Mon, 24 Mar 2014 15:30:14 +0100
8035406: Improve data structure for Code Cache remembered sets
Summary: Change the code cache remembered sets data structure from a GrowableArray to a chunked list of nmethods. This makes the data structure more amenable to parallelization, and decreases freeing time.
Reviewed-by: mgerdin, brutisso
1 /*
2 * Copyright (c) 2001, 2013, 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/g1Log.hpp"
30 #include "gc_implementation/g1/g1MMUTracker.hpp"
31 #include "gc_implementation/g1/vm_operations_g1.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "runtime/vmThread.hpp"
35 // ======= Concurrent Mark Thread ========
37 // The CM thread is created when the G1 garbage collector is used
39 SurrogateLockerThread*
40 ConcurrentMarkThread::_slt = NULL;
42 ConcurrentMarkThread::ConcurrentMarkThread(ConcurrentMark* cm) :
43 ConcurrentGCThread(),
44 _cm(cm),
45 _started(false),
46 _in_progress(false),
47 _vtime_accum(0.0),
48 _vtime_mark_accum(0.0) {
49 create_and_start();
50 }
52 class CMCheckpointRootsFinalClosure: public VoidClosure {
54 ConcurrentMark* _cm;
55 public:
57 CMCheckpointRootsFinalClosure(ConcurrentMark* cm) :
58 _cm(cm) {}
60 void do_void(){
61 _cm->checkpointRootsFinal(false); // !clear_all_soft_refs
62 }
63 };
65 class CMCleanUp: public VoidClosure {
66 ConcurrentMark* _cm;
67 public:
69 CMCleanUp(ConcurrentMark* cm) :
70 _cm(cm) {}
72 void do_void(){
73 _cm->cleanup();
74 }
75 };
79 void ConcurrentMarkThread::run() {
80 initialize_in_thread();
81 _vtime_start = os::elapsedVTime();
82 wait_for_universe_init();
84 G1CollectedHeap* g1h = G1CollectedHeap::heap();
85 G1CollectorPolicy* g1_policy = g1h->g1_policy();
86 G1MMUTracker *mmu_tracker = g1_policy->mmu_tracker();
87 Thread *current_thread = Thread::current();
89 while (!_should_terminate) {
90 // wait until started is set.
91 sleepBeforeNextCycle();
92 {
93 ResourceMark rm;
94 HandleMark hm;
95 double cycle_start = os::elapsedVTime();
97 // We have to ensure that we finish scanning the root regions
98 // before the next GC takes place. To ensure this we have to
99 // make sure that we do not join the STS until the root regions
100 // have been scanned. If we did then it's possible that a
101 // subsequent GC could block us from joining the STS and proceed
102 // without the root regions have been scanned which would be a
103 // correctness issue.
105 double scan_start = os::elapsedTime();
106 if (!cm()->has_aborted()) {
107 if (G1Log::fine()) {
108 gclog_or_tty->date_stamp(PrintGCDateStamps);
109 gclog_or_tty->stamp(PrintGCTimeStamps);
110 gclog_or_tty->print_cr("[GC concurrent-root-region-scan-start]");
111 }
113 _cm->scanRootRegions();
115 double scan_end = os::elapsedTime();
116 if (G1Log::fine()) {
117 gclog_or_tty->date_stamp(PrintGCDateStamps);
118 gclog_or_tty->stamp(PrintGCTimeStamps);
119 gclog_or_tty->print_cr("[GC concurrent-root-region-scan-end, %1.7lf secs]",
120 scan_end - scan_start);
121 }
122 }
124 double mark_start_sec = os::elapsedTime();
125 if (G1Log::fine()) {
126 gclog_or_tty->date_stamp(PrintGCDateStamps);
127 gclog_or_tty->stamp(PrintGCTimeStamps);
128 gclog_or_tty->print_cr("[GC concurrent-mark-start]");
129 }
131 int iter = 0;
132 do {
133 iter++;
134 if (!cm()->has_aborted()) {
135 _cm->markFromRoots();
136 }
138 double mark_end_time = os::elapsedVTime();
139 double mark_end_sec = os::elapsedTime();
140 _vtime_mark_accum += (mark_end_time - cycle_start);
141 if (!cm()->has_aborted()) {
142 if (g1_policy->adaptive_young_list_length()) {
143 double now = os::elapsedTime();
144 double remark_prediction_ms = g1_policy->predict_remark_time_ms();
145 jlong sleep_time_ms = mmu_tracker->when_ms(now, remark_prediction_ms);
146 os::sleep(current_thread, sleep_time_ms, false);
147 }
149 if (G1Log::fine()) {
150 gclog_or_tty->date_stamp(PrintGCDateStamps);
151 gclog_or_tty->stamp(PrintGCTimeStamps);
152 gclog_or_tty->print_cr("[GC concurrent-mark-end, %1.7lf secs]",
153 mark_end_sec - mark_start_sec);
154 }
156 CMCheckpointRootsFinalClosure final_cl(_cm);
157 VM_CGC_Operation op(&final_cl, "GC remark", true /* needs_pll */);
158 VMThread::execute(&op);
159 }
160 if (cm()->restart_for_overflow()) {
161 if (G1TraceMarkStackOverflow) {
162 gclog_or_tty->print_cr("Restarting conc marking because of MS overflow "
163 "in remark (restart #%d).", iter);
164 }
165 if (G1Log::fine()) {
166 gclog_or_tty->date_stamp(PrintGCDateStamps);
167 gclog_or_tty->stamp(PrintGCTimeStamps);
168 gclog_or_tty->print_cr("[GC concurrent-mark-restart-for-overflow]");
169 }
170 }
171 } while (cm()->restart_for_overflow());
173 double end_time = os::elapsedVTime();
174 // Update the total virtual time before doing this, since it will try
175 // to measure it to get the vtime for this marking. We purposely
176 // neglect the presumably-short "completeCleanup" phase here.
177 _vtime_accum = (end_time - _vtime_start);
179 if (!cm()->has_aborted()) {
180 if (g1_policy->adaptive_young_list_length()) {
181 double now = os::elapsedTime();
182 double cleanup_prediction_ms = g1_policy->predict_cleanup_time_ms();
183 jlong sleep_time_ms = mmu_tracker->when_ms(now, cleanup_prediction_ms);
184 os::sleep(current_thread, sleep_time_ms, false);
185 }
187 CMCleanUp cl_cl(_cm);
188 VM_CGC_Operation op(&cl_cl, "GC cleanup", false /* needs_pll */);
189 VMThread::execute(&op);
190 } else {
191 // We don't want to update the marking status if a GC pause
192 // is already underway.
193 _sts.join();
194 g1h->set_marking_complete();
195 _sts.leave();
196 }
198 // Check if cleanup set the free_regions_coming flag. If it
199 // hasn't, we can just skip the next step.
200 if (g1h->free_regions_coming()) {
201 // The following will finish freeing up any regions that we
202 // found to be empty during cleanup. We'll do this part
203 // without joining the suspendible set. If an evacuation pause
204 // takes place, then we would carry on freeing regions in
205 // case they are needed by the pause. If a Full GC takes
206 // place, it would wait for us to process the regions
207 // reclaimed by cleanup.
209 double cleanup_start_sec = os::elapsedTime();
210 if (G1Log::fine()) {
211 gclog_or_tty->date_stamp(PrintGCDateStamps);
212 gclog_or_tty->stamp(PrintGCTimeStamps);
213 gclog_or_tty->print_cr("[GC concurrent-cleanup-start]");
214 }
216 // Now do the concurrent cleanup operation.
217 _cm->completeCleanup();
219 // Notify anyone who's waiting that there are no more free
220 // regions coming. We have to do this before we join the STS
221 // (in fact, we should not attempt to join the STS in the
222 // interval between finishing the cleanup pause and clearing
223 // the free_regions_coming flag) otherwise we might deadlock:
224 // a GC worker could be blocked waiting for the notification
225 // whereas this thread will be blocked for the pause to finish
226 // while it's trying to join the STS, which is conditional on
227 // the GC workers finishing.
228 g1h->reset_free_regions_coming();
230 double cleanup_end_sec = os::elapsedTime();
231 if (G1Log::fine()) {
232 gclog_or_tty->date_stamp(PrintGCDateStamps);
233 gclog_or_tty->stamp(PrintGCTimeStamps);
234 gclog_or_tty->print_cr("[GC concurrent-cleanup-end, %1.7lf secs]",
235 cleanup_end_sec - cleanup_start_sec);
236 }
237 }
238 guarantee(cm()->cleanup_list_is_empty(),
239 "at this point there should be no regions on the cleanup list");
241 // There is a tricky race before recording that the concurrent
242 // cleanup has completed and a potential Full GC starting around
243 // the same time. We want to make sure that the Full GC calls
244 // abort() on concurrent mark after
245 // record_concurrent_mark_cleanup_completed(), since abort() is
246 // the method that will reset the concurrent mark state. If we
247 // end up calling record_concurrent_mark_cleanup_completed()
248 // after abort() then we might incorrectly undo some of the work
249 // abort() did. Checking the has_aborted() flag after joining
250 // the STS allows the correct ordering of the two methods. There
251 // are two scenarios:
252 //
253 // a) If we reach here before the Full GC, the fact that we have
254 // joined the STS means that the Full GC cannot start until we
255 // leave the STS, so record_concurrent_mark_cleanup_completed()
256 // will complete before abort() is called.
257 //
258 // b) If we reach here during the Full GC, we'll be held up from
259 // joining the STS until the Full GC is done, which means that
260 // abort() will have completed and has_aborted() will return
261 // true to prevent us from calling
262 // record_concurrent_mark_cleanup_completed() (and, in fact, it's
263 // not needed any more as the concurrent mark state has been
264 // already reset).
265 _sts.join();
266 if (!cm()->has_aborted()) {
267 g1_policy->record_concurrent_mark_cleanup_completed();
268 }
269 _sts.leave();
271 if (cm()->has_aborted()) {
272 if (G1Log::fine()) {
273 gclog_or_tty->date_stamp(PrintGCDateStamps);
274 gclog_or_tty->stamp(PrintGCTimeStamps);
275 gclog_or_tty->print_cr("[GC concurrent-mark-abort]");
276 }
277 }
279 // We now want to allow clearing of the marking bitmap to be
280 // suspended by a collection pause.
281 _sts.join();
282 _cm->clearNextBitmap();
283 _sts.leave();
284 }
286 // Update the number of full collections that have been
287 // completed. This will also notify the FullGCCount_lock in case a
288 // Java thread is waiting for a full GC to happen (e.g., it
289 // called System.gc() with +ExplicitGCInvokesConcurrent).
290 _sts.join();
291 g1h->increment_old_marking_cycles_completed(true /* concurrent */);
292 g1h->register_concurrent_cycle_end();
293 _sts.leave();
294 }
295 assert(_should_terminate, "just checking");
297 terminate();
298 }
301 void ConcurrentMarkThread::yield() {
302 _sts.yield("Concurrent Mark");
303 }
305 void ConcurrentMarkThread::stop() {
306 // it is ok to take late safepoints here, if needed
307 MutexLockerEx mu(Terminator_lock);
308 _should_terminate = true;
309 while (!_has_terminated) {
310 Terminator_lock->wait();
311 }
312 }
314 void ConcurrentMarkThread::print() const {
315 print_on(tty);
316 }
318 void ConcurrentMarkThread::print_on(outputStream* st) const {
319 st->print("\"G1 Main Concurrent Mark GC Thread\" ");
320 Thread::print_on(st);
321 st->cr();
322 }
324 void ConcurrentMarkThread::sleepBeforeNextCycle() {
325 // We join here because we don't want to do the "shouldConcurrentMark()"
326 // below while the world is otherwise stopped.
327 assert(!in_progress(), "should have been cleared");
329 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
330 while (!started()) {
331 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
332 }
333 set_in_progress();
334 clear_started();
335 }
337 // Note: As is the case with CMS - this method, although exported
338 // by the ConcurrentMarkThread, which is a non-JavaThread, can only
339 // be called by a JavaThread. Currently this is done at vm creation
340 // time (post-vm-init) by the main/Primordial (Java)Thread.
341 // XXX Consider changing this in the future to allow the CM thread
342 // itself to create this thread?
343 void ConcurrentMarkThread::makeSurrogateLockerThread(TRAPS) {
344 assert(UseG1GC, "SLT thread needed only for concurrent GC");
345 assert(THREAD->is_Java_thread(), "must be a Java thread");
346 assert(_slt == NULL, "SLT already created");
347 _slt = SurrogateLockerThread::make(THREAD);
348 }