#pragma once

#include <atomic>
#include <vector>
#include <thread>
#include <unordered_map>
#include <shared_mutex>

#include <semaphore.h>

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

#include <dml/dml.hpp>

namespace offcache {
    // execution policy selects in which way the data is supposed to be cached
    // and returned with the following behaviour is guaranteed in addition to the
    // returned value being valid:
    // Immediate:        return as fast as possible
    //                   may return cached data, can return data in RAM
    //                   will trigger caching of the data provided
    // ImmediateNoCache: return as fast as possible and never trigger caching
    //                   same as Immediate but will not trigger caching
    // Relaxed:          no rapid return needed, take time
    //                   will trigger caching and may only return
    //                   once the caching is successful but can still
    //                   provide data in RAM
    enum class ExecutionPolicy {
        Relaxed, Immediate, ImmediateNoCache
    };

    // the cache task structure will be used to submit and
    // control a cache element, while providing source pointer
    // and size in bytes for submission
    //
    // then the submitting thread may wait on the atomic "result"
    // which will be notified by the cache worker upon processing
    // after which the atomic-bool-ptr active will also become valid
    struct CacheTask {
        uint8_t* data_;
        size_t size_;
        uint8_t* result_ = nullptr;
        uint8_t* maybe_result_ = nullptr;
        std::atomic<bool> active_ { true };
        std::atomic<bool> valid_ { false };
        std::vector<dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>>> handlers_;
    };

    // singleton which holds the cache workers
    // and is the place where work will be submited
    class CacheCoordinator {
    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);

    private:
        std::shared_mutex cache_mutex_;

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

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

        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;

        void SubmitTask(CacheTask* task);

        CacheTask* CreateTask(const uint8_t *data, const size_t size) const;

        void DestroyTask(CacheTask* task) const;

    public:
        void Init(CachePolicy* cache_policy_function, CopyPolicy* copy_policy_function);

        // function to perform data access through the cache
        // behaviour depends on the chosen execution policy
        // Immediate and ImmediateNoCache return a cache task
        // with guaranteed-valid result value where Relaxed
        // policy does not come with this guarantee.
        CacheTask* Access(uint8_t* data, const size_t size, const ExecutionPolicy policy);

        // waits upon completion of caching
        static void WaitOnCompletion(CacheTask* task);

        // invalidates the given pointer
        // afterwards the reference to the
        // cache task object may be forgotten
        static void SignalDataUnused(CacheTask* task);

        // returns the location of the cached data
        // which may or may not be valid
        static uint8_t* GetDataLocation(CacheTask* task);

        void Flush();
    };
}

inline void offcache::CacheCoordinator::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();
}

inline offcache::CacheTask* offcache::CacheCoordinator::Access(uint8_t* data, const size_t size, const ExecutionPolicy policy) {
    // 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
    {
        std::shared_lock<std::shared_mutex> lock(cache_mutex_);

        const auto search = cache_state_.find(data);

        if (search != cache_state_.end()) {
            if (search->second->size_ == size) {
                search->second->active_.store(true);
                // TODO: check for completed status depending on execution policy
                return search->second;
            }
            else {
                DestroyTask(search->second);
                cache_state_.erase(search);
            }
        }
    }

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

    CacheTask* task = CreateTask(data, size);

    if (policy == ExecutionPolicy::Immediate) {
        // in intermediate mode the returned task
        // object is guaranteed to be valid and therefore
        // its resulting location must be validated
        // after which we submit the task
        // maybe_result is then set by submit

        task->result_ = data;
        SubmitTask(task);
        return task;
    }
    else if (policy == ExecutionPolicy::ImmediateNoCache) {
        // for immediatenocache we just validate
        // the generated task and return it
        // we must also set maybe_result in case
        // someone waits on this

        task->result_ = data;
        task->maybe_result_ = data;
        return task;
    }
    else if (policy == ExecutionPolicy::Relaxed) {
        // for relaxed no valid task must be returned
        // and therefore we just submit and then give
        // the possible invalid task back with only
        // maybe_result set by submission
        
        SubmitTask(task);
        return task;
    }
    else {
        // this should not be reached
    }
}

inline void offcache::CacheCoordinator::SubmitTask(CacheTask* task) {
    // 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

    int data_node = -1;
    get_mempolicy(&data_node, NULL, 0, (void*)task->data_, MPOL_F_NODE | MPOL_F_ADDR);

    // querry cache policy function for the destination numa node

    const uint32_t dst_node = cache_policy_function_(current_node, data_node, task->size_);

    // 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

    uint8_t* dst = numa_alloc_onnode(task->size_, dst_node);

    if (dst == nullptr) {
        Flush();

        dst = numa_alloc_onnode(task->size_, dst_node);

        if (dst == nullptr) {
            return;
        }
    }

    task->maybe_result_ = dst;

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

    const std::vector<int> executing_nodes = copy_policy_function_(dst_node, data_node);
    const size_t task_count = executing_nodes.size();

    // at this point the task may be added to the cache structure
    // due to the task being initialized with the valid flag set to false

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

        const auto state = cache_state_.insert({task->data_, 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 must perform an abort
        // TODO: abort is not the only way to handle this situation

        if (!state.second) {
            // abort by doing the following steps
            // (1) free the allocated memory, (2) remove the "maybe result" as
            // we will not run the caching operation, (3) clear the sub tasks
            // for the very same reason, (4) set the result to the RAM-location

            numa_free(dst, task->size_);
            task->maybe_result_ = nullptr;
            task->result_ = task->data_;
            return;
        }
    }

    // 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

    const int 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->data_ + local_offset; 
        uint8_t* local_dst = dst + local_offset;

        const auto handler = ExecuteCopy(local_src, local_dst, local_size, executing_nodes[i]);
        task->handlers_.emplace_back(handler);
    }

    // set the valid flag of the task as all handlers
    // required for completion signal are registered

    task->valid_.store(true);
    task->valid_.notify_all();

    // restore the previous nodemask

    numa_run_on_node_mask(nodemask);
}

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

    numa_run_on_node(node);

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

inline offcache::CacheTask* offcache::CacheCoordinator::CreateTask(const uint8_t* data, const size_t size) const {
    CacheTask* task = new CacheTask();
    task->data_ = data;
    task->size_ = size;
    return task;
}

inline void offcache::CacheCoordinator::DestroyTask(CacheTask* task) const {
    numa_free(task->result_, task->size_);
    delete task;
}

inline void offcache::CacheCoordinator::WaitOnCompletion(CacheTask* task) {
    task->valid_.wait(false);

    for (auto& handler : task->handlers_) {
        auto result = handler.get();
        // TODO: handle the returned status code
    }

    task->handlers_.clear();
}

inline uint8_t* offcache::CacheCoordinator::GetDataLocation(CacheTask* task) {
    return task->result_;
}

inline void offcache::CacheCoordinator::SignalDataUnused(CacheTask* task) {
    task->active_.store(false);
}

inline void offcache::CacheCoordinator::Flush() {
    // TODO: there probably is a better way to implement this flush

    {
        std::unique_lock<std::shared_mutex> lock(cache_mutex_);
        
        auto it = cache_state_.begin();

        while (it != cache_state_.end()) {
            if (it->second->active_.load() == false) {
                DestroyTask(it->second);
                cache_state_.erase(it);
                it = cache_state_.begin();
            }
            else {
                it++;
            }
        }
    }
}