Thu, 12 Jan 2012 00:06:47 -0800
6484965: G1: piggy-back liveness accounting phase on marking
Summary: Remove the separate counting phase of concurrent marking by tracking the amount of marked bytes and the cards spanned by marked objects in marking task/worker thread local data structures, which are updated as individual objects are marked.
Reviewed-by: brutisso, tonyp
1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
27 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
28 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
29 #include "gc_implementation/g1/g1MMUTracker.hpp"
30 #include "gc_implementation/g1/vm_operations_g1.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "runtime/vmThread.hpp"
34 // ======= Concurrent Mark Thread ========
36 // The CM thread is created when the G1 garbage collector is used
38 SurrogateLockerThread*
39 ConcurrentMarkThread::_slt = NULL;
41 ConcurrentMarkThread::ConcurrentMarkThread(ConcurrentMark* cm) :
42 ConcurrentGCThread(),
43 _cm(cm),
44 _started(false),
45 _in_progress(false),
46 _vtime_accum(0.0),
47 _vtime_mark_accum(0.0) {
48 create_and_start();
49 }
51 class CMCheckpointRootsFinalClosure: public VoidClosure {
53 ConcurrentMark* _cm;
54 public:
56 CMCheckpointRootsFinalClosure(ConcurrentMark* cm) :
57 _cm(cm) {}
59 void do_void(){
60 _cm->checkpointRootsFinal(false); // !clear_all_soft_refs
61 }
62 };
64 class CMCleanUp: public VoidClosure {
65 ConcurrentMark* _cm;
66 public:
68 CMCleanUp(ConcurrentMark* cm) :
69 _cm(cm) {}
71 void do_void(){
72 _cm->cleanup();
73 }
74 };
78 void ConcurrentMarkThread::run() {
79 initialize_in_thread();
80 _vtime_start = os::elapsedVTime();
81 wait_for_universe_init();
83 G1CollectedHeap* g1h = G1CollectedHeap::heap();
84 G1CollectorPolicy* g1_policy = g1h->g1_policy();
85 G1MMUTracker *mmu_tracker = g1_policy->mmu_tracker();
86 Thread *current_thread = Thread::current();
88 while (!_should_terminate) {
89 // wait until started is set.
90 sleepBeforeNextCycle();
91 {
92 ResourceMark rm;
93 HandleMark hm;
94 double cycle_start = os::elapsedVTime();
95 double mark_start_sec = os::elapsedTime();
96 char verbose_str[128];
98 if (PrintGC) {
99 gclog_or_tty->date_stamp(PrintGCDateStamps);
100 gclog_or_tty->stamp(PrintGCTimeStamps);
101 gclog_or_tty->print_cr("[GC concurrent-mark-start]");
102 }
104 int iter = 0;
105 do {
106 iter++;
107 if (!cm()->has_aborted()) {
108 _cm->markFromRoots();
109 }
111 double mark_end_time = os::elapsedVTime();
112 double mark_end_sec = os::elapsedTime();
113 _vtime_mark_accum += (mark_end_time - cycle_start);
114 if (!cm()->has_aborted()) {
115 if (g1_policy->adaptive_young_list_length()) {
116 double now = os::elapsedTime();
117 double remark_prediction_ms = g1_policy->predict_remark_time_ms();
118 jlong sleep_time_ms = mmu_tracker->when_ms(now, remark_prediction_ms);
119 os::sleep(current_thread, sleep_time_ms, false);
120 }
122 if (PrintGC) {
123 gclog_or_tty->date_stamp(PrintGCDateStamps);
124 gclog_or_tty->stamp(PrintGCTimeStamps);
125 gclog_or_tty->print_cr("[GC concurrent-mark-end, %1.7lf sec]",
126 mark_end_sec - mark_start_sec);
127 }
129 CMCheckpointRootsFinalClosure final_cl(_cm);
130 sprintf(verbose_str, "GC remark");
131 VM_CGC_Operation op(&final_cl, verbose_str);
132 VMThread::execute(&op);
133 }
134 if (cm()->restart_for_overflow() &&
135 G1TraceMarkStackOverflow) {
136 gclog_or_tty->print_cr("Restarting conc marking because of MS overflow "
137 "in remark (restart #%d).", iter);
138 }
140 if (cm()->restart_for_overflow()) {
141 if (PrintGC) {
142 gclog_or_tty->date_stamp(PrintGCDateStamps);
143 gclog_or_tty->stamp(PrintGCTimeStamps);
144 gclog_or_tty->print_cr("[GC concurrent-mark-restart-for-overflow]");
145 }
146 }
147 } while (cm()->restart_for_overflow());
149 double end_time = os::elapsedVTime();
150 // Update the total virtual time before doing this, since it will try
151 // to measure it to get the vtime for this marking. We purposely
152 // neglect the presumably-short "completeCleanup" phase here.
153 _vtime_accum = (end_time - _vtime_start);
155 if (!cm()->has_aborted()) {
156 if (g1_policy->adaptive_young_list_length()) {
157 double now = os::elapsedTime();
158 double cleanup_prediction_ms = g1_policy->predict_cleanup_time_ms();
159 jlong sleep_time_ms = mmu_tracker->when_ms(now, cleanup_prediction_ms);
160 os::sleep(current_thread, sleep_time_ms, false);
161 }
163 CMCleanUp cl_cl(_cm);
164 sprintf(verbose_str, "GC cleanup");
165 VM_CGC_Operation op(&cl_cl, verbose_str);
166 VMThread::execute(&op);
167 } else {
168 // We don't want to update the marking status if a GC pause
169 // is already underway.
170 _sts.join();
171 g1h->set_marking_complete();
172 _sts.leave();
173 }
175 // Check if cleanup set the free_regions_coming flag. If it
176 // hasn't, we can just skip the next step.
177 if (g1h->free_regions_coming()) {
178 // The following will finish freeing up any regions that we
179 // found to be empty during cleanup. We'll do this part
180 // without joining the suspendible set. If an evacuation pause
181 // takes place, then we would carry on freeing regions in
182 // case they are needed by the pause. If a Full GC takes
183 // place, it would wait for us to process the regions
184 // reclaimed by cleanup.
186 double cleanup_start_sec = os::elapsedTime();
187 if (PrintGC) {
188 gclog_or_tty->date_stamp(PrintGCDateStamps);
189 gclog_or_tty->stamp(PrintGCTimeStamps);
190 gclog_or_tty->print_cr("[GC concurrent-cleanup-start]");
191 }
193 // Now do the concurrent cleanup operation.
194 _cm->completeCleanup();
196 // Notify anyone who's waiting that there are no more free
197 // regions coming. We have to do this before we join the STS
198 // (in fact, we should not attempt to join the STS in the
199 // interval between finishing the cleanup pause and clearing
200 // the free_regions_coming flag) otherwise we might deadlock:
201 // a GC worker could be blocked waiting for the notification
202 // whereas this thread will be blocked for the pause to finish
203 // while it's trying to join the STS, which is conditional on
204 // the GC workers finishing.
205 g1h->reset_free_regions_coming();
207 double cleanup_end_sec = os::elapsedTime();
208 if (PrintGC) {
209 gclog_or_tty->date_stamp(PrintGCDateStamps);
210 gclog_or_tty->stamp(PrintGCTimeStamps);
211 gclog_or_tty->print_cr("[GC concurrent-cleanup-end, %1.7lf]",
212 cleanup_end_sec - cleanup_start_sec);
213 }
214 }
215 guarantee(cm()->cleanup_list_is_empty(),
216 "at this point there should be no regions on the cleanup list");
218 // There is a tricky race before recording that the concurrent
219 // cleanup has completed and a potential Full GC starting around
220 // the same time. We want to make sure that the Full GC calls
221 // abort() on concurrent mark after
222 // record_concurrent_mark_cleanup_completed(), since abort() is
223 // the method that will reset the concurrent mark state. If we
224 // end up calling record_concurrent_mark_cleanup_completed()
225 // after abort() then we might incorrectly undo some of the work
226 // abort() did. Checking the has_aborted() flag after joining
227 // the STS allows the correct ordering of the two methods. There
228 // are two scenarios:
229 //
230 // a) If we reach here before the Full GC, the fact that we have
231 // joined the STS means that the Full GC cannot start until we
232 // leave the STS, so record_concurrent_mark_cleanup_completed()
233 // will complete before abort() is called.
234 //
235 // b) If we reach here during the Full GC, we'll be held up from
236 // joining the STS until the Full GC is done, which means that
237 // abort() will have completed and has_aborted() will return
238 // true to prevent us from calling
239 // record_concurrent_mark_cleanup_completed() (and, in fact, it's
240 // not needed any more as the concurrent mark state has been
241 // already reset).
242 _sts.join();
243 if (!cm()->has_aborted()) {
244 g1_policy->record_concurrent_mark_cleanup_completed();
245 }
246 _sts.leave();
248 if (cm()->has_aborted()) {
249 if (PrintGC) {
250 gclog_or_tty->date_stamp(PrintGCDateStamps);
251 gclog_or_tty->stamp(PrintGCTimeStamps);
252 gclog_or_tty->print_cr("[GC concurrent-mark-abort]");
253 }
254 }
256 // We now want to allow clearing of the marking bitmap to be
257 // suspended by a collection pause.
258 _sts.join();
259 _cm->clearNextBitmap();
260 _sts.leave();
261 }
263 // Update the number of full collections that have been
264 // completed. This will also notify the FullGCCount_lock in case a
265 // Java thread is waiting for a full GC to happen (e.g., it
266 // called System.gc() with +ExplicitGCInvokesConcurrent).
267 _sts.join();
268 g1h->increment_full_collections_completed(true /* concurrent */);
269 _sts.leave();
270 }
271 assert(_should_terminate, "just checking");
273 terminate();
274 }
277 void ConcurrentMarkThread::yield() {
278 _sts.yield("Concurrent Mark");
279 }
281 void ConcurrentMarkThread::stop() {
282 // it is ok to take late safepoints here, if needed
283 MutexLockerEx mu(Terminator_lock);
284 _should_terminate = true;
285 while (!_has_terminated) {
286 Terminator_lock->wait();
287 }
288 }
290 void ConcurrentMarkThread::print() const {
291 print_on(tty);
292 }
294 void ConcurrentMarkThread::print_on(outputStream* st) const {
295 st->print("\"G1 Main Concurrent Mark GC Thread\" ");
296 Thread::print_on(st);
297 st->cr();
298 }
300 void ConcurrentMarkThread::sleepBeforeNextCycle() {
301 // We join here because we don't want to do the "shouldConcurrentMark()"
302 // below while the world is otherwise stopped.
303 assert(!in_progress(), "should have been cleared");
305 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
306 while (!started()) {
307 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
308 }
309 set_in_progress();
310 clear_started();
311 }
313 // Note: As is the case with CMS - this method, although exported
314 // by the ConcurrentMarkThread, which is a non-JavaThread, can only
315 // be called by a JavaThread. Currently this is done at vm creation
316 // time (post-vm-init) by the main/Primordial (Java)Thread.
317 // XXX Consider changing this in the future to allow the CM thread
318 // itself to create this thread?
319 void ConcurrentMarkThread::makeSurrogateLockerThread(TRAPS) {
320 assert(UseG1GC, "SLT thread needed only for concurrent GC");
321 assert(THREAD->is_Java_thread(), "must be a Java thread");
322 assert(_slt == NULL, "SLT already created");
323 _slt = SurrogateLockerThread::make(THREAD);
324 }