1 /*
2 * Copyright (c) 2021, 2024, 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 */
24
25 #include "precompiled.hpp"
26
27 #include "classfile/classLoaderDataGraph.inline.hpp"
28 #include "classfile/javaClasses.inline.hpp"
29 #include "compiler/oopMap.hpp"
30 #include "gc/g1/g1Allocator.hpp"
31 #include "gc/g1/g1CardSetMemory.hpp"
32 #include "gc/g1/g1CollectedHeap.inline.hpp"
33 #include "gc/g1/g1CollectionSetCandidates.inline.hpp"
34 #include "gc/g1/g1CollectorState.hpp"
35 #include "gc/g1/g1ConcurrentMark.hpp"
36 #include "gc/g1/g1GCPhaseTimes.hpp"
37 #include "gc/g1/g1EvacFailureRegions.inline.hpp"
38 #include "gc/g1/g1EvacInfo.hpp"
39 #include "gc/g1/g1HeapRegionPrinter.hpp"
40 #include "gc/g1/g1MonitoringSupport.hpp"
41 #include "gc/g1/g1ParScanThreadState.inline.hpp"
42 #include "gc/g1/g1Policy.hpp"
43 #include "gc/g1/g1RedirtyCardsQueue.hpp"
44 #include "gc/g1/g1RegionPinCache.inline.hpp"
45 #include "gc/g1/g1RemSet.hpp"
46 #include "gc/g1/g1RootProcessor.hpp"
47 #include "gc/g1/g1Trace.hpp"
48 #include "gc/g1/g1YoungCollector.hpp"
49 #include "gc/g1/g1YoungGCAllocationFailureInjector.hpp"
50 #include "gc/g1/g1YoungGCPostEvacuateTasks.hpp"
51 #include "gc/g1/g1YoungGCPreEvacuateTasks.hpp"
52 #include "gc/shared/concurrentGCBreakpoints.hpp"
53 #include "gc/shared/gcTraceTime.inline.hpp"
54 #include "gc/shared/gcTimer.hpp"
55 #include "gc/shared/gc_globals.hpp"
56 #include "gc/shared/referenceProcessor.hpp"
57 #include "gc/shared/weakProcessor.inline.hpp"
58 #include "gc/shared/workerPolicy.hpp"
59 #include "gc/shared/workerThread.hpp"
60 #include "jfr/jfrEvents.hpp"
61 #include "memory/resourceArea.hpp"
62 #include "runtime/threads.hpp"
63 #include "utilities/ticks.hpp"
64
65 // GCTraceTime wrapper that constructs the message according to GC pause type and
66 // GC cause.
67 // The code relies on the fact that GCTraceTimeWrapper stores the string passed
68 // initially as a reference only, so that we can modify it as needed.
69 class G1YoungGCTraceTime {
70 G1YoungCollector* _collector;
71
72 G1GCPauseType _pause_type;
73 GCCause::Cause _pause_cause;
74
75 static const uint MaxYoungGCNameLength = 128;
76 char _young_gc_name_data[MaxYoungGCNameLength];
77
78 GCTraceTime(Info, gc) _tt;
79
80 const char* update_young_gc_name() {
81 char evacuation_failed_string[48];
82 evacuation_failed_string[0] = '\0';
83
84 if (_collector->evacuation_failed()) {
85 snprintf(evacuation_failed_string,
86 ARRAY_SIZE(evacuation_failed_string),
87 " (Evacuation Failure: %s%s%s)",
88 _collector->evacuation_alloc_failed() ? "Allocation" : "",
89 _collector->evacuation_alloc_failed() && _collector->evacuation_pinned() ? " / " : "",
90 _collector->evacuation_pinned() ? "Pinned" : "");
91 }
92 snprintf(_young_gc_name_data,
93 MaxYoungGCNameLength,
94 "Pause Young (%s) (%s)%s",
95 G1GCPauseTypeHelper::to_string(_pause_type),
96 GCCause::to_string(_pause_cause),
97 evacuation_failed_string);
98 return _young_gc_name_data;
99 }
100
101 public:
102 G1YoungGCTraceTime(G1YoungCollector* collector, GCCause::Cause cause) :
103 _collector(collector),
104 // Take snapshot of current pause type at start as it may be modified during gc.
105 // The strings for all Concurrent Start pauses are the same, so the parameter
106 // does not matter here.
107 _pause_type(_collector->collector_state()->young_gc_pause_type(false /* concurrent_operation_is_full_mark */)),
108 _pause_cause(cause),
109 // Fake a "no cause" and manually add the correct string in update_young_gc_name()
110 // to make the string look more natural.
111 _tt(update_young_gc_name(), nullptr, GCCause::_no_gc, true) {
112 }
113
114 ~G1YoungGCTraceTime() {
115 update_young_gc_name();
116 }
117 };
118
119 class G1YoungGCNotifyPauseMark : public StackObj {
120 G1YoungCollector* _collector;
121
122 public:
123 G1YoungGCNotifyPauseMark(G1YoungCollector* collector) : _collector(collector) {
124 G1CollectedHeap::heap()->policy()->record_young_gc_pause_start();
125 }
126
127 ~G1YoungGCNotifyPauseMark() {
128 G1CollectedHeap::heap()->policy()->record_young_gc_pause_end(_collector->evacuation_failed());
129 }
130 };
131
132 class G1YoungGCJFRTracerMark : public G1JFRTracerMark {
133 G1EvacInfo _evacuation_info;
134
135 G1NewTracer* tracer() const { return (G1NewTracer*)_tracer; }
136
137 public:
138
139 G1EvacInfo* evacuation_info() { return &_evacuation_info; }
140
141 G1YoungGCJFRTracerMark(STWGCTimer* gc_timer_stw, G1NewTracer* gc_tracer_stw, GCCause::Cause cause) :
142 G1JFRTracerMark(gc_timer_stw, gc_tracer_stw), _evacuation_info() { }
143
144 void report_pause_type(G1GCPauseType type) {
145 tracer()->report_young_gc_pause(type);
146 }
147
148 ~G1YoungGCJFRTracerMark() {
149 G1CollectedHeap* g1h = G1CollectedHeap::heap();
150
151 tracer()->report_evacuation_info(&_evacuation_info);
152 tracer()->report_tenuring_threshold(g1h->policy()->tenuring_threshold());
153 }
154 };
155
156 class G1YoungGCVerifierMark : public StackObj {
157 G1YoungCollector* _collector;
158 G1HeapVerifier::G1VerifyType _type;
159
160 static G1HeapVerifier::G1VerifyType young_collection_verify_type() {
161 G1CollectorState* state = G1CollectedHeap::heap()->collector_state();
162 if (state->in_concurrent_start_gc()) {
163 return G1HeapVerifier::G1VerifyConcurrentStart;
164 } else if (state->in_young_only_phase()) {
165 return G1HeapVerifier::G1VerifyYoungNormal;
166 } else {
167 return G1HeapVerifier::G1VerifyMixed;
168 }
169 }
170
171 public:
172 G1YoungGCVerifierMark(G1YoungCollector* collector) : _collector(collector), _type(young_collection_verify_type()) {
173 G1CollectedHeap::heap()->verify_before_young_collection(_type);
174 }
175
176 ~G1YoungGCVerifierMark() {
177 // Inject evacuation failure tag into type if needed.
178 G1HeapVerifier::G1VerifyType type = _type;
179 if (_collector->evacuation_failed()) {
180 type = (G1HeapVerifier::G1VerifyType)(type | G1HeapVerifier::G1VerifyYoungEvacFail);
181 }
182 G1CollectedHeap::heap()->verify_after_young_collection(type);
183 }
184 };
185
186 G1Allocator* G1YoungCollector::allocator() const {
187 return _g1h->allocator();
188 }
189
190 G1CollectionSet* G1YoungCollector::collection_set() const {
191 return _g1h->collection_set();
192 }
193
194 G1CollectorState* G1YoungCollector::collector_state() const {
195 return _g1h->collector_state();
196 }
197
198 G1ConcurrentMark* G1YoungCollector::concurrent_mark() const {
199 return _g1h->concurrent_mark();
200 }
201
202 STWGCTimer* G1YoungCollector::gc_timer_stw() const {
203 return _g1h->gc_timer_stw();
204 }
205
206 G1NewTracer* G1YoungCollector::gc_tracer_stw() const {
207 return _g1h->gc_tracer_stw();
208 }
209
210 G1Policy* G1YoungCollector::policy() const {
211 return _g1h->policy();
212 }
213
214 G1GCPhaseTimes* G1YoungCollector::phase_times() const {
215 return _g1h->phase_times();
216 }
217
218 G1MonitoringSupport* G1YoungCollector::monitoring_support() const {
219 return _g1h->monitoring_support();
220 }
221
222 G1RemSet* G1YoungCollector::rem_set() const {
223 return _g1h->rem_set();
224 }
225
226 G1ScannerTasksQueueSet* G1YoungCollector::task_queues() const {
227 return _g1h->task_queues();
228 }
229
230 G1SurvivorRegions* G1YoungCollector::survivor_regions() const {
231 return _g1h->survivor();
232 }
233
234 ReferenceProcessor* G1YoungCollector::ref_processor_stw() const {
235 return _g1h->ref_processor_stw();
236 }
237
238 WorkerThreads* G1YoungCollector::workers() const {
239 return _g1h->workers();
240 }
241
242 G1YoungGCAllocationFailureInjector* G1YoungCollector::allocation_failure_injector() const {
243 return _g1h->allocation_failure_injector();
244 }
245
246
247 void G1YoungCollector::wait_for_root_region_scanning() {
248 Ticks start = Ticks::now();
249 // We have to wait until the CM threads finish scanning the
250 // root regions as it's the only way to ensure that all the
251 // objects on them have been correctly scanned before we start
252 // moving them during the GC.
253 bool waited = concurrent_mark()->wait_until_root_region_scan_finished();
254 Tickspan wait_time;
255 if (waited) {
256 wait_time = (Ticks::now() - start);
257 }
258 phase_times()->record_root_region_scan_wait_time(wait_time.seconds() * MILLIUNITS);
259 }
260
261 class G1PrintCollectionSetClosure : public HeapRegionClosure {
262 public:
263 virtual bool do_heap_region(G1HeapRegion* r) {
264 G1HeapRegionPrinter::cset(r);
265 return false;
266 }
267 };
268
269 void G1YoungCollector::calculate_collection_set(G1EvacInfo* evacuation_info, double target_pause_time_ms) {
270 // Forget the current allocation region (we might even choose it to be part
271 // of the collection set!) before finalizing the collection set.
272 allocator()->release_mutator_alloc_regions();
273
274 collection_set()->finalize_initial_collection_set(target_pause_time_ms, survivor_regions());
275 evacuation_info->set_collection_set_regions(collection_set()->region_length() +
276 collection_set()->optional_region_length());
277
278 concurrent_mark()->verify_no_collection_set_oops();
279
280 if (G1HeapRegionPrinter::is_active()) {
281 G1PrintCollectionSetClosure cl;
282 collection_set()->iterate(&cl);
283 collection_set()->iterate_optional(&cl);
284 }
285 }
286
287 class G1PrepareEvacuationTask : public WorkerTask {
288 class G1PrepareRegionsClosure : public HeapRegionClosure {
289 G1CollectedHeap* _g1h;
290 G1PrepareEvacuationTask* _parent_task;
291 uint _worker_humongous_total;
292 uint _worker_humongous_candidates;
293
294 G1MonotonicArenaMemoryStats _card_set_stats;
295
296 void sample_card_set_size(G1HeapRegion* hr) {
297 // Sample card set sizes for young gen and humongous before GC: this makes
298 // the policy to give back memory to the OS keep the most recent amount of
299 // memory for these regions.
300 if (hr->is_young() || hr->is_starts_humongous()) {
301 _card_set_stats.add(hr->rem_set()->card_set_memory_stats());
302 }
303 }
304
305 bool humongous_region_is_candidate(G1HeapRegion* region) const {
306 assert(region->is_starts_humongous(), "Must start a humongous object");
307
308 oop obj = cast_to_oop(region->bottom());
309
310 // Dead objects cannot be eager reclaim candidates. Due to class
311 // unloading it is unsafe to query their classes so we return early.
312 if (_g1h->is_obj_dead(obj, region)) {
313 return false;
314 }
315
316 // If we do not have a complete remembered set for the region, then we can
317 // not be sure that we have all references to it.
318 if (!region->rem_set()->is_complete()) {
319 return false;
320 }
321 // We also cannot collect the humongous object if it is pinned.
322 if (region->has_pinned_objects()) {
323 return false;
324 }
325 // Candidate selection must satisfy the following constraints
326 // while concurrent marking is in progress:
327 //
328 // * In order to maintain SATB invariants, an object must not be
329 // reclaimed if it was allocated before the start of marking and
330 // has not had its references scanned. Such an object must have
331 // its references (including type metadata) scanned to ensure no
332 // live objects are missed by the marking process. Objects
333 // allocated after the start of concurrent marking don't need to
334 // be scanned.
335 //
336 // * An object must not be reclaimed if it is on the concurrent
337 // mark stack. Objects allocated after the start of concurrent
338 // marking are never pushed on the mark stack.
339 //
340 // Nominating only objects allocated after the start of concurrent
341 // marking is sufficient to meet both constraints. This may miss
342 // some objects that satisfy the constraints, but the marking data
343 // structures don't support efficiently performing the needed
344 // additional tests or scrubbing of the mark stack.
345 //
346 // However, we presently only nominate is_typeArray() objects.
347 // A humongous object containing references induces remembered
348 // set entries on other regions. In order to reclaim such an
349 // object, those remembered sets would need to be cleaned up.
350 //
351 // We also treat is_typeArray() objects specially, allowing them
352 // to be reclaimed even if allocated before the start of
353 // concurrent mark. For this we rely on mark stack insertion to
354 // exclude is_typeArray() objects, preventing reclaiming an object
355 // that is in the mark stack. We also rely on the metadata for
356 // such objects to be built-in and so ensured to be kept live.
357 // Frequent allocation and drop of large binary blobs is an
358 // important use case for eager reclaim, and this special handling
359 // may reduce needed headroom.
360
361 return obj->is_typeArray() &&
362 _g1h->is_potential_eager_reclaim_candidate(region);
363 }
364
365 public:
366 G1PrepareRegionsClosure(G1CollectedHeap* g1h, G1PrepareEvacuationTask* parent_task) :
367 _g1h(g1h),
368 _parent_task(parent_task),
369 _worker_humongous_total(0),
370 _worker_humongous_candidates(0) { }
371
372 ~G1PrepareRegionsClosure() {
373 _parent_task->add_humongous_candidates(_worker_humongous_candidates);
374 _parent_task->add_humongous_total(_worker_humongous_total);
375 }
376
377 virtual bool do_heap_region(G1HeapRegion* hr) {
378 // First prepare the region for scanning
379 _g1h->rem_set()->prepare_region_for_scan(hr);
380
381 sample_card_set_size(hr);
382
383 // Now check if region is a humongous candidate
384 if (!hr->is_starts_humongous()) {
385 _g1h->register_region_with_region_attr(hr);
386 return false;
387 }
388
389 uint index = hr->hrm_index();
390 if (humongous_region_is_candidate(hr)) {
391 _g1h->register_humongous_candidate_region_with_region_attr(index);
392 _worker_humongous_candidates++;
393 // We will later handle the remembered sets of these regions.
394 } else {
395 _g1h->register_region_with_region_attr(hr);
396 }
397 log_debug(gc, humongous)("Humongous region %u (object size %zu @ " PTR_FORMAT ") remset %zu code roots %zu "
398 "marked %d pinned count %zu reclaim candidate %d type array %d",
399 index,
400 cast_to_oop(hr->bottom())->size() * HeapWordSize,
401 p2i(hr->bottom()),
402 hr->rem_set()->occupied(),
403 hr->rem_set()->code_roots_list_length(),
404 _g1h->concurrent_mark()->mark_bitmap()->is_marked(hr->bottom()),
405 hr->pinned_count(),
406 _g1h->is_humongous_reclaim_candidate(index),
407 cast_to_oop(hr->bottom())->is_typeArray()
408 );
409 _worker_humongous_total++;
410
411 return false;
412 }
413
414 G1MonotonicArenaMemoryStats card_set_stats() const {
415 return _card_set_stats;
416 }
417 };
418
419 G1CollectedHeap* _g1h;
420 HeapRegionClaimer _claimer;
421 volatile uint _humongous_total;
422 volatile uint _humongous_candidates;
423
424 G1MonotonicArenaMemoryStats _all_card_set_stats;
425
426 public:
427 G1PrepareEvacuationTask(G1CollectedHeap* g1h) :
428 WorkerTask("Prepare Evacuation"),
429 _g1h(g1h),
430 _claimer(_g1h->workers()->active_workers()),
431 _humongous_total(0),
432 _humongous_candidates(0) { }
433
434 void work(uint worker_id) {
435 G1PrepareRegionsClosure cl(_g1h, this);
436 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_claimer, worker_id);
437
438 MutexLocker x(G1RareEvent_lock, Mutex::_no_safepoint_check_flag);
439 _all_card_set_stats.add(cl.card_set_stats());
440 }
441
442 void add_humongous_candidates(uint candidates) {
443 Atomic::add(&_humongous_candidates, candidates);
444 }
445
446 void add_humongous_total(uint total) {
447 Atomic::add(&_humongous_total, total);
448 }
449
450 uint humongous_candidates() {
451 return _humongous_candidates;
452 }
453
454 uint humongous_total() {
455 return _humongous_total;
456 }
457
458 const G1MonotonicArenaMemoryStats all_card_set_stats() const {
459 return _all_card_set_stats;
460 }
461 };
462
463 Tickspan G1YoungCollector::run_task_timed(WorkerTask* task) {
464 Ticks start = Ticks::now();
465 workers()->run_task(task);
466 return Ticks::now() - start;
467 }
468
469 void G1YoungCollector::set_young_collection_default_active_worker_threads(){
470 uint active_workers = WorkerPolicy::calc_active_workers(workers()->max_workers(),
471 workers()->active_workers(),
472 Threads::number_of_non_daemon_threads());
473 active_workers = workers()->set_active_workers(active_workers);
474 log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers()->max_workers());
475 }
476
477 void G1YoungCollector::pre_evacuate_collection_set(G1EvacInfo* evacuation_info) {
478 // Flush various data in thread-local buffers to be able to determine the collection
479 // set
480 {
481 Ticks start = Ticks::now();
482 G1PreEvacuateCollectionSetBatchTask cl;
483 G1CollectedHeap::heap()->run_batch_task(&cl);
484 phase_times()->record_pre_evacuate_prepare_time_ms((Ticks::now() - start).seconds() * 1000.0);
485 }
486
487 // Needs log buffers flushed.
488 calculate_collection_set(evacuation_info, policy()->max_pause_time_ms());
489
490 if (collector_state()->in_concurrent_start_gc()) {
491 concurrent_mark()->pre_concurrent_start(_gc_cause);
492 }
493
494 // Please see comment in g1CollectedHeap.hpp and
495 // G1CollectedHeap::ref_processing_init() to see how
496 // reference processing currently works in G1.
497 ref_processor_stw()->start_discovery(false /* always_clear */);
498
499 _evac_failure_regions.pre_collection(_g1h->max_reserved_regions());
500
501 _g1h->gc_prologue(false);
502
503 // Initialize the GC alloc regions.
504 allocator()->init_gc_alloc_regions(evacuation_info);
505
506 {
507 Ticks start = Ticks::now();
508 rem_set()->prepare_for_scan_heap_roots();
509 phase_times()->record_prepare_heap_roots_time_ms((Ticks::now() - start).seconds() * 1000.0);
510 }
511
512 {
513 G1PrepareEvacuationTask g1_prep_task(_g1h);
514 Tickspan task_time = run_task_timed(&g1_prep_task);
515
516 _g1h->set_young_gen_card_set_stats(g1_prep_task.all_card_set_stats());
517 _g1h->set_humongous_stats(g1_prep_task.humongous_total(), g1_prep_task.humongous_candidates());
518
519 phase_times()->record_register_regions(task_time.seconds() * 1000.0);
520 }
521
522 assert(_g1h->verifier()->check_region_attr_table(), "Inconsistency in the region attributes table.");
523
524 #if COMPILER2_OR_JVMCI
525 DerivedPointerTable::clear();
526 #endif
527
528 allocation_failure_injector()->arm_if_needed();
529 }
530
531 class G1ParEvacuateFollowersClosure : public VoidClosure {
532 double _start_term;
533 double _term_time;
534 size_t _term_attempts;
535
536 void start_term_time() { _term_attempts++; _start_term = os::elapsedTime(); }
537 void end_term_time() { _term_time += (os::elapsedTime() - _start_term); }
538
539 G1CollectedHeap* _g1h;
540 G1ParScanThreadState* _par_scan_state;
541 G1ScannerTasksQueueSet* _queues;
542 TaskTerminator* _terminator;
543 G1GCPhaseTimes::GCParPhases _phase;
544
545 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
546 G1ScannerTasksQueueSet* queues() { return _queues; }
547 TaskTerminator* terminator() { return _terminator; }
548
549 inline bool offer_termination() {
550 EventGCPhaseParallel event;
551 G1ParScanThreadState* const pss = par_scan_state();
552 start_term_time();
553 const bool res = (terminator() == nullptr) ? true : terminator()->offer_termination();
554 end_term_time();
555 event.commit(GCId::current(), pss->worker_id(), G1GCPhaseTimes::phase_name(G1GCPhaseTimes::Termination));
556 return res;
557 }
558
559 public:
560 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
561 G1ParScanThreadState* par_scan_state,
562 G1ScannerTasksQueueSet* queues,
563 TaskTerminator* terminator,
564 G1GCPhaseTimes::GCParPhases phase)
565 : _start_term(0.0), _term_time(0.0), _term_attempts(0),
566 _g1h(g1h), _par_scan_state(par_scan_state),
567 _queues(queues), _terminator(terminator), _phase(phase) {}
568
569 void do_void() {
570 EventGCPhaseParallel event;
571 G1ParScanThreadState* const pss = par_scan_state();
572 pss->trim_queue();
573 event.commit(GCId::current(), pss->worker_id(), G1GCPhaseTimes::phase_name(_phase));
574 do {
575 EventGCPhaseParallel event;
576 pss->steal_and_trim_queue(queues());
577 event.commit(GCId::current(), pss->worker_id(), G1GCPhaseTimes::phase_name(_phase));
578 } while (!offer_termination());
579 }
580
581 double term_time() const { return _term_time; }
582 size_t term_attempts() const { return _term_attempts; }
583 };
584
585 class G1EvacuateRegionsBaseTask : public WorkerTask {
586 protected:
587 G1CollectedHeap* _g1h;
588 G1ParScanThreadStateSet* _per_thread_states;
589 G1ScannerTasksQueueSet* _task_queues;
590 TaskTerminator _terminator;
591 uint _num_workers;
592
593 void evacuate_live_objects(G1ParScanThreadState* pss,
594 uint worker_id,
595 G1GCPhaseTimes::GCParPhases objcopy_phase,
596 G1GCPhaseTimes::GCParPhases termination_phase) {
597 G1GCPhaseTimes* p = _g1h->phase_times();
598
599 Ticks start = Ticks::now();
600 G1ParEvacuateFollowersClosure cl(_g1h, pss, _task_queues, &_terminator, objcopy_phase);
601 cl.do_void();
602
603 assert(pss->queue_is_empty(), "should be empty");
604
605 Tickspan evac_time = (Ticks::now() - start);
606 p->record_or_add_time_secs(objcopy_phase, worker_id, evac_time.seconds() - cl.term_time());
607
608 if (termination_phase == G1GCPhaseTimes::Termination) {
609 p->record_time_secs(termination_phase, worker_id, cl.term_time());
610 p->record_thread_work_item(termination_phase, worker_id, cl.term_attempts());
611 } else {
612 p->record_or_add_time_secs(termination_phase, worker_id, cl.term_time());
613 p->record_or_add_thread_work_item(termination_phase, worker_id, cl.term_attempts());
614 }
615 assert(pss->trim_ticks().value() == 0,
616 "Unexpected partial trimming during evacuation value " JLONG_FORMAT,
617 pss->trim_ticks().value());
618 }
619
620 virtual void start_work(uint worker_id) { }
621
622 virtual void end_work(uint worker_id) { }
623
624 virtual void scan_roots(G1ParScanThreadState* pss, uint worker_id) = 0;
625
626 virtual void evacuate_live_objects(G1ParScanThreadState* pss, uint worker_id) = 0;
627
628 public:
629 G1EvacuateRegionsBaseTask(const char* name,
630 G1ParScanThreadStateSet* per_thread_states,
631 G1ScannerTasksQueueSet* task_queues,
632 uint num_workers) :
633 WorkerTask(name),
634 _g1h(G1CollectedHeap::heap()),
635 _per_thread_states(per_thread_states),
636 _task_queues(task_queues),
637 _terminator(num_workers, _task_queues),
638 _num_workers(num_workers)
639 { }
640
641 void work(uint worker_id) {
642 start_work(worker_id);
643
644 {
645 ResourceMark rm;
646
647 G1ParScanThreadState* pss = _per_thread_states->state_for_worker(worker_id);
648 pss->set_ref_discoverer(_g1h->ref_processor_stw());
649
650 scan_roots(pss, worker_id);
651 evacuate_live_objects(pss, worker_id);
652 }
653
654 end_work(worker_id);
655 }
656 };
657
658 class G1EvacuateRegionsTask : public G1EvacuateRegionsBaseTask {
659 G1RootProcessor* _root_processor;
660 bool _has_optional_evacuation_work;
661
662 void scan_roots(G1ParScanThreadState* pss, uint worker_id) {
663 _root_processor->evacuate_roots(pss, worker_id);
664 _g1h->rem_set()->scan_heap_roots(pss, worker_id, G1GCPhaseTimes::ScanHR, G1GCPhaseTimes::ObjCopy, _has_optional_evacuation_work);
665 _g1h->rem_set()->scan_collection_set_regions(pss, worker_id, G1GCPhaseTimes::ScanHR, G1GCPhaseTimes::CodeRoots, G1GCPhaseTimes::ObjCopy);
666 }
667
668 void evacuate_live_objects(G1ParScanThreadState* pss, uint worker_id) {
669 G1EvacuateRegionsBaseTask::evacuate_live_objects(pss, worker_id, G1GCPhaseTimes::ObjCopy, G1GCPhaseTimes::Termination);
670 }
671
672 void start_work(uint worker_id) {
673 _g1h->phase_times()->record_time_secs(G1GCPhaseTimes::GCWorkerStart, worker_id, Ticks::now().seconds());
674 }
675
676 void end_work(uint worker_id) {
677 _g1h->phase_times()->record_time_secs(G1GCPhaseTimes::GCWorkerEnd, worker_id, Ticks::now().seconds());
678 }
679
680 public:
681 G1EvacuateRegionsTask(G1CollectedHeap* g1h,
682 G1ParScanThreadStateSet* per_thread_states,
683 G1ScannerTasksQueueSet* task_queues,
684 G1RootProcessor* root_processor,
685 uint num_workers,
686 bool has_optional_evacuation_work) :
687 G1EvacuateRegionsBaseTask("G1 Evacuate Regions", per_thread_states, task_queues, num_workers),
688 _root_processor(root_processor),
689 _has_optional_evacuation_work(has_optional_evacuation_work)
690 { }
691 };
692
693 void G1YoungCollector::evacuate_initial_collection_set(G1ParScanThreadStateSet* per_thread_states,
694 bool has_optional_evacuation_work) {
695 G1GCPhaseTimes* p = phase_times();
696
697 {
698 Ticks start = Ticks::now();
699 rem_set()->merge_heap_roots(true /* initial_evacuation */);
700 p->record_merge_heap_roots_time((Ticks::now() - start).seconds() * 1000.0);
701 }
702
703 Tickspan task_time;
704 const uint num_workers = workers()->active_workers();
705
706 Ticks start_processing = Ticks::now();
707 {
708 G1RootProcessor root_processor(_g1h, num_workers);
709 G1EvacuateRegionsTask g1_par_task(_g1h,
710 per_thread_states,
711 task_queues(),
712 &root_processor,
713 num_workers,
714 has_optional_evacuation_work);
715 task_time = run_task_timed(&g1_par_task);
716 // Closing the inner scope will execute the destructor for the
717 // G1RootProcessor object. By subtracting the WorkerThreads task from the total
718 // time of this scope, we get the "NMethod List Cleanup" time. This list is
719 // constructed during "STW two-phase nmethod root processing", see more in
720 // nmethod.hpp
721 }
722 Tickspan total_processing = Ticks::now() - start_processing;
723
724 p->record_initial_evac_time(task_time.seconds() * 1000.0);
725 p->record_or_add_nmethod_list_cleanup_time((total_processing - task_time).seconds() * 1000.0);
726
727 rem_set()->complete_evac_phase(has_optional_evacuation_work);
728 }
729
730 class G1EvacuateOptionalRegionsTask : public G1EvacuateRegionsBaseTask {
731
732 void scan_roots(G1ParScanThreadState* pss, uint worker_id) {
733 _g1h->rem_set()->scan_heap_roots(pss, worker_id, G1GCPhaseTimes::OptScanHR, G1GCPhaseTimes::OptObjCopy, true /* remember_already_scanned_cards */);
734 _g1h->rem_set()->scan_collection_set_regions(pss, worker_id, G1GCPhaseTimes::OptScanHR, G1GCPhaseTimes::OptCodeRoots, G1GCPhaseTimes::OptObjCopy);
735 }
736
737 void evacuate_live_objects(G1ParScanThreadState* pss, uint worker_id) {
738 G1EvacuateRegionsBaseTask::evacuate_live_objects(pss, worker_id, G1GCPhaseTimes::OptObjCopy, G1GCPhaseTimes::OptTermination);
739 }
740
741 public:
742 G1EvacuateOptionalRegionsTask(G1ParScanThreadStateSet* per_thread_states,
743 G1ScannerTasksQueueSet* queues,
744 uint num_workers) :
745 G1EvacuateRegionsBaseTask("G1 Evacuate Optional Regions", per_thread_states, queues, num_workers) {
746 }
747 };
748
749 void G1YoungCollector::evacuate_next_optional_regions(G1ParScanThreadStateSet* per_thread_states) {
750 // To access the protected constructor/destructor
751 class G1MarkScope : public MarkScope { };
752
753 Tickspan task_time;
754
755 Ticks start_processing = Ticks::now();
756 {
757 G1MarkScope code_mark_scope;
758 G1EvacuateOptionalRegionsTask task(per_thread_states, task_queues(), workers()->active_workers());
759 task_time = run_task_timed(&task);
760 // See comment in evacuate_initial_collection_set() for the reason of the scope.
761 }
762 Tickspan total_processing = Ticks::now() - start_processing;
763
764 G1GCPhaseTimes* p = phase_times();
765 p->record_or_add_nmethod_list_cleanup_time((total_processing - task_time).seconds() * 1000.0);
766 }
767
768 void G1YoungCollector::evacuate_optional_collection_set(G1ParScanThreadStateSet* per_thread_states) {
769 const double collection_start_time_ms = phase_times()->cur_collection_start_sec() * 1000.0;
770
771 while (!evacuation_alloc_failed() && collection_set()->optional_region_length() > 0) {
772
773 double time_used_ms = os::elapsedTime() * 1000.0 - collection_start_time_ms;
774 double time_left_ms = MaxGCPauseMillis - time_used_ms;
775
776 if (time_left_ms < 0 ||
777 !collection_set()->finalize_optional_for_evacuation(time_left_ms * policy()->optional_evacuation_fraction())) {
778 log_trace(gc, ergo, cset)("Skipping evacuation of %u optional regions, no more regions can be evacuated in %.3fms",
779 collection_set()->optional_region_length(), time_left_ms);
780 break;
781 }
782
783 {
784 Ticks start = Ticks::now();
785 rem_set()->merge_heap_roots(false /* initial_evacuation */);
786 phase_times()->record_or_add_optional_merge_heap_roots_time((Ticks::now() - start).seconds() * 1000.0);
787 }
788
789 {
790 Ticks start = Ticks::now();
791 evacuate_next_optional_regions(per_thread_states);
792 phase_times()->record_or_add_optional_evac_time((Ticks::now() - start).seconds() * 1000.0);
793 }
794
795 rem_set()->complete_evac_phase(true /* has_more_than_one_evacuation_phase */);
796 }
797
798 collection_set()->abandon_optional_collection_set(per_thread_states);
799 }
800
801 // Non Copying Keep Alive closure
802 class G1KeepAliveClosure: public OopClosure {
803 G1CollectedHeap*_g1h;
804 public:
805 G1KeepAliveClosure(G1CollectedHeap* g1h) :_g1h(g1h) {}
806 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
807 void do_oop(oop* p) {
808 oop obj = *p;
809 assert(obj != nullptr, "the caller should have filtered out null values");
810
811 const G1HeapRegionAttr region_attr =_g1h->region_attr(obj);
812 if (!region_attr.is_in_cset_or_humongous_candidate()) {
813 return;
814 }
815 if (region_attr.is_in_cset()) {
816 assert(obj->is_forwarded(), "invariant" );
817 *p = obj->forwardee();
818 } else {
819 assert(!obj->is_forwarded(), "invariant" );
820 assert(region_attr.is_humongous_candidate(),
821 "Only allowed G1HeapRegionAttr state is IsHumongous, but is %d", region_attr.type());
822 _g1h->set_humongous_is_live(obj);
823 }
824 }
825 };
826
827 // Copying Keep Alive closure - can be called from both
828 // serial and parallel code as long as different worker
829 // threads utilize different G1ParScanThreadState instances
830 // and different queues.
831 class G1CopyingKeepAliveClosure: public OopClosure {
832 G1CollectedHeap* _g1h;
833 G1ParScanThreadState* _par_scan_state;
834
835 public:
836 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
837 G1ParScanThreadState* pss):
838 _g1h(g1h),
839 _par_scan_state(pss)
840 {}
841
842 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
843 virtual void do_oop( oop* p) { do_oop_work(p); }
844
845 template <class T> void do_oop_work(T* p) {
846 oop obj = RawAccess<>::oop_load(p);
847
848 if (_g1h->is_in_cset_or_humongous_candidate(obj)) {
849 // If the referent object has been forwarded (either copied
850 // to a new location or to itself in the event of an
851 // evacuation failure) then we need to update the reference
852 // field and, if both reference and referent are in the G1
853 // heap, update the RSet for the referent.
854 //
855 // If the referent has not been forwarded then we have to keep
856 // it alive by policy. Therefore we have copy the referent.
857 //
858 // When the queue is drained (after each phase of reference processing)
859 // the object and it's followers will be copied, the reference field set
860 // to point to the new location, and the RSet updated.
861 _par_scan_state->push_on_queue(ScannerTask(p));
862 }
863 }
864 };
865
866 class G1STWRefProcProxyTask : public RefProcProxyTask {
867 G1CollectedHeap& _g1h;
868 G1ParScanThreadStateSet& _pss;
869 TaskTerminator _terminator;
870 G1ScannerTasksQueueSet& _task_queues;
871
872 // Special closure for enqueuing discovered fields: during enqueue the card table
873 // may not be in shape to properly handle normal barrier calls (e.g. card marks
874 // in regions that failed evacuation, scribbling of various values by card table
875 // scan code). Additionally the regular barrier enqueues into the "global"
876 // DCQS, but during GC we need these to-be-refined entries in the GC local queue
877 // so that after clearing the card table, the redirty cards phase will properly
878 // mark all dirty cards to be picked up by refinement.
879 class G1EnqueueDiscoveredFieldClosure : public EnqueueDiscoveredFieldClosure {
880 G1CollectedHeap* _g1h;
881 G1ParScanThreadState* _pss;
882
883 public:
884 G1EnqueueDiscoveredFieldClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : _g1h(g1h), _pss(pss) { }
885
886 void enqueue(HeapWord* discovered_field_addr, oop value) override {
887 assert(_g1h->is_in(discovered_field_addr), PTR_FORMAT " is not in heap ", p2i(discovered_field_addr));
888 // Store the value first, whatever it is.
889 RawAccess<>::oop_store(discovered_field_addr, value);
890 if (value == nullptr) {
891 return;
892 }
893 _pss->write_ref_field_post(discovered_field_addr, value);
894 }
895 };
896
897 public:
898 G1STWRefProcProxyTask(uint max_workers, G1CollectedHeap& g1h, G1ParScanThreadStateSet& pss, G1ScannerTasksQueueSet& task_queues)
899 : RefProcProxyTask("G1STWRefProcProxyTask", max_workers),
900 _g1h(g1h),
901 _pss(pss),
902 _terminator(max_workers, &task_queues),
903 _task_queues(task_queues) {}
904
905 void work(uint worker_id) override {
906 assert(worker_id < _max_workers, "sanity");
907 uint index = (_tm == RefProcThreadModel::Single) ? 0 : worker_id;
908
909 G1ParScanThreadState* pss = _pss.state_for_worker(index);
910 pss->set_ref_discoverer(nullptr);
911
912 G1STWIsAliveClosure is_alive(&_g1h);
913 G1CopyingKeepAliveClosure keep_alive(&_g1h, pss);
914 G1EnqueueDiscoveredFieldClosure enqueue(&_g1h, pss);
915 G1ParEvacuateFollowersClosure complete_gc(&_g1h, pss, &_task_queues, _tm == RefProcThreadModel::Single ? nullptr : &_terminator, G1GCPhaseTimes::ObjCopy);
916 _rp_task->rp_work(worker_id, &is_alive, &keep_alive, &enqueue, &complete_gc);
917
918 // We have completed copying any necessary live referent objects.
919 assert(pss->queue_is_empty(), "both queue and overflow should be empty");
920 }
921
922 void prepare_run_task_hook() override {
923 _terminator.reset_for_reuse(_queue_count);
924 }
925 };
926
927 void G1YoungCollector::process_discovered_references(G1ParScanThreadStateSet* per_thread_states) {
928 Ticks start = Ticks::now();
929
930 ReferenceProcessor* rp = ref_processor_stw();
931 assert(rp->discovery_enabled(), "should have been enabled");
932
933 uint no_of_gc_workers = workers()->active_workers();
934 rp->set_active_mt_degree(no_of_gc_workers);
935
936 G1STWRefProcProxyTask task(rp->max_num_queues(), *_g1h, *per_thread_states, *task_queues());
937 ReferenceProcessorPhaseTimes& pt = *phase_times()->ref_phase_times();
938 ReferenceProcessorStats stats = rp->process_discovered_references(task, pt);
939
940 gc_tracer_stw()->report_gc_reference_stats(stats);
941
942 _g1h->make_pending_list_reachable();
943
944 phase_times()->record_ref_proc_time((Ticks::now() - start).seconds() * MILLIUNITS);
945 }
946
947 void G1YoungCollector::post_evacuate_cleanup_1(G1ParScanThreadStateSet* per_thread_states) {
948 Ticks start = Ticks::now();
949 {
950 G1PostEvacuateCollectionSetCleanupTask1 cl(per_thread_states, &_evac_failure_regions);
951 _g1h->run_batch_task(&cl);
952 }
953 phase_times()->record_post_evacuate_cleanup_task_1_time((Ticks::now() - start).seconds() * 1000.0);
954 }
955
956 void G1YoungCollector::post_evacuate_cleanup_2(G1ParScanThreadStateSet* per_thread_states,
957 G1EvacInfo* evacuation_info) {
958 Ticks start = Ticks::now();
959 {
960 G1PostEvacuateCollectionSetCleanupTask2 cl(per_thread_states, evacuation_info, &_evac_failure_regions);
961 _g1h->run_batch_task(&cl);
962 }
963 phase_times()->record_post_evacuate_cleanup_task_2_time((Ticks::now() - start).seconds() * 1000.0);
964 }
965
966 void G1YoungCollector::enqueue_candidates_as_root_regions() {
967 assert(collector_state()->in_concurrent_start_gc(), "must be");
968
969 G1CollectionSetCandidates* candidates = collection_set()->candidates();
970 for (G1HeapRegion* r : *candidates) {
971 _g1h->concurrent_mark()->add_root_region(r);
972 }
973 }
974
975 void G1YoungCollector::post_evacuate_collection_set(G1EvacInfo* evacuation_info,
976 G1ParScanThreadStateSet* per_thread_states) {
977 G1GCPhaseTimes* p = phase_times();
978
979 // Process any discovered reference objects - we have
980 // to do this _before_ we retire the GC alloc regions
981 // as we may have to copy some 'reachable' referent
982 // objects (and their reachable sub-graphs) that were
983 // not copied during the pause.
984 process_discovered_references(per_thread_states);
985
986 G1STWIsAliveClosure is_alive(_g1h);
987 G1KeepAliveClosure keep_alive(_g1h);
988
989 WeakProcessor::weak_oops_do(workers(), &is_alive, &keep_alive, p->weak_phase_times());
990
991 allocator()->release_gc_alloc_regions(evacuation_info);
992
993 post_evacuate_cleanup_1(per_thread_states);
994
995 post_evacuate_cleanup_2(per_thread_states, evacuation_info);
996
997 // Regions in the collection set candidates are roots for the marking (they are
998 // not marked through considering they are very likely to be reclaimed soon.
999 // They need to be enqueued explicitly compared to survivor regions.
1000 if (collector_state()->in_concurrent_start_gc()) {
1001 enqueue_candidates_as_root_regions();
1002 }
1003
1004 _evac_failure_regions.post_collection();
1005
1006 assert_used_and_recalculate_used_equal(_g1h);
1007
1008 _g1h->rebuild_free_region_list();
1009
1010 _g1h->record_obj_copy_mem_stats();
1011
1012 evacuation_info->set_bytes_used(_g1h->bytes_used_during_gc());
1013
1014 _g1h->prepare_for_mutator_after_young_collection();
1015
1016 _g1h->gc_epilogue(false);
1017
1018 _g1h->expand_heap_after_young_collection();
1019 }
1020
1021 bool G1YoungCollector::evacuation_failed() const {
1022 return _evac_failure_regions.has_regions_evac_failed();
1023 }
1024
1025 bool G1YoungCollector::evacuation_pinned() const {
1026 return _evac_failure_regions.has_regions_evac_pinned();
1027 }
1028
1029 bool G1YoungCollector::evacuation_alloc_failed() const {
1030 return _evac_failure_regions.has_regions_alloc_failed();
1031 }
1032
1033 G1YoungCollector::G1YoungCollector(GCCause::Cause gc_cause) :
1034 _g1h(G1CollectedHeap::heap()),
1035 _gc_cause(gc_cause),
1036 _concurrent_operation_is_full_mark(false),
1037 _evac_failure_regions()
1038 {
1039 }
1040
1041 void G1YoungCollector::collect() {
1042 // Do timing/tracing/statistics/pre- and post-logging/verification work not
1043 // directly related to the collection. They should not be accounted for in
1044 // collection work timing.
1045
1046 // The G1YoungGCTraceTime message depends on collector state, so must come after
1047 // determining collector state.
1048 G1YoungGCTraceTime tm(this, _gc_cause);
1049
1050 // JFR
1051 G1YoungGCJFRTracerMark jtm(gc_timer_stw(), gc_tracer_stw(), _gc_cause);
1052 // JStat/MXBeans
1053 G1YoungGCMonitoringScope ms(monitoring_support(),
1054 !collection_set()->candidates()->is_empty() /* all_memory_pools_affected */);
1055 // Create the heap printer before internal pause timing to have
1056 // heap information printed as last part of detailed GC log.
1057 G1HeapPrinterMark hpm(_g1h);
1058 // Young GC internal pause timing
1059 G1YoungGCNotifyPauseMark npm(this);
1060
1061 // Verification may use the workers, so they must be set up before.
1062 // Individual parallel phases may override this.
1063 set_young_collection_default_active_worker_threads();
1064
1065 // Wait for root region scan here to make sure that it is done before any
1066 // use of the STW workers to maximize cpu use (i.e. all cores are available
1067 // just to do that).
1068 wait_for_root_region_scanning();
1069
1070 G1YoungGCVerifierMark vm(this);
1071 {
1072 // Actual collection work starts and is executed (only) in this scope.
1073
1074 // Young GC internal collection timing. The elapsed time recorded in the
1075 // policy for the collection deliberately elides verification (and some
1076 // other trivial setup above).
1077 policy()->record_young_collection_start();
1078
1079 pre_evacuate_collection_set(jtm.evacuation_info());
1080
1081 G1ParScanThreadStateSet per_thread_states(_g1h,
1082 workers()->active_workers(),
1083 collection_set(),
1084 &_evac_failure_regions);
1085
1086 bool may_do_optional_evacuation = collection_set()->optional_region_length() != 0;
1087 // Actually do the work...
1088 evacuate_initial_collection_set(&per_thread_states, may_do_optional_evacuation);
1089
1090 if (may_do_optional_evacuation) {
1091 evacuate_optional_collection_set(&per_thread_states);
1092 }
1093 post_evacuate_collection_set(jtm.evacuation_info(), &per_thread_states);
1094
1095 // Refine the type of a concurrent mark operation now that we did the
1096 // evacuation, eventually aborting it.
1097 _concurrent_operation_is_full_mark = policy()->concurrent_operation_is_full_mark("Revise IHOP");
1098
1099 // Need to report the collection pause now since record_collection_pause_end()
1100 // modifies it to the next state.
1101 jtm.report_pause_type(collector_state()->young_gc_pause_type(_concurrent_operation_is_full_mark));
1102
1103 policy()->record_young_collection_end(_concurrent_operation_is_full_mark, evacuation_alloc_failed());
1104 }
1105 TASKQUEUE_STATS_ONLY(_g1h->task_queues()->print_and_reset_taskqueue_stats("Oop Queue");)
1106 }