source: libcfa/src/concurrency/actor.hfa @ b2f3880

#pragma once

#include <locks.hfa>
#include <limits.hfa>
#include <kernel.hfa>
#include <iofwd.hfa>

#ifdef __CFA_DEBUG__
#define CFA_DEBUG( stmt ) stmt
#else
#define CFA_DEBUG( stmt )
#endif // CFA_DEBUG

// Define the default number of processors created in the executor. Must be greater than 0.
#define __DEFAULT_EXECUTOR_PROCESSORS__ 2

// Define the default number of threads created in the executor. Must be greater than 0.
#define __DEFAULT_EXECUTOR_WORKERS__ 2

// Define the default number of executor request-queues (mailboxes) written to by actors and serviced by the
// actor-executor threads. Must be greater than 0.
#define __DEFAULT_EXECUTOR_RQUEUES__ 4

// Define if executor is created in a separate cluster
#define __DEFAULT_EXECUTOR_SEPCLUS__ false

#define __DEFAULT_EXECUTOR_BUFSIZE__ 10

#define __STEAL 0 // workstealing toggle. Disjoint from toggles above

// workstealing heuristic selection (only set one to be 1)
// #define RAND 0
#define SEARCH 1

// show stats
// #define STATS

// forward decls
struct actor;
struct message;
struct executor;

enum Allocation { Nodelete, Delete, Destroy, Finished }; // allocation status

typedef Allocation (*__receive_fn)(actor &, message &);
struct request {
    actor * receiver;
    message * msg;
    __receive_fn fn;
    bool stop;
};

static inline void ?{}( request & this ) { this.stop = true; } // default ctor makes a sentinel
static inline void ?{}( request & this, actor * receiver, message * msg, __receive_fn fn ) {
    this.receiver = receiver;
    this.msg = msg;
    this.fn = fn;
    this.stop = false;
}
static inline void ?{}( request & this, request & copy ) {
    this.receiver = copy.receiver;
    this.msg = copy.msg;
    this.fn = copy.fn;
    this.stop = copy.stop;
}

// Vector-like data structure that supports O(1) queue operations with no bound on size
// assumes gulping behaviour (once a remove occurs, removes happen until empty beforw next insert)
struct copy_queue {
    request * buffer;
    size_t count, buffer_size, index, utilized, last_size;
};
static inline void ?{}( copy_queue & this ) {}
static inline void ?{}( copy_queue & this, size_t buf_size ) with(this) { 
    buffer_size = buf_size;
    buffer = aalloc( buffer_size );
    count = 0;
    utilized = 0;
    index = 0;
    last_size = 0;
}
static inline void ^?{}( copy_queue & this ) with(this) { adelete(buffer); }

static inline void insert( copy_queue & this, request & elem ) with(this) {
    if ( count >= buffer_size ) { // increase arr size
        last_size = buffer_size;
        buffer_size = 2 * buffer_size;
        buffer = realloc( buffer, sizeof( request ) * buffer_size );
        /* paranoid */ verify( buffer );
    }
    memcpy( &buffer[count], &elem, sizeof(request) );
    count++;
}

// once you start removing you need to remove all elements
// it is not supported to call insert() before the array is fully empty
static inline request & remove( copy_queue & this ) with(this) {
    if ( count > 0 ) {
        count--;
        size_t old_idx = index;
        index = count == 0 ? 0 : index + 1;
        return buffer[old_idx];
    }
    request * ret = 0p;
    return *0p;
}

// try to reclaim some memory if less than half of buffer is utilized
static inline void reclaim( copy_queue & this ) with(this) {
    if ( utilized >= last_size || buffer_size <= 4 ) { utilized = 0; return; }
    utilized = 0;
    buffer_size--;
    buffer = realloc( buffer, sizeof( request ) * buffer_size ); // try to reclaim some memory
}

static inline bool isEmpty( copy_queue & this ) with(this) { return count == 0; }

struct work_queue {
    __spinlock_t mutex_lock;
    copy_queue * owned_queue;       // copy queue allocated and cleaned up by this work_queue
    copy_queue * c_queue;           // current queue
    volatile bool being_processed;  // flag to prevent concurrent processing
    #ifdef STATS
    unsigned int id;
    size_t missed;                  // transfers skipped due to being_processed flag being up
    #endif
}; // work_queue
static inline void ?{}( work_queue & this, size_t buf_size, unsigned int i ) with(this) { 
    owned_queue = alloc();      // allocated separately to avoid false sharing
    (*owned_queue){ buf_size };
    c_queue = owned_queue;
    being_processed = false;
    #ifdef STATS
    id = i;
    missed = 0;
    #endif
}

// clean up copy_queue owned by this work_queue
static inline void ^?{}( work_queue & this ) with(this) { delete( owned_queue ); }

static inline void insert( work_queue & this, request & elem ) with(this) {
    lock( mutex_lock __cfaabi_dbg_ctx2 );
    insert( *c_queue, elem );
    unlock( mutex_lock );
} // insert

static inline void transfer( work_queue & this, copy_queue ** transfer_to ) with(this) {
    lock( mutex_lock __cfaabi_dbg_ctx2 );
    #ifdef __STEAL

    // check if queue is being processed elsewhere
    if ( unlikely( being_processed ) ) {
        #ifdef STATS
        missed++;
        #endif
        unlock( mutex_lock );
        return;
    }

    being_processed = c_queue->count != 0;
    #endif // __STEAL

    c_queue->utilized = c_queue->count;

    // swap copy queue ptrs
    copy_queue * temp = *transfer_to;
    *transfer_to = c_queue;
    c_queue = temp;
    unlock( mutex_lock );
} // transfer

// needed since some info needs to persist past worker lifetimes
struct worker_info {
    volatile unsigned long long stamp;
    #ifdef STATS
    size_t stolen_from, try_steal, stolen, failed_swaps, msgs_stolen;
    unsigned long long processed;
    size_t gulps;
    #endif
};
static inline void ?{}( worker_info & this ) {
    #ifdef STATS
    this.stolen_from = 0;
    this.try_steal = 0;                             // attempts to steal
    this.stolen = 0;                                // successful steals
    this.processed = 0;                             // requests processed
    this.gulps = 0;                                 // number of gulps
    this.failed_swaps = 0;                          // steal swap failures
    this.msgs_stolen = 0;                           // number of messages stolen
    #endif
    this.stamp = rdtscl();
}

// #ifdef STATS
// unsigned int * stolen_arr;
// unsigned int * replaced_queue;
// #endif
thread worker {
    work_queue ** request_queues;
    copy_queue * current_queue;
    executor * executor_;
    unsigned int start, range;
    int id;
};

#ifdef STATS
// aggregate counters for statistics
size_t __total_tries = 0, __total_stolen = 0, __total_workers, __all_gulps = 0,
    __total_failed_swaps = 0, __all_processed = 0, __num_actors_stats = 0, __all_msgs_stolen = 0;
#endif
static inline void ?{}( worker & this, cluster & clu, work_queue ** request_queues, copy_queue * current_queue, executor * executor_,
    unsigned int start, unsigned int range, int id ) {
    ((thread &)this){ clu };
    this.request_queues = request_queues;           // array of all queues
    this.current_queue = current_queue;             // currently gulped queue (start with empty queue to use in swap later)
    this.executor_ = executor_;                     // pointer to current executor
    this.start = start;                             // start of worker's subrange of request_queues
    this.range = range;                             // size of worker's subrange of request_queues
    this.id = id;                                   // worker's id and index in array of workers
}

static bool no_steal = false;
struct executor {
    cluster * cluster;                                                      // if workers execute on separate cluster
        processor ** processors;                                            // array of virtual processors adding parallelism for workers
        work_queue * request_queues;                                // master array of work request queues
    copy_queue * local_queues;                      // array of all worker local queues to avoid deletion race
        work_queue ** worker_req_queues;                // secondary array of work queues to allow for swapping
    worker ** workers;                                                          // array of workers executing work requests
    worker_info * w_infos;                          // array of info about each worker 
        unsigned int nprocessors, nworkers, nrqueues;   // number of processors/threads/request queues
        bool seperate_clus;                                                             // use same or separate cluster for executor
}; // executor

// #ifdef STATS
// __spinlock_t out_lock;
// #endif
static inline void ^?{}( worker & mutex this ) with(this) { 
    #ifdef STATS
    __atomic_add_fetch(&__all_gulps, executor_->w_infos[id].gulps,__ATOMIC_SEQ_CST);
    __atomic_add_fetch(&__all_processed, executor_->w_infos[id].processed,__ATOMIC_SEQ_CST);
    __atomic_add_fetch(&__all_msgs_stolen, executor_->w_infos[id].msgs_stolen,__ATOMIC_SEQ_CST);
    __atomic_add_fetch(&__total_tries, executor_->w_infos[id].try_steal, __ATOMIC_SEQ_CST);
    __atomic_add_fetch(&__total_stolen, executor_->w_infos[id].stolen, __ATOMIC_SEQ_CST);
    __atomic_add_fetch(&__total_failed_swaps, executor_->w_infos[id].failed_swaps, __ATOMIC_SEQ_CST);

    // per worker steal stats (uncomment alongside the lock above this routine to print)
    // lock( out_lock __cfaabi_dbg_ctx2 );
    // printf("Worker id: %d, processed: %llu messages, attempted %lu, stole: %lu, stolen from: %lu\n", id, processed, try_steal, stolen, __atomic_add_fetch(&executor_->w_infos[id].stolen_from, 0, __ATOMIC_SEQ_CST) );
    // int count = 0;
    // int count2 = 0;
    // for ( i; range ) {
    //     if ( replaced_queue[start + i] > 0 ){
    //         count++;
    //         // printf("%d: %u, ",i, replaced_queue[i]);
    //     }
    //     if (__atomic_add_fetch(&stolen_arr[start + i],0,__ATOMIC_SEQ_CST) > 0)
    //         count2++;
    // }
    // printf("swapped with: %d of %u indices\n", count, executor_->nrqueues / executor_->nworkers );
    // printf("%d of %u indices were stolen\n", count2, executor_->nrqueues / executor_->nworkers );
    // unlock( out_lock );
    #endif
}

static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers, unsigned int nrqueues, bool seperate_clus, size_t buf_size ) with(this) {
    if ( nrqueues < nworkers ) abort( "nrqueues needs to be >= nworkers\n" );
    this.nprocessors = nprocessors;
    this.nworkers = nworkers;
    this.nrqueues = nrqueues;
    this.seperate_clus = seperate_clus;

    if ( nworkers == nrqueues )
        no_steal = true;
    
    #ifdef STATS
    // stolen_arr = aalloc( nrqueues );
    // replaced_queue = aalloc( nrqueues );
    __total_workers = nworkers;
    #endif

    if ( seperate_clus ) {
        cluster = alloc();
        (*cluster){};
    } else cluster = active_cluster();

    request_queues = aalloc( nrqueues );
    worker_req_queues = aalloc( nrqueues );
    for ( i; nrqueues ) {
        request_queues[i]{ buf_size, i };
        worker_req_queues[i] = &request_queues[i];
    }
    
    processors = aalloc( nprocessors );
    for ( i; nprocessors ) 
        (*(processors[i] = alloc())){ *cluster };

    local_queues = aalloc( nworkers );
    workers = aalloc( nworkers );
    w_infos = aalloc( nworkers );
    unsigned int reqPerWorker = nrqueues / nworkers, extras = nrqueues % nworkers;

    for ( i; nworkers ) {
        w_infos[i]{};
        local_queues[i]{ buf_size };
    }

    for ( unsigned int i = 0, start = 0, range; i < nworkers; i += 1, start += range ) {
        range = reqPerWorker + ( i < extras ? 1 : 0 );
        (*(workers[i] = alloc())){ *cluster, worker_req_queues, &local_queues[i], &this, start, range, i };
    } // for
}
static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers, unsigned int nrqueues, bool seperate_clus ) { this{ nprocessors, nworkers, nrqueues, seperate_clus, __DEFAULT_EXECUTOR_BUFSIZE__ }; }
static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers, unsigned int nrqueues ) { this{ nprocessors, nworkers, nrqueues, __DEFAULT_EXECUTOR_SEPCLUS__ }; }
static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers ) { this{ nprocessors, nworkers, __DEFAULT_EXECUTOR_RQUEUES__ }; }
static inline void ?{}( executor & this, unsigned int nprocessors ) { this{ nprocessors, __DEFAULT_EXECUTOR_WORKERS__ }; }
static inline void ?{}( executor & this ) { this{ __DEFAULT_EXECUTOR_PROCESSORS__ }; }

static inline void ^?{}( executor & this ) with(this) {
    #ifdef __STEAL
    request sentinels[nrqueues];
    for ( unsigned int i = 0; i < nrqueues; i++ ) {
        insert( request_queues[i], sentinels[i] );              // force eventually termination
    } // for
    #else
    request sentinels[nworkers];
    unsigned int reqPerWorker = nrqueues / nworkers, extras = nrqueues % nworkers;
    for ( unsigned int i = 0, step = 0, range; i < nworkers; i += 1, step += range ) {
        range = reqPerWorker + ( i < extras ? 1 : 0 );
        insert( request_queues[step], sentinels[i] );           // force eventually termination
    } // for
    #endif

    for ( i; nworkers )
        delete( workers[i] );

    for ( i; nprocessors ) {
        delete( processors[i] );
    } // for

    #ifdef STATS
    size_t misses = 0;
    for ( i; nrqueues ) {
        misses += worker_req_queues[i]->missed;
    }
    // adelete( stolen_arr );
    // adelete( replaced_queue );
    #endif

    adelete( workers );
    adelete( w_infos );
    adelete( local_queues );
    adelete( request_queues );
    adelete( worker_req_queues );
    adelete( processors );
    if ( seperate_clus ) delete( cluster );

    #ifdef STATS
    printf("    Actor System Stats:\n");
    printf("\tActors Created:\t\t\t\t%lu\n\tMessages Sent:\t\t\t\t%lu\n", __num_actors_stats, __all_processed);
    size_t avg_gulps = __all_gulps == 0 ? 0 : __all_processed / __all_gulps;
    printf("\tGulps:\t\t\t\t\t%lu\n\tAverage Gulp Size:\t\t\t%lu\n\tMissed gulps:\t\t\t\t%lu\n", __all_gulps, avg_gulps, misses);
    printf("\tSteal attempts:\t\t\t\t%lu\n\tSteals:\t\t\t\t\t%lu\n\tSteal failures (no candidates):\t\t%lu\n\tSteal failures (failed swaps):\t\t%lu\n", 
        __total_tries, __total_stolen, __total_tries - __total_stolen - __total_failed_swaps, __total_failed_swaps);
    size_t avg_steal = __total_stolen == 0 ? 0 : __all_msgs_stolen / __total_stolen;
    printf("\tMessages stolen:\t\t\t%lu\n\tAverage steal size:\t\t\t%lu\n", __all_msgs_stolen, avg_steal);
    #endif
        
}

// this is a static field of executor but have to forward decl for get_next_ticket
static size_t __next_ticket = 0; 

static inline size_t __get_next_ticket( executor & this ) with(this) {
    #ifdef __CFA_DEBUG__
    size_t temp = __atomic_fetch_add( &__next_ticket, 1, __ATOMIC_SEQ_CST) % nrqueues;

    // reserve MAX for dead actors
    if ( unlikely( temp == MAX ) ) temp = __atomic_fetch_add( &__next_ticket, 1, __ATOMIC_SEQ_CST) % nrqueues;
    return temp;
    #else
    return __atomic_fetch_add( &__next_ticket, 1, __ATOMIC_RELAXED) % nrqueues;
    #endif
} // tickets

// TODO: update globals in this file to be static fields once the static fields project is done
static executor * __actor_executor_ = 0p;
static bool __actor_executor_passed = false;            // was an executor passed to start_actor_system
static size_t __num_actors_ = 0;                                // number of actor objects in system
static struct thread$ * __actor_executor_thd = 0p;              // used to wake executor after actors finish
struct actor {
    size_t ticket;                              // executor-queue handle
    Allocation allocation_;                                         // allocation action
};

static inline void ?{}( actor & this ) with(this) {
    // Once an actor is allocated it must be sent a message or the actor system cannot stop. Hence, its receive
    // member must be called to end it
    verifyf( __actor_executor_, "Creating actor before calling start_actor_system() can cause undefined behaviour.\n" ); 
    allocation_ = Nodelete;
    ticket = __get_next_ticket( *__actor_executor_ );
    __atomic_fetch_add( &__num_actors_, 1, __ATOMIC_RELAXED );
    #ifdef STATS
    __atomic_fetch_add( &__num_actors_stats, 1, __ATOMIC_SEQ_CST );
    #endif
}
static inline void ^?{}( actor & this ) {}

static inline void check_actor( actor & this ) {
    if ( this.allocation_ != Nodelete ) {
        switch( this.allocation_ ) {
            case Delete: delete( &this ); break;
            case Destroy:
                CFA_DEBUG( this.ticket = MAX; );        // mark as terminated
                ^?{}(this);
                break;
            case Finished:
                CFA_DEBUG( this.ticket = MAX; );        // mark as terminated
                break;
            default: ;                                                          // stop warning
        }

        if ( unlikely( __atomic_add_fetch( &__num_actors_, -1, __ATOMIC_RELAXED ) == 0 ) ) { // all actors have terminated
            unpark( __actor_executor_thd );
        }
    }
}

struct message {
    Allocation allocation_;                     // allocation action
};

static inline void ?{}( message & this ) { this.allocation_ = Nodelete; }
static inline void ?{}( message & this, Allocation allocation ) {
    memcpy( &this.allocation_, &allocation, sizeof(allocation) ); // optimization to elide ctor
    verifyf( this.allocation_ != Finished, "The Finished Allocation status is not supported for message types.\n");
}
static inline void ^?{}( message & this ) {
    CFA_DEBUG( if ( this.allocation_ == Nodelete ) printf("A message at location %p was allocated but never sent.\n", &this); )
}

static inline void check_message( message & this ) {
    #ifdef __CFA_DEBUG__
    Allocation temp = this.allocation_;
    this.allocation_ = Finished;
    switch ( temp )
    #else
    switch ( this.allocation_ )
    #endif
    {                                           // analyze message status
        case Nodelete: break;
        case Delete: delete( &this ); break;
        case Destroy: ^?{}(this); break;
        case Finished: break;
    } // switch
}
static inline void set_allocation( message & this, Allocation state ) { this.allocation_ = state; }

static inline void deliver_request( request & this ) {
    this.receiver->allocation_ = this.fn( *this.receiver, *this.msg );
    check_message( *this.msg );
    check_actor( *this.receiver );
}

// tries to atomically swap two queues and returns 0p if the swap failed
// returns ptr to newly owned queue if swap succeeds
static inline work_queue * try_swap_queues( worker & this, unsigned int victim_idx, unsigned int my_idx ) with(this) {
    work_queue * my_queue = request_queues[my_idx];
    work_queue * other_queue = request_queues[victim_idx];

    // if either queue is 0p then they are in the process of being stolen
    if ( other_queue == 0p ) return 0p;

    // try to set our queue ptr to be 0p. If it fails someone moved our queue so return false
    if ( !__atomic_compare_exchange_n( &request_queues[my_idx], &my_queue, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) )
        return 0p;

    // try to set other queue ptr to be our queue ptr. If it fails someone moved the other queue so fix up then return false
    if ( !__atomic_compare_exchange_n( &request_queues[victim_idx], &other_queue, my_queue, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) ) {
        /* paranoid */ verify( request_queues[my_idx] == 0p );
        request_queues[my_idx] = my_queue; // reset my queue ptr back to appropriate val
        return 0p;
    }

    // we have successfully swapped and since our queue is 0p no one will touch it so write back new queue ptr non atomically
    request_queues[my_idx] = other_queue; // last write does not need to be atomic
    return other_queue;
}

// once a worker to steal from has been chosen, choose queue to steal from
static inline void choose_queue( worker & this, unsigned int victim_id, unsigned int swap_idx ) with(this) {
    // have to calculate victim start and range since victim may be deleted before us in shutdown
    const unsigned int queues_per_worker = executor_->nrqueues / executor_->nworkers;
    const unsigned int extras = executor_->nrqueues % executor_->nworkers;
    unsigned int vic_start, vic_range;
    if ( extras > victim_id  ) {
        vic_range = queues_per_worker + 1;
        vic_start = vic_range * victim_id;
    } else {
        vic_start = extras + victim_id * queues_per_worker;
        vic_range = queues_per_worker;
    }
    unsigned int start_idx = prng( vic_range );

    unsigned int tries = 0;
    work_queue * curr_steal_queue;

    for ( unsigned int i = start_idx; tries < vic_range; i = (i + 1) % vic_range ) {
        tries++;
        curr_steal_queue = request_queues[ i + vic_start ];
        // avoid empty queues and queues that are being operated on
        if ( curr_steal_queue == 0p || curr_steal_queue->being_processed || isEmpty( *curr_steal_queue->c_queue ) )
            continue;

        #ifdef STATS
        curr_steal_queue = try_swap_queues( this, i + vic_start, swap_idx );
        if ( curr_steal_queue ) {
            executor_->w_infos[id].msgs_stolen += curr_steal_queue->c_queue->count;
            executor_->w_infos[id].stolen++;
            // __atomic_add_fetch(&executor_->w_infos[victim_id].stolen_from, 1, __ATOMIC_RELAXED);
            // replaced_queue[swap_idx]++;
            // __atomic_add_fetch(&stolen_arr[ i + vic_start ], 1, __ATOMIC_RELAXED);
        } else {
            executor_->w_infos[id].failed_swaps++;
        }
        #else
        curr_steal_queue = try_swap_queues( this, i + vic_start, swap_idx );
        #endif // STATS

        return;
    }

    return;
}

// choose a worker to steal from
static inline void steal_work( worker & this, unsigned int swap_idx ) with(this) {
    #if RAND
    unsigned int victim = prng( executor_->nworkers );
    if ( victim == id ) victim = ( victim + 1 ) % executor_->nworkers;
    choose_queue( this, victim, swap_idx );
    #elif SEARCH
    unsigned long long min = MAX; // smaller timestamp means longer since service
    int min_id = 0; // use ints not uints to avoid integer underflow without hacky math
    int n_workers = executor_->nworkers;
    unsigned long long curr_stamp;
    int scount = 1;
    for ( int i = (id + 1) % n_workers; scount < n_workers; i = (i + 1) % n_workers, scount++ ) {
        curr_stamp = executor_->w_infos[i].stamp;
        if ( curr_stamp < min ) {
            min = curr_stamp;
            min_id = i;
        }
    } 
    choose_queue( this, min_id, swap_idx );
    #endif
}

void main( worker & this ) with(this) {
    // #ifdef STATS
    // for ( i; executor_->nrqueues ) {
    //     replaced_queue[i] = 0;
    //     __atomic_store_n( &stolen_arr[i], 0, __ATOMIC_SEQ_CST );
    // }
    // #endif

    // threshold of empty queues we see before we go stealing
    const unsigned int steal_threshold = 2 * range;

    // Store variable data here instead of worker struct to avoid any potential false sharing
    unsigned int empty_count = 0;
    request & req;
    work_queue * curr_work_queue;

    Exit:
    for ( unsigned int i = 0;; i = (i + 1) % range ) { // cycle through set of request buffers
        curr_work_queue = request_queues[i + start];
        
        // check if queue is empty before trying to gulp it
        if ( isEmpty( *curr_work_queue->c_queue ) ) {
            #ifdef __STEAL
            empty_count++;
            if ( empty_count < steal_threshold ) continue;
            #else
            continue;
            #endif
        }
        transfer( *curr_work_queue, &current_queue );
        #ifdef STATS
        executor_->w_infos[id].gulps++;
        #endif // STATS
        #ifdef __STEAL
        if ( isEmpty( *current_queue ) ) {
            if ( unlikely( no_steal ) ) continue;
            empty_count++;
            if ( empty_count < steal_threshold ) continue;
            empty_count = 0;

            __atomic_store_n( &executor_->w_infos[id].stamp, rdtscl(), __ATOMIC_RELAXED );
            
            #ifdef STATS
            executor_->w_infos[id].try_steal++;
            #endif // STATS
            
            steal_work( this, start + prng( range ) );
            continue;
        }
        #endif // __STEAL
        while ( ! isEmpty( *current_queue ) ) {
            #ifdef STATS
            executor_->w_infos[id].processed++;
            #endif
            &req = &remove( *current_queue );
            if ( !&req ) continue;
            if ( req.stop ) break Exit;
            deliver_request( req );
        }
        #ifdef __STEAL
        curr_work_queue->being_processed = false; // set done processing
        empty_count = 0; // we found work so reset empty counter
        #endif
        
        // potentially reclaim some of the current queue's vector space if it is unused
        reclaim( *current_queue );
    } // for
}

static inline void send( executor & this, request & req, unsigned long int ticket ) with(this) {
    insert( request_queues[ticket], req);
}

static inline void send( actor & this, request & req ) {
    verifyf( this.ticket != (unsigned long int)MAX, "Attempted to send message to deleted/dead actor\n" );
    send( *__actor_executor_, req, this.ticket );
}

static inline void __reset_stats() {
    #ifdef STATS
    __total_tries = 0;
    __total_stolen = 0;
    __all_gulps = 0;
    __total_failed_swaps = 0;
    __all_processed = 0;
    __num_actors_stats = 0;
    __all_msgs_stolen = 0;
    #endif
}

static inline void start_actor_system( size_t num_thds ) {
    __reset_stats();
    __actor_executor_thd = active_thread();
    __actor_executor_ = alloc();
    (*__actor_executor_){ 0, num_thds, num_thds == 1 ? 1 : num_thds * 16 };
}

// TODO: potentially revisit getting number of processors
//  ( currently the value stored in active_cluster()->procs.total is often stale 
//  and doesn't reflect how many procs are allocated )
// static inline void start_actor_system() { start_actor_system( active_cluster()->procs.total ); }
static inline void start_actor_system() { start_actor_system( 1 ); }

static inline void start_actor_system( executor & this ) {
    __reset_stats();
    __actor_executor_thd = active_thread();
    __actor_executor_ = &this;
    __actor_executor_passed = true;
}

static inline void stop_actor_system() {
    park( ); // will receive signal when actor system is finished

    if ( !__actor_executor_passed ) delete( __actor_executor_ );
    __actor_executor_ = 0p;
    __actor_executor_thd = 0p;
    __next_ticket = 0;
    __actor_executor_passed = false;
}

// Default messages to send to any actor to change status
// assigned at creation to __base_msg_finished to avoid unused message warning
message __base_msg_finished @= { .allocation_ : Finished };
struct __DeleteMsg { inline message; } DeleteMsg = __base_msg_finished;
struct __DestroyMsg { inline message; } DestroyMsg = __base_msg_finished;
struct __FinishedMsg { inline message; } FinishedMsg = __base_msg_finished;

Allocation receive( actor & this, __DeleteMsg & msg ) { return Delete; }
Allocation receive( actor & this, __DestroyMsg & msg ) { return Destroy; }
Allocation receive( actor & this, __FinishedMsg & msg ) { return Finished; }