concurrent_hash_map.hpp

/*
 * Copyright 2019-2020, Intel Corporation
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 *     * Redistributions of source code must retain the above copyright
 *       notice, this list of conditions and the following disclaimer.
 *
 *     * Redistributions in binary form must reproduce the above copyright
 *       notice, this list of conditions and the following disclaimer in
 *       the documentation and/or other materials provided with the
 *       distribution.
 *
 *     * Neither the name of the copyright holder nor the names of its
 *       contributors may be used to endorse or promote products derived
 *       from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
#ifndef PMEMOBJ_CONCURRENT_HASH_MAP_HPP
#define PMEMOBJ_CONCURRENT_HASH_MAP_HPP
 
#include <libpmemobj++/detail/atomic_backoff.hpp>
#include <libpmemobj++/detail/common.hpp>
#include <libpmemobj++/detail/template_helpers.hpp>
 
#include <libpmemobj++/make_persistent.hpp>
#include <libpmemobj++/mutex.hpp>
#include <libpmemobj++/p.hpp>
#include <libpmemobj++/persistent_ptr.hpp>
#include <libpmemobj++/transaction.hpp>
 
#include <libpmemobj++/detail/persistent_pool_ptr.hpp>
#include <libpmemobj++/shared_mutex.hpp>
 
#include <atomic>
#include <cassert>
#include <functional>
#include <initializer_list>
#include <iterator> // for std::distance
#include <memory>
#include <mutex>
#include <thread>
#include <type_traits>
#include <utility>
 
namespace std
{
template <typename T>
struct hash<pmem::obj::p<T>> {
    size_t
    operator()(const pmem::obj::p<T> &x) const
    {
        return hash<T>()(x.get_ro());
    }
};
} /* namespace std */
 
namespace pmem
{
namespace obj
{
 
namespace concurrent_hash_map_internal
{
template <typename SharedMutexT>
class shared_mutex_scoped_lock {
    using rw_mutex_type = SharedMutexT;
 
public:
    shared_mutex_scoped_lock(const shared_mutex_scoped_lock &) = delete;
    shared_mutex_scoped_lock &
    operator=(const shared_mutex_scoped_lock &) = delete;
 
    shared_mutex_scoped_lock() : mutex(nullptr), is_writer(false)
    {
    }
 
    shared_mutex_scoped_lock(rw_mutex_type &m, bool write = true)
        : mutex(nullptr)
    {
        acquire(m, write);
    }
 
    ~shared_mutex_scoped_lock()
    {
        if (mutex)
            release();
    }
 
    void
    acquire(rw_mutex_type &m, bool write = true)
    {
        is_writer = write;
        mutex = &m;
        if (write)
            mutex->lock();
        else
            mutex->lock_shared();
    }
 
    void
    release()
    {
        assert(mutex);
        rw_mutex_type *m = mutex;
        mutex = nullptr;
        if (is_writer) {
            m->unlock();
        } else {
            m->unlock_shared();
        }
    }
 
    bool
    try_acquire(rw_mutex_type &m, bool write = true)
    {
        assert(!mutex);
        bool result;
        is_writer = write;
        result = write ? m.try_lock() : m.try_lock_shared();
        if (result)
            mutex = &m;
        return result;
    }
 
protected:
    rw_mutex_type *mutex;
    bool is_writer;
}; /* class shared_mutex_scoped_lock */
 
template <typename ScopedLockType>
using scoped_lock_upgrade_to_writer =
    decltype(std::declval<ScopedLockType>().upgrade_to_writer());
 
template <typename ScopedLockType>
using scoped_lock_has_upgrade_to_writer =
    detail::supports<ScopedLockType, scoped_lock_upgrade_to_writer>;
 
template <typename ScopedLockType>
using scoped_lock_downgrade_to_reader =
    decltype(std::declval<ScopedLockType>().downgrade_to_reader());
 
template <typename ScopedLockType>
using scoped_lock_has_downgrade_to_reader =
    detail::supports<ScopedLockType, scoped_lock_downgrade_to_reader>;
 
template <typename ScopedLockType,
      bool = scoped_lock_has_upgrade_to_writer<ScopedLockType>::value
          &&scoped_lock_has_downgrade_to_reader<ScopedLockType>::value>
class scoped_lock_traits {
public:
    using scope_lock_type = ScopedLockType;
 
    static bool
    initial_rw_state(bool write)
    {
        /* For upgradeable locks, initial state is always read */
        return false;
    }
 
    static bool
    upgrade_to_writer(scope_lock_type &lock)
    {
        return lock.upgrade_to_writer();
    }
 
    static bool
    downgrade_to_reader(scope_lock_type &lock)
    {
        return lock.downgrade_to_reader();
    }
};
 
template <typename ScopedLockType>
class scoped_lock_traits<ScopedLockType, false> {
public:
    using scope_lock_type = ScopedLockType;
 
    static bool
    initial_rw_state(bool write)
    {
        /* For non-upgradeable locks, we take lock in required mode
         * immediately */
        return write;
    }
 
    static bool
    upgrade_to_writer(scope_lock_type &lock)
    {
        /* This overload is for locks which do not support upgrade
         * operation. For those locks, upgrade_to_writer should not be
         * called when holding a read lock */
        return true;
    }
 
    static bool
    downgrade_to_reader(scope_lock_type &lock)
    {
        /* This overload is for locks which do not support downgrade
         * operation. For those locks, downgrade_to_reader should never
         * be called */
        assert(false);
 
        return false;
    }
};
}
 
template <typename Key, typename T, typename Hash = std::hash<Key>,
      typename KeyEqual = std::equal_to<Key>,
      typename MutexType = pmem::obj::shared_mutex,
      typename ScopedLockType = concurrent_hash_map_internal::
          shared_mutex_scoped_lock<MutexType>>
class concurrent_hash_map;
 
namespace concurrent_hash_map_internal
{
/* Helper method which throws an exception when called in a tx */
static inline void
check_outside_tx()
{
    if (pmemobj_tx_stage() != TX_STAGE_NONE)
        throw pmem::transaction_scope_error(
            "Function called inside transaction scope.");
}
 
template <typename Hash>
using transparent_key_equal = typename Hash::transparent_key_equal;
 
template <typename Hash>
using has_transparent_key_equal = detail::supports<Hash, transparent_key_equal>;
 
template <typename Hash, typename Pred,
      bool = has_transparent_key_equal<Hash>::value>
struct key_equal_type {
    using type = typename Hash::transparent_key_equal;
};
 
template <typename Hash, typename Pred>
struct key_equal_type<Hash, Pred, false> {
    using type = Pred;
};
 
template <typename Mutex, typename ScopedLockType>
void
assert_not_locked(Mutex &mtx)
{
#ifndef NDEBUG
    ScopedLockType scoped_lock;
    assert(scoped_lock.try_acquire(mtx));
    scoped_lock.release();
#else
    (void)mtx;
#endif
}
 
template <typename Key, typename T, typename MutexType, typename ScopedLockType>
struct hash_map_node {
    using mutex_t = MutexType;
 
    using scoped_t = ScopedLockType;
 
    using value_type = std::pair<const Key, T>;
 
    using node_ptr_t = detail::persistent_pool_ptr<
        hash_map_node<Key, T, mutex_t, scoped_t>>;
 
    node_ptr_t next;
 
    mutex_t mutex;
 
    value_type item;
 
    hash_map_node() : next(OID_NULL)
    {
    }
 
    hash_map_node(const node_ptr_t &_next) : next(_next)
    {
    }
 
    hash_map_node(node_ptr_t &&_next) : next(std::move(_next))
    {
    }
 
    hash_map_node(const node_ptr_t &_next, const Key &key)
        : next(_next), item(key, T())
    {
    }
 
    hash_map_node(const node_ptr_t &_next, const Key &key, const T &t)
        : next(_next), item(key, t)
    {
    }
 
    hash_map_node(const node_ptr_t &_next, value_type &&i)
        : next(_next), item(std::move(i))
    {
    }
 
    template <typename... Args>
    hash_map_node(node_ptr_t &&_next, Args &&... args)
        : next(std::forward<node_ptr_t>(_next)),
          item(std::forward<Args>(args)...)
    {
    }
 
    hash_map_node(const node_ptr_t &_next, const value_type &i)
        : next(_next), item(i)
    {
    }
 
    hash_map_node(const hash_map_node &) = delete;
 
    hash_map_node &operator=(const hash_map_node &) = delete;
}; /* struct node */
 
template <typename Bucket>
class segment_traits {
public:
    using segment_index_t = size_t;
    using size_type = size_t;
    using bucket_type = Bucket;
 
protected:
    constexpr static size_type max_allocation_size = PMEMOBJ_MAX_ALLOC_SIZE;
 
    constexpr static segment_index_t first_big_block = 27;
    /* TODO: avoid hardcoded value; need constexpr  similar to:
     * Log2(max_allocation_size / sizeof(bucket_type)) */
 
    constexpr static size_type big_block_size = size_type(1)
        << first_big_block;
 
    /* Block size in bytes cannot exceed max_allocation_size */
    static_assert((big_block_size * sizeof(bucket_type)) <
                  max_allocation_size,
              "Block size exceeds max_allocation_size");
 
    constexpr static segment_index_t
    first_block_in_segment(segment_index_t seg)
    {
        return seg < first_big_block
            ? seg
            : (first_big_block +
               (segment_index_t(1) << (seg - first_big_block)) - 1);
    }
 
    constexpr static size_type
    blocks_in_segment(segment_index_t seg)
    {
        return seg < first_big_block
            ? segment_index_t(1)
            : segment_index_t(1) << (seg - first_big_block);
    }
 
    constexpr static size_type
    block_size(segment_index_t b)
    {
        return b < first_big_block ? segment_size(b ? b : 1)
                       : big_block_size;
    }
 
public:
    constexpr static segment_index_t embedded_segments = 1;
 
    constexpr static size_type embedded_buckets = 1 << embedded_segments;
 
    constexpr static segment_index_t number_of_segments = 32;
 
    static const size_type first_block = 8;
 
    constexpr static segment_index_t
    number_of_blocks()
    {
        return first_block_in_segment(number_of_segments);
    }
 
    static segment_index_t
    segment_index_of(size_type index)
    {
        return segment_index_t(detail::Log2(index | 1));
    }
 
    constexpr static segment_index_t
    segment_base(segment_index_t k)
    {
        return (segment_index_t(1) << k) & ~segment_index_t(1);
    }
 
    constexpr static size_type
    segment_size(segment_index_t k)
    {
        return size_type(1) << k; // fake value for k == 0
    }
    static_assert(
        embedded_segments < first_big_block,
        "Number of embedded segments cannot exceed max_allocation_size");
}; /* End of class segment_traits */
 
template <typename BlockTable, typename SegmentTraits, bool is_const>
class segment_facade_impl : public SegmentTraits {
private:
    using traits_type = SegmentTraits;
    using traits_type::block_size;
    using traits_type::blocks_in_segment;
    using traits_type::embedded_buckets;
    using traits_type::embedded_segments;
    using traits_type::first_block;
    using traits_type::first_block_in_segment;
    using traits_type::segment_base;
    using traits_type::segment_size;
 
public:
    using table_reference =
        typename std::conditional<is_const, const BlockTable &,
                      BlockTable &>::type;
 
    using table_pointer =
        typename std::conditional<is_const, const BlockTable *,
                      BlockTable *>::type;
 
    using bucket_type = typename traits_type::bucket_type;
    using segment_index_t = typename traits_type::segment_index_t;
    using size_type = typename traits_type::size_type;
 
    segment_facade_impl(table_reference table, segment_index_t s)
        : my_table(&table), my_seg(s)
    {
        assert(my_seg < traits_type::number_of_segments);
    }
 
    segment_facade_impl(const segment_facade_impl &src)
        : my_table(src.my_table), my_seg(src.my_seg)
    {
    }
 
    segment_facade_impl(segment_facade_impl &&src) = default;
 
    segment_facade_impl &
    operator=(const segment_facade_impl &src)
    {
        my_table = src.my_table;
        my_seg = src.my_seg;
        return *this;
    }
 
    segment_facade_impl &
    operator=(segment_facade_impl &&src)
    {
        my_table = src.my_table;
        my_seg = src.my_seg;
        return *this;
    }
 
    bucket_type &operator[](size_type i) const
    {
        assert(i < size());
 
        segment_index_t table_block = first_block_in_segment(my_seg);
        size_type b_size = block_size(table_block);
 
        table_block += i / b_size;
        i = i % b_size;
 
        return (*my_table)[table_block][static_cast<std::ptrdiff_t>(i)];
    }
 
    segment_facade_impl &
    operator++()
    {
        ++my_seg;
        return *this;
    }
 
    segment_facade_impl
    operator++(int)
    {
        segment_facade_impl tmp = *this;
        ++(*this);
        return tmp;
    }
 
    segment_facade_impl &
    operator--()
    {
        --my_seg;
        return *this;
    }
 
    segment_facade_impl
    operator--(int)
    {
        segment_facade_impl tmp = *this;
        --(*this);
        return tmp;
    }
 
    segment_facade_impl &
    operator+=(segment_index_t off)
    {
        my_seg += off;
        return *this;
    }
 
    segment_facade_impl &
    operator-=(segment_index_t off)
    {
        my_seg -= off;
        return *this;
    }
 
    segment_facade_impl
    operator+(segment_index_t off) const
    {
        return segment_facade_impl(*(this->my_table),
                       this->my_seg + off);
    }
 
    segment_facade_impl
    operator-(segment_index_t off) const
    {
        return segment_facade_impl(*(this->my_table),
                       this->my_seg - off);
    }
 
    void
    enable(pool_base &pop)
    {
        assert(my_seg >= embedded_segments);
 
        if (my_seg < first_block) {
            enable_first_block(pop);
        } else {
            enable_big_segment(pop);
        }
    }
 
    void
    disable()
    {
        assert(my_seg >= embedded_segments);
 
        if (my_seg < first_block) {
            if (my_seg == embedded_segments) {
                size_type sz = segment_size(first_block) -
                    embedded_buckets;
                delete_persistent<bucket_type[]>(
                    (*my_table)[my_seg], sz);
            }
            (*my_table)[my_seg] = nullptr;
        } else {
            block_range blocks = segment_blocks(my_seg);
 
            for (segment_index_t b = blocks.first;
                 b < blocks.second; ++b) {
                if ((*my_table)[b] != nullptr) {
                    delete_persistent<bucket_type[]>(
                        (*my_table)[b], block_size(b));
                    (*my_table)[b] = nullptr;
                }
            }
        }
    }
 
    constexpr size_type
    size() const
    {
        return segment_size(my_seg ? my_seg : 1);
    }
 
    bool
    is_valid() const
    {
        block_range blocks = segment_blocks(my_seg);
 
        for (segment_index_t b = blocks.first; b < blocks.second; ++b) {
            if ((*my_table)[b] == nullptr)
                return false;
        }
 
        return true;
    }
 
private:
    using block_range = std::pair<segment_index_t, segment_index_t>;
 
    static block_range
    segment_blocks(segment_index_t seg)
    {
        segment_index_t begin = first_block_in_segment(seg);
 
        return block_range(begin, begin + blocks_in_segment(seg));
    }
 
    void
    enable_first_block(pool_base &pop)
    {
        assert(my_seg == embedded_segments);
        {
            transaction::manual tx(pop);
 
            size_type sz =
                segment_size(first_block) - embedded_buckets;
            (*my_table)[my_seg] =
                make_persistent<bucket_type[]>(sz);
 
            persistent_ptr<bucket_type> base =
                (*my_table)[embedded_segments].raw();
 
            for (segment_index_t s = my_seg + 1; s < first_block;
                 ++s) {
                std::ptrdiff_t off =
                    static_cast<std::ptrdiff_t>(
                        segment_base(s) -
                        segment_base(my_seg));
 
                (*my_table)[s] = (base + off).raw();
            }
 
            transaction::commit();
        }
    }
 
    void
    enable_big_segment(pool_base &pop)
    {
        block_range blocks = segment_blocks(my_seg);
        {
            transaction::manual tx(pop);
 
            for (segment_index_t b = blocks.first;
                 b < blocks.second; ++b) {
                assert((*my_table)[b] == nullptr);
                (*my_table)[b] = make_persistent<bucket_type[]>(
                    block_size(b));
            }
 
            transaction::commit();
        }
    }
 
    table_pointer my_table;
 
    segment_index_t my_seg;
}; /* End of class segment_facade_impl */
 
template <typename Key, typename T, typename MutexType, typename ScopedLockType>
class hash_map_base {
public:
    using mutex_t = MutexType;
    using scoped_t = ScopedLockType;
 
    using size_type = size_t;
 
    using hashcode_type = size_t;
 
    using node = hash_map_node<Key, T, mutex_t, scoped_t>;
 
    using node_ptr_t = detail::persistent_pool_ptr<node>;
 
    struct bucket {
        using mutex_t = MutexType;
        using scoped_t = ScopedLockType;
 
        mutex_t mutex;
 
        p<std::atomic<uint64_t>> rehashed;
 
        node_ptr_t node_list;
 
        bucket() : node_list(nullptr)
        {
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
            VALGRIND_HG_DISABLE_CHECKING(&rehashed,
                             sizeof(rehashed));
#endif
            rehashed.get_rw() = false;
        }
 
        bool
        is_rehashed(std::memory_order order)
        {
            return rehashed.get_ro().load(order);
        }
 
        void
        set_rehashed(std::memory_order order)
        {
            rehashed.get_rw().store(true, order);
        }
 
        bucket(const bucket &) = delete;
 
        bucket &operator=(const bucket &) = delete;
    }; /* End of struct bucket */
 
    using segment_traits_t = segment_traits<bucket>;
 
    using segment_index_t = typename segment_traits_t::segment_index_t;
 
    static const size_type embedded_buckets =
        segment_traits_t::embedded_buckets;
 
    static const size_type first_block = segment_traits_t::first_block;
 
    constexpr static size_type block_table_size =
        segment_traits_t::number_of_blocks();
 
    using segment_ptr_t = persistent_ptr<bucket[]>;
 
    using bucket_ptr_t = persistent_ptr<bucket>;
 
    using blocks_table_t = segment_ptr_t[block_table_size];
 
    using segment_enable_mutex_t = pmem::obj::mutex;
 
    struct features {
        uint32_t compat;
        uint32_t incompat;
    };
 
    /* --------------------------------------------------------- */
 
    p<uint64_t> my_pool_uuid;
 
    features layout_features;
 
    std::aligned_storage<sizeof(size_t), sizeof(size_t)>::type
        my_mask_reserved;
 
    /* my_mask always restored on restart. */
    p<std::atomic<hashcode_type>> my_mask;
 
    std::aligned_storage<32, 8>::type padding1;
 
    blocks_table_t my_table;
 
    /* It must be in separate cache line from my_mask due to performance
     * effects */
    p<std::atomic<size_type>> my_size;
 
    std::aligned_storage<24, 8>::type padding2;
 
    std::aligned_storage<64, 8>::type reserved;
 
    segment_enable_mutex_t my_segment_enable_mutex;
 
    bucket my_embedded_segment[embedded_buckets];
 
    /* --------------------------------------------------------- */
 
    static constexpr features
    header_features()
    {
        return {0, 0};
    }
 
    const std::atomic<hashcode_type> &
    mask() const noexcept
    {
        return my_mask.get_ro();
    }
 
    std::atomic<hashcode_type> &
    mask() noexcept
    {
        return my_mask.get_rw();
    }
 
    using const_segment_facade_t =
        segment_facade_impl<blocks_table_t, segment_traits_t, true>;
 
    using segment_facade_t =
        segment_facade_impl<blocks_table_t, segment_traits_t, false>;
 
    hash_map_base()
    {
        static_assert(
            sizeof(size_type) == sizeof(std::atomic<size_type>),
            "std::atomic should have the same layout as underlying integral type");
 
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
        VALGRIND_HG_DISABLE_CHECKING(&my_size, sizeof(my_size));
        VALGRIND_HG_DISABLE_CHECKING(&my_mask, sizeof(my_mask));
#endif
        layout_features = header_features();
 
        my_size.get_rw() = 0;
        PMEMoid oid = pmemobj_oid(this);
 
        assert(!OID_IS_NULL(oid));
 
        my_pool_uuid = oid.pool_uuid_lo;
 
        pool_base pop = get_pool_base();
        /* enable embedded segments */
        for (size_type i = 0; i < segment_traits_t::embedded_segments;
             ++i) {
            my_table[i] =
                pmemobj_oid(my_embedded_segment +
                        segment_traits_t::segment_base(i));
            segment_facade_t seg(my_table, i);
            mark_rehashed<false>(pop, seg);
        }
    }
 
    void
    calculate_mask()
    {
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
        VALGRIND_HG_DISABLE_CHECKING(&my_size, sizeof(my_size));
        VALGRIND_HG_DISABLE_CHECKING(&my_mask, sizeof(my_mask));
#endif
#if LIBPMEMOBJ_CPP_VG_PMEMCHECK_ENABLED
        VALGRIND_PMC_REMOVE_PMEM_MAPPING(&my_mask, sizeof(my_mask));
#endif
 
        hashcode_type m = embedded_buckets - 1;
 
        const_segment_facade_t segment(
            my_table, segment_traits_t::embedded_segments);
 
        while (segment.is_valid()) {
            m += segment.size();
            ++segment;
        }
 
        mask().store(m, std::memory_order_relaxed);
    }
 
    void
    restore_size(size_type actual_size)
    {
        my_size.get_rw().store(actual_size, std::memory_order_relaxed);
        pool_base pop = get_pool_base();
        pop.persist(my_size);
    }
 
    template <bool Flush = true>
    void
    mark_rehashed(pool_base &pop, segment_facade_t &segment)
    {
        for (size_type i = 0; i < segment.size(); ++i) {
            bucket *b = &(segment[i]);
 
            assert_not_locked<mutex_t, scoped_t>(b->mutex);
 
            b->set_rehashed(std::memory_order_relaxed);
        }
 
        if (Flush) {
            /* Flush in separate loop to avoid read-after-flush */
            for (size_type i = 0; i < segment.size(); ++i) {
                bucket *b = &(segment[i]);
                pop.flush(b->rehashed);
            }
 
            pop.drain();
        }
    }
 
    void
    enable_segment(segment_index_t k, bool is_initial = false)
    {
        assert(k);
 
        pool_base pop = get_pool_base();
        size_type sz;
 
        if (k >= first_block) {
            segment_facade_t new_segment(my_table, k);
 
            sz = new_segment.size();
            if (!new_segment.is_valid())
                new_segment.enable(pop);
 
            if (is_initial) {
                mark_rehashed(pop, new_segment);
            }
 
            /* double it to get entire capacity of the container */
            sz <<= 1;
        } else {
            /* the first block */
            assert(k == segment_traits_t::embedded_segments);
 
            for (segment_index_t i = k; i < first_block; ++i) {
                segment_facade_t new_segment(my_table, i);
 
                if (!new_segment.is_valid())
                    new_segment.enable(pop);
 
                if (is_initial) {
                    mark_rehashed(pop, new_segment);
                }
            }
 
            sz = segment_traits_t::segment_size(first_block);
        }
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
        ANNOTATE_HAPPENS_BEFORE(&my_mask);
#endif
        mask().store(sz - 1, std::memory_order_release);
    }
 
    bucket *
    get_bucket(hashcode_type h) const
    {
        segment_index_t s = segment_traits_t::segment_index_of(h);
 
        h -= segment_traits_t::segment_base(s);
 
        const_segment_facade_t segment(my_table, s);
 
        assert(segment.is_valid());
 
        return &(segment[h]);
    }
 
    inline bool
    check_mask_race(hashcode_type h, hashcode_type &m) const
    {
        hashcode_type m_now, m_old = m;
 
        m_now = mask().load(std::memory_order_acquire);
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
        ANNOTATE_HAPPENS_AFTER(&(this->my_mask));
#endif
 
        if (m_old != m_now)
            return check_rehashing_collision(h, m_old, m = m_now);
 
        return false;
    }
 
    bool
    check_rehashing_collision(hashcode_type h, hashcode_type m_old,
                  hashcode_type m) const
    {
        assert(m_old != m);
 
        if ((h & m_old) != (h & m)) {
            /* mask changed for this hashcode, rare event condition
             * above proves that 'h' has some other bits set beside
             * 'm_old', find next applicable mask after m_old */
 
            for (++m_old; !(h & m_old); m_old <<= 1)
                ;
 
            m_old = (m_old << 1) - 1; /* get full mask from a bit */
 
            assert((m_old & (m_old + 1)) == 0 && m_old <= m);
 
            /* check whether it is rehashing/ed */
            bucket *b = get_bucket(h & m_old);
            return b->is_rehashed(std::memory_order_acquire);
        }
 
        return false;
    }
 
    template <typename Node, typename... Args>
    size_type
    insert_new_node(bucket *b, detail::persistent_pool_ptr<Node> &new_node,
            Args &&... args)
    {
        pool_base pop = get_pool_base();
 
        pmem::obj::transaction::run(pop, [&] {
            new_node = pmem::obj::make_persistent<Node>(
                b->node_list, std::forward<Args>(args)...);
            b->node_list = new_node; /* bucket is locked */
        });
 
        /* prefix form is to enforce allocation after the first item
         * inserted */
        size_t sz = ++(my_size.get_rw());
        pop.persist(&my_size, sizeof(my_size));
 
        return sz;
    }
 
    bool
    check_growth(hashcode_type m, size_type sz)
    {
        if (sz >= m) {
            segment_index_t new_seg =
                static_cast<segment_index_t>(detail::Log2(
                    m +
                    1)); /* optimized segment_index_of */
 
            assert(segment_facade_t(my_table, new_seg - 1)
                       .is_valid());
 
            std::unique_lock<segment_enable_mutex_t> lock(
                my_segment_enable_mutex, std::try_to_lock);
 
            if (lock) {
                if (mask().load(std::memory_order_relaxed) ==
                    m) {
                    /* Otherwise, other thread enable this
                     * segment */
                    enable_segment(new_seg);
 
                    return true;
                }
            }
        }
 
        return false;
    }
 
    void
    reserve(size_type buckets)
    {
        if (buckets == 0)
            return;
 
        --buckets;
 
        bool is_initial = (my_size.get_ro() == 0);
 
        for (size_type m = mask(); buckets > m; m = mask())
            enable_segment(
                segment_traits_t::segment_index_of(m + 1),
                is_initial);
    }
 
    void
    internal_swap(hash_map_base<Key, T, mutex_t, scoped_t> &table)
    {
        pool_base p = get_pool_base();
        {
            transaction::manual tx(p);
 
            this->my_pool_uuid.swap(table.my_pool_uuid);
            /* Swap my_mask */
            this->mask() = table.mask().exchange(
                this->mask(), std::memory_order_relaxed);
 
            /* Swap my_size */
            this->my_size.get_rw() =
                table.my_size.get_rw().exchange(
                    this->my_size.get_ro(),
                    std::memory_order_relaxed);
 
            for (size_type i = 0; i < embedded_buckets; ++i)
                this->my_embedded_segment[i].node_list.swap(
                    table.my_embedded_segment[i].node_list);
 
            for (size_type i = segment_traits_t::embedded_segments;
                 i < block_table_size; ++i)
                this->my_table[i].swap(table.my_table[i]);
 
            transaction::commit();
        }
    }
 
    pool_base
    get_pool_base()
    {
        PMEMobjpool *pop =
            pmemobj_pool_by_oid(PMEMoid{my_pool_uuid, 0});
 
        return pool_base(pop);
    }
}; /* End of class hash_map_base */
 
template <typename Container, bool is_const>
class hash_map_iterator {
public:
    using iterator_category = std::forward_iterator_tag;
    using difference_type = ptrdiff_t;
    using map_type = Container;
    using value_type = typename map_type::value_type;
    using node = typename map_type::node;
    using bucket = typename map_type::bucket;
    using map_ptr = typename std::conditional<is_const, const map_type *,
                          map_type *>::type;
    using reference =
        typename std::conditional<is_const,
                      typename map_type::const_reference,
                      typename map_type::reference>::type;
    using pointer =
        typename std::conditional<is_const,
                      typename map_type::const_pointer,
                      typename map_type::pointer>::type;
 
    template <typename C, bool M, bool U>
    friend bool operator==(const hash_map_iterator<C, M> &i,
                   const hash_map_iterator<C, U> &j);
 
    template <typename C, bool M, bool U>
    friend bool operator!=(const hash_map_iterator<C, M> &i,
                   const hash_map_iterator<C, U> &j);
 
    friend class hash_map_iterator<map_type, true>;
 
#if !defined(_MSC_VER) || defined(__INTEL_COMPILER)
private:
    template <typename Key, typename T, typename Hash, typename KeyEqual,
          typename MutexType, typename ScopedLockType>
    friend class ::pmem::obj::concurrent_hash_map;
#else
public: /* workaround */
#endif
    hash_map_iterator(map_ptr map, size_t index)
        : my_map(map), my_index(index), my_bucket(nullptr), my_node(nullptr)
    {
        if (my_index <= my_map->mask()) {
            bucket_accessor acc(my_map, my_index);
            my_bucket = acc.get();
            my_node = static_cast<node *>(
                my_bucket->node_list.get(my_map->my_pool_uuid));
 
            if (!my_node) {
                advance_to_next_bucket();
            }
        }
    }
 
public:
    hash_map_iterator() = default;
 
    hash_map_iterator(const hash_map_iterator &other)
        : my_map(other.my_map),
          my_index(other.my_index),
          my_bucket(other.my_bucket),
          my_node(other.my_node)
    {
    }
 
    template <typename U = void,
          typename = typename std::enable_if<is_const, U>::type>
    hash_map_iterator(const hash_map_iterator<map_type, false> &other)
        : my_map(other.my_map),
          my_index(other.my_index),
          my_bucket(other.my_bucket),
          my_node(other.my_node)
    {
    }
 
    hash_map_iterator &operator=(const hash_map_iterator &it) = default;
 
    reference operator*() const
    {
        assert(my_node);
        return my_node->item;
    }
 
    pointer operator->() const
    {
        return &operator*();
    }
 
    hash_map_iterator &
    operator++()
    {
        my_node = static_cast<node *>(
            my_node->next.get((my_map->my_pool_uuid)));
 
        if (!my_node)
            advance_to_next_bucket();
 
        return *this;
    }
 
    hash_map_iterator
    operator++(int)
    {
        hash_map_iterator old(*this);
        operator++();
        return old;
    }
 
private:
    map_ptr my_map = nullptr;
 
    size_t my_index = 0;
 
    bucket *my_bucket = nullptr;
 
    node *my_node = nullptr;
 
    class bucket_accessor {
    public:
        bucket_accessor(map_ptr m, size_t index)
        {
            my_bucket = m->get_bucket(index);
        }
 
        bucket *
        get() const
        {
            return my_bucket;
        }
 
    private:
        bucket *my_bucket;
    };
 
    void
    advance_to_next_bucket()
    {
        size_t k = my_index + 1;
 
        assert(my_bucket);
 
        while (k <= my_map->mask()) {
            bucket_accessor acc(my_map, k);
            my_bucket = acc.get();
 
            if (my_bucket->node_list) {
                my_node = static_cast<node *>(
                    my_bucket->node_list.get(
                        my_map->my_pool_uuid));
 
                my_index = k;
 
                return;
            }
 
            ++k;
        }
 
        my_bucket = 0;
        my_node = 0;
        my_index = k;
    }
};
 
template <typename Container, bool M, bool U>
bool
operator==(const hash_map_iterator<Container, M> &i,
       const hash_map_iterator<Container, U> &j)
{
    return i.my_node == j.my_node && i.my_map == j.my_map;
}
 
template <typename Container, bool M, bool U>
bool
operator!=(const hash_map_iterator<Container, M> &i,
       const hash_map_iterator<Container, U> &j)
{
    return i.my_node != j.my_node || i.my_map != j.my_map;
}
} /* namespace concurrent_hash_map_internal */
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
class concurrent_hash_map
    : protected concurrent_hash_map_internal::hash_map_base<Key, T, MutexType,
                                ScopedLockType> {
    template <typename Container, bool is_const>
    friend class concurrent_hash_map_internal::hash_map_iterator;
 
public:
    using size_type = typename concurrent_hash_map_internal::hash_map_base<
        Key, T, MutexType, ScopedLockType>::size_type;
    using hashcode_type =
        typename concurrent_hash_map_internal::hash_map_base<
            Key, T, MutexType, ScopedLockType>::hashcode_type;
    using key_type = Key;
    using mapped_type = T;
    using value_type = typename concurrent_hash_map_internal::hash_map_base<
        Key, T, MutexType, ScopedLockType>::node::value_type;
    using difference_type = ptrdiff_t;
    using pointer = value_type *;
    using const_pointer = const value_type *;
    using reference = value_type &;
    using const_reference = const value_type &;
    using iterator = concurrent_hash_map_internal::hash_map_iterator<
        concurrent_hash_map, false>;
    using const_iterator = concurrent_hash_map_internal::hash_map_iterator<
        concurrent_hash_map, true>;
    using hasher = Hash;
    using key_equal = typename concurrent_hash_map_internal::key_equal_type<
        Hash, KeyEqual>::type;
 
protected:
    using mutex_t = MutexType;
    using scoped_t = ScopedLockType;
    /*
     * Explicitly use methods and types from template base class
     */
    using hash_map_base =
        concurrent_hash_map_internal::hash_map_base<Key, T, mutex_t,
                                scoped_t>;
    using hash_map_base::calculate_mask;
    using hash_map_base::check_growth;
    using hash_map_base::check_mask_race;
    using hash_map_base::embedded_buckets;
    using hash_map_base::get_bucket;
    using hash_map_base::get_pool_base;
    using hash_map_base::header_features;
    using hash_map_base::insert_new_node;
    using hash_map_base::internal_swap;
    using hash_map_base::layout_features;
    using hash_map_base::mask;
    using hash_map_base::reserve;
    using node = typename hash_map_base::node;
    using node_mutex_t = typename node::mutex_t;
    using node_ptr_t = typename hash_map_base::node_ptr_t;
    using bucket = typename hash_map_base::bucket;
    using bucket_lock_type = typename bucket::scoped_t;
    using segment_index_t = typename hash_map_base::segment_index_t;
    using segment_traits_t = typename hash_map_base::segment_traits_t;
    using segment_facade_t = typename hash_map_base::segment_facade_t;
    using scoped_lock_traits_type =
        concurrent_hash_map_internal::scoped_lock_traits<scoped_t>;
 
    friend class const_accessor;
    using persistent_node_ptr_t = detail::persistent_pool_ptr<node>;
 
    void
    delete_node(const node_ptr_t &n)
    {
        delete_persistent<node>(
            detail::static_persistent_pool_pointer_cast<node>(n)
                .get_persistent_ptr(this->my_pool_uuid));
    }
 
    template <typename K>
    persistent_node_ptr_t
    search_bucket(const K &key, bucket *b) const
    {
        assert(b->is_rehashed(std::memory_order_relaxed));
 
        persistent_node_ptr_t n =
            detail::static_persistent_pool_pointer_cast<node>(
                b->node_list);
 
        while (n &&
               !key_equal{}(key,
                    n.get(this->my_pool_uuid)->item.first)) {
            n = detail::static_persistent_pool_pointer_cast<node>(
                n.get(this->my_pool_uuid)->next);
        }
 
        return n;
    }
 
    class bucket_accessor : public bucket_lock_type {
        bucket *my_b;
 
    public:
        bucket_accessor(concurrent_hash_map *base,
                const hashcode_type h, bool writer = false)
        {
            acquire(base, h, writer);
        }
 
        inline void
        acquire(concurrent_hash_map *base, const hashcode_type h,
            bool writer = false)
        {
            my_b = base->get_bucket(h);
 
            if (my_b->is_rehashed(std::memory_order_acquire) ==
                    false &&
                bucket_lock_type::try_acquire(this->my_b->mutex,
                              /*write=*/true)) {
                if (my_b->is_rehashed(
                        std::memory_order_relaxed) ==
                    false) {
                    /* recursive rehashing */
                    base->rehash_bucket<false>(my_b, h);
                }
            } else {
                bucket_lock_type::acquire(my_b->mutex, writer);
            }
 
            assert(my_b->is_rehashed(std::memory_order_relaxed));
        }
 
        bool
        is_writer() const
        {
            return bucket_lock_type::is_writer;
        }
 
        bucket *
        get() const
        {
            return my_b;
        }
 
        bucket *operator->() const
        {
            return this->get();
        }
    };
 
    class serial_bucket_accessor {
        bucket *my_b;
 
    public:
        serial_bucket_accessor(concurrent_hash_map *base,
                       const hashcode_type h,
                       bool writer = false)
        {
            acquire(base, h, writer);
        }
 
        /*
         * Find a bucket by masked hashcode, optionally rehash
         */
        inline void
        acquire(concurrent_hash_map *base, const hashcode_type h,
            bool writer = false)
        {
            my_b = base->get_bucket(h);
 
            if (my_b->is_rehashed(std::memory_order_relaxed) ==
                false) {
                /* recursive rehashing */
                base->rehash_bucket<true>(my_b, h);
            }
 
            assert(my_b->is_rehashed(std::memory_order_relaxed));
        }
 
        bool
        is_writer() const
        {
            return true;
        }
 
        bucket *
        get() const
        {
            return my_b;
        }
 
        bucket *operator->() const
        {
            return this->get();
        }
    };
 
    hashcode_type
    get_hash_code(node_ptr_t &n)
    {
        return hasher{}(
            detail::static_persistent_pool_pointer_cast<node>(n)(
                this->my_pool_uuid)
                ->item.first);
    }
 
    template <bool serial>
    void
    rehash_bucket(bucket *b_new, const hashcode_type h)
    {
        using accessor_type = typename std::conditional<
            serial, serial_bucket_accessor, bucket_accessor>::type;
 
        using scoped_lock_traits_type =
            concurrent_hash_map_internal::scoped_lock_traits<
                accessor_type>;
 
        /* First two bucket should be always rehashed */
        assert(h > 1);
 
        pool_base pop = get_pool_base();
        node_ptr_t *p_new = &(b_new->node_list);
 
        /* This condition is only true when there was a failure just
         * before setting rehashed flag */
        if (*p_new != nullptr) {
            assert(!b_new->is_rehashed(std::memory_order_relaxed));
 
            b_new->set_rehashed(std::memory_order_relaxed);
            pop.persist(b_new->rehashed);
 
            return;
        }
 
        /* get parent mask from the topmost bit */
        hashcode_type mask = (1u << detail::Log2(h)) - 1;
        assert((h & mask) < h);
        accessor_type b_old(
            this, h & mask,
            scoped_lock_traits_type::initial_rw_state(true));
 
        pmem::obj::transaction::run(pop, [&] {
            /* get full mask for new bucket */
            mask = (mask << 1) | 1;
            assert((mask & (mask + 1)) == 0 && (h & mask) == h);
 
        restart:
            for (node_ptr_t *p_old = &(b_old->node_list),
                    n = *p_old;
                 n; n = *p_old) {
                hashcode_type c = get_hash_code(n);
#ifndef NDEBUG
                hashcode_type bmask = h & (mask >> 1);
 
                bmask = bmask == 0
                    ? 1 /* minimal mask of parent bucket */
                    : (1u << (detail::Log2(bmask) + 1)) - 1;
 
                assert((c & bmask) == (h & bmask));
#endif
 
                if ((c & mask) == h) {
                    if (!b_old.is_writer() &&
                        !scoped_lock_traits_type::
                            upgrade_to_writer(b_old)) {
                        goto restart;
                        /* node ptr can be invalid due
                         * to concurrent erase */
                    }
 
                    /* Add to new b_new */
                    *p_new = n;
 
                    /* exclude from b_old */
                    *p_old = n(this->my_pool_uuid)->next;
 
                    p_new = &(n(this->my_pool_uuid)->next);
                } else {
                    /* iterate to next item */
                    p_old = &(n(this->my_pool_uuid)->next);
                }
            }
 
            *p_new = nullptr;
        });
 
        /* mark rehashed */
        b_new->set_rehashed(std::memory_order_release);
        pop.persist(b_new->rehashed);
    }
 
    void
    check_features()
    {
        if (layout_features.incompat != header_features().incompat)
            throw pmem::layout_error(
                "Incompat flags mismatch, for more details go to: https://pmem.io/pmdk/cpp_obj/ \n");
    }
 
public:
    class accessor;
    class const_accessor
        : protected node::scoped_t /*which derived from no_copy*/ {
        friend class concurrent_hash_map<Key, T, Hash, KeyEqual,
                         mutex_t, scoped_t>;
        friend class accessor;
        using node_ptr_t = pmem::obj::persistent_ptr<node>;
        using node::scoped_t::try_acquire;
 
    public:
        using value_type =
            const typename concurrent_hash_map::value_type;
 
        bool
        empty() const
        {
            return !my_node;
        }
 
        void
        release()
        {
            concurrent_hash_map_internal::check_outside_tx();
 
            if (my_node) {
                node::scoped_t::release();
                my_node = 0;
            }
        }
 
        const_reference operator*() const
        {
            assert(my_node);
 
            return my_node->item;
        }
 
        const_pointer operator->() const
        {
            return &operator*();
        }
 
        const_accessor() : my_node(OID_NULL)
        {
            concurrent_hash_map_internal::check_outside_tx();
        }
 
        ~const_accessor()
        {
            my_node = OID_NULL; // scoped lock's release() is called
                        // in its destructor
        }
 
    protected:
        node_ptr_t my_node;
 
        hashcode_type my_hash;
    };
 
    class accessor : public const_accessor {
    public:
        using value_type = typename concurrent_hash_map::value_type;
 
        reference operator*() const
        {
            assert(this->my_node);
 
            return this->my_node->item;
        }
 
        pointer operator->() const
        {
            return &operator*();
        }
    };
 
    concurrent_hash_map() : hash_map_base()
    {
        runtime_initialize(true);
    }
 
    concurrent_hash_map(size_type n) : hash_map_base()
    {
        runtime_initialize(true);
 
        reserve(n);
    }
 
    concurrent_hash_map(const concurrent_hash_map &table) : hash_map_base()
    {
        runtime_initialize(true);
 
        reserve(table.size());
 
        internal_copy(table);
    }
 
    concurrent_hash_map(concurrent_hash_map &&table) : hash_map_base()
    {
        runtime_initialize(true);
 
        swap(table);
    }
 
    template <typename I>
    concurrent_hash_map(I first, I last)
    {
        runtime_initialize(true);
 
        reserve(static_cast<size_type>(std::distance(first, last)));
 
        internal_copy(first, last);
    }
 
    concurrent_hash_map(std::initializer_list<value_type> il)
    {
        runtime_initialize(true);
 
        reserve(il.size());
 
        internal_copy(il.begin(), il.end());
    }
 
    void
    runtime_initialize(bool graceful_shutdown = false)
    {
        check_features();
 
        calculate_mask();
 
        if (!graceful_shutdown) {
            auto actual_size =
                std::distance(this->begin(), this->end());
            assert(actual_size >= 0);
            this->restore_size(size_type(actual_size));
        } else {
            assert(this->size() ==
                   size_type(std::distance(this->begin(),
                               this->end())));
        }
    }
 
    concurrent_hash_map &
    operator=(const concurrent_hash_map &table)
    {
        if (this != &table) {
            clear();
            internal_copy(table);
        }
 
        return *this;
    }
 
    concurrent_hash_map &
    operator=(std::initializer_list<value_type> il)
    {
        clear();
 
        reserve(il.size());
 
        internal_copy(il.begin(), il.end());
 
        return *this;
    }
 
    void rehash(size_type n = 0);
 
    void clear();
 
    ~concurrent_hash_map()
    {
        clear();
    }
 
    //------------------------------------------------------------------------
    // STL support - not thread-safe methods
    //------------------------------------------------------------------------
 
    iterator
    begin()
    {
        return iterator(this, 0);
    }
 
    iterator
    end()
    {
        return iterator(this, mask() + 1);
    }
 
    const_iterator
    begin() const
    {
        return const_iterator(this, 0);
    }
 
    const_iterator
    end() const
    {
        return const_iterator(this, mask() + 1);
    }
 
    size_type
    size() const
    {
        return this->my_size.get_ro();
    }
 
    bool
    empty() const
    {
        return this->my_size.get_ro() == 0;
    }
 
    size_type
    max_size() const
    {
        return (~size_type(0)) / sizeof(node);
    }
 
    size_type
    bucket_count() const
    {
        return mask() + 1;
    }
 
    void swap(concurrent_hash_map &table);
 
    //------------------------------------------------------------------------
    // concurrent map operations
    //------------------------------------------------------------------------
 
    size_type
    count(const Key &key) const
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        return const_cast<concurrent_hash_map *>(this)->internal_find(
            key, nullptr, false);
    }
 
    template <typename K,
          typename = typename std::enable_if<
              concurrent_hash_map_internal::
                  has_transparent_key_equal<hasher>::value,
              K>::type>
    size_type
    count(const K &key) const
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        return const_cast<concurrent_hash_map *>(this)->internal_find(
            key, nullptr, false);
    }
 
    bool
    find(const_accessor &result, const Key &key) const
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return const_cast<concurrent_hash_map *>(this)->internal_find(
            key, &result, false);
    }
 
    template <typename K,
          typename = typename std::enable_if<
              concurrent_hash_map_internal::
                  has_transparent_key_equal<hasher>::value,
              K>::type>
    bool
    find(const_accessor &result, const K &key) const
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return const_cast<concurrent_hash_map *>(this)->internal_find(
            key, &result, false);
    }
 
    bool
    find(accessor &result, const Key &key)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_find(key, &result, true);
    }
 
    template <typename K,
          typename = typename std::enable_if<
              concurrent_hash_map_internal::
                  has_transparent_key_equal<hasher>::value,
              K>::type>
    bool
    find(accessor &result, const K &key)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_find(key, &result, true);
    }
    bool
    insert(const_accessor &result, const Key &key)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_insert(key, &result, false, key);
    }
 
    bool
    insert(accessor &result, const Key &key)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_insert(key, &result, true, key);
    }
 
    bool
    insert(const_accessor &result, const value_type &value)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_insert(value.first, &result, false, value);
    }
 
    bool
    insert(accessor &result, const value_type &value)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_insert(value.first, &result, true, value);
    }
 
    bool
    insert(const value_type &value)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        return internal_insert(value.first, nullptr, false, value);
    }
 
    bool
    insert(const_accessor &result, value_type &&value)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_insert(value.first, &result, false,
                       std::move(value));
    }
 
    bool
    insert(accessor &result, value_type &&value)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        result.release();
 
        return internal_insert(value.first, &result, true,
                       std::move(value));
    }
 
    bool
    insert(value_type &&value)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        return internal_insert(value.first, nullptr, false,
                       std::move(value));
    }
 
    template <typename I>
    void
    insert(I first, I last)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        for (; first != last; ++first)
            insert(*first);
    }
 
    void
    insert(std::initializer_list<value_type> il)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        insert(il.begin(), il.end());
    }
 
    bool
    erase(const Key &key)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        return internal_erase(key);
    }
 
    template <typename K,
          typename = typename std::enable_if<
              concurrent_hash_map_internal::
                  has_transparent_key_equal<hasher>::value,
              K>::type>
    bool
    erase(const K &key)
    {
        concurrent_hash_map_internal::check_outside_tx();
 
        return internal_erase(key);
    }
 
protected:
    /*
     * Try to acquire the mutex for read or write.
     *
     * If acquiring succeeds returns true, otherwise retries for few times.
     * If acquiring fails after all attempts returns false.
     */
    bool try_acquire_item(const_accessor *result, node_mutex_t &mutex,
                  bool write);
 
    template <typename K>
    bool internal_find(const K &key, const_accessor *result, bool write);
 
    template <typename K, typename... Args>
    bool internal_insert(const K &key, const_accessor *result, bool write,
                 Args &&... args);
 
    /* Obtain pointer to node and lock bucket */
    template <bool Bucket_rw_lock, typename K>
    persistent_node_ptr_t
    get_node(const K &key, bucket_accessor &b)
    {
        /* find a node */
        auto n = search_bucket(key, b.get());
 
        if (!n) {
            if (Bucket_rw_lock && !b.is_writer() &&
                !scoped_lock_traits_type::upgrade_to_writer(b)) {
                /* Rerun search_list, in case another
                 * thread inserted the item during the
                 * upgrade. */
                n = search_bucket(key, b.get());
                if (n) {
                    /* unfortunately, it did */
                    scoped_lock_traits_type::
                        downgrade_to_reader(b);
                    return n;
                }
            }
        }
 
        return n;
    }
 
    template <typename K>
    bool internal_erase(const K &key);
 
    void clear_segment(segment_index_t s);
 
    void internal_copy(const concurrent_hash_map &source);
 
    template <typename I>
    void internal_copy(I first, I last);
 
}; // class concurrent_hash_map
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
bool
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
            ScopedLockType>::try_acquire_item(const_accessor *result,
                              node_mutex_t &mutex,
                              bool write)
{
    /* acquire the item */
    if (!result->try_acquire(mutex, write)) {
        for (detail::atomic_backoff backoff(true);;) {
            if (result->try_acquire(mutex, write))
                break;
 
            if (!backoff.bounded_pause())
                return false;
        }
    }
 
    return true;
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
template <typename K>
bool
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
            ScopedLockType>::internal_find(const K &key,
                           const_accessor *result,
                           bool write)
{
    assert(!result || !result->my_node);
 
    hashcode_type m = mask().load(std::memory_order_acquire);
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
    ANNOTATE_HAPPENS_AFTER(&(this->my_mask));
#endif
 
    assert((m & (m + 1)) == 0);
 
    hashcode_type const h = hasher{}(key);
 
    persistent_node_ptr_t node;
 
    while (true) {
        /* get bucket and acquire the lock */
        bucket_accessor b(
            this, h & m,
            scoped_lock_traits_type::initial_rw_state(false));
        node = get_node<false>(key, b);
 
        if (!node) {
            /* Element was possibly relocated, try again */
            if (check_mask_race(h, m)) {
                b.release();
                continue;
            } else {
                return false;
            }
        }
 
        /* No need to acquire the item or item acquired */
        if (!result ||
            try_acquire_item(
                result, node.get(this->my_pool_uuid)->mutex, write))
            break;
 
        /* the wait takes really long, restart the
         * operation */
        b.release();
 
        std::this_thread::yield();
 
        m = mask().load(std::memory_order_acquire);
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
        ANNOTATE_HAPPENS_AFTER(&(this->my_mask));
#endif
    }
 
    if (result) {
        result->my_node = node.get_persistent_ptr(this->my_pool_uuid);
        result->my_hash = h;
    }
 
    return true;
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
template <typename K, typename... Args>
bool
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
            ScopedLockType>::internal_insert(const K &key,
                             const_accessor *result,
                             bool write,
                             Args &&... args)
{
    assert(!result || !result->my_node);
 
    hashcode_type m = mask().load(std::memory_order_acquire);
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
    ANNOTATE_HAPPENS_AFTER(&(this->my_mask));
#endif
 
    assert((m & (m + 1)) == 0);
 
    hashcode_type const h = hasher{}(key);
 
    persistent_node_ptr_t node;
    size_t new_size = 0;
    bool inserted = false;
 
    while (true) {
        /* get bucket and acquire the lock */
        bucket_accessor b(
            this, h & m,
            scoped_lock_traits_type::initial_rw_state(true));
        node = get_node<true>(key, b);
 
        if (!node) {
            /* Element was possibly relocated, try again */
            if (check_mask_race(h, m)) {
                b.release();
                continue;
            }
 
            /* insert and set flag to grow the container */
            new_size = insert_new_node(b.get(), node,
                           std::forward<Args>(args)...);
            inserted = true;
        }
 
        /* No need to acquire the item or item acquired */
        if (!result ||
            try_acquire_item(
                result, node.get(this->my_pool_uuid)->mutex, write))
            break;
 
        /* the wait takes really long, restart the
         * operation */
        b.release();
 
        std::this_thread::yield();
 
        m = mask().load(std::memory_order_acquire);
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
        ANNOTATE_HAPPENS_AFTER(&(this->my_mask));
#endif
    }
 
    if (result) {
        result->my_node = node.get_persistent_ptr(this->my_pool_uuid);
        result->my_hash = h;
    }
 
    check_growth(m, new_size);
 
    return inserted;
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
template <typename K>
bool
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
            ScopedLockType>::internal_erase(const K &key)
{
    node_ptr_t n;
    hashcode_type const h = hasher{}(key);
    hashcode_type m = mask().load(std::memory_order_acquire);
#if LIBPMEMOBJ_CPP_VG_HELGRIND_ENABLED
    ANNOTATE_HAPPENS_AFTER(&(this->my_mask));
#endif
 
    pool_base pop = get_pool_base();
 
restart : {
    /* lock scope */
    /* get bucket */
    bucket_accessor b(this, h & m,
              scoped_lock_traits_type::initial_rw_state(true));
 
search:
    node_ptr_t *p = &b->node_list;
    n = *p;
 
    while (n &&
           !key_equal{}(key,
                detail::static_persistent_pool_pointer_cast<node>(
                    n)(this->my_pool_uuid)
                    ->item.first)) {
        p = &n(this->my_pool_uuid)->next;
        n = *p;
    }
 
    if (!n) {
        /* not found, but mask could be changed */
        if (check_mask_race(h, m))
            goto restart;
 
        return false;
    } else if (!b.is_writer() &&
           !scoped_lock_traits_type::upgrade_to_writer(b)) {
        if (check_mask_race(h, m)) /* contended upgrade, check mask */
            goto restart;
 
        goto search;
    }
 
    persistent_ptr<node> del = n(this->my_pool_uuid);
 
    {
        /* We cannot remove this element immediately because
         * other threads might work with this element via
         * accessors. The item_locker required to wait while
         * other threads use the node. */
        const_accessor acc;
        if (!try_acquire_item(&acc, del->mutex, true)) {
            /* the wait takes really long, restart the operation */
            b.release();
 
            std::this_thread::yield();
 
            m = mask().load(std::memory_order_acquire);
 
            goto restart;
        }
    }
 
    /* Only one thread can delete it due to write lock on the bucket */
    transaction::run(pop, [&] {
        *p = del->next;
        delete_node(del);
    });
 
    --(this->my_size.get_rw());
    pop.persist(this->my_size);
}
 
    return true;
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
void
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType, ScopedLockType>::swap(
    concurrent_hash_map<Key, T, Hash, KeyEqual, mutex_t, scoped_t> &table)
{
    internal_swap(table);
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
void
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType, ScopedLockType>::rehash(
    size_type sz)
{
    concurrent_hash_map_internal::check_outside_tx();
 
    reserve(sz);
    hashcode_type m = mask();
 
    /* only the last segment should be scanned for rehashing size or first
     * index of the last segment */
    hashcode_type b = (m + 1) >> 1;
 
    /* zero or power of 2 */
    assert((b & (b - 1)) == 0);
 
    for (; b <= m; ++b) {
        bucket *bp = get_bucket(b);
 
        concurrent_hash_map_internal::assert_not_locked<mutex_t,
                                scoped_t>(
            bp->mutex);
        /* XXX Need to investigate if this statement is needed */
        if (bp->is_rehashed(std::memory_order_relaxed) == false)
            rehash_bucket<true>(bp, b);
    }
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
void
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType, ScopedLockType>::clear()
{
    hashcode_type m = mask();
 
    assert((m & (m + 1)) == 0);
 
#ifndef NDEBUG
    /* check consistency */
    for (segment_index_t b = 0; b <= m; ++b) {
        bucket *bp = get_bucket(b);
        concurrent_hash_map_internal::assert_not_locked<mutex_t,
                                scoped_t>(
            bp->mutex);
    }
#endif
 
    pool_base pop = get_pool_base();
    { /* transaction scope */
 
        transaction::manual tx(pop);
 
        this->my_size.get_rw() = 0;
        segment_index_t s = segment_traits_t::segment_index_of(m);
 
        assert(s + 1 == this->block_table_size ||
               !segment_facade_t(this->my_table, s + 1).is_valid());
 
        do {
            clear_segment(s);
        } while (s-- > 0);
 
        mask().store(embedded_buckets - 1, std::memory_order_relaxed);
 
        transaction::commit();
    }
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
void
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
            ScopedLockType>::clear_segment(segment_index_t s)
{
    segment_facade_t segment(this->my_table, s);
 
    assert(segment.is_valid());
 
    size_type sz = segment.size();
    for (segment_index_t i = 0; i < sz; ++i) {
        for (node_ptr_t n = segment[i].node_list; n;
             n = segment[i].node_list) {
            segment[i].node_list = n(this->my_pool_uuid)->next;
            delete_node(n);
        }
    }
 
    if (s >= segment_traits_t::embedded_segments)
        segment.disable();
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
void
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType, ScopedLockType>::
	internal_copy(const concurrent_hash_map &source)
{
    reserve(source.my_size.get_ro());
    internal_copy(source.begin(), source.end());
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
template <typename I>
void
concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
            ScopedLockType>::internal_copy(I first, I last)
{
    hashcode_type m = mask();
 
    for (; first != last; ++first) {
        hashcode_type h = hasher{}(first->first);
        bucket *b = get_bucket(h & m);
 
        assert(b->is_rehashed(std::memory_order_relaxed));
 
        detail::persistent_pool_ptr<node> p;
        insert_new_node(b, p, *first);
    }
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
inline bool
operator==(const concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
                     ScopedLockType> &a,
       const concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
                     ScopedLockType> &b)
{
    if (a.size() != b.size())
        return false;
 
    typename concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
                     ScopedLockType>::const_iterator
        i(a.begin()),
        i_end(a.end());
 
    typename concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
                     ScopedLockType>::const_iterator j,
        j_end(b.end());
 
    for (; i != i_end; ++i) {
        j = b.equal_range(i->first).first;
 
        if (j == j_end || !(i->second == j->second))
            return false;
    }
 
    return true;
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
inline bool
operator!=(const concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
                     ScopedLockType> &a,
       const concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType,
                     ScopedLockType> &b)
{
    return !(a == b);
}
 
template <typename Key, typename T, typename Hash, typename KeyEqual,
      typename MutexType, typename ScopedLockType>
inline void
swap(concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType, ScopedLockType> &a,
     concurrent_hash_map<Key, T, Hash, KeyEqual, MutexType, ScopedLockType> &b)
{
    a.swap(b);
}
 
} /* namespace obj */
} /* namespace pmem */
 
#endif /* PMEMOBJ_CONCURRENT_HASH_MAP_HPP */