/****************************************************************************** * Icinga 2 * * Copyright (C) 2012-present Icinga Development Team (http://www.icinga.org) * * * * This program is free software; you can redistribute it and/or * * modify it under the terms of the GNU General Public License * * as published by the Free Software Foundation; either version 2 * * of the License, or (at your option) any later version. * * * * This program is distributed in the hope that it will be useful, * * but WITHOUT ANY WARRANTY; without even the implied warranty of * * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * * GNU General Public License for more details. * * * * You should have received a copy of the GNU General Public License * * along with this program; if not, write to the Free Software Foundation * * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. * ******************************************************************************/ #include "base/threadpool.h" #include "base/logger_fwd.h" #include "base/convert.h" #include "base/debug.h" #include "base/utility.h" #include "base/application.h" #include "base/exception.h" #include #include #include #include using namespace icinga; int ThreadPool::m_NextID = 1; ThreadPool::ThreadPool(int max_threads) : m_ID(m_NextID++), m_MaxThreads(max_threads), m_Stopped(false) { if (m_MaxThreads != -1 && m_MaxThreads < sizeof(m_Queues) / sizeof(m_Queues[0])) m_MaxThreads = sizeof(m_Queues) / sizeof(m_Queues[0]); Start(); } ThreadPool::~ThreadPool(void) { Stop(); Join(true); } void ThreadPool::Start(void) { for (int i = 0; i < sizeof(m_Queues) / sizeof(m_Queues[0]); i++) m_Queues[i].SpawnWorker(m_ThreadGroup); m_ThreadGroup.create_thread(boost::bind(&ThreadPool::ManagerThreadProc, this)); m_ThreadGroup.create_thread(boost::bind(&ThreadPool::StatsThreadProc, this)); } void ThreadPool::Stop(void) { for (int i = 0; i < sizeof(m_Queues) / sizeof(m_Queues[0]); i++) { boost::mutex::scoped_lock lock(m_Queues[i].Mutex); m_Queues[i].Stopped = true; m_Queues[i].CV.notify_all(); } boost::mutex::scoped_lock lock(m_MgmtMutex); m_Stopped = true; m_MgmtCV.notify_all(); } /** * Waits for all worker threads to finish. */ void ThreadPool::Join(bool wait_for_stop) { if (wait_for_stop) { m_ThreadGroup.join_all(); return; } for (int i = 0; i < sizeof(m_Queues) / sizeof(m_Queues[0]); i++) { boost::mutex::scoped_lock lock(m_Queues[i].Mutex); while (!m_Queues[i].Items.empty()) m_Queues[i].CVStarved.wait(lock); } } /** * Waits for work items and processes them. */ void ThreadPool::WorkerThread::ThreadProc(Queue& queue) { std::ostringstream idbuf; idbuf << "Q #" << &queue << " W #" << this; Utility::SetThreadName(idbuf.str()); for (;;) { WorkItem wi; { boost::mutex::scoped_lock lock(queue.Mutex); UpdateUtilization(ThreadIdle); while (queue.Items.empty() && !queue.Stopped && !Zombie) { if (queue.Items.empty()) queue.CVStarved.notify_all(); queue.CV.wait(lock); } if (Zombie) break; if (queue.Items.empty() && queue.Stopped) break; wi = queue.Items.front(); queue.Items.pop_front(); UpdateUtilization(ThreadBusy); } double st = Utility::GetTime();; #ifdef _DEBUG # ifdef RUSAGE_THREAD struct rusage usage_start, usage_end; (void) getrusage(RUSAGE_THREAD, &usage_start); # endif /* RUSAGE_THREAD */ #endif /* _DEBUG */ try { wi.Callback(); } catch (const std::exception& ex) { std::ostringstream msgbuf; msgbuf << "Exception thrown in event handler: " << std::endl << DiagnosticInformation(ex); Log(LogCritical, "base", msgbuf.str()); } catch (...) { Log(LogCritical, "base", "Exception of unknown type thrown in event handler."); } double et = Utility::GetTime(); double latency = st - wi.Timestamp; { boost::mutex::scoped_lock lock(queue.Mutex); queue.WaitTime += latency; queue.ServiceTime += et - st; queue.TaskCount++; } #ifdef _DEBUG # ifdef RUSAGE_THREAD (void) getrusage(RUSAGE_THREAD, &usage_end); double duser = (usage_end.ru_utime.tv_sec - usage_start.ru_utime.tv_sec) + (usage_end.ru_utime.tv_usec - usage_start.ru_utime.tv_usec) / 1000000.0; double dsys = (usage_end.ru_stime.tv_sec - usage_start.ru_stime.tv_sec) + (usage_end.ru_stime.tv_usec - usage_start.ru_stime.tv_usec) / 1000000.0; double dwait = (et - st) - (duser + dsys); int dminfaults = usage_end.ru_minflt - usage_start.ru_minflt; int dmajfaults = usage_end.ru_majflt - usage_start.ru_majflt; int dvctx = usage_end.ru_nvcsw - usage_start.ru_nvcsw; int divctx = usage_end.ru_nivcsw - usage_start.ru_nivcsw; # endif /* RUSAGE_THREAD */ if (et - st > 0.5) { std::ostringstream msgbuf; # ifdef RUSAGE_THREAD msgbuf << "Event call took user:" << duser << "s, system:" << dsys << "s, wait:" << dwait << "s, minor_faults:" << dminfaults << ", major_faults:" << dmajfaults << ", voluntary_csw:" << dvctx << ", involuntary_csw:" << divctx; # else msgbuf << "Event call took " << (et - st) << "s"; # endif /* RUSAGE_THREAD */ Log(LogWarning, "base", msgbuf.str()); } #endif /* _DEBUG */ } boost::mutex::scoped_lock lock(queue.Mutex); UpdateUtilization(ThreadDead); Zombie = false; } /** * Appends a work item to the work queue. Work items will be processed in FIFO order. * * @param callback The callback function for the work item. * @returns true if the item was queued, false otherwise. */ bool ThreadPool::Post(const ThreadPool::WorkFunction& callback) { WorkItem wi; wi.Callback = callback; wi.Timestamp = Utility::GetTime(); Queue& queue = m_Queues[Utility::Random() % (sizeof(m_Queues) / sizeof(m_Queues[0]))]; { boost::mutex::scoped_lock lock(queue.Mutex); if (queue.Stopped) return false; queue.Items.push_back(wi); queue.CV.notify_one(); } return true; } void ThreadPool::ManagerThreadProc(void) { std::ostringstream idbuf; idbuf << "TP #" << m_ID << " Manager"; Utility::SetThreadName(idbuf.str()); for (;;) { size_t total_pending = 0, total_alive = 0; double total_avg_latency = 0; double total_utilization = 0; { boost::mutex::scoped_lock lock(m_MgmtMutex); if (!m_Stopped) m_MgmtCV.timed_wait(lock, boost::posix_time::seconds(5)); if (m_Stopped) break; } for (int i = 0; i < sizeof(m_Queues) / sizeof(m_Queues[0]); i++) { size_t pending, alive = 0; double avg_latency; double utilization = 0; Queue& queue = m_Queues[i]; boost::mutex::scoped_lock lock(queue.Mutex); pending = queue.Items.size(); for (size_t i = 0; i < sizeof(queue.Threads) / sizeof(queue.Threads[0]); i++) { if (queue.Threads[i].State != ThreadDead && !queue.Threads[i].Zombie) { alive++; utilization += queue.Threads[i].Utilization * 100; } } utilization /= alive; if (queue.TaskCount > 0) avg_latency = queue.WaitTime / (queue.TaskCount * 1.0); else avg_latency = 0; if (utilization < 60 || utilization > 80 || alive < 8) { double wthreads = ceil((utilization * alive) / 80.0); int tthreads = wthreads - alive; /* Don't ever kill the last threads. */ if (alive + tthreads < 2) tthreads = 2 - alive; /* Don't kill more than 8 threads at once. */ if (tthreads < -8) tthreads = -8; /* Spawn more workers if there are outstanding work items. */ if (tthreads > 0 && pending > 0) tthreads = 8; if (m_MaxThreads != -1 && (alive + tthreads) * (sizeof(m_Queues) / sizeof(m_Queues[0])) > m_MaxThreads) tthreads = m_MaxThreads / (sizeof(m_Queues) / sizeof(m_Queues[0])) - alive; std::ostringstream msgbuf; msgbuf << "Thread pool; current: " << alive << "; adjustment: " << tthreads; Log(LogDebug, "base", msgbuf.str()); for (int i = 0; i < -tthreads; i++) queue.KillWorker(m_ThreadGroup); for (int i = 0; i < tthreads; i++) queue.SpawnWorker(m_ThreadGroup); } queue.WaitTime = 0; queue.ServiceTime = 0; queue.TaskCount = 0; total_pending += pending; total_alive += alive; total_avg_latency += avg_latency; total_utilization += utilization; } std::ostringstream msgbuf; msgbuf << "Pool #" << m_ID << ": Pending tasks: " << total_pending << "; Average latency: " << (long)(total_avg_latency * 1000 / (sizeof(m_Queues) / sizeof(m_Queues[0]))) << "ms" << "; Threads: " << total_alive << "; Pool utilization: " << (total_utilization / (sizeof(m_Queues) / sizeof(m_Queues[0]))) << "%"; Log(LogInformation, "base", msgbuf.str()); } } /** * Note: Caller must hold m_Mutex */ void ThreadPool::Queue::SpawnWorker(boost::thread_group& group) { for (size_t i = 0; i < sizeof(Threads) / sizeof(Threads[0]); i++) { if (Threads[i].State == ThreadDead) { Log(LogDebug, "base", "Spawning worker thread."); Threads[i] = WorkerThread(ThreadIdle); Threads[i].Thread = group.create_thread(boost::bind(&ThreadPool::WorkerThread::ThreadProc, boost::ref(Threads[i]), boost::ref(*this))); break; } } } /** * Note: Caller must hold Mutex. */ void ThreadPool::Queue::KillWorker(boost::thread_group& group) { for (size_t i = 0; i < sizeof(Threads) / sizeof(Threads[0]); i++) { if (Threads[i].State == ThreadIdle && !Threads[i].Zombie) { Log(LogDebug, "base", "Killing worker thread."); group.remove_thread(Threads[i].Thread); Threads[i].Thread->detach(); delete Threads[i].Thread; Threads[i].Zombie = true; CV.notify_all(); break; } } } void ThreadPool::StatsThreadProc(void) { std::ostringstream idbuf; idbuf << "TP #" << m_ID << " Stats"; Utility::SetThreadName(idbuf.str()); for (;;) { { boost::mutex::scoped_lock lock(m_MgmtMutex); if (!m_Stopped) m_MgmtCV.timed_wait(lock, boost::posix_time::milliseconds(250)); if (m_Stopped) break; } for (int i = 0; i < sizeof(m_Queues) / sizeof(m_Queues[0]); i++) { Queue& queue = m_Queues[i]; boost::mutex::scoped_lock lock(queue.Mutex); for (size_t i = 0; i < sizeof(queue.Threads) / sizeof(queue.Threads[0]); i++) queue.Threads[i].UpdateUtilization(); } } } /** * Note: Caller must hold queue Mutex. */ void ThreadPool::WorkerThread::UpdateUtilization(ThreadState state) { double utilization; switch (State) { case ThreadDead: return; case ThreadIdle: utilization = 0; break; case ThreadBusy: utilization = 1; break; default: ASSERT(0); } double now = Utility::GetTime(); double time = now - LastUpdate; const double avg_time = 5.0; if (time > avg_time) time = avg_time; Utilization = (Utilization * (avg_time - time) + utilization * time) / avg_time; LastUpdate = now; if (state != ThreadUnspecified) State = state; }