task_control.cpp 14.1 KB
Newer Older
gejun's avatar
gejun committed
1 2 3 4 5 6 7 8 9
// bthread - A M:N threading library to make applications more concurrent.
// Copyright (c) 2012 Baidu.com, Inc. All Rights Reserved

// Author: Ge,Jun (gejun@baidu.com)
// Date: Tue Jul 10 17:40:58 CST 2012

#include "base/scoped_lock.h"             // BAIDU_SCOPED_LOCK
#include "base/errno.h"                   // berror
#include "base/logging.h"
10
#include "base/third_party/murmurhash3/murmurhash3.h"
gejun's avatar
gejun committed
11 12 13 14 15 16 17
#include "bthread/sys_futex.h"            // futex_wake_private
#include "bthread/interrupt_pthread.h"
#include "bthread/processor.h"            // cpu_relax
#include "bthread/task_group.h"           // TaskGroup
#include "bthread/task_control.h"
#include "bthread/timer_thread.h"         // global_timer_thread
#include <gflags/gflags.h>
18
#include "bthread/log.h"
gejun's avatar
gejun committed
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86

DEFINE_int32(task_group_delete_delay, 1,
             "delay deletion of TaskGroup for so many seconds");
DEFINE_int32(task_group_runqueue_capacity, 4096,
             "capacity of runqueue in each TaskGroup");
DEFINE_int32(task_group_yield_before_idle, 0,
             "TaskGroup yields so many times before idle");

namespace bthread {

extern BAIDU_THREAD_LOCAL TaskGroup* tls_task_group;
void (*g_worker_startfn)() = NULL;

// May be called in other modules to run startfn in non-worker pthreads.
void run_worker_startfn() {
    if (g_worker_startfn) {
        g_worker_startfn();
    }
}

void* TaskControl::worker_thread(void* arg) {
    run_worker_startfn();    
#ifdef BAIDU_INTERNAL
    logging::ComlogInitializer comlog_initializer;
#endif
    
    TaskControl* c = static_cast<TaskControl*>(arg);
    TaskGroup* g = c->create_group();
    TaskStatistics stat;
    if (NULL == g) {
        LOG(ERROR) << "Fail to create TaskGroup in pthread=" << pthread_self();
        return NULL;
    }
    BT_VLOG << "Created worker=" << pthread_self()
            << " bthread=" << g->main_tid();

    tls_task_group = g;
    c->_nworkers << 1;
    g->run_main_task();

    stat = g->main_stat();
    BT_VLOG << "Destroying worker=" << pthread_self() << " bthread="
            << g->main_tid() << " idle=" << stat.cputime_ns / 1000000.0
            << "ms uptime=" << g->current_uptime_ns() / 1000000.0 << "ms";
    tls_task_group = NULL;
    g->destroy_self();
    c->_nworkers << -1;
    return NULL;
}

TaskGroup* TaskControl::create_group() {
    TaskGroup* g = new (std::nothrow) TaskGroup(this);
    if (NULL == g) {
        LOG(FATAL) << "Fail to new TaskGroup";
        return NULL;
    }
    if (g->init(FLAGS_task_group_runqueue_capacity) != 0) {
        LOG(ERROR) << "Fail to init TaskGroup";
        delete g;
        return NULL;
    }
    if (_add_group(g) != 0) {
        delete g;
        return NULL;
    }
    return g;
}

gejun's avatar
gejun committed
87
static void print_rq_sizes_in_the_tc(std::ostream &os, void *arg) {
gejun's avatar
gejun committed
88
    TaskControl *tc = (TaskControl *)arg;
gejun's avatar
gejun committed
89
    tc->print_rq_sizes(os);
gejun's avatar
gejun committed
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
}

static double get_cumulated_worker_time_from_this(void *arg) {
    return static_cast<TaskControl*>(arg)->get_cumulated_worker_time();
}

static int64_t get_cumulated_switch_count_from_this(void *arg) {
    return static_cast<TaskControl*>(arg)->get_cumulated_switch_count();
}

static int64_t get_cumulated_signal_count_from_this(void *arg) {
    return static_cast<TaskControl*>(arg)->get_cumulated_signal_count();
}

TaskControl::TaskControl()
    // NOTE: all fileds must be initialized before the vars.
    : _ngroup(0)
    , _groups((TaskGroup**)calloc(BTHREAD_MAX_CONCURRENCY, sizeof(TaskGroup*)))
    , _stop(false)
    , _concurrency(0)
    , _nworkers("bthread_worker_count")
gejun's avatar
gejun committed
111
    , _pending_time(NULL)
gejun's avatar
gejun committed
112 113 114
      // Delay exposure of following two vars because they rely on TC which
      // is not initialized yet.
    , _cumulated_worker_time(get_cumulated_worker_time_from_this, this)
gejun's avatar
gejun committed
115
    , _worker_usage_second(&_cumulated_worker_time, 1)
gejun's avatar
gejun committed
116 117 118 119
    , _cumulated_switch_count(get_cumulated_switch_count_from_this, this)
    , _switch_per_second(&_cumulated_switch_count)
    , _cumulated_signal_count(get_cumulated_signal_count_from_this, this)
    , _signal_per_second(&_cumulated_signal_count)
gejun's avatar
gejun committed
120
    , _status(print_rq_sizes_in_the_tc, this)
gejun's avatar
gejun committed
121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
    , _nbthreads("bthread_count")
{
    // calloc shall set memory to zero
    CHECK(_groups) << "Fail to create array of groups";
}

int TaskControl::init(int concurrency) {
    if (_concurrency != 0) {
        LOG(ERROR) << "Already initialized";
        return -1;
    }
    if (concurrency <= 0) {
        LOG(ERROR) << "Invalid concurrency=" << concurrency;
        return -1;
    }
    _concurrency = concurrency;

    // Make sure TimerThread is ready.
    if (get_or_create_global_timer_thread() == NULL) {
        LOG(ERROR) << "Fail to get global_timer_thread";
        return -1;
    }
    
    _workers.resize(_concurrency);   
    for (int i = 0; i < _concurrency; ++i) {
        const int rc = pthread_create(&_workers[i], NULL, worker_thread, this);
        if (rc) {
            LOG(ERROR) << "Fail to create _workers[" << i << "], " << berror(rc);
            return -1;
        }
    }
gejun's avatar
gejun committed
152
    _worker_usage_second.expose("bthread_worker_usage");
gejun's avatar
gejun committed
153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
    _switch_per_second.expose("bthread_switch_second");
    _signal_per_second.expose("bthread_signal_second");
    _status.expose("bthread_group_status");

    // Wait for at least one group is added so that choose_one_group()
    // never returns NULL.
    // TODO: Handle the case that worker quits before add_group
    while (_ngroup == 0) {
        usleep(100);  // TODO: Elaborate
    }
    return 0;
}

int TaskControl::add_workers(int num) {
    if (num <= 0) {
        return 0;
    }
    try {
        _workers.resize(_concurrency + num);
    } catch (...) {
        return 0;
    }
    const int old_concurency = _concurrency.load(base::memory_order_relaxed);
    for (int i = 0; i < num; ++i) {
        // Worker will add itself to _idle_workers, so we have to add
        // _concurrency before create a worker.
        _concurrency.fetch_add(1);
        const int rc = pthread_create(
                &_workers[i + old_concurency], NULL, worker_thread, this);
        if (rc) {
            LOG(WARNING) << "Fail to create _workers[" << i + old_concurency
                         << "], " << berror(rc);
            _concurrency.fetch_sub(1, base::memory_order_release);
            break;
        }
    }
    // Cannot fail
    _workers.resize(_concurrency.load(base::memory_order_relaxed));
    return _concurrency.load(base::memory_order_relaxed) - old_concurency;
}

TaskGroup* TaskControl::choose_one_group() {
    const size_t ngroup = _ngroup.load(base::memory_order_relaxed);
    if (ngroup != 0) {
        return _groups[base::fast_rand_less_than(ngroup)];
    }
    CHECK(false) << "Impossible: ngroup is 0";
    return NULL;
}

extern int stop_and_join_epoll_threads();

void TaskControl::stop_and_join() {
    // Close epoll threads so that worker threads are not waiting on epoll(
    // which cannot be woken up by signal_task below)
    CHECK_EQ(0, stop_and_join_epoll_threads());

    // Stop workers
    {
        BAIDU_SCOPED_LOCK(_modify_group_mutex);
        _stop = true;
        _ngroup.exchange(0, base::memory_order_relaxed); 
    }
216 217 218
    for (int i = 0; i < PARKING_LOT_NUM; ++i) {
        _pl[i].stop();
    }
gejun's avatar
gejun committed
219 220 221 222 223 224 225 226 227 228 229 230 231
    // Interrupt blocking operations.
    for (size_t i = 0; i < _workers.size(); ++i) {
        interrupt_pthread(_workers[i]);
    }
    // Join workers
    for (size_t i = 0; i < _workers.size(); ++i) {
        pthread_join(_workers[i], NULL);
    }
}

TaskControl::~TaskControl() {
    // NOTE: g_task_control is not destructed now because the situation
    //       is extremely racy.
gejun's avatar
gejun committed
232 233
    delete _pending_time.exchange(NULL, base::memory_order_relaxed);
    _worker_usage_second.hide();
gejun's avatar
gejun committed
234 235
    _switch_per_second.hide();
    _signal_per_second.hide();
gejun's avatar
gejun committed
236
    _status.hide();
gejun's avatar
gejun committed
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
    
    stop_and_join();

    free(_groups);
    _groups = NULL;
}

int TaskControl::_add_group(TaskGroup* g) {
    if (__builtin_expect(NULL == g, 0)) {
        return -1;
    }
    std::unique_lock<base::Mutex> mu(_modify_group_mutex);
    if (_stop) {
        return -1;
    }
    size_t ngroup = _ngroup.load(base::memory_order_relaxed);
    if (ngroup < (size_t)BTHREAD_MAX_CONCURRENCY) {
        _groups[ngroup] = g;
        _ngroup.store(ngroup + 1, base::memory_order_release);
    }
    mu.unlock();
    // See the comments in _destroy_group
    // TODO: Not needed anymore since non-worker pthread cannot have TaskGroup
    signal_task(65536);
    return 0;
}

void TaskControl::delete_task_group(void* arg) {
    delete(TaskGroup*)arg;
}

int TaskControl::_destroy_group(TaskGroup* g) {
    if (NULL == g) {
        LOG(ERROR) << "Param[g] is NULL";
        return -1;
    }
    if (g->_control != this) {
        LOG(ERROR) << "TaskGroup=" << g
                   << " does not belong to this TaskControl=" << this;
        return -1;
    }
    bool erased = false;
    {
        BAIDU_SCOPED_LOCK(_modify_group_mutex);
        const size_t ngroup = _ngroup.load(base::memory_order_relaxed);
        for (size_t i = 0; i < ngroup; ++i) {
            if (_groups[i] == g) {
                // No need for atomic_thread_fence because lock did it.
                _groups[i] = _groups[ngroup - 1];
                // Change _ngroup and keep _groups unchanged at last so that:
                //  - If steal_task sees the newest _ngroup, it would not touch
                //    _groups[ngroup -1]
                //  - If steal_task sees old _ngroup and is still iterating on
                //    _groups, it would not miss _groups[ngroup - 1] which was 
                //    swapped to _groups[i]. Although adding new group would
                //    overwrite it, since we do signal_task in _add_group(),
                //    we think the pending tasks of _groups[ngroup - 1] would
                //    not miss.
                _ngroup.store(ngroup - 1, base::memory_order_release);
                //_groups[ngroup - 1] = NULL;
                erased = true;
                break;
            }
        }
    }

    // Can't delete g immediately because for performance consideration,
    // we don't lock _modify_group_mutex in steal_task which may
    // access the removed group concurrently. We use simple strategy here:
    // Schedule a function which deletes the TaskGroup after
    // FLAGS_task_group_delete_delay seconds
    if (erased) {
        get_global_timer_thread()->schedule(
            delete_task_group, g,
            base::microseconds_from_now(FLAGS_task_group_delete_delay * 1000000L));
    }
    return 0;
}

bool TaskControl::steal_task(bthread_t* tid, size_t* seed, size_t offset) {
    // 1: Acquiring fence is paired with releasing fence in _add_group to
    // avoid accessing uninitialized slot of _groups.
    const size_t ngroup = _ngroup.load(base::memory_order_acquire/*1*/);
    if (0 == ngroup) {
        return false;
    }

    // NOTE: Don't return inside `for' iteration since we need to update |seed|
    bool stolen = false;
    size_t s = *seed;
    for (size_t i = 0; i < ngroup; ++i, s += offset) {
        TaskGroup* g = _groups[s % ngroup];
        // g is possibly NULL because of concurrent _destroy_group
330 331 332 333 334 335 336 337 338
        if (g) {
            if (g->_rq.steal(tid)) {
                stolen = true;
                break;
            }
            if (g->_remote_rq.pop(tid)) {
                stolen = true;
                break;
            }
gejun's avatar
gejun committed
339 340 341 342 343 344 345 346 347 348
        }
    }
    *seed = s;
    return stolen;
}

void TaskControl::signal_task(int num_task) {
    if (num_task <= 0) {
        return;
    }
349 350 351 352 353 354 355
    // TODO(gejun): Current algorithm does not guarantee enough threads will
    // be created to match caller's requests. But in another side, there's also
    // many useless signalings according to current impl. Capping the concurrency
    // is a good balance between performance and timeliness of scheduling.
    if (num_task > 2) {
        num_task = 2;
    }
gejun's avatar
gejun committed
356
    int start_index = base::fmix64(pthread_self()) % PARKING_LOT_NUM;
357 358 359 360 361 362 363 364
    num_task -= _pl[start_index].signal(1);
    if (num_task > 0) {
        for (int i = 1; i < PARKING_LOT_NUM && num_task > 0; ++i) {
            if (++start_index >= PARKING_LOT_NUM) {
                start_index = 0;
            }
            num_task -= _pl[start_index].signal(1);
        }
gejun's avatar
gejun committed
365 366 367
    }
}

gejun's avatar
gejun committed
368
void TaskControl::print_rq_sizes(std::ostream& os) {
gejun's avatar
gejun committed
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412
    const size_t ngroup = _ngroup.load(base::memory_order_relaxed);
    DEFINE_SMALL_ARRAY(int, nums, ngroup, 128);
    {
        BAIDU_SCOPED_LOCK(_modify_group_mutex);
        // ngroup > _ngroup: nums[_ngroup ... ngroup-1] = 0
        // ngroup < _ngroup: just ignore _groups[_ngroup ... ngroup-1]
        for (size_t i = 0; i < ngroup; ++i) {
            nums[i] = (_groups[i] ? _groups[i]->_rq.volatile_size() : 0);
        }
    }
    for (size_t i = 0; i < ngroup; ++i) {
        os << nums[i] << ' ';
    }
}

double TaskControl::get_cumulated_worker_time() {
    int64_t cputime_ns = 0;
    BAIDU_SCOPED_LOCK(_modify_group_mutex);
    const size_t ngroup = _ngroup.load(base::memory_order_relaxed);
    for (size_t i = 0; i < ngroup; ++i) {
        if (_groups[i]) {
            cputime_ns += _groups[i]->_cumulated_cputime_ns;
        }
    }
    return cputime_ns / 1000000000.0;
}

int64_t TaskControl::get_cumulated_switch_count() {
    int64_t c = 0;
    BAIDU_SCOPED_LOCK(_modify_group_mutex);
    const size_t ngroup = _ngroup.load(base::memory_order_relaxed);
    for (size_t i = 0; i < ngroup; ++i) {
        if (_groups[i]) {
            c += _groups[i]->_nswitch;
        }
    }
    return c;
}

int64_t TaskControl::get_cumulated_signal_count() {
    int64_t c = 0;
    BAIDU_SCOPED_LOCK(_modify_group_mutex);
    const size_t ngroup = _ngroup.load(base::memory_order_relaxed);
    for (size_t i = 0; i < ngroup; ++i) {
413 414 415
        TaskGroup* g = _groups[i];
        if (g) {
            c += g->_nsignaled + g->_remote_nsignaled;
gejun's avatar
gejun committed
416 417 418 419 420
        }
    }
    return c;
}

gejun's avatar
gejun committed
421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
bvar::LatencyRecorder* TaskControl::create_exposed_pending_time() {
    bool is_creator = false;
    _pending_time_mutex.lock();
    bvar::LatencyRecorder* pt = _pending_time.load(base::memory_order_consume);
    if (!pt) {
        pt = new bvar::LatencyRecorder;
        _pending_time.store(pt, base::memory_order_release);
        is_creator = true;
    }
    _pending_time_mutex.unlock();
    if (is_creator) {
        pt->expose("bthread_creation");
    }
    return pt;
}

gejun's avatar
gejun committed
437
}  // namespace bthread