hazard_eras.hpp

//
// Copyright (c) 2018-2020 Manuel Pöter.
// Licensed under the MIT License. See LICENSE file in the project root for full license information.
//
 
#ifndef HAZARD_ERAS_IMPL
  #error "This is an impl file and must not be included directly!"
#endif
 
#include <xenium/aligned_object.hpp>
#include <xenium/detail/port.hpp>
 
#include <algorithm>
#include <new>
#include <vector>
 
#ifdef _MSC_VER
  #pragma warning(push)
  #pragma warning(disable : 4324)  // structure was padded due to alignment specifier
  #pragma warning(disable : 26495) // uninitialized member variable
#endif
 
namespace xenium::reclamation {
 
template <class Traits>
template <class T, class MarkedPtr>
hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::guard_ptr(const MarkedPtr& p) : base(p), he() {
  if (this->ptr.get() != nullptr) {
    auto era = era_clock.load(std::memory_order_relaxed);
    he = local_thread_data().alloc_hazard_era(era);
  }
}
 
template <class Traits>
template <class T, class MarkedPtr>
hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::guard_ptr(const guard_ptr& p) : base(p.ptr), he(p.he) {
  if (he) {
    he->add_guard();
  }
}
 
template <class Traits>
template <class T, class MarkedPtr>
hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::guard_ptr(guard_ptr&& p) noexcept : base(p.ptr), he(p.he) {
  p.ptr.reset();
  p.he = nullptr;
}
 
template <class Traits>
template <class T, class MarkedPtr>
auto hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::operator=(const guard_ptr& p) noexcept -> guard_ptr& {
  if (&p == this) {
    return *this;
  }
 
  reset();
  this->ptr = p.ptr;
  he = p.he;
  if (he) {
    he->add_guard();
  }
  return *this;
}
 
template <class Traits>
template <class T, class MarkedPtr>
auto hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::operator=(guard_ptr&& p) noexcept -> guard_ptr& {
  if (&p == this) {
    return *this;
  }
 
  reset();
  this->ptr = std::move(p.ptr);
  this->he = p.he;
  p.ptr.reset();
  p.he = nullptr;
  return *this;
}
 
template <class Traits>
template <class T, class MarkedPtr>
void hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::acquire(const concurrent_ptr<T>& p, std::memory_order order) {
  if (order == std::memory_order_relaxed || order == std::memory_order_consume) {
    // we have to use memory_order_acquire (or something stricter) to ensure that
    // the era_clock.load cannot return an outdated value.
    order = std::memory_order_acquire;
  }
  era_t prev_era = he == nullptr ? 0 : he->get_era();
  for (;;) {
    // (1) - this load operation synchronizes-with any release operation on p.
    // we have to use acquire here to ensure that the subsequent era_clock.load
    // sees a value >= p.construction_era
    auto value = p.load(order);
 
    auto era = era_clock.load(std::memory_order_relaxed);
    if (era == prev_era) {
      this->ptr = value;
      return;
    }
 
    if (he != nullptr) {
      assert(he->get_era() != era);
      if (he->guards() == 1) {
        // we are the only guard using this HE instance -> reuse it and simply update the era
        he->set_era(era);
        prev_era = era;
        continue;
      }
      he->release_guard();
      he = nullptr;
    }
    assert(he == nullptr);
    he = local_thread_data().alloc_hazard_era(era);
    prev_era = era;
  }
}
 
template <class Traits>
template <class T, class MarkedPtr>
bool hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::acquire_if_equal(const concurrent_ptr<T>& p,
                                                                    const MarkedPtr& expected,
                                                                    std::memory_order order) {
  if (order == std::memory_order_relaxed || order == std::memory_order_consume) {
    // we have to use memory_order_acquire (or something stricter) to ensure that
    // the era_clock.load cannot return an outdated value.
    order = std::memory_order_acquire;
  }
 
  // (2) - this load operation synchronizes-with any release operation on p.
  // we have to use acquire here to ensure that the subsequent era_clock.load
  // sees a value >= p.construction_era
  auto p1 = p.load(order);
  if (p1 == nullptr || p1 != expected) {
    reset();
    return p1 == expected;
  }
 
  const auto era = era_clock.load(std::memory_order_relaxed);
  if (he != nullptr && he->guards() == 1) {
    he->set_era(era);
  } else {
    if (he != nullptr) {
      he->release_guard();
    }
 
    he = local_thread_data().alloc_hazard_era(era);
  }
 
  this->ptr = p.load(std::memory_order_relaxed);
  if (this->ptr != p1) {
    reset();
    return false;
  }
  return true;
}
 
template <class Traits>
template <class T, class MarkedPtr>
void hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::reset() noexcept {
  local_thread_data().release_hazard_era(he);
  assert(this->he == nullptr);
  this->ptr.reset();
}
 
template <class Traits>
template <class T, class MarkedPtr>
void hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::do_swap(guard_ptr& g) noexcept {
  std::swap(he, g.he);
}
 
template <class Traits>
template <class T, class MarkedPtr>
void hazard_eras<Traits>::guard_ptr<T, MarkedPtr>::reclaim(Deleter d) noexcept {
  auto* p = this->ptr.get();
  reset();
  p->set_deleter(std::move(d));
  // (3) - this release fetch-add synchronizes-with the seq-cst fence (5)
  p->retirement_era = era_clock.fetch_add(1, std::memory_order_release);
 
  if (local_thread_data().add_retired_node(p) >= allocation_strategy::retired_nodes_threshold()) {
    local_thread_data().scan();
  }
}
 
namespace detail {
  template <class Strategy, class Derived>
  struct alignas(64) basic_he_thread_control_block :
      detail::thread_block_list<Derived, detail::deletable_object_with_eras>::entry,
      aligned_object<basic_he_thread_control_block<Strategy, Derived>> {
    using era_t = uint64_t;
 
    ~basic_he_thread_control_block() { assert(last_hazard_era != nullptr); }
 
    struct hazard_era {
      void set_era(era_t era) {
        assert(era != 0);
        // (4) - this release-store synchronizes-with the acquire-fence (10)
        value.store(marked_ptr(reinterpret_cast<void**>(era << 1)), std::memory_order_release);
        // This release is required because when acquire/acquire_if_equal is called on a
        // guard_ptr with with an active HE entry, set_era is called without an intermediate
        // call to set_link, i.e., the protected era is updated. This ensures the required
        // happens-before relation between releasing a guard_ptr to a node and reclamation
        // of that node.
 
        // (5) - this seq_cst-fence enforces a total order with the seq_cst-fence (9)
        //       and synchronizes-with the release-fetch-add (3)
        std::atomic_thread_fence(std::memory_order_seq_cst);
      }
 
      [[nodiscard]] era_t get_era() const {
        era_t result = 0;
        bool success = try_get_era(result);
        (void)success;
        assert(success);
        assert(result != 0);
        return result;
      }
 
      [[nodiscard]] uint64_t guards() const { return guard_cnt; }
      uint64_t add_guard() { return ++guard_cnt; }
      uint64_t release_guard() { return --guard_cnt; }
 
      bool try_get_era(era_t& result) const {
        // TSan does not support explicit fences, so we cannot rely on the acquire-fence (10)
        // but have to perform an acquire-load here to avoid false positives.
        constexpr auto memory_order = TSAN_MEMORY_ORDER(std::memory_order_acquire, std::memory_order_relaxed);
        auto v = value.load(memory_order);
        if (v.mark() == 0) {
          result = reinterpret_cast<era_t>(v.get()) >> 1;
          return true;
        }
        return false; // value contains a link
      }
 
      void set_link(hazard_era* link) {
        assert(guard_cnt == 0);
        // (6) - this release store synchronizes-with the acquire fence (10)
        value.store(marked_ptr(reinterpret_cast<void**>(link), 1), std::memory_order_release);
      }
      [[nodiscard]] hazard_era* get_link() const {
        assert(is_link());
        return reinterpret_cast<hazard_era*>(value.load(std::memory_order_relaxed).get());
      }
 
      [[nodiscard]] bool is_link() const { return value.load(std::memory_order_relaxed).mark() != 0; }
 
    private:
      using marked_ptr = xenium::marked_ptr<void*, 1>;
      // since we use the hazard era array to build our internal linked list of hazard eras we set
      // the LSB to signal that this is an internal pointer and not a pointer to a protected object.
      std::atomic<marked_ptr> value{nullptr};
      // the number of guard_ptrs that reference this instance and therefore observe the same era
      uint64_t guard_cnt = 0;
    };
 
    using hint = hazard_era*;
 
    void initialize(hint& hint) {
      Strategy::number_of_active_hes.fetch_add(self().number_of_hes(), std::memory_order_relaxed);
      hint = initialize_block(self());
    }
 
    void abandon() {
      Strategy::number_of_active_hes.fetch_sub(self().number_of_hes(), std::memory_order_relaxed);
      detail::thread_block_list<Derived, detail::deletable_object_with_eras>::entry::abandon();
    }
 
    hazard_era* alloc_hazard_era(hint& hint, era_t era) {
      if (last_hazard_era && last_era == era) {
        last_hazard_era->add_guard();
        return last_hazard_era;
      }
      auto result = hint;
      if (result == nullptr) {
        result = self().need_more_hes();
      }
 
      hint = result->get_link();
      result->set_era(era);
      result->add_guard();
 
      last_hazard_era = result;
      last_era = era;
      return result;
    }
 
    void release_hazard_era(hazard_era*& he, hint& hint) {
      assert(he != nullptr);
      if (he->release_guard() == 0) {
        if (he == last_hazard_era) {
          last_hazard_era = nullptr;
        }
 
        he->set_link(hint);
        hint = he;
      }
      he = nullptr;
    }
 
  protected:
    Derived& self() { return static_cast<Derived&>(*this); }
 
    hazard_era* begin() { return &eras[0]; }
    hazard_era* end() { return &eras[Strategy::K]; }
    [[nodiscard]] const hazard_era* begin() const { return &eras[0]; }
    [[nodiscard]] const hazard_era* end() const { return &eras[Strategy::K]; }
 
    template <typename T>
    static hazard_era* initialize_block(T& block) {
      auto begin = block.begin();
      auto end = block.end() - 1; // the last element is handled specially, so loop only over n-1 entries
      for (auto it = begin; it != end;) {
        auto next = it + 1;
        it->set_link(next);
        it = next;
      }
      end->set_link(block.initialize_next_block());
      return begin;
    }
 
    static void
      gather_protected_eras(std::vector<era_t>& protected_eras, const hazard_era* begin, const hazard_era* end) {
      for (auto it = begin; it != end; ++it) {
        era_t era;
        if (it->try_get_era(era)) {
          protected_eras.push_back(era);
        }
      }
    }
 
    hazard_era* last_hazard_era;
    era_t last_era;
    hazard_era eras[Strategy::K];
  };
 
  template <class Strategy>
  struct static_he_thread_control_block :
      basic_he_thread_control_block<Strategy, static_he_thread_control_block<Strategy>> {
    using base = basic_he_thread_control_block<Strategy, static_he_thread_control_block>;
    using hazard_era = typename base::hazard_era;
    using era_t = typename base::era_t;
    friend base;
 
    void gather_protected_eras(std::vector<era_t>& protected_eras) const {
      base::gather_protected_eras(protected_eras, this->begin(), this->end());
    }
 
  private:
    hazard_era* need_more_hes() { throw bad_hazard_era_alloc("hazard era pool exceeded"); }
    [[nodiscard]] constexpr size_t number_of_hes() const { return Strategy::K; }
    [[nodiscard]] constexpr hazard_era* initialize_next_block() const { return nullptr; }
  };
 
  template <class Strategy>
  struct dynamic_he_thread_control_block :
      basic_he_thread_control_block<Strategy, dynamic_he_thread_control_block<Strategy>> {
    using base = basic_he_thread_control_block<Strategy, dynamic_he_thread_control_block>;
    using hazard_era = typename base::hazard_era;
    using era_t = typename base::era_t;
    friend base;
 
    void gather_protected_eras(std::vector<era_t>& protected_eras) const {
      gather_protected_eras(*this, protected_eras);
    }
 
  private:
    struct alignas(64) hazard_eras_block : aligned_object<hazard_eras_block> {
      explicit hazard_eras_block(size_t size) : size(size) {
        for (auto it = begin(); it != end(); ++it) {
          new (it) hazard_era;
        }
      }
 
      hazard_era* begin() { return reinterpret_cast<hazard_era*>(this + 1); }
      hazard_era* end() { return begin() + size; }
 
      [[nodiscard]] const hazard_era* begin() const { return reinterpret_cast<const hazard_era*>(this + 1); }
      [[nodiscard]] const hazard_era* end() const { return begin() + size; }
 
      [[nodiscard]] const hazard_eras_block* next_block() const { return next; }
      hazard_era* initialize_next_block() { return next ? base::initialize_block(*next) : nullptr; }
 
      hazard_eras_block* next = nullptr;
      const size_t size;
    };
 
    [[nodiscard]] const hazard_eras_block* next_block() const {
      // (7) - this acquire-load synchronizes-with the release-store (8)
      return he_block.load(std::memory_order_acquire);
    }
    [[nodiscard]] size_t number_of_hes() const { return total_number_of_hes; }
    hazard_era* need_more_hes() { return allocate_new_hazard_eras_block(); }
 
    hazard_era* initialize_next_block() {
      auto block = he_block.load(std::memory_order_relaxed);
      return block ? base::initialize_block(*block) : nullptr;
    }
 
    template <typename T>
    static void gather_protected_eras(const T& block, std::vector<era_t>& protected_eras) {
      base::gather_protected_eras(protected_eras, block.begin(), block.end());
 
      const auto* next = block.next_block();
      if (next) {
        gather_protected_eras(*next, protected_eras);
      }
    }
 
    hazard_era* allocate_new_hazard_eras_block() {
      size_t hes = std::max(static_cast<size_t>(Strategy::K), total_number_of_hes / 2);
      total_number_of_hes += hes;
      Strategy::number_of_active_hes.fetch_add(hes, std::memory_order_relaxed);
 
      size_t buffer_size = sizeof(hazard_eras_block) + hes * sizeof(hazard_era);
      void* buffer = hazard_eras_block::operator new(buffer_size);
      auto block = ::new (buffer) hazard_eras_block(hes);
      auto result = this->initialize_block(*block);
      block->next = he_block.load(std::memory_order_relaxed);
      // (8) - this release-store synchronizes-with the acquire-load (7)
      he_block.store(block, std::memory_order_release);
      return result;
    }
 
    size_t total_number_of_hes = Strategy::K;
    std::atomic<hazard_eras_block*> he_block;
  };
} // namespace detail
 
template <class Traits>
struct alignas(64) hazard_eras<Traits>::thread_data : aligned_object<thread_data> {
  using HE = typename thread_control_block::hazard_era*;
 
  ~thread_data() {
    if (retire_list != nullptr) {
      scan();
      if (retire_list != nullptr) {
        global_thread_block_list.abandon_retired_nodes(retire_list);
      }
      retire_list = nullptr;
    }
 
    if (control_block != nullptr) {
      global_thread_block_list.release_entry(control_block);
      control_block = nullptr;
    }
  }
 
  HE alloc_hazard_era(era_t era) {
    ensure_has_control_block();
    return control_block->alloc_hazard_era(hint, era);
  }
 
  void release_hazard_era(HE& he) {
    if (he) {
      assert(control_block != nullptr);
      control_block->release_hazard_era(he, hint);
    }
  }
 
  std::size_t add_retired_node(detail::deletable_object_with_eras* p) {
    p->next = retire_list;
    retire_list = p;
    return ++number_of_retired_nodes;
  }
 
  void scan() {
    std::vector<era_t> protected_eras;
    protected_eras.reserve(allocation_strategy::number_of_active_hazard_eras());
 
    // (9) - this seq_cst-fence enforces a total order with the seq_cst-fence (5)
    std::atomic_thread_fence(std::memory_order_seq_cst);
 
    auto adopted_nodes = global_thread_block_list.adopt_abandoned_retired_nodes();
 
    std::for_each(
      global_thread_block_list.begin(), global_thread_block_list.end(), [&protected_eras](const auto& entry) {
        // TSan does not support explicit fences, so we cannot rely on the acquire-fence (10)
        // but have to perform an acquire-load here to avoid false positives.
        constexpr auto memory_order = TSAN_MEMORY_ORDER(std::memory_order_acquire, std::memory_order_relaxed);
        if (entry.is_active(memory_order)) {
          entry.gather_protected_eras(protected_eras);
        }
      });
 
    // (10) - this acquire-fence synchronizes-with the release-store (4, 6)
    std::atomic_thread_fence(std::memory_order_acquire);
 
    std::sort(protected_eras.begin(), protected_eras.end());
    auto last = std::unique(protected_eras.begin(), protected_eras.end());
    protected_eras.erase(last, protected_eras.end());
 
    auto* list = retire_list;
    retire_list = nullptr;
    number_of_retired_nodes = 0;
    reclaim_nodes(list, protected_eras);
    reclaim_nodes(adopted_nodes, protected_eras);
  }
 
private:
  void ensure_has_control_block() {
    if (control_block == nullptr) {
      control_block = global_thread_block_list.acquire_inactive_entry();
      control_block->initialize(hint);
      control_block->activate();
    }
  }
 
  void reclaim_nodes(detail::deletable_object_with_eras* list, const std::vector<era_t>& protected_eras) {
    while (list != nullptr) {
      auto* cur = list;
      list = list->next;
 
      auto era_it = std::lower_bound(protected_eras.begin(), protected_eras.end(), cur->construction_era);
      if (era_it == protected_eras.end() || *era_it > cur->retirement_era) {
        cur->delete_self();
      } else {
        add_retired_node(cur);
      }
    }
  }
 
  detail::deletable_object_with_eras* retire_list = nullptr;
  std::size_t number_of_retired_nodes = 0;
  typename thread_control_block::hint hint;
 
  thread_control_block* control_block = nullptr;
 
  friend class hazard_eras;
  ALLOCATION_COUNTER(hazard_eras);
};
 
template <class Traits>
inline typename hazard_eras<Traits>::thread_data& hazard_eras<Traits>::local_thread_data() {
  // workaround for a Clang-8 issue that causes multiple re-initializations of thread_local variables
  static thread_local thread_data local_thread_data;
  return local_thread_data;
}
 
#ifdef TRACK_ALLOCATIONS
template <class Traits>
inline void hazard_eras<Traits>::count_allocation() {
  local_thread_data().allocation_counter.count_allocation();
}
 
template <class Traits>
inline void hazard_eras<Traits>::count_reclamation() {
  local_thread_data().allocation_counter.count_reclamation();
}
#endif
} // namespace xenium::reclamation
 
#ifdef _MSC_VER
  #pragma warning(pop)
#endif