) {
          return WaitResult::PUSH;
        }

        if (UNLIKELY(head.isShutdown())) {
          return WaitResult::SHUTDOWN;
        }

        auto& node = idxToNode(idx);
        node.next = head.isNodeIdx() ? head.idx() : 0;
        if (head_->compare_exchange_strong(head, head.withPush(idx))) {
          // push succeeded
          return WaitResult::PUSH;
        }
      }
    }
    // retry
  }
};

} // namespace detail

template <template <typename> class Atom, class BatonType>
struct LifoSemImpl : public detail::LifoSemBase<BatonType, Atom> {
  constexpr explicit LifoSemImpl(uint32_t v = 0)
      : detail::LifoSemBase<BatonType, Atom>(v) {}
};

} // namespace folly