bachelor-thesis/offloading-cacher/cache.hpp

#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 {
    inline bool CheckFlag(const uint64_t value, const uint64_t flag) {
        return (value & flag) != 0;
    }
    inline uint64_t UnsetFlag(const uint64_t value, const uint64_t flag) {
        return value & (~flag);
    }
    inline uint64_t SetFlag(const uint64_t value, const uint64_t flag) {
        return value | flag;
    }

    constexpr uint64_t FLAG_WAIT_WEAK = 0b1ULL << 63;
    constexpr uint64_t FLAG_HANDLE_PF = 0b1ULL << 62;
    constexpr uint64_t FLAG_ACCESS_WEAK = 0b1ULL << 61;
    constexpr uint64_t FLAG_DEFAULT = 0ULL;

    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:
        static constexpr uint64_t maxptr = 0xffff'ffff'ffff'ffff;

        // set to false if we do not own the cache pointer
        bool delete_ = false;

        // 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
        // contract: only access when being in sole posession of handlers
        uint8_t** incomplete_cache_;

        // flags inherited from parent cache
        uint64_t flags_ = 0;

        // dml handler vector pointer which is used
        // to wait on caching task completion
        std::atomic<dml_handler*>* handler_;

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

        size_t GetSize() const { return size_; }
        uint8_t* GetSource() const { return src_; }
        int32_t GetRefCount() const { return active_->load(); }
        void SetTaskHandlerAndCache(uint8_t* cache, dml_handler* handler);
        void SetCacheToSource() { cache_->store(src_); delete_ = false; }

        // initializes the class after which it is thread safe
        // but may only be destroyed safely after setting handlers
        void Init();

        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 { return cache_->load(); }

        void SetFlags(const uint64_t flags) { flags_ = flags; }
        uint64_t GetFlags() const { return flags_; }
    };

    /*
     * 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:
        // flags to store options duh

        uint64_t flags_ = 0;

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

        struct LockedNodeCacheState {
            std::shared_mutex cache_mutex_;
            std::unordered_map<uint8_t*, CacheData> node_cache_state_;
        };

        std::unordered_map<uint8_t, LockedNodeCacheState*> 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();
        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, behaviour depends
        // on flags, by default will also perform prefetch, otherwise with
        // FLAG_ACCESS_WEAK set will not perform prefetch and instead return
        // a cache entry with the data source as cache location on cache miss,
        // this flag must be set for each invocation, the flags set for the
        // entire cache will not be evaluated for this
        std::unique_ptr<CacheData> Access(uint8_t* data, const size_t size, const uint64_t flags = FLAG_DEFAULT);

        // 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);

        void SetFlags(const uint64_t flags) { flags_ = flags; }
        uint64_t GetFlags() { return flags_; }
    };
}

inline void dsacache::Cache::Clear() {
    for (auto& nc : cache_state_) {
        std::unique_lock<std::shared_mutex> lock(nc.second->cache_mutex_);
        nc.second->node_cache_state_.clear();
    }
}

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)) {
            void* block = numa_alloc_onnode(sizeof(LockedNodeCacheState), node);
            auto* state = new(block)LockedNodeCacheState;
            cache_state_.insert({node,state});
        }
    }
}

inline std::unique_ptr<dsacache::CacheData> dsacache::Cache::Access(uint8_t* data, const size_t size, const uint64_t flags) {
    // get destination numa node for the cache

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

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

    // 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, copying our current flags

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

    // when the ACCESS_WEAK flag is set for the flags parameter (!)
    // and we have reached this point, there was no cache entry
    // present for the requested data and therefore we abort
    // but to keep validity, we return the previously created
    // CacheData struct, setting the cache variable to the
    // data source location

    if (CheckFlag(flags, FLAG_ACCESS_WEAK)) {
        std::cerr << "[!] CacheAccess with WEAK set encountered miss!" << std::endl;
        task->SetCacheToSource();
        return std::move(task);
    }

    // the following operation adds the task to the cache state
    // which requires unique locking of the current nodes entry

    {
        LockedNodeCacheState* local_cache_state = cache_state_[dst_node];

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

        const auto state = local_cache_state->node_cache_state_.emplace(task->GetSource(), *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) {
            return std::move(std::make_unique<CacheData>(state.first->second));
        }

        // initialize the task now for thread safety
        // as we are now sure that we will submit work
        // to it and will not delete it beforehand

        task->Init();
    }

    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) {
        // 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) {
            return nullptr;
        }
    }

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

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

    return dst;
}

inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, const int src_node) {
    static thread_local int last_node_index = -1;

    // stores the last node used for the local thread so we can achieve some
    // load balancing which locally might look like round robin, but considering
    // that one source thread may see different results for "executing_nodes" with
    // different sizes, and that multiple threads will submit, in reality we
    // achieve a "wild-west-style" load balance here

    uint8_t* dst = AllocOnNode(task->GetSize(), dst_node);

    if (dst == nullptr) {
        return;
    }

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

    const std::vector<int> executing_nodes = copy_policy_function_(dst_node, src_node, task->GetSize());

    // use our load balancing method and determine node for this task

    last_node_index = ++last_node_index % executing_nodes.size();
    const int node = executing_nodes[last_node_index];

    // submit the copy and attach it to the task entry

    auto* handler = new CacheData::dml_handler();
    *handler = ExecuteCopy(task->GetSource(), dst, task->GetSize(), node);
    task->SetTaskHandlerAndCache(dst, handler);
}

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);

    if (CheckFlag(flags_, FLAG_HANDLE_PF)) {
        return dml::submit<dml::hardware>(
                dml::mem_copy.block_on_fault(), srcv, dstv,
                dml::execution_interface<dml::hardware,std::allocator<uint8_t>>(), node
        );
    }
    else {
        return dml::submit<dml::hardware>(
                dml::mem_copy, srcv, dstv,
                dml::execution_interface<dml::hardware,std::allocator<uint8_t>>(), node
        );
    }
}

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.GetRefCount() <= 1) {
                // 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
    // node == -1 means that cache on all nodes should be flushed

    if (node == -1) {
        for (auto& nc : cache_state_) {
            std::unique_lock<std::shared_mutex> lock(nc.second->cache_mutex_);
            FlushNode(nc.second->node_cache_state_);
        }
    }
    else {
        std::unique_lock<std::shared_mutex> lock(cache_state_[node]->cache_mutex_);
        FlushNode(cache_state_[node]->node_cache_state_);
    }
}

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

    LockedNodeCacheState* local_cache_state = cache_state_[dst_node];

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

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

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

    const auto search = local_cache_state->node_cache_state_.find(src);

    // if the data is in our structure we continue

    if (search != local_cache_state->node_cache_state_.end()) {

        // now check whether the sizes match

        if (search->second.GetSize() >= 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

            local_cache_state->node_cache_state_.erase(search);
        }
    }

    return nullptr;
}

void dsacache::Cache::Invalidate(uint8_t* data) {
    // as the cache is modified we must obtain a unique writers lock
    // loop through all per-node-caches available

    for (auto node : cache_state_) {
        std::unique_lock<std::shared_mutex> lock(node.second->cache_mutex_);

        // search for an entry for the given data pointer

        auto search = node.second->node_cache_state_.find(data);

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

            node.second->node_cache_state_.erase(search);
        }
    }
}

inline dsacache::Cache::~Cache() {
    for (auto node : cache_state_) {
        node.second->~LockedNodeCacheState();
        numa_free(reinterpret_cast<void*>(node.second), sizeof(LockedNodeCacheState));
    }
}

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

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);

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

    incomplete_cache_ = other.incomplete_cache_;
    handler_ = other.handler_;
}

inline dsacache::CacheData::~CacheData() {
    // 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) {
        // on deletion we must ensure that all offloaded
        // operations have completed successfully

        WaitOnCompletion();

        // only then can we deallocate the memory

        Deallocate();

        delete active_;
        delete cache_;
        delete handler_;
        delete incomplete_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
    // and check for incomplete cache if no deallocation
    // takes place for the retrieved local cache

    uint8_t* cache_local = cache_->exchange(nullptr);
    if (cache_local != nullptr && delete_) numa_free(cache_local, size_);
    else if (*incomplete_cache_ != nullptr) numa_free(*incomplete_cache_, size_);
    else;
}

inline void dsacache::CacheData::WaitOnCompletion() {
    // first check if waiting is even neccessary as a valid
    // cache pointer signals that no waiting is to be performed

    if (cache_->load() != nullptr) {
        return;
    }

    // then check if the handlers are available

    handler_->wait(nullptr);

    // exchange the global handlers pointer with nullptr to have a local
    // copy - this signals that this thread is the sole owner and therefore
    // responsible for waiting for them. we can not set to nullptr here but
    // set to maximum of 64-bit in order to prevent deadlocks from the above
    // waiting construct

    dml_handler* local_handler = handler_->exchange(reinterpret_cast<dml_handler*>(maxptr));

    // ensure that no other thread snatched the handlers before us
    // and in case one did, wait again and then return

    if (local_handler == nullptr || local_handler == reinterpret_cast<dml_handler*>(maxptr)) {
        cache_->wait(nullptr);
        return;
    }

    // at this point we are responsible for waiting for the handlers
    // and handling any error that comes through them gracefully

    if (CheckFlag(flags_, FLAG_WAIT_WEAK) && !local_handler->is_finished()) {
        handler_->store(local_handler);
        return;
    }

    // perform the wait

    auto result = local_handler->get();

    // at this point handlers has been waited for
    // and therefore may be decomissioned

    delete local_handler;

    // if the copy tasks failed we abort the whole task
    // otherwise the cache will be set to valid now

    if (result.status != dml::status_code::ok) {
        cache_->store(src_);
        numa_free(*incomplete_cache_, size_);
        delete_ = false;
        *incomplete_cache_ = nullptr;
    }
    else {
        cache_->store(*incomplete_cache_);
    }

    // notify all waiting threads so they wake up quickly

    cache_->notify_all();
    handler_->notify_all();
}

void dsacache::CacheData::SetTaskHandlerAndCache(uint8_t* cache, dml_handler* handler) {
    *incomplete_cache_ = cache;
    handler_->store(handler);
    handler_->notify_one();
}

void dsacache::CacheData::Init() {
    cache_->store(nullptr);
    delete_ = true;
}