#pragma once

#include <iostream>

#include <unordered_map>
#include <shared_mutex>
#include <mutex>
#include <memory>

#include <sched.h>
#include <numa.h>
#include <numaif.h>

#include <dml/dml.hpp>

namespace dml {
    inline const std::string StatusCodeToString(const dml::status_code code) {
        switch (code) {
            case dml::status_code::ok: return "ok";
            case dml::status_code::false_predicate: return "false predicate";
            case dml::status_code::partial_completion: return "partial completion";
            case dml::status_code::nullptr_error: return "nullptr error";
            case dml::status_code::bad_size: return "bad size";
            case dml::status_code::bad_length: return "bad length";
            case dml::status_code::inconsistent_size: return "inconsistent size";
            case dml::status_code::dualcast_bad_padding: return "dualcast bad padding";
            case dml::status_code::bad_alignment: return "bad alignment";
            case dml::status_code::buffers_overlapping: return "buffers overlapping";
            case dml::status_code::delta_delta_empty: return "delta delta empty";
            case dml::status_code::batch_overflow: return "batch overflow";
            case dml::status_code::execution_failed: return "execution failed";
            case dml::status_code::unsupported_operation: return "unsupported operation";
            case dml::status_code::queue_busy: return "queue busy";
            case dml::status_code::error: return "unknown error";
            case dml::status_code::config_error: return "config error";
            default: return "unhandled error";
        }
    }
}

namespace dsacache {
    class Cache;

    /*
     * Class Description:
     * Holds all required information on one cache entry and is used
     * both internally by the Cache and externally by the user.
     *
     * Important Usage Notes:
     * The pointer is only updated in WaitOnCompletion() which
     * therefore must be called by the user at some point in order
     * to use the cached data. Using this class as T for
     * std::shared_ptr<T> is not recommended as references are
     * already counted internally.
     *
     * Cache Lifetime:
     * As long as the instance is referenced, the pointer it stores
     * is guaranteed to be either nullptr or pointing to a valid copy.
     *
     * Implementation Detail:
     * Performs self-reference counting with a shared atomic integer.
     * Therefore on creating a copy the reference count is increased
     * and with the destructor it is deacresed. If the last copy is
     * destroyed the actual underlying data is freed and all shared
     * variables deleted.
     *
     * Notes on Thread Safety:
     * Class is thread safe in any possible state and performs
     * reference counting and deallocation itself entirely atomically.
     */

    class CacheData {
    public:
        using dml_handler = dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>>;

    private:
        // data source and size of the block
        uint8_t* src_;
        size_t size_;

        // global reference counting object
        std::atomic<int32_t>* active_;

        // global cache-location pointer
        std::atomic<uint8_t*>* cache_;

        // object-local incomplete cache location pointer
        // which is only available in the first instance
        uint8_t* incomplete_cache_;

        // dml handler vector pointer which is only
        // available in the first instance
        std::unique_ptr<std::vector<dml_handler>> handlers_;

        // deallocates the global cache-location
        // and invalidates it
        void Deallocate();

        // checks whether there are at least two
        // valid references to this object which
        // is done as the cache always has one
        // internally to any living instance
        bool Active() const;

        friend Cache;

    public:
        CacheData(uint8_t* data, const size_t size);
        CacheData(const CacheData& other);
        ~CacheData();

        // waits on completion of caching operations
        // for this task and is safe to be called in
        // any state of the object
        void WaitOnCompletion();

        // returns the cache data location for this
        // instance which is valid as long as the
        // instance is alive - !!! this may also
        // yield a nullptr !!!
        uint8_t* GetDataLocation() const;
    };

    /*
     * Class Description:
     * Class will handle access to data through internal copies.
     * These are obtained via work submission to the Intel DSA which takes
     * care of asynchronously duplicating the data. The user will define
     * where these copies lie and which system nodes will perform the copy.
     * This is done through policy functions set during initialization.
     *
     * Placement Policy:
     * The Placement Policy Function decides on which node a particular
     * entry is to be placed, given the current executing node and the
     * data source node and data size. This in turn means that for one
     * datum, multiple cached copies may exist at one time.
     *
     * Cache Lifetime:
     * When accessing the cache, a CacheData-object will be returned.
     * As long as this object lives, the pointer which it holds is
     * guaranteed to be either nullptr or a valid copy. When destroyed
     * the entry is marked for deletion which is only carried out
     * when system memory pressure drives an automated cache flush.
     *
     * Restrictions:
     * - Overlapping Pointers may lead to undefined behaviour during
     *   manual cache invalidation which should not be used if you
     *   intend to have these types of pointers
     * - Cache Invalidation may only be performed manually and gives
     *   no ordering guarantees. Therefore, it is the users responsibility
     *   to ensure that results after invalidation have been generated
     *   using the latest state of data. The cache is best suited
     *   to static data.
     *
     * Notes on Thread Safety:
     * - Cache is completely thread-safe after initialization
     * - CacheData-class will handle deallocation of data itself by
     *   performing self-reference-counting atomically and only
     *   deallocating if the last reference is destroyed
     * - The internal cache state has one lock which is either
     *   acquired shared for reading the state (upon accessing an already
     *   cached element) or unique (accessing a new element, flushing, invalidating)
     * - Waiting on copy completion is done over an atomic-wait in copies
     *   of the original CacheData-instance
     * - Overall this class may experience performance issues due to the use
     *   of locking (in any configuration), lock contention (worsens with higher
     *   core count, node count and utilization) and atomics (worse in the same
     *   situations as lock contention)
     *
     * Improving Performance:
     * When data is never shared between threads or memory size for the cache is
     * not an issue you may consider having one Cache-instance per thread and removing
     * the lock in Cache and modifying the reference counting and waiting mechanisms
     * of CacheData accordingly (although this is high effort and will yield little due
     * to the atomics not being shared among cores/nodes).
     * Otherwise, one Cache-instance per node could also be considered. This will allow
     * the placement policy function to be barebones and reduces the lock contention and
     * synchronization impact of the atomic variables.
     */

    class Cache {
    public:
        // cache policy is defined as a type here to allow flexible usage of the cacher
        // given a numa destination node (where the data will be needed), the numa source
        // node (current location of the data) and the data size, this function should
        // return optimal cache placement
        // dst node and returned value can differ if the system, for example, has HBM
        // attached accessible directly to node n under a different node id m
        typedef int (CachePolicy)(const int numa_dst_node, const int numa_src_node, const size_t data_size);

        // copy policy specifies the copy-executing nodes for a given task
        // which allows flexibility in assignment for optimizing raw throughput
        // or choosing a conservative usage policy
        typedef std::vector<int> (CopyPolicy)(const int numa_dst_node, const int numa_src_node, const size_t data_size);

    private:
        // mutex for accessing the cache state map

        std::shared_mutex cache_mutex_;

        // map from [dst-numa-node,map2]
        // map2 from [data-ptr,cache-structure]

        std::unordered_map<uint8_t, std::unordered_map<uint8_t*, CacheData>> cache_state_;

        CachePolicy* cache_policy_function_ = nullptr;
        CopyPolicy* copy_policy_function_ = nullptr;

        // function used to submit a copy task on a specific node to the dml
        // engine on that node - will change the current threads node assignment
        // to achieve this so take care to restore this
        dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>> ExecuteCopy(
                const uint8_t* src, uint8_t* dst, const size_t size, const int node
        ) const;

        // allocates the required memory on the destination node
        // and then submits task to the dml library for processing
        // and attaches the handlers to the cache data structure
        void SubmitTask(CacheData* task, const int dst_node, const int src_node);

        // querries the policy functions for the given data and size
        // to obtain destination cache node, also returns the datas
        // source node for further usage
        // output may depend on the calling threads node assignment
        // as this is set as the "optimal placement" node
        void GetCacheNode(uint8_t* src, const size_t size, int* OUT_DST_NODE, int* OUT_SRC_NODE) const;

        // allocates memory of size "size" on the numa node "node"
        // and returns nullptr if this is not possible, also may
        // try to flush the cache of the requested node to
        // alleviate encountered shortage
        uint8_t* AllocOnNode(const size_t size, const int node);

        // checks whether the cache contains an entry for
        // the given data in the given memory node and
        // returns it, otherwise returns nullptr
        std::unique_ptr<CacheData> GetFromCache(uint8_t* src, const size_t size, const int dst_node);

    public:
        Cache() = default;
        Cache(const Cache& other) = delete;

        // initializes the cache with the two policy functions
        // only after this is it safe to use in a threaded environment
        void Init(CachePolicy* cache_policy_function, CopyPolicy* copy_policy_function);

        // function to perform data access through the cache
        std::unique_ptr<CacheData> Access(uint8_t* data, const size_t size);

        // flushes the cache of inactive entries
        // if node is -1 then the whole cache is
        // checked and otherwise the specified
        // node - no checks on node validity
        void Flush(const int node = -1);

        // forces out all entries from the
        // cache and therefore will also "forget"
        // still-in-use entries, these will still
        // be properly deleted, but the cache
        // will be fresh - use for testing
        void Clear();

        void Invalidate(uint8_t* data);
    };
}

inline void dsacache::Cache::Clear() {
    std::unique_lock<std::shared_mutex> lock(cache_mutex_);

    cache_state_.clear();

    Init(cache_policy_function_, copy_policy_function_);
}

inline void dsacache::Cache::Init(CachePolicy* cache_policy_function, CopyPolicy* copy_policy_function) {
    cache_policy_function_ = cache_policy_function;
    copy_policy_function_ = copy_policy_function;

    // initialize numa library

    numa_available();

    // obtain all available nodes
    // and those we may allocate
    // memory on

    const int nodes_max = numa_num_configured_nodes();
    const bitmask* valid_nodes = numa_get_mems_allowed();

    // prepare the cache state with entries
    // for all given nodes

    for (int node = 0; node < nodes_max; node++) {
        if (numa_bitmask_isbitset(valid_nodes, node)) {
            cache_state_.insert({node,{}});
        }
    }
}

inline std::unique_ptr<dsacache::CacheData> dsacache::Cache::Access(uint8_t* data, const size_t size) {
    // get destination numa node for the cache

    int dst_node = -1;
    int src_node = -1;

    GetCacheNode(data, size, &dst_node, &src_node);

    // TODO: at this point it could be beneficial to check whether
    // TODO: the given destination node is present as an entry
    // TODO: in the cache state to see if it is valid

    // check whether the data is already cached

    std::unique_ptr<CacheData> task = GetFromCache(data, size, dst_node);

    if (task != nullptr) {
        return std::move(task);
    }

    // at this point the requested data is not present in cache
    // and we create a caching task for it

    task = std::make_unique<CacheData>(data, size);

    {
        std::unique_lock<std::shared_mutex> lock(cache_mutex_);

        const auto state = cache_state_[dst_node].emplace(task->src_, *task);

        // if state.second is false then no insertion took place
        // which means that concurrently whith this thread
        // some other thread must have accessed the same
        // resource in which case we return the other
        // threads data cache structure

        if (!state.second) {
            std::cout << "[!] Found another cache instance for 0x" << std::hex << (uint64_t)task->src_ << std::dec << std::endl;
            return std::move(std::make_unique<CacheData>(state.first->second));
        }
    }

    SubmitTask(task.get(), dst_node, src_node);

    return std::move(task);
}

inline uint8_t* dsacache::Cache::AllocOnNode(const size_t size, const int node) {
    // allocate data on this node and flush the unused parts of the
    // cache if the operation fails and retry once
    // TODO: smarter flush strategy could keep some stuff cached

    // check currently free memory to see if the data fits

    long long int free_space = 0;
    numa_node_size64(node, &free_space);

    if (free_space < size) {
        std::cout << "[!] Memory shortage when allocating " << size << "B on node " << node << std::endl;

        // dst node lacks memory space so we flush the cache for this
        // node hoping to free enough currently unused entries to make
        // the second allocation attempt successful

        Flush(node);

        // re-test by getting the free space and checking again

        numa_node_size64(node, &free_space);

        if (free_space < size) {
            std::cout << "[x] Memory shortage after flush when allocating " << size << "B on node " << node << std::endl;

            return nullptr;
        }
    }

    uint8_t* dst = reinterpret_cast<uint8_t*>(numa_alloc_onnode(size, node));

    if (dst == nullptr) {
        std::cout << "[x] Allocation try failed for " << size << "B on node " << node << std::endl;

        return nullptr;
    }

    return dst;
}

inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, const int src_node) {
    uint8_t* dst = AllocOnNode(task->size_, dst_node);

    if (dst == nullptr) {
        std::cout << "[x] Allocation failed so we can not cache" << std::endl;
        return;
    }

    task->incomplete_cache_ = dst;

    // querry copy policy function for the nodes to use for the copy

    const std::vector<int> executing_nodes = copy_policy_function_(dst_node, src_node, task->size_);
    const size_t task_count = executing_nodes.size();

    // each task will copy one fair part of the total size
    // and in case the total size is not a factor of the
    // given task count the last node must copy the remainder

    const size_t size = task->size_ / task_count;
    const size_t last_size = size + task->size_ % task_count;

    // save the current numa node mask to restore later
    // as executing the copy task will place this thread
    // on a different node

    bitmask* nodemask = numa_get_run_node_mask();

    for (uint32_t i = 0; i < task_count; i++) {
        const size_t local_size = i + 1 == task_count ? size : last_size;
        const size_t local_offset = i * size;
        const uint8_t* local_src = task->src_ + local_offset;
        uint8_t* local_dst = dst + local_offset;

        task->handlers_->emplace_back(ExecuteCopy(local_src, local_dst, local_size, executing_nodes[i]));
    }

    // restore the previous nodemask

    numa_run_on_node_mask(nodemask);
    numa_free_nodemask(nodemask);
}

inline dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>> dsacache::Cache::ExecuteCopy(
        const uint8_t* src, uint8_t* dst, const size_t size, const int node
) const {
    dml::const_data_view srcv = dml::make_view(src, size);
    dml::data_view dstv = dml::make_view(dst, size);

    numa_run_on_node(node);

    return dml::submit<dml::automatic>(dml::mem_copy.block_on_fault(), srcv, dstv);
}

inline void dsacache::Cache::GetCacheNode(uint8_t* src, const size_t size, int* OUT_DST_NODE, int* OUT_SRC_NODE) const {
    // obtain numa node of current thread to determine where the data is needed

    const int current_cpu = sched_getcpu();
    const int current_node = numa_node_of_cpu(current_cpu);

    // obtain node that the given data pointer is allocated on

    *OUT_SRC_NODE = -1;
    get_mempolicy(OUT_SRC_NODE, NULL, 0, (void*)src, MPOL_F_NODE | MPOL_F_ADDR);

    // querry cache policy function for the destination numa node

    *OUT_DST_NODE = cache_policy_function_(current_node, *OUT_SRC_NODE, size);
}

inline void dsacache::Cache::Flush(const int node) {
    // this lambda is used because below we have two code paths that
    // flush nodes, either one single or all successively

    const auto FlushNode = [](std::unordered_map<uint8_t*,CacheData>& map) {
        // begin at the front of the map

        auto it = map.begin();

        // loop until we reach the end of the map

        while (it != map.end()) {
            // if the iterator points to an inactive element
            // then we may erase it

            if (it->second.Active() == false) {
                // erase the iterator from the map

                map.erase(it);

                // as the erasure invalidated out iterator
                // we must start at the beginning again

                it = map.begin();
            }
            else {
                // if element is active just move over to the next one

                it++;
            }
        }
    };

    {
        // we require exclusive lock as we modify the cache state

        std::unique_lock<std::shared_mutex> lock(cache_mutex_);

        // node == -1 means that cache on all nodes should be flushed

        if (node == -1) {
            for (auto& nc : cache_state_) {
                FlushNode(nc.second);
            }
        }
        else {
            FlushNode(cache_state_[node]);
        }
    }
}

inline std::unique_ptr<dsacache::CacheData> dsacache::Cache::GetFromCache(uint8_t* src, const size_t size, const int dst_node) {
    // the best situation is if this data is already cached
    // which we check in an unnamed block in which the cache
    // is locked for reading to prevent another thread
    // from marking the element we may find as unused and
    // clearing it

    // lock the cache state in shared-mode because we read

    std::shared_lock<std::shared_mutex> lock(cache_mutex_);

    // search for the data in our cache state structure at the given node

    const auto search = cache_state_[dst_node].find(src);

    // if the data is in our structure we continue

    if (search != cache_state_[dst_node].end()) {

        // now check whether the sizes match

        if (search->second.size_ >= size) {
            // return a unique copy of the entry which uses the object
            // lifetime and destructor to safely handle deallocation

            return std::move(std::make_unique<CacheData>(search->second));
        }
        else {
            // if the sizes missmatch then we clear the current entry from cache
            // which will cause its deletion only after the last possible outside
            // reference is also destroyed

            cache_state_[dst_node].erase(search);
        }
    }

    return nullptr;
}

void dsacache::Cache::Invalidate(uint8_t* data) {
    // as the cache is modified we must obtain a unique writers lock

    std::unique_lock<std::shared_mutex> lock(cache_mutex_);

    // loop through all per-node-caches available

    for (auto node : cache_state_) {
        // search for an entry for the given data pointer

        auto search = node.second.find(data);

        if (search != node.second.end()) {
            // if the data is represented in-cache
            // then it will be erased to re-trigger
            // caching on next access

            node.second.erase(search);
        }
    }
}

inline dsacache::CacheData::CacheData(uint8_t* data, const size_t size) {
    src_ = data;
    size_ = size;
    active_ = new std::atomic<int32_t>(1);
    cache_ = new std::atomic<uint8_t*>();
    incomplete_cache_ = nullptr;
    handlers_ = std::make_unique<std::vector<dml_handler>>();
}

inline dsacache::CacheData::CacheData(const dsacache::CacheData& other) {
    // we copy the ptr to the global atomic reference counter
    // and increase the amount of active references

    active_ = other.active_;
    const int current_active = active_->fetch_add(1);

    // source and size will be copied too
    // as well as the reference to the global
    // atomic cache pointer

    src_ = other.src_;
    size_ = other.size_;
    cache_ = other.cache_;

    // incomplete cache and handlers will not
    // be copied because only the first instance
    // will wait on the completion of handlers

    incomplete_cache_ = nullptr;
    handlers_ = nullptr;
}

inline dsacache::CacheData::~CacheData() {
    // if this is the first instance of this cache structure
    // and it has not been waited on and is now being destroyed
    // we must wait on completion here to ensure the cache
    // remains in a valid state

    if (handlers_ != nullptr) {
        WaitOnCompletion();
    }

    // due to fetch_sub returning the preivously held value
    // we must subtract one locally to get the current value

    const int32_t v = active_->fetch_sub(1) - 1;

    // if the returned value is zero or lower
    // then we must execute proper deletion
    // as this was the last reference

    if (v <= 0) {
        Deallocate();

        delete active_;
        delete cache_;
    }
}

inline void dsacache::CacheData::Deallocate() {
    // although deallocate should only be called from
    // a safe context to do so, it can not hurt to
    // defensively perform the operation atomically

    uint8_t* cache_local = cache_->exchange(nullptr);
    if (cache_local != nullptr) numa_free(cache_local, size_);

    // if the cache was never waited for then incomplete_cache_
    // may still contain a valid pointer which has to be freed

    if (incomplete_cache_ != nullptr) numa_free(incomplete_cache_, size_);
}

inline uint8_t* dsacache::CacheData::GetDataLocation() const {
    return cache_->load();
}

inline bool dsacache::CacheData::Active() const {
    // this entry is active if more than one
    // reference exists to it, as the Cache
    // will always keep one internally until
    // the entry is cleared from cache

    return active_->load() > 1;
}

inline void dsacache::CacheData::WaitOnCompletion() {
    // the cache data entry can be in two states
    // either it is the original one which has not
    // been waited for in which case the handlers
    // are non-null or it is not

    if (handlers_ == nullptr) {
        // when no handlers are attached to this cache entry we wait on a
        // value change for the cache structure from nullptr to non-null
        // which will either go through immediately if the cache is valid
        // already or wait until the handler-owning thread notifies us

        cache_->wait(nullptr);
    }
    else {
        // when the handlers are non-null there are some DSA task handlers
        // available on which we must wait here

        // abort is set if any operation encountered an error

        bool abort = false;

        for (auto& handler : *handlers_) {
            auto result = handler.get();

            if (result.status != dml::status_code::ok) {
                std::cerr << "[x] Encountered bad status code for operation: " << dml::StatusCodeToString(result.status) << std::endl;

                // if one of the copy tasks failed we abort the whole task
                // after all operations are completed on it

                abort = true;
            }
        }

        // the handlers are cleared after all have completed

        handlers_ = nullptr;

        // now we act depending on whether an abort has been
        // called for which signals operation incomplete

        if (abort) {
            // store nullptr in the cache location

            cache_->store(nullptr);

            // then free the now incomplete cache

            // TODO: it would be possible to salvage the
            // TODO: operation at this point but this
            // TODO: is quite complicated so we just abort

            numa_free(incomplete_cache_, size_);
        }
        else {
            // incomplete cache is now safe to use and therefore we
            // swap it with the global cache state of this entry
            // and notify potentially waiting threads

            cache_->store(incomplete_cache_);
        }

        // as a last step all waiting threads must
        // be notified (copies of this will wait on value
        // change of the cache) and the incomplete cache
        // is cleared to nullptr as it is not incomplete

        cache_->notify_all();
        incomplete_cache_ = nullptr;
    }
}