OpenJDK / portola / portola
changeset 47792:2a5b7592ca11
Merge
author | tschatzl |
---|---|
date | Mon, 06 Nov 2017 15:33:54 +0100 |
parents | 12c0996f50f4 71c9ee6b7cb3 |
children | 3dcd54513db1 |
files | src/hotspot/share/gc/g1/concurrentG1Refine.cpp src/hotspot/share/gc/g1/concurrentG1Refine.hpp src/hotspot/share/gc/g1/concurrentG1RefineThread.cpp src/hotspot/share/gc/g1/concurrentG1RefineThread.hpp |
diffstat | 15 files changed, 762 insertions(+), 762 deletions(-) [+] |
line wrap: on
line diff
--- a/src/hotspot/share/gc/g1/concurrentG1Refine.cpp Mon Nov 06 14:31:22 2017 +0000 +++ /dev/null Thu Jan 01 00:00:00 1970 +0000 @@ -1,370 +0,0 @@ -/* - * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. - * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. - * - * This code is free software; you can redistribute it and/or modify it - * under the terms of the GNU General Public License version 2 only, as - * published by the Free Software Foundation. - * - * This code 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 - * version 2 for more details (a copy is included in the LICENSE file that - * accompanied this code). - * - * You should have received a copy of the GNU General Public License version - * 2 along with this work; if not, write to the Free Software Foundation, - * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. - * - * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA - * or visit www.oracle.com if you need additional information or have any - * questions. - * - */ - -#include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" -#include "gc/g1/concurrentG1RefineThread.hpp" -#include "gc/g1/g1YoungRemSetSamplingThread.hpp" -#include "logging/log.hpp" -#include "runtime/java.hpp" -#include "runtime/thread.hpp" -#include "utilities/debug.hpp" -#include "utilities/globalDefinitions.hpp" -#include "utilities/pair.hpp" -#include <math.h> - -// Arbitrary but large limits, to simplify some of the zone calculations. -// The general idea is to allow expressions like -// MIN2(x OP y, max_XXX_zone) -// without needing to check for overflow in "x OP y", because the -// ranges for x and y have been restricted. -STATIC_ASSERT(sizeof(LP64_ONLY(jint) NOT_LP64(jshort)) <= (sizeof(size_t)/2)); -const size_t max_yellow_zone = LP64_ONLY(max_jint) NOT_LP64(max_jshort); -const size_t max_green_zone = max_yellow_zone / 2; -const size_t max_red_zone = INT_MAX; // For dcqs.set_max_completed_queue. -STATIC_ASSERT(max_yellow_zone <= max_red_zone); - -// Range check assertions for green zone values. -#define assert_zone_constraints_g(green) \ - do { \ - size_t azc_g_green = (green); \ - assert(azc_g_green <= max_green_zone, \ - "green exceeds max: " SIZE_FORMAT, azc_g_green); \ - } while (0) - -// Range check assertions for green and yellow zone values. -#define assert_zone_constraints_gy(green, yellow) \ - do { \ - size_t azc_gy_green = (green); \ - size_t azc_gy_yellow = (yellow); \ - assert_zone_constraints_g(azc_gy_green); \ - assert(azc_gy_yellow <= max_yellow_zone, \ - "yellow exceeds max: " SIZE_FORMAT, azc_gy_yellow); \ - assert(azc_gy_green <= azc_gy_yellow, \ - "green (" SIZE_FORMAT ") exceeds yellow (" SIZE_FORMAT ")", \ - azc_gy_green, azc_gy_yellow); \ - } while (0) - -// Range check assertions for green, yellow, and red zone values. -#define assert_zone_constraints_gyr(green, yellow, red) \ - do { \ - size_t azc_gyr_green = (green); \ - size_t azc_gyr_yellow = (yellow); \ - size_t azc_gyr_red = (red); \ - assert_zone_constraints_gy(azc_gyr_green, azc_gyr_yellow); \ - assert(azc_gyr_red <= max_red_zone, \ - "red exceeds max: " SIZE_FORMAT, azc_gyr_red); \ - assert(azc_gyr_yellow <= azc_gyr_red, \ - "yellow (" SIZE_FORMAT ") exceeds red (" SIZE_FORMAT ")", \ - azc_gyr_yellow, azc_gyr_red); \ - } while (0) - -// Logging tag sequence for refinement control updates. -#define CTRL_TAGS gc, ergo, refine - -// For logging zone values, ensuring consistency of level and tags. -#define LOG_ZONES(...) log_debug( CTRL_TAGS )(__VA_ARGS__) - -// Package for pair of refinement thread activation and deactivation -// thresholds. The activation and deactivation levels are resp. the first -// and second values of the pair. -typedef Pair<size_t, size_t> Thresholds; -inline size_t activation_level(const Thresholds& t) { return t.first; } -inline size_t deactivation_level(const Thresholds& t) { return t.second; } - -static Thresholds calc_thresholds(size_t green_zone, - size_t yellow_zone, - uint worker_i) { - double yellow_size = yellow_zone - green_zone; - double step = yellow_size / ConcurrentG1Refine::thread_num(); - if (worker_i == 0) { - // Potentially activate worker 0 more aggressively, to keep - // available buffers near green_zone value. When yellow_size is - // large we don't want to allow a full step to accumulate before - // doing any processing, as that might lead to significantly more - // than green_zone buffers to be processed by update_rs. - step = MIN2(step, ParallelGCThreads / 2.0); - } - size_t activate_offset = static_cast<size_t>(ceil(step * (worker_i + 1))); - size_t deactivate_offset = static_cast<size_t>(floor(step * worker_i)); - return Thresholds(green_zone + activate_offset, - green_zone + deactivate_offset); -} - -ConcurrentG1Refine::ConcurrentG1Refine(size_t green_zone, - size_t yellow_zone, - size_t red_zone, - size_t min_yellow_zone_size) : - _threads(NULL), - _sample_thread(NULL), - _n_worker_threads(thread_num()), - _green_zone(green_zone), - _yellow_zone(yellow_zone), - _red_zone(red_zone), - _min_yellow_zone_size(min_yellow_zone_size) -{ - assert_zone_constraints_gyr(green_zone, yellow_zone, red_zone); -} - -static size_t calc_min_yellow_zone_size() { - size_t step = G1ConcRefinementThresholdStep; - uint n_workers = ConcurrentG1Refine::thread_num(); - if ((max_yellow_zone / step) < n_workers) { - return max_yellow_zone; - } else { - return step * n_workers; - } -} - -static size_t calc_init_green_zone() { - size_t green = G1ConcRefinementGreenZone; - if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) { - green = ParallelGCThreads; - } - return MIN2(green, max_green_zone); -} - -static size_t calc_init_yellow_zone(size_t green, size_t min_size) { - size_t config = G1ConcRefinementYellowZone; - size_t size = 0; - if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) { - size = green * 2; - } else if (green < config) { - size = config - green; - } - size = MAX2(size, min_size); - size = MIN2(size, max_yellow_zone); - return MIN2(green + size, max_yellow_zone); -} - -static size_t calc_init_red_zone(size_t green, size_t yellow) { - size_t size = yellow - green; - if (!FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) { - size_t config = G1ConcRefinementRedZone; - if (yellow < config) { - size = MAX2(size, config - yellow); - } - } - return MIN2(yellow + size, max_red_zone); -} - -ConcurrentG1Refine* ConcurrentG1Refine::create(jint* ecode) { - size_t min_yellow_zone_size = calc_min_yellow_zone_size(); - size_t green_zone = calc_init_green_zone(); - size_t yellow_zone = calc_init_yellow_zone(green_zone, min_yellow_zone_size); - size_t red_zone = calc_init_red_zone(green_zone, yellow_zone); - - LOG_ZONES("Initial Refinement Zones: " - "green: " SIZE_FORMAT ", " - "yellow: " SIZE_FORMAT ", " - "red: " SIZE_FORMAT ", " - "min yellow size: " SIZE_FORMAT, - green_zone, yellow_zone, red_zone, min_yellow_zone_size); - - ConcurrentG1Refine* cg1r = new ConcurrentG1Refine(green_zone, - yellow_zone, - red_zone, - min_yellow_zone_size); - - if (cg1r == NULL) { - *ecode = JNI_ENOMEM; - vm_shutdown_during_initialization("Could not create ConcurrentG1Refine"); - return NULL; - } - - cg1r->_threads = NEW_C_HEAP_ARRAY_RETURN_NULL(ConcurrentG1RefineThread*, cg1r->_n_worker_threads, mtGC); - if (cg1r->_threads == NULL) { - *ecode = JNI_ENOMEM; - vm_shutdown_during_initialization("Could not allocate an array for ConcurrentG1RefineThread"); - return NULL; - } - - uint worker_id_offset = DirtyCardQueueSet::num_par_ids(); - - ConcurrentG1RefineThread *next = NULL; - for (uint i = cg1r->_n_worker_threads - 1; i != UINT_MAX; i--) { - Thresholds thresholds = calc_thresholds(green_zone, yellow_zone, i); - ConcurrentG1RefineThread* t = - new ConcurrentG1RefineThread(cg1r, - next, - worker_id_offset, - i, - activation_level(thresholds), - deactivation_level(thresholds)); - assert(t != NULL, "Conc refine should have been created"); - if (t->osthread() == NULL) { - *ecode = JNI_ENOMEM; - vm_shutdown_during_initialization("Could not create ConcurrentG1RefineThread"); - return NULL; - } - - assert(t->cg1r() == cg1r, "Conc refine thread should refer to this"); - cg1r->_threads[i] = t; - next = t; - } - - cg1r->_sample_thread = new G1YoungRemSetSamplingThread(); - if (cg1r->_sample_thread->osthread() == NULL) { - *ecode = JNI_ENOMEM; - vm_shutdown_during_initialization("Could not create G1YoungRemSetSamplingThread"); - return NULL; - } - - *ecode = JNI_OK; - return cg1r; -} - -void ConcurrentG1Refine::stop() { - for (uint i = 0; i < _n_worker_threads; i++) { - _threads[i]->stop(); - } - _sample_thread->stop(); -} - -void ConcurrentG1Refine::update_thread_thresholds() { - for (uint i = 0; i < _n_worker_threads; i++) { - Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, i); - _threads[i]->update_thresholds(activation_level(thresholds), - deactivation_level(thresholds)); - } -} - -ConcurrentG1Refine::~ConcurrentG1Refine() { - for (uint i = 0; i < _n_worker_threads; i++) { - delete _threads[i]; - } - FREE_C_HEAP_ARRAY(ConcurrentG1RefineThread*, _threads); - - delete _sample_thread; -} - -void ConcurrentG1Refine::threads_do(ThreadClosure *tc) { - worker_threads_do(tc); - tc->do_thread(_sample_thread); -} - -void ConcurrentG1Refine::worker_threads_do(ThreadClosure * tc) { - for (uint i = 0; i < _n_worker_threads; i++) { - tc->do_thread(_threads[i]); - } -} - -uint ConcurrentG1Refine::thread_num() { - return G1ConcRefinementThreads; -} - -void ConcurrentG1Refine::print_worker_threads_on(outputStream* st) const { - for (uint i = 0; i < _n_worker_threads; ++i) { - _threads[i]->print_on(st); - st->cr(); - } - _sample_thread->print_on(st); - st->cr(); -} - -static size_t calc_new_green_zone(size_t green, - double update_rs_time, - size_t update_rs_processed_buffers, - double goal_ms) { - // Adjust green zone based on whether we're meeting the time goal. - // Limit to max_green_zone. - const double inc_k = 1.1, dec_k = 0.9; - if (update_rs_time > goal_ms) { - if (green > 0) { - green = static_cast<size_t>(green * dec_k); - } - } else if (update_rs_time < goal_ms && - update_rs_processed_buffers > green) { - green = static_cast<size_t>(MAX2(green * inc_k, green + 1.0)); - green = MIN2(green, max_green_zone); - } - return green; -} - -static size_t calc_new_yellow_zone(size_t green, size_t min_yellow_size) { - size_t size = green * 2; - size = MAX2(size, min_yellow_size); - return MIN2(green + size, max_yellow_zone); -} - -static size_t calc_new_red_zone(size_t green, size_t yellow) { - return MIN2(yellow + (yellow - green), max_red_zone); -} - -void ConcurrentG1Refine::update_zones(double update_rs_time, - size_t update_rs_processed_buffers, - double goal_ms) { - log_trace( CTRL_TAGS )("Updating Refinement Zones: " - "update_rs time: %.3fms, " - "update_rs buffers: " SIZE_FORMAT ", " - "update_rs goal time: %.3fms", - update_rs_time, - update_rs_processed_buffers, - goal_ms); - - _green_zone = calc_new_green_zone(_green_zone, - update_rs_time, - update_rs_processed_buffers, - goal_ms); - _yellow_zone = calc_new_yellow_zone(_green_zone, _min_yellow_zone_size); - _red_zone = calc_new_red_zone(_green_zone, _yellow_zone); - - assert_zone_constraints_gyr(_green_zone, _yellow_zone, _red_zone); - LOG_ZONES("Updated Refinement Zones: " - "green: " SIZE_FORMAT ", " - "yellow: " SIZE_FORMAT ", " - "red: " SIZE_FORMAT, - _green_zone, _yellow_zone, _red_zone); -} - -void ConcurrentG1Refine::adjust(double update_rs_time, - size_t update_rs_processed_buffers, - double goal_ms) { - DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); - - if (G1UseAdaptiveConcRefinement) { - update_zones(update_rs_time, update_rs_processed_buffers, goal_ms); - update_thread_thresholds(); - - // Change the barrier params - if (_n_worker_threads == 0) { - // Disable dcqs notification when there are no threads to notify. - dcqs.set_process_completed_threshold(INT_MAX); - } else { - // Worker 0 is the primary; wakeup is via dcqs notification. - STATIC_ASSERT(max_yellow_zone <= INT_MAX); - size_t activate = _threads[0]->activation_threshold(); - dcqs.set_process_completed_threshold((int)activate); - } - dcqs.set_max_completed_queue((int)red_zone()); - } - - size_t curr_queue_size = dcqs.completed_buffers_num(); - if (curr_queue_size >= yellow_zone()) { - dcqs.set_completed_queue_padding(curr_queue_size); - } else { - dcqs.set_completed_queue_padding(0); - } - dcqs.notify_if_necessary(); -}
--- a/src/hotspot/share/gc/g1/concurrentG1Refine.hpp Mon Nov 06 14:31:22 2017 +0000 +++ /dev/null Thu Jan 01 00:00:00 1970 +0000 @@ -1,107 +0,0 @@ -/* - * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. - * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. - * - * This code is free software; you can redistribute it and/or modify it - * under the terms of the GNU General Public License version 2 only, as - * published by the Free Software Foundation. - * - * This code 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 - * version 2 for more details (a copy is included in the LICENSE file that - * accompanied this code). - * - * You should have received a copy of the GNU General Public License version - * 2 along with this work; if not, write to the Free Software Foundation, - * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. - * - * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA - * or visit www.oracle.com if you need additional information or have any - * questions. - * - */ - -#ifndef SHARE_VM_GC_G1_CONCURRENTG1REFINE_HPP -#define SHARE_VM_GC_G1_CONCURRENTG1REFINE_HPP - -#include "memory/allocation.hpp" -#include "utilities/globalDefinitions.hpp" - -// Forward decl -class CardTableEntryClosure; -class ConcurrentG1RefineThread; -class G1YoungRemSetSamplingThread; -class outputStream; -class ThreadClosure; - -class ConcurrentG1Refine: public CHeapObj<mtGC> { - G1YoungRemSetSamplingThread* _sample_thread; - - ConcurrentG1RefineThread** _threads; - uint _n_worker_threads; - /* - * The value of the update buffer queue length falls into one of 3 zones: - * green, yellow, red. If the value is in [0, green) nothing is - * done, the buffers are left unprocessed to enable the caching effect of the - * dirtied cards. In the yellow zone [green, yellow) the concurrent refinement - * threads are gradually activated. In [yellow, red) all threads are - * running. If the length becomes red (max queue length) the mutators start - * processing the buffers. - * - * There are some interesting cases (when G1UseAdaptiveConcRefinement - * is turned off): - * 1) green = yellow = red = 0. In this case the mutator will process all - * buffers. Except for those that are created by the deferred updates - * machinery during a collection. - * 2) green = 0. Means no caching. Can be a good way to minimize the - * amount of time spent updating rsets during a collection. - */ - size_t _green_zone; - size_t _yellow_zone; - size_t _red_zone; - size_t _min_yellow_zone_size; - - ConcurrentG1Refine(size_t green_zone, - size_t yellow_zone, - size_t red_zone, - size_t min_yellow_zone_size); - - // Update green/yellow/red zone values based on how well goals are being met. - void update_zones(double update_rs_time, - size_t update_rs_processed_buffers, - double goal_ms); - - // Update thread thresholds to account for updated zone values. - void update_thread_thresholds(); - - public: - ~ConcurrentG1Refine(); - - // Returns ConcurrentG1Refine instance if succeeded to create/initialize ConcurrentG1Refine and ConcurrentG1RefineThread. - // Otherwise, returns NULL with error code. - static ConcurrentG1Refine* create(jint* ecode); - - void stop(); - - void adjust(double update_rs_time, size_t update_rs_processed_buffers, double goal_ms); - - // Iterate over all concurrent refinement threads - void threads_do(ThreadClosure *tc); - - // Iterate over all worker refinement threads - void worker_threads_do(ThreadClosure * tc); - - // The RS sampling thread has nothing to do with refinement, but is here for now. - G1YoungRemSetSamplingThread * sampling_thread() const { return _sample_thread; } - - static uint thread_num(); - - void print_worker_threads_on(outputStream* st) const; - - size_t green_zone() const { return _green_zone; } - size_t yellow_zone() const { return _yellow_zone; } - size_t red_zone() const { return _red_zone; } -}; - -#endif // SHARE_VM_GC_G1_CONCURRENTG1REFINE_HPP
--- a/src/hotspot/share/gc/g1/concurrentG1RefineThread.cpp Mon Nov 06 14:31:22 2017 +0000 +++ /dev/null Thu Jan 01 00:00:00 1970 +0000 @@ -1,174 +0,0 @@ -/* - * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. - * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. - * - * This code is free software; you can redistribute it and/or modify it - * under the terms of the GNU General Public License version 2 only, as - * published by the Free Software Foundation. - * - * This code 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 - * version 2 for more details (a copy is included in the LICENSE file that - * accompanied this code). - * - * You should have received a copy of the GNU General Public License version - * 2 along with this work; if not, write to the Free Software Foundation, - * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. - * - * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA - * or visit www.oracle.com if you need additional information or have any - * questions. - * - */ - -#include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" -#include "gc/g1/concurrentG1RefineThread.hpp" -#include "gc/g1/g1CollectedHeap.inline.hpp" -#include "gc/g1/g1RemSet.hpp" -#include "gc/shared/suspendibleThreadSet.hpp" -#include "logging/log.hpp" -#include "memory/resourceArea.hpp" -#include "runtime/handles.inline.hpp" -#include "runtime/mutexLocker.hpp" - -ConcurrentG1RefineThread:: -ConcurrentG1RefineThread(ConcurrentG1Refine* cg1r, ConcurrentG1RefineThread *next, - uint worker_id_offset, uint worker_id, - size_t activate, size_t deactivate) : - ConcurrentGCThread(), - _worker_id_offset(worker_id_offset), - _worker_id(worker_id), - _active(false), - _next(next), - _monitor(NULL), - _cg1r(cg1r), - _vtime_accum(0.0), - _activation_threshold(activate), - _deactivation_threshold(deactivate) -{ - - // Each thread has its own monitor. The i-th thread is responsible for signaling - // to thread i+1 if the number of buffers in the queue exceeds a threshold for this - // thread. Monitors are also used to wake up the threads during termination. - // The 0th (primary) worker is notified by mutator threads and has a special monitor. - if (!is_primary()) { - _monitor = new Monitor(Mutex::nonleaf, "Refinement monitor", true, - Monitor::_safepoint_check_never); - } else { - _monitor = DirtyCardQ_CBL_mon; - } - - // set name - set_name("G1 Refine#%d", worker_id); - create_and_start(); -} - -void ConcurrentG1RefineThread::update_thresholds(size_t activate, - size_t deactivate) { - assert(deactivate < activate, "precondition"); - _activation_threshold = activate; - _deactivation_threshold = deactivate; -} - -void ConcurrentG1RefineThread::wait_for_completed_buffers() { - MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); - while (!should_terminate() && !is_active()) { - _monitor->wait(Mutex::_no_safepoint_check_flag); - } -} - -bool ConcurrentG1RefineThread::is_active() { - DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); - return is_primary() ? dcqs.process_completed_buffers() : _active; -} - -void ConcurrentG1RefineThread::activate() { - MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); - if (!is_primary()) { - set_active(true); - } else { - DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); - dcqs.set_process_completed(true); - } - _monitor->notify(); -} - -void ConcurrentG1RefineThread::deactivate() { - MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); - if (!is_primary()) { - set_active(false); - } else { - DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); - dcqs.set_process_completed(false); - } -} - -void ConcurrentG1RefineThread::run_service() { - _vtime_start = os::elapsedVTime(); - - while (!should_terminate()) { - // Wait for work - wait_for_completed_buffers(); - if (should_terminate()) { - break; - } - - size_t buffers_processed = 0; - DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); - log_debug(gc, refine)("Activated %d, on threshold: " SIZE_FORMAT ", current: " SIZE_FORMAT, - _worker_id, _activation_threshold, dcqs.completed_buffers_num()); - - { - SuspendibleThreadSetJoiner sts_join; - - while (!should_terminate()) { - if (sts_join.should_yield()) { - sts_join.yield(); - continue; // Re-check for termination after yield delay. - } - - size_t curr_buffer_num = dcqs.completed_buffers_num(); - // If the number of the buffers falls down into the yellow zone, - // that means that the transition period after the evacuation pause has ended. - if (dcqs.completed_queue_padding() > 0 && curr_buffer_num <= cg1r()->yellow_zone()) { - dcqs.set_completed_queue_padding(0); - } - - // Check if we need to activate the next thread. - if ((_next != NULL) && - !_next->is_active() && - (curr_buffer_num > _next->_activation_threshold)) { - _next->activate(); - } - - // Process the next buffer, if there are enough left. - if (!dcqs.refine_completed_buffer_concurrently(_worker_id + _worker_id_offset, _deactivation_threshold)) { - break; // Deactivate, number of buffers fell below threshold. - } - ++buffers_processed; - } - } - - deactivate(); - log_debug(gc, refine)("Deactivated %d, off threshold: " SIZE_FORMAT - ", current: " SIZE_FORMAT ", processed: " SIZE_FORMAT, - _worker_id, _deactivation_threshold, - dcqs.completed_buffers_num(), - buffers_processed); - - if (os::supports_vtime()) { - _vtime_accum = (os::elapsedVTime() - _vtime_start); - } else { - _vtime_accum = 0.0; - } - } - - log_debug(gc, refine)("Stopping %d", _worker_id); -} - -void ConcurrentG1RefineThread::stop_service() { - MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); - _monitor->notify(); -}
--- a/src/hotspot/share/gc/g1/concurrentG1RefineThread.hpp Mon Nov 06 14:31:22 2017 +0000 +++ /dev/null Thu Jan 01 00:00:00 1970 +0000 @@ -1,85 +0,0 @@ -/* - * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. - * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. - * - * This code is free software; you can redistribute it and/or modify it - * under the terms of the GNU General Public License version 2 only, as - * published by the Free Software Foundation. - * - * This code 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 - * version 2 for more details (a copy is included in the LICENSE file that - * accompanied this code). - * - * You should have received a copy of the GNU General Public License version - * 2 along with this work; if not, write to the Free Software Foundation, - * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. - * - * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA - * or visit www.oracle.com if you need additional information or have any - * questions. - * - */ - -#ifndef SHARE_VM_GC_G1_CONCURRENTG1REFINETHREAD_HPP -#define SHARE_VM_GC_G1_CONCURRENTG1REFINETHREAD_HPP - -#include "gc/g1/dirtyCardQueue.hpp" -#include "gc/shared/concurrentGCThread.hpp" - -// Forward Decl. -class CardTableEntryClosure; -class ConcurrentG1Refine; - -// One or more G1 Concurrent Refinement Threads may be active if concurrent -// refinement is in progress. -class ConcurrentG1RefineThread: public ConcurrentGCThread { - friend class VMStructs; - friend class G1CollectedHeap; - - double _vtime_start; // Initial virtual time. - double _vtime_accum; // Accumulated virtual time. - uint _worker_id; - uint _worker_id_offset; - - // The refinement threads collection is linked list. A predecessor can activate a successor - // when the number of the rset update buffer crosses a certain threshold. A successor - // would self-deactivate when the number of the buffers falls below the threshold. - bool _active; - ConcurrentG1RefineThread* _next; - Monitor* _monitor; - ConcurrentG1Refine* _cg1r; - - // This thread's activation/deactivation thresholds - size_t _activation_threshold; - size_t _deactivation_threshold; - - void wait_for_completed_buffers(); - - void set_active(bool x) { _active = x; } - bool is_active(); - void activate(); - void deactivate(); - - bool is_primary() { return (_worker_id == 0); } - - void run_service(); - void stop_service(); - -public: - // Constructor - ConcurrentG1RefineThread(ConcurrentG1Refine* cg1r, ConcurrentG1RefineThread* next, - uint worker_id_offset, uint worker_id, - size_t activate, size_t deactivate); - - void update_thresholds(size_t activate, size_t deactivate); - size_t activation_threshold() const { return _activation_threshold; } - - // Total virtual time so far. - double vtime_accum() { return _vtime_accum; } - - ConcurrentG1Refine* cg1r() { return _cg1r; } -}; - -#endif // SHARE_VM_GC_G1_CONCURRENTG1REFINETHREAD_HPP
--- a/src/hotspot/share/gc/g1/g1CollectedHeap.cpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/g1CollectedHeap.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -29,14 +29,14 @@ #include "code/codeCache.hpp" #include "code/icBuffer.hpp" #include "gc/g1/bufferingOopClosure.hpp" -#include "gc/g1/concurrentG1Refine.hpp" -#include "gc/g1/concurrentG1RefineThread.hpp" #include "gc/g1/concurrentMarkThread.inline.hpp" #include "gc/g1/g1Allocator.inline.hpp" #include "gc/g1/g1CollectedHeap.inline.hpp" #include "gc/g1/g1CollectionSet.hpp" #include "gc/g1/g1CollectorPolicy.hpp" #include "gc/g1/g1CollectorState.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" +#include "gc/g1/g1ConcurrentRefineThread.hpp" #include "gc/g1/g1EvacStats.inline.hpp" #include "gc/g1/g1FullGCScope.hpp" #include "gc/g1/g1GCPhaseTimes.hpp" @@ -54,6 +54,7 @@ #include "gc/g1/g1SerialFullCollector.hpp" #include "gc/g1/g1StringDedup.hpp" #include "gc/g1/g1YCTypes.hpp" +#include "gc/g1/g1YoungRemSetSamplingThread.hpp" #include "gc/g1/heapRegion.inline.hpp" #include "gc/g1/heapRegionRemSet.hpp" #include "gc/g1/heapRegionSet.inline.hpp" @@ -1541,6 +1542,7 @@ G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* collector_policy) : CollectedHeap(), + _young_gen_sampling_thread(NULL), _collector_policy(collector_policy), _gc_timer_stw(new (ResourceObj::C_HEAP, mtGC) STWGCTimer()), _gc_tracer_stw(new (ResourceObj::C_HEAP, mtGC) G1NewTracer()), @@ -1554,7 +1556,7 @@ _bot(NULL), _hot_card_cache(NULL), _g1_rem_set(NULL), - _cg1r(NULL), + _cr(NULL), _g1mm(NULL), _preserved_marks_set(true /* in_c_heap */), _secondary_free_list("Secondary Free List", new SecondaryFreeRegionListMtSafeChecker()), @@ -1633,10 +1635,19 @@ jint G1CollectedHeap::initialize_concurrent_refinement() { jint ecode = JNI_OK; - _cg1r = ConcurrentG1Refine::create(&ecode); + _cr = G1ConcurrentRefine::create(&ecode); return ecode; } +jint G1CollectedHeap::initialize_young_gen_sampling_thread() { + _young_gen_sampling_thread = new G1YoungRemSetSamplingThread(); + if (_young_gen_sampling_thread->osthread() == NULL) { + vm_shutdown_during_initialization("Could not create G1YoungRemSetSamplingThread"); + return JNI_ENOMEM; + } + return JNI_OK; +} + jint G1CollectedHeap::initialize() { CollectedHeap::pre_initialize(); os::enable_vtime(); @@ -1789,10 +1800,15 @@ return ecode; } + ecode = initialize_young_gen_sampling_thread(); + if (ecode != JNI_OK) { + return ecode; + } + JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, DirtyCardQ_FL_lock, - (int)concurrent_g1_refine()->yellow_zone(), - (int)concurrent_g1_refine()->red_zone(), + (int)concurrent_refine()->yellow_zone(), + (int)concurrent_refine()->red_zone(), Shared_DirtyCardQ_lock, NULL, // fl_owner true); // init_free_ids @@ -1836,7 +1852,8 @@ // Stop all concurrent threads. We do this to make sure these threads // do not continue to execute and access resources (e.g. logging) // that are destroyed during shutdown. - _cg1r->stop(); + _cr->stop(); + _young_gen_sampling_thread->stop(); _cmThread->stop(); if (G1StringDedup::is_enabled()) { G1StringDedup::stop(); @@ -2436,7 +2453,8 @@ _cmThread->print_on(st); st->cr(); _cm->print_worker_threads_on(st); - _cg1r->print_worker_threads_on(st); // also prints the sample thread + _cr->print_threads_on(st); + _young_gen_sampling_thread->print_on(st); if (G1StringDedup::is_enabled()) { G1StringDedup::print_worker_threads_on(st); } @@ -2446,7 +2464,8 @@ workers()->threads_do(tc); tc->do_thread(_cmThread); _cm->threads_do(tc); - _cg1r->threads_do(tc); // also iterates over the sample thread + _cr->threads_do(tc); + tc->do_thread(_young_gen_sampling_thread); if (G1StringDedup::is_enabled()) { G1StringDedup::threads_do(tc); }
--- a/src/hotspot/share/gc/g1/g1CollectedHeap.hpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/g1CollectedHeap.hpp Mon Nov 06 15:33:54 2017 +0100 @@ -73,10 +73,11 @@ class G1Policy; class G1HotCardCache; class G1RemSet; +class G1YoungRemSetSamplingThread; class HeapRegionRemSetIterator; class G1ConcurrentMark; class ConcurrentMarkThread; -class ConcurrentG1Refine; +class G1ConcurrentRefine; class GenerationCounters; class STWGCTimer; class G1NewTracer; @@ -142,6 +143,8 @@ friend class G1CheckCSetFastTableClosure; private: + G1YoungRemSetSamplingThread* _young_gen_sampling_thread; + WorkGang* _workers; G1CollectorPolicy* _collector_policy; @@ -553,6 +556,8 @@ // during GC into global variables. void merge_per_thread_state_info(G1ParScanThreadStateSet* per_thread_states); public: + G1YoungRemSetSamplingThread* sampling_thread() const { return _young_gen_sampling_thread; } + WorkGang* workers() const { return _workers; } G1Allocator* allocator() { @@ -806,7 +811,7 @@ ConcurrentMarkThread* _cmThread; // The concurrent refiner. - ConcurrentG1Refine* _cg1r; + G1ConcurrentRefine* _cr; // The parallel task queues RefToScanQueueSet *_task_queues; @@ -959,6 +964,7 @@ private: jint initialize_concurrent_refinement(); + jint initialize_young_gen_sampling_thread(); public: // Initialize the G1CollectedHeap to have the initial and // maximum sizes and remembered and barrier sets @@ -1389,7 +1395,7 @@ // Refinement - ConcurrentG1Refine* concurrent_g1_refine() const { return _cg1r; } + G1ConcurrentRefine* concurrent_refine() const { return _cr; } // Optimized nmethod scanning support routines
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/hotspot/share/gc/g1/g1ConcurrentRefine.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -0,0 +1,351 @@ +/* + * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code 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 + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#include "precompiled.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" +#include "gc/g1/g1ConcurrentRefineThread.hpp" +#include "logging/log.hpp" +#include "runtime/java.hpp" +#include "runtime/thread.hpp" +#include "utilities/debug.hpp" +#include "utilities/globalDefinitions.hpp" +#include "utilities/pair.hpp" +#include <math.h> + +// Arbitrary but large limits, to simplify some of the zone calculations. +// The general idea is to allow expressions like +// MIN2(x OP y, max_XXX_zone) +// without needing to check for overflow in "x OP y", because the +// ranges for x and y have been restricted. +STATIC_ASSERT(sizeof(LP64_ONLY(jint) NOT_LP64(jshort)) <= (sizeof(size_t)/2)); +const size_t max_yellow_zone = LP64_ONLY(max_jint) NOT_LP64(max_jshort); +const size_t max_green_zone = max_yellow_zone / 2; +const size_t max_red_zone = INT_MAX; // For dcqs.set_max_completed_queue. +STATIC_ASSERT(max_yellow_zone <= max_red_zone); + +// Range check assertions for green zone values. +#define assert_zone_constraints_g(green) \ + do { \ + size_t azc_g_green = (green); \ + assert(azc_g_green <= max_green_zone, \ + "green exceeds max: " SIZE_FORMAT, azc_g_green); \ + } while (0) + +// Range check assertions for green and yellow zone values. +#define assert_zone_constraints_gy(green, yellow) \ + do { \ + size_t azc_gy_green = (green); \ + size_t azc_gy_yellow = (yellow); \ + assert_zone_constraints_g(azc_gy_green); \ + assert(azc_gy_yellow <= max_yellow_zone, \ + "yellow exceeds max: " SIZE_FORMAT, azc_gy_yellow); \ + assert(azc_gy_green <= azc_gy_yellow, \ + "green (" SIZE_FORMAT ") exceeds yellow (" SIZE_FORMAT ")", \ + azc_gy_green, azc_gy_yellow); \ + } while (0) + +// Range check assertions for green, yellow, and red zone values. +#define assert_zone_constraints_gyr(green, yellow, red) \ + do { \ + size_t azc_gyr_green = (green); \ + size_t azc_gyr_yellow = (yellow); \ + size_t azc_gyr_red = (red); \ + assert_zone_constraints_gy(azc_gyr_green, azc_gyr_yellow); \ + assert(azc_gyr_red <= max_red_zone, \ + "red exceeds max: " SIZE_FORMAT, azc_gyr_red); \ + assert(azc_gyr_yellow <= azc_gyr_red, \ + "yellow (" SIZE_FORMAT ") exceeds red (" SIZE_FORMAT ")", \ + azc_gyr_yellow, azc_gyr_red); \ + } while (0) + +// Logging tag sequence for refinement control updates. +#define CTRL_TAGS gc, ergo, refine + +// For logging zone values, ensuring consistency of level and tags. +#define LOG_ZONES(...) log_debug( CTRL_TAGS )(__VA_ARGS__) + +// Package for pair of refinement thread activation and deactivation +// thresholds. The activation and deactivation levels are resp. the first +// and second values of the pair. +typedef Pair<size_t, size_t> Thresholds; +inline size_t activation_level(const Thresholds& t) { return t.first; } +inline size_t deactivation_level(const Thresholds& t) { return t.second; } + +static Thresholds calc_thresholds(size_t green_zone, + size_t yellow_zone, + uint worker_i) { + double yellow_size = yellow_zone - green_zone; + double step = yellow_size / G1ConcurrentRefine::thread_num(); + if (worker_i == 0) { + // Potentially activate worker 0 more aggressively, to keep + // available buffers near green_zone value. When yellow_size is + // large we don't want to allow a full step to accumulate before + // doing any processing, as that might lead to significantly more + // than green_zone buffers to be processed by update_rs. + step = MIN2(step, ParallelGCThreads / 2.0); + } + size_t activate_offset = static_cast<size_t>(ceil(step * (worker_i + 1))); + size_t deactivate_offset = static_cast<size_t>(floor(step * worker_i)); + return Thresholds(green_zone + activate_offset, + green_zone + deactivate_offset); +} + +G1ConcurrentRefine::G1ConcurrentRefine(size_t green_zone, + size_t yellow_zone, + size_t red_zone, + size_t min_yellow_zone_size) : + _threads(NULL), + _n_worker_threads(thread_num()), + _green_zone(green_zone), + _yellow_zone(yellow_zone), + _red_zone(red_zone), + _min_yellow_zone_size(min_yellow_zone_size) +{ + assert_zone_constraints_gyr(green_zone, yellow_zone, red_zone); +} + +static size_t calc_min_yellow_zone_size() { + size_t step = G1ConcRefinementThresholdStep; + uint n_workers = G1ConcurrentRefine::thread_num(); + if ((max_yellow_zone / step) < n_workers) { + return max_yellow_zone; + } else { + return step * n_workers; + } +} + +static size_t calc_init_green_zone() { + size_t green = G1ConcRefinementGreenZone; + if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) { + green = ParallelGCThreads; + } + return MIN2(green, max_green_zone); +} + +static size_t calc_init_yellow_zone(size_t green, size_t min_size) { + size_t config = G1ConcRefinementYellowZone; + size_t size = 0; + if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) { + size = green * 2; + } else if (green < config) { + size = config - green; + } + size = MAX2(size, min_size); + size = MIN2(size, max_yellow_zone); + return MIN2(green + size, max_yellow_zone); +} + +static size_t calc_init_red_zone(size_t green, size_t yellow) { + size_t size = yellow - green; + if (!FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) { + size_t config = G1ConcRefinementRedZone; + if (yellow < config) { + size = MAX2(size, config - yellow); + } + } + return MIN2(yellow + size, max_red_zone); +} + +G1ConcurrentRefine* G1ConcurrentRefine::create(jint* ecode) { + size_t min_yellow_zone_size = calc_min_yellow_zone_size(); + size_t green_zone = calc_init_green_zone(); + size_t yellow_zone = calc_init_yellow_zone(green_zone, min_yellow_zone_size); + size_t red_zone = calc_init_red_zone(green_zone, yellow_zone); + + LOG_ZONES("Initial Refinement Zones: " + "green: " SIZE_FORMAT ", " + "yellow: " SIZE_FORMAT ", " + "red: " SIZE_FORMAT ", " + "min yellow size: " SIZE_FORMAT, + green_zone, yellow_zone, red_zone, min_yellow_zone_size); + + G1ConcurrentRefine* cr = new G1ConcurrentRefine(green_zone, + yellow_zone, + red_zone, + min_yellow_zone_size); + + if (cr == NULL) { + *ecode = JNI_ENOMEM; + vm_shutdown_during_initialization("Could not create G1ConcurrentRefine"); + return NULL; + } + + cr->_threads = NEW_C_HEAP_ARRAY_RETURN_NULL(G1ConcurrentRefineThread*, cr->_n_worker_threads, mtGC); + if (cr->_threads == NULL) { + *ecode = JNI_ENOMEM; + vm_shutdown_during_initialization("Could not allocate an array for G1ConcurrentRefineThread"); + return NULL; + } + + uint worker_id_offset = DirtyCardQueueSet::num_par_ids(); + + G1ConcurrentRefineThread *next = NULL; + for (uint i = cr->_n_worker_threads - 1; i != UINT_MAX; i--) { + Thresholds thresholds = calc_thresholds(green_zone, yellow_zone, i); + G1ConcurrentRefineThread* t = + new G1ConcurrentRefineThread(cr, + next, + worker_id_offset, + i, + activation_level(thresholds), + deactivation_level(thresholds)); + assert(t != NULL, "Conc refine should have been created"); + if (t->osthread() == NULL) { + *ecode = JNI_ENOMEM; + vm_shutdown_during_initialization("Could not create G1ConcurrentRefineThread"); + return NULL; + } + + assert(t->cr() == cr, "Conc refine thread should refer to this"); + cr->_threads[i] = t; + next = t; + } + + *ecode = JNI_OK; + return cr; +} + +void G1ConcurrentRefine::stop() { + for (uint i = 0; i < _n_worker_threads; i++) { + _threads[i]->stop(); + } +} + +void G1ConcurrentRefine::update_thread_thresholds() { + for (uint i = 0; i < _n_worker_threads; i++) { + Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, i); + _threads[i]->update_thresholds(activation_level(thresholds), + deactivation_level(thresholds)); + } +} + +G1ConcurrentRefine::~G1ConcurrentRefine() { + for (uint i = 0; i < _n_worker_threads; i++) { + delete _threads[i]; + } + FREE_C_HEAP_ARRAY(G1ConcurrentRefineThread*, _threads); +} + +void G1ConcurrentRefine::threads_do(ThreadClosure *tc) { + for (uint i = 0; i < _n_worker_threads; i++) { + tc->do_thread(_threads[i]); + } +} + +uint G1ConcurrentRefine::thread_num() { + return G1ConcRefinementThreads; +} + +void G1ConcurrentRefine::print_threads_on(outputStream* st) const { + for (uint i = 0; i < _n_worker_threads; ++i) { + _threads[i]->print_on(st); + st->cr(); + } +} + +static size_t calc_new_green_zone(size_t green, + double update_rs_time, + size_t update_rs_processed_buffers, + double goal_ms) { + // Adjust green zone based on whether we're meeting the time goal. + // Limit to max_green_zone. + const double inc_k = 1.1, dec_k = 0.9; + if (update_rs_time > goal_ms) { + if (green > 0) { + green = static_cast<size_t>(green * dec_k); + } + } else if (update_rs_time < goal_ms && + update_rs_processed_buffers > green) { + green = static_cast<size_t>(MAX2(green * inc_k, green + 1.0)); + green = MIN2(green, max_green_zone); + } + return green; +} + +static size_t calc_new_yellow_zone(size_t green, size_t min_yellow_size) { + size_t size = green * 2; + size = MAX2(size, min_yellow_size); + return MIN2(green + size, max_yellow_zone); +} + +static size_t calc_new_red_zone(size_t green, size_t yellow) { + return MIN2(yellow + (yellow - green), max_red_zone); +} + +void G1ConcurrentRefine::update_zones(double update_rs_time, + size_t update_rs_processed_buffers, + double goal_ms) { + log_trace( CTRL_TAGS )("Updating Refinement Zones: " + "update_rs time: %.3fms, " + "update_rs buffers: " SIZE_FORMAT ", " + "update_rs goal time: %.3fms", + update_rs_time, + update_rs_processed_buffers, + goal_ms); + + _green_zone = calc_new_green_zone(_green_zone, + update_rs_time, + update_rs_processed_buffers, + goal_ms); + _yellow_zone = calc_new_yellow_zone(_green_zone, _min_yellow_zone_size); + _red_zone = calc_new_red_zone(_green_zone, _yellow_zone); + + assert_zone_constraints_gyr(_green_zone, _yellow_zone, _red_zone); + LOG_ZONES("Updated Refinement Zones: " + "green: " SIZE_FORMAT ", " + "yellow: " SIZE_FORMAT ", " + "red: " SIZE_FORMAT, + _green_zone, _yellow_zone, _red_zone); +} + +void G1ConcurrentRefine::adjust(double update_rs_time, + size_t update_rs_processed_buffers, + double goal_ms) { + DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); + + if (G1UseAdaptiveConcRefinement) { + update_zones(update_rs_time, update_rs_processed_buffers, goal_ms); + update_thread_thresholds(); + + // Change the barrier params + if (_n_worker_threads == 0) { + // Disable dcqs notification when there are no threads to notify. + dcqs.set_process_completed_threshold(INT_MAX); + } else { + // Worker 0 is the primary; wakeup is via dcqs notification. + STATIC_ASSERT(max_yellow_zone <= INT_MAX); + size_t activate = _threads[0]->activation_threshold(); + dcqs.set_process_completed_threshold((int)activate); + } + dcqs.set_max_completed_queue((int)red_zone()); + } + + size_t curr_queue_size = dcqs.completed_buffers_num(); + if (curr_queue_size >= yellow_zone()) { + dcqs.set_completed_queue_padding(curr_queue_size); + } else { + dcqs.set_completed_queue_padding(0); + } + dcqs.notify_if_necessary(); +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/hotspot/share/gc/g1/g1ConcurrentRefine.hpp Mon Nov 06 15:33:54 2017 +0100 @@ -0,0 +1,98 @@ +/* + * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code 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 + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#ifndef SHARE_VM_GC_G1_G1CONCURRENTREFINE_HPP +#define SHARE_VM_GC_G1_G1CONCURRENTREFINE_HPP + +#include "memory/allocation.hpp" +#include "utilities/globalDefinitions.hpp" + +// Forward decl +class CardTableEntryClosure; +class G1ConcurrentRefineThread; +class outputStream; +class ThreadClosure; + +class G1ConcurrentRefine : public CHeapObj<mtGC> { + G1ConcurrentRefineThread** _threads; + uint _n_worker_threads; + /* + * The value of the update buffer queue length falls into one of 3 zones: + * green, yellow, red. If the value is in [0, green) nothing is + * done, the buffers are left unprocessed to enable the caching effect of the + * dirtied cards. In the yellow zone [green, yellow) the concurrent refinement + * threads are gradually activated. In [yellow, red) all threads are + * running. If the length becomes red (max queue length) the mutators start + * processing the buffers. + * + * There are some interesting cases (when G1UseAdaptiveConcRefinement + * is turned off): + * 1) green = yellow = red = 0. In this case the mutator will process all + * buffers. Except for those that are created by the deferred updates + * machinery during a collection. + * 2) green = 0. Means no caching. Can be a good way to minimize the + * amount of time spent updating rsets during a collection. + */ + size_t _green_zone; + size_t _yellow_zone; + size_t _red_zone; + size_t _min_yellow_zone_size; + + G1ConcurrentRefine(size_t green_zone, + size_t yellow_zone, + size_t red_zone, + size_t min_yellow_zone_size); + + // Update green/yellow/red zone values based on how well goals are being met. + void update_zones(double update_rs_time, + size_t update_rs_processed_buffers, + double goal_ms); + + // Update thread thresholds to account for updated zone values. + void update_thread_thresholds(); + + public: + ~G1ConcurrentRefine(); + + // Returns a G1ConcurrentRefine instance if succeeded to create/initialize G1ConcurrentRefine and G1ConcurrentRefineThreads. + // Otherwise, returns NULL with error code. + static G1ConcurrentRefine* create(jint* ecode); + + void stop(); + + void adjust(double update_rs_time, size_t update_rs_processed_buffers, double goal_ms); + + // Iterate over all concurrent refinement threads applying the given closure. + void threads_do(ThreadClosure *tc); + + static uint thread_num(); + + void print_threads_on(outputStream* st) const; + + size_t green_zone() const { return _green_zone; } + size_t yellow_zone() const { return _yellow_zone; } + size_t red_zone() const { return _red_zone; } +}; + +#endif // SHARE_VM_GC_G1_G1CONCURRENTREFINE_HPP
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/hotspot/share/gc/g1/g1ConcurrentRefineThread.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -0,0 +1,176 @@ +/* + * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code 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 + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#include "precompiled.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" +#include "gc/g1/g1ConcurrentRefineThread.hpp" +#include "gc/g1/g1CollectedHeap.inline.hpp" +#include "gc/g1/g1RemSet.hpp" +#include "gc/shared/suspendibleThreadSet.hpp" +#include "logging/log.hpp" +#include "memory/resourceArea.hpp" +#include "runtime/handles.inline.hpp" +#include "runtime/mutexLocker.hpp" + +G1ConcurrentRefineThread::G1ConcurrentRefineThread(G1ConcurrentRefine* cr, + G1ConcurrentRefineThread *next, + uint worker_id_offset, + uint worker_id, + size_t activate, + size_t deactivate) : + ConcurrentGCThread(), + _worker_id_offset(worker_id_offset), + _worker_id(worker_id), + _active(false), + _next(next), + _monitor(NULL), + _cr(cr), + _vtime_accum(0.0), + _activation_threshold(activate), + _deactivation_threshold(deactivate) +{ + + // Each thread has its own monitor. The i-th thread is responsible for signaling + // to thread i+1 if the number of buffers in the queue exceeds a threshold for this + // thread. Monitors are also used to wake up the threads during termination. + // The 0th (primary) worker is notified by mutator threads and has a special monitor. + if (!is_primary()) { + _monitor = new Monitor(Mutex::nonleaf, "Refinement monitor", true, + Monitor::_safepoint_check_never); + } else { + _monitor = DirtyCardQ_CBL_mon; + } + + // set name + set_name("G1 Refine#%d", worker_id); + create_and_start(); +} + +void G1ConcurrentRefineThread::update_thresholds(size_t activate, + size_t deactivate) { + assert(deactivate < activate, "precondition"); + _activation_threshold = activate; + _deactivation_threshold = deactivate; +} + +void G1ConcurrentRefineThread::wait_for_completed_buffers() { + MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); + while (!should_terminate() && !is_active()) { + _monitor->wait(Mutex::_no_safepoint_check_flag); + } +} + +bool G1ConcurrentRefineThread::is_active() { + DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); + return is_primary() ? dcqs.process_completed_buffers() : _active; +} + +void G1ConcurrentRefineThread::activate() { + MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); + if (!is_primary()) { + set_active(true); + } else { + DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); + dcqs.set_process_completed(true); + } + _monitor->notify(); +} + +void G1ConcurrentRefineThread::deactivate() { + MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); + if (!is_primary()) { + set_active(false); + } else { + DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); + dcqs.set_process_completed(false); + } +} + +void G1ConcurrentRefineThread::run_service() { + _vtime_start = os::elapsedVTime(); + + while (!should_terminate()) { + // Wait for work + wait_for_completed_buffers(); + if (should_terminate()) { + break; + } + + size_t buffers_processed = 0; + DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); + log_debug(gc, refine)("Activated %d, on threshold: " SIZE_FORMAT ", current: " SIZE_FORMAT, + _worker_id, _activation_threshold, dcqs.completed_buffers_num()); + + { + SuspendibleThreadSetJoiner sts_join; + + while (!should_terminate()) { + if (sts_join.should_yield()) { + sts_join.yield(); + continue; // Re-check for termination after yield delay. + } + + size_t curr_buffer_num = dcqs.completed_buffers_num(); + // If the number of the buffers falls down into the yellow zone, + // that means that the transition period after the evacuation pause has ended. + if (dcqs.completed_queue_padding() > 0 && curr_buffer_num <= cr()->yellow_zone()) { + dcqs.set_completed_queue_padding(0); + } + + // Check if we need to activate the next thread. + if ((_next != NULL) && + !_next->is_active() && + (curr_buffer_num > _next->_activation_threshold)) { + _next->activate(); + } + + // Process the next buffer, if there are enough left. + if (!dcqs.refine_completed_buffer_concurrently(_worker_id + _worker_id_offset, _deactivation_threshold)) { + break; // Deactivate, number of buffers fell below threshold. + } + ++buffers_processed; + } + } + + deactivate(); + log_debug(gc, refine)("Deactivated %d, off threshold: " SIZE_FORMAT + ", current: " SIZE_FORMAT ", processed: " SIZE_FORMAT, + _worker_id, _deactivation_threshold, + dcqs.completed_buffers_num(), + buffers_processed); + + if (os::supports_vtime()) { + _vtime_accum = (os::elapsedVTime() - _vtime_start); + } else { + _vtime_accum = 0.0; + } + } + + log_debug(gc, refine)("Stopping %d", _worker_id); +} + +void G1ConcurrentRefineThread::stop_service() { + MutexLockerEx x(_monitor, Mutex::_no_safepoint_check_flag); + _monitor->notify(); +}
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/hotspot/share/gc/g1/g1ConcurrentRefineThread.hpp Mon Nov 06 15:33:54 2017 +0100 @@ -0,0 +1,85 @@ +/* + * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code 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 + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#ifndef SHARE_VM_GC_G1_G1CONCURRENTREFINETHREAD_HPP +#define SHARE_VM_GC_G1_G1CONCURRENTREFINETHREAD_HPP + +#include "gc/g1/dirtyCardQueue.hpp" +#include "gc/shared/concurrentGCThread.hpp" + +// Forward Decl. +class CardTableEntryClosure; +class G1ConcurrentRefine; + +// One or more G1 Concurrent Refinement Threads may be active if concurrent +// refinement is in progress. +class G1ConcurrentRefineThread: public ConcurrentGCThread { + friend class VMStructs; + friend class G1CollectedHeap; + + double _vtime_start; // Initial virtual time. + double _vtime_accum; // Accumulated virtual time. + uint _worker_id; + uint _worker_id_offset; + + // The refinement threads collection is linked list. A predecessor can activate a successor + // when the number of the rset update buffer crosses a certain threshold. A successor + // would self-deactivate when the number of the buffers falls below the threshold. + bool _active; + G1ConcurrentRefineThread* _next; + Monitor* _monitor; + G1ConcurrentRefine* _cr; + + // This thread's activation/deactivation thresholds + size_t _activation_threshold; + size_t _deactivation_threshold; + + void wait_for_completed_buffers(); + + void set_active(bool x) { _active = x; } + bool is_active(); + void activate(); + void deactivate(); + + bool is_primary() { return (_worker_id == 0); } + + void run_service(); + void stop_service(); + +public: + // Constructor + G1ConcurrentRefineThread(G1ConcurrentRefine* cr, G1ConcurrentRefineThread* next, + uint worker_id_offset, uint worker_id, + size_t activate, size_t deactivate); + + void update_thresholds(size_t activate, size_t deactivate); + size_t activation_threshold() const { return _activation_threshold; } + + // Total virtual time so far. + double vtime_accum() { return _vtime_accum; } + + G1ConcurrentRefine* cr() { return _cr; } +}; + +#endif // SHARE_VM_GC_G1_G1CONCURRENTREFINETHREAD_HPP
--- a/src/hotspot/share/gc/g1/g1DefaultPolicy.cpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/g1DefaultPolicy.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -23,12 +23,12 @@ */ #include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" #include "gc/g1/concurrentMarkThread.inline.hpp" #include "gc/g1/g1Analytics.hpp" #include "gc/g1/g1CollectedHeap.inline.hpp" #include "gc/g1/g1CollectionSet.hpp" #include "gc/g1/g1ConcurrentMark.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" #include "gc/g1/g1DefaultPolicy.hpp" #include "gc/g1/g1HotCardCache.hpp" #include "gc/g1/g1IHOPControl.hpp" @@ -745,7 +745,7 @@ } else { update_rs_time_goal_ms -= scan_hcc_time_ms; } - _g1->concurrent_g1_refine()->adjust(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, + _g1->concurrent_refine()->adjust(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms, phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS), update_rs_time_goal_ms);
--- a/src/hotspot/share/gc/g1/g1RemSet.cpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/g1RemSet.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -23,10 +23,10 @@ */ #include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" #include "gc/g1/dirtyCardQueue.hpp" #include "gc/g1/g1BlockOffsetTable.inline.hpp" #include "gc/g1/g1CollectedHeap.inline.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" #include "gc/g1/g1FromCardCache.hpp" #include "gc/g1/g1GCPhaseTimes.hpp" #include "gc/g1/g1HotCardCache.hpp" @@ -298,7 +298,7 @@ } uint G1RemSet::num_par_rem_sets() { - return MAX2(DirtyCardQueueSet::num_par_ids() + ConcurrentG1Refine::thread_num(), ParallelGCThreads); + return MAX2(DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::thread_num(), ParallelGCThreads); } void G1RemSet::initialize(size_t capacity, uint max_regions) {
--- a/src/hotspot/share/gc/g1/g1RemSetSummary.cpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/g1RemSetSummary.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -23,9 +23,9 @@ */ #include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" -#include "gc/g1/concurrentG1RefineThread.hpp" #include "gc/g1/g1CollectedHeap.inline.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" +#include "gc/g1/g1ConcurrentRefineThread.hpp" #include "gc/g1/g1RemSet.inline.hpp" #include "gc/g1/g1RemSetSummary.hpp" #include "gc/g1/g1YoungRemSetSamplingThread.hpp" @@ -45,7 +45,7 @@ } virtual void do_thread(Thread* t) { - ConcurrentG1RefineThread* crt = (ConcurrentG1RefineThread*) t; + G1ConcurrentRefineThread* crt = (G1ConcurrentRefineThread*) t; _summary->set_rs_thread_vtime(_counter, crt->vtime_accum()); _counter++; } @@ -59,12 +59,13 @@ _num_coarsenings = HeapRegionRemSet::n_coarsenings(); - ConcurrentG1Refine * cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); + G1CollectedHeap* g1h = G1CollectedHeap::heap(); + G1ConcurrentRefine* cg1r = g1h->concurrent_refine(); if (_rs_threads_vtimes != NULL) { GetRSThreadVTimeClosure p(this); - cg1r->worker_threads_do(&p); + cg1r->threads_do(&p); } - set_sampling_thread_vtime(cg1r->sampling_thread()->vtime_accum()); + set_sampling_thread_vtime(g1h->sampling_thread()->vtime_accum()); } void G1RemSetSummary::set_rs_thread_vtime(uint thread, double value) { @@ -85,7 +86,7 @@ _num_processed_buf_mutator(0), _num_processed_buf_rs_threads(0), _num_coarsenings(0), - _num_vtimes(ConcurrentG1Refine::thread_num()), + _num_vtimes(G1ConcurrentRefine::thread_num()), _rs_threads_vtimes(NEW_C_HEAP_ARRAY(double, _num_vtimes, mtGC)), _sampling_thread_vtime(0.0f) { @@ -98,7 +99,7 @@ _num_processed_buf_mutator(0), _num_processed_buf_rs_threads(0), _num_coarsenings(0), - _num_vtimes(ConcurrentG1Refine::thread_num()), + _num_vtimes(G1ConcurrentRefine::thread_num()), _rs_threads_vtimes(NEW_C_HEAP_ARRAY(double, _num_vtimes, mtGC)), _sampling_thread_vtime(0.0f) { update();
--- a/src/hotspot/share/gc/g1/heapRegionManager.cpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/heapRegionManager.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -23,8 +23,8 @@ */ #include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" #include "gc/g1/g1CollectedHeap.inline.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" #include "gc/g1/heapRegion.hpp" #include "gc/g1/heapRegionManager.inline.hpp" #include "gc/g1/heapRegionSet.inline.hpp"
--- a/src/hotspot/share/gc/g1/heapRegionRemSet.cpp Mon Nov 06 14:31:22 2017 +0000 +++ b/src/hotspot/share/gc/g1/heapRegionRemSet.cpp Mon Nov 06 15:33:54 2017 +0100 @@ -23,9 +23,9 @@ */ #include "precompiled.hpp" -#include "gc/g1/concurrentG1Refine.hpp" #include "gc/g1/g1BlockOffsetTable.inline.hpp" #include "gc/g1/g1CollectedHeap.inline.hpp" +#include "gc/g1/g1ConcurrentRefine.hpp" #include "gc/g1/g1CardLiveData.inline.hpp" #include "gc/g1/heapRegionManager.inline.hpp" #include "gc/g1/heapRegionRemSet.hpp"