arena.h

Go to the documentation of this file.

/*
    Copyright (c) 2005-2020 Intel Corporation

    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License.
*/

#ifndef _TBB_arena_H
#define _TBB_arena_H

#include "tbb/tbb_stddef.h"
#include "tbb/atomic.h"

#include "tbb/tbb_machine.h"

#include "scheduler_common.h"
#include "intrusive_list.h"
#if __TBB_PREVIEW_CRITICAL_TASKS && __TBB_CPF_BUILD
#include "task_stream_extended.h"
#else
#include "task_stream.h"
#endif
#include "../rml/include/rml_tbb.h"
#include "mailbox.h"
#include "observer_proxy.h"
#include "market.h"
#include "governor.h"
#include "concurrent_monitor.h"

#if __TBB_PREVIEW_RESUMABLE_TASKS
#include "tbb/spin_mutex.h"
#endif

namespace tbb {

class task_group_context;
class allocate_root_with_context_proxy;

namespace internal {

#if __TBB_NUMA_SUPPORT
class numa_binding_observer;
#endif /*__TBB_NUMA_SUPPORT*/

#if __TBB_PREVIEW_RESUMABLE_TASKS
class arena_co_cache {
    generic_scheduler** my_co_scheduler_cache;
    unsigned my_head;
    unsigned my_max_index;
    tbb::spin_mutex my_co_cache_mutex;

    unsigned next_index() {
        return ( my_head == my_max_index ) ? 0 : my_head + 1;
    }

    unsigned prev_index() {
        return ( my_head == 0 ) ? my_max_index : my_head - 1;
    }

    bool internal_empty() {
        return my_co_scheduler_cache[prev_index()] == NULL;
    }

    void internal_scheduler_cleanup(generic_scheduler* to_cleanup) {
        to_cleanup->my_arena_slot = NULL;
        // Needed by cleanup_worker function, as well as arena slot clearing
        governor::assume_scheduler(to_cleanup);
        generic_scheduler::cleanup_worker(to_cleanup, true);
    }

public:
    void init(unsigned cache_capacity) {
        size_t alloc_size = cache_capacity * sizeof(generic_scheduler*);
        my_co_scheduler_cache = (generic_scheduler**)NFS_Allocate(1, alloc_size, NULL);
        memset( my_co_scheduler_cache, 0, alloc_size );
        my_head = 0;
        my_max_index = cache_capacity - 1;
    }

    void cleanup() {
        generic_scheduler* current = governor::local_scheduler_if_initialized();
        while (generic_scheduler* to_cleanup = pop()) {
            internal_scheduler_cleanup(to_cleanup);
        }
        governor::assume_scheduler(current);
        NFS_Free(my_co_scheduler_cache);
    }

    void push(generic_scheduler* s) {
        generic_scheduler* to_cleanup = NULL;
        {
            tbb::spin_mutex::scoped_lock lock(my_co_cache_mutex);
            // Check if we are replacing some existing buffer entrance
            if (my_co_scheduler_cache[my_head] != NULL) {
                to_cleanup = my_co_scheduler_cache[my_head];
            }
            // Store the cached value
            my_co_scheduler_cache[my_head] = s;
            // Move head index to the next slot
            my_head = next_index();
        }
        // Cleanup replaced buffer if any
        if (to_cleanup) {
            generic_scheduler* current = governor::local_scheduler_if_initialized();
            internal_scheduler_cleanup(to_cleanup);
            governor::assume_scheduler(current);
        }
    }

    generic_scheduler* pop() {
        tbb::spin_mutex::scoped_lock lock(my_co_cache_mutex);
        // No cached coroutine
        if (internal_empty()) return NULL;
        // Move head index to the currently available value
        my_head = prev_index();
        // Retrieve the value from the buffer
        generic_scheduler* to_return = my_co_scheduler_cache[my_head];
        // Clear the previous entrance value
        my_co_scheduler_cache[my_head] = NULL;
        return to_return;
    }
};
#endif // __TBB_PREVIEW_RESUMABLE_TASKS


struct arena_base : padded<intrusive_list_node> {
    unsigned my_num_workers_allotted;   // heavy use in stealing loop


    atomic<unsigned> my_references;     // heavy use in stealing loop

#if __TBB_TASK_PRIORITY
    volatile intptr_t my_top_priority;  // heavy use in stealing loop
#endif /* !__TBB_TASK_PRIORITY */

    atomic<unsigned> my_limit;          // heavy use in stealing loop


#if __TBB_PREVIEW_CRITICAL_TASKS && __TBB_CPF_BUILD
    task_stream<num_priority_levels, front_accessor> my_task_stream; // heavy use in stealing loop
#else
    task_stream<num_priority_levels> my_task_stream; // heavy use in stealing loop
#endif

#if __TBB_PREVIEW_CRITICAL_TASKS

    // used on the hot path of the task dispatch loop
    task_stream<1, back_nonnull_accessor> my_critical_task_stream;
#endif

    unsigned my_max_num_workers;

    int my_num_workers_requested;


    tbb::atomic<uintptr_t> my_pool_state;

#if __TBB_ARENA_OBSERVER
    observer_list my_observers;
#endif

#if __TBB_NUMA_SUPPORT
    numa_binding_observer* my_numa_binding_observer;
#endif /*__TBB_NUMA_SUPPORT*/

#if __TBB_TASK_PRIORITY
    intptr_t my_bottom_priority;


    uintptr_t my_reload_epoch;

    task* my_orphaned_tasks;

    tbb::atomic<uintptr_t> my_abandonment_epoch;


    tbb::atomic<intptr_t> my_skipped_fifo_priority;
#endif /* !__TBB_TASK_PRIORITY */

    // Below are rarely modified members

    market* my_market;

    uintptr_t my_aba_epoch;

#if !__TBB_FP_CONTEXT
    cpu_ctl_env my_cpu_ctl_env;
#endif

#if __TBB_TASK_GROUP_CONTEXT

    task_group_context* my_default_ctx;
#endif /* __TBB_TASK_GROUP_CONTEXT */

    unsigned my_num_slots;

    unsigned my_num_reserved_slots;

#if __TBB_ENQUEUE_ENFORCED_CONCURRENCY
    // arena needs an extra worker despite the arena limit
    bool my_local_concurrency_mode;
    // arena needs an extra worker despite a global limit
    bool my_global_concurrency_mode;
#endif /* __TBB_ENQUEUE_ENFORCED_CONCURRENCY */

    concurrent_monitor my_exit_monitors;

#if __TBB_PREVIEW_RESUMABLE_TASKS
    arena_co_cache my_co_cache;
#endif

#if TBB_USE_ASSERT
    uintptr_t my_guard;
#endif /* TBB_USE_ASSERT */
}; // struct arena_base

class arena: public padded<arena_base>
{
    void restore_priority_if_need();
public:
    typedef padded<arena_base> base_type;

    enum new_work_type {
        work_spawned,
        wakeup,
        work_enqueued
    };

    arena ( market&, unsigned max_num_workers, unsigned num_reserved_slots );

    static arena& allocate_arena( market&, unsigned num_slots, unsigned num_reserved_slots );

    static int unsigned num_arena_slots ( unsigned num_slots ) {
        return max(2u, num_slots);
    }

    static int allocation_size ( unsigned num_slots ) {
        return sizeof(base_type) + num_slots * (sizeof(mail_outbox) + sizeof(arena_slot));
    }

    mail_outbox& mailbox( affinity_id id ) {
        __TBB_ASSERT( 0<id, "affinity id must be positive integer" );
        __TBB_ASSERT( id <= my_num_slots, "affinity id out of bounds" );

        return ((mail_outbox*)this)[-(int)id];
    }

    void free_arena ();

    typedef uintptr_t pool_state_t;

    static const pool_state_t SNAPSHOT_EMPTY = 0;

    static const pool_state_t SNAPSHOT_FULL = pool_state_t(-1);

    static const unsigned ref_external_bits = 12; // up to 4095 external and 1M workers

    static const unsigned ref_external = 1;
    static const unsigned ref_worker   = 1<<ref_external_bits;

    static bool is_busy_or_empty( pool_state_t s ) { return s < SNAPSHOT_FULL; }

    unsigned num_workers_active() const {
        return my_references >> ref_external_bits;
    }

    bool is_recall_requested() const {
        return num_workers_active() > my_num_workers_allotted;
    }

    template<arena::new_work_type work_type> void advertise_new_work();


    bool is_out_of_work();

    void enqueue_task( task&, intptr_t, FastRandom & );

    void process( generic_scheduler& );

    template<unsigned ref_param>
    inline void on_thread_leaving ( );

#if __TBB_STATISTICS
    void dump_arena_statistics ();
#endif /* __TBB_STATISTICS */

#if __TBB_TASK_PRIORITY

    inline bool may_have_tasks ( generic_scheduler*, bool& tasks_present, bool& dequeuing_possible );

    void orphan_offloaded_tasks ( generic_scheduler& s );
#endif /* __TBB_TASK_PRIORITY */

#if __TBB_COUNT_TASK_NODES
    intptr_t workers_task_node_count();
#endif

    bool has_enqueued_tasks();

    static const size_t out_of_arena = ~size_t(0);
    template <bool as_worker>
    size_t occupy_free_slot( generic_scheduler& s );
    size_t occupy_free_slot_in_range( generic_scheduler& s, size_t lower, size_t upper );

    arena_slot my_slots[1];
}; // class arena

template<unsigned ref_param>
inline void arena::on_thread_leaving ( ) {
    //
    // Implementation of arena destruction synchronization logic contained various
    // bugs/flaws at the different stages of its evolution, so below is a detailed
    // description of the issues taken into consideration in the framework of the
    // current design.
    //
    // In case of using fire-and-forget tasks (scheduled via task::enqueue())
    // master thread is allowed to leave its arena before all its work is executed,
    // and market may temporarily revoke all workers from this arena. Since revoked
    // workers never attempt to reset arena state to EMPTY and cancel its request
    // to RML for threads, the arena object is destroyed only when both the last
    // thread is leaving it and arena's state is EMPTY (that is its master thread
    // left and it does not contain any work).
    // Thus resetting arena to EMPTY state (as earlier TBB versions did) should not
    // be done here (or anywhere else in the master thread to that matter); doing so
    // can result either in arena's premature destruction (at least without
    // additional costly checks in workers) or in unnecessary arena state changes
    // (and ensuing workers migration).
    //
    // A worker that checks for work presence and transitions arena to the EMPTY
    // state (in snapshot taking procedure arena::is_out_of_work()) updates
    // arena::my_pool_state first and only then arena::my_num_workers_requested.
    // So the check for work absence must be done against the latter field.
    //
    // In a time window between decrementing the active threads count and checking
    // if there is an outstanding request for workers. New worker thread may arrive,
    // finish remaining work, set arena state to empty, and leave decrementing its
    // refcount and destroying. Then the current thread will destroy the arena
    // the second time. To preclude it a local copy of the outstanding request
    // value can be stored before decrementing active threads count.
    //
    // But this technique may cause two other problem. When the stored request is
    // zero, it is possible that arena still has threads and they can generate new
    // tasks and thus re-establish non-zero requests. Then all the threads can be
    // revoked (as described above) leaving this thread the last one, and causing
    // it to destroy non-empty arena.
    //
    // The other problem takes place when the stored request is non-zero. Another
    // thread may complete the work, set arena state to empty, and leave without
    // arena destruction before this thread decrements the refcount. This thread
    // cannot destroy the arena either. Thus the arena may be "orphaned".
    //
    // In both cases we cannot dereference arena pointer after the refcount is
    // decremented, as our arena may already be destroyed.
    //
    // If this is the master thread, the market is protected by refcount to it.
    // In case of workers market's liveness is ensured by the RML connection
    // rundown protocol, according to which the client (i.e. the market) lives
    // until RML server notifies it about connection termination, and this
    // notification is fired only after all workers return into RML.
    //
    // Thus if we decremented refcount to zero we ask the market to check arena
    // state (including the fact if it is alive) under the lock.
    //
    uintptr_t aba_epoch = my_aba_epoch;
    market* m = my_market;
    __TBB_ASSERT(my_references >= ref_param, "broken arena reference counter");
#if __TBB_STATISTICS_EARLY_DUMP
    // While still holding a reference to the arena, compute how many external references are left.
    // If just one, dump statistics.
    if ( modulo_power_of_two(my_references,ref_worker)==ref_param ) // may only be true with ref_external
        GATHER_STATISTIC( dump_arena_statistics() );
#endif
#if __TBB_ENQUEUE_ENFORCED_CONCURRENCY
    // When there is no workers someone must free arena, as
    // without workers, no one calls is_out_of_work().
    // Skip workerless arenas because they have no demand for workers.
    // TODO: consider more strict conditions for the cleanup,
    // because it can create the demand of workers,
    // but the arena can be already empty (and so ready for destroying)
    // TODO: Fix the race: while we check soft limit and it might be changed.
    if( ref_param==ref_external && my_num_slots != my_num_reserved_slots
        && 0 == m->my_num_workers_soft_limit && !my_global_concurrency_mode ) {
        bool is_out = false;
        for (int i=0; i<num_priority_levels; i++) {
            is_out = is_out_of_work();
            if (is_out)
                break;
        }
        // We expect, that in worst case it's enough to have num_priority_levels-1
        // calls to restore priorities and yet another is_out_of_work() to conform
        // that no work was found. But as market::set_active_num_workers() can be called
        // concurrently, can't guarantee last is_out_of_work() return true.
    }
#endif
    if ( (my_references -= ref_param ) == 0 )
        m->try_destroy_arena( this, aba_epoch );
}

template<arena::new_work_type work_type> void arena::advertise_new_work() {
    if( work_type == work_enqueued ) {
#if __TBB_ENQUEUE_ENFORCED_CONCURRENCY
        if ( as_atomic(my_market->my_num_workers_soft_limit) == 0 && as_atomic(my_global_concurrency_mode) == false )
            my_market->enable_mandatory_concurrency(this);

        if ( my_max_num_workers == 0 && my_num_reserved_slots == 1 ) {
            __TBB_ASSERT(!my_local_concurrency_mode, NULL);
            my_local_concurrency_mode = true;
            my_pool_state = SNAPSHOT_FULL;
            my_max_num_workers = 1;
            my_market->adjust_demand(*this, my_max_num_workers);
            return;
        }
#endif /* __TBB_ENQUEUE_ENFORCED_CONCURRENCY */
        // Local memory fence here and below is required to avoid missed wakeups; see the comment below.
        // Starvation resistant tasks require concurrency, so missed wakeups are unacceptable.
        atomic_fence();
    }
    else if( work_type == wakeup ) {
        __TBB_ASSERT(my_max_num_workers!=0, "Unexpected worker wakeup request");
        atomic_fence();
    }
    // Double-check idiom that, in case of spawning, is deliberately sloppy about memory fences.
    // Technically, to avoid missed wakeups, there should be a full memory fence between the point we
    // released the task pool (i.e. spawned task) and read the arena's state.  However, adding such a
    // fence might hurt overall performance more than it helps, because the fence would be executed
    // on every task pool release, even when stealing does not occur.  Since TBB allows parallelism,
    // but never promises parallelism, the missed wakeup is not a correctness problem.
    pool_state_t snapshot = my_pool_state;
    if( is_busy_or_empty(snapshot) ) {
        // Attempt to mark as full.  The compare_and_swap below is a little unusual because the
        // result is compared to a value that can be different than the comparand argument.
        if( my_pool_state.compare_and_swap( SNAPSHOT_FULL, snapshot )==SNAPSHOT_EMPTY ) {
            if( snapshot!=SNAPSHOT_EMPTY ) {
                // This thread read "busy" into snapshot, and then another thread transitioned
                // my_pool_state to "empty" in the meantime, which caused the compare_and_swap above
                // to fail.  Attempt to transition my_pool_state from "empty" to "full".
                if( my_pool_state.compare_and_swap( SNAPSHOT_FULL, SNAPSHOT_EMPTY )!=SNAPSHOT_EMPTY ) {
                    // Some other thread transitioned my_pool_state from "empty", and hence became
                    // responsible for waking up workers.
                    return;
                }
            }
            // This thread transitioned pool from empty to full state, and thus is responsible for
            // telling the market that there is work to do.
#if __TBB_ENQUEUE_ENFORCED_CONCURRENCY
            if( work_type == work_spawned ) {
                if( my_local_concurrency_mode ) {
                    __TBB_ASSERT(my_max_num_workers==1, "");
                    __TBB_ASSERT(!governor::local_scheduler()->is_worker(), "");
                    // There was deliberate oversubscription on 1 core for sake of starvation-resistant tasks.
                    // Now a single active thread (must be the master) supposedly starts a new parallel region
                    // with relaxed sequential semantics, and oversubscription should be avoided.
                    // Demand for workers has been decreased to 0 during SNAPSHOT_EMPTY, so just keep it.
                    my_max_num_workers = 0;
                    my_local_concurrency_mode = false;
                    return;
                }
                if ( as_atomic(my_global_concurrency_mode) == true )
                    my_market->mandatory_concurrency_disable( this );
            }
#endif /* __TBB_ENQUEUE_ENFORCED_CONCURRENCY */
            // TODO: investigate adjusting of arena's demand by a single worker.
            my_market->adjust_demand( *this, my_max_num_workers );
        }
    }
}

} // namespace internal
} // namespace tbb

#endif /* _TBB_arena_H */