1 /*
2 * Copyright (c) 2017, 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 #include "classfile/classLoaderDataGraph.hpp"
27 #include "gc/g1/g1CollectedHeap.hpp"
28 #include "gc/g1/g1FullCollector.inline.hpp"
29 #include "gc/g1/g1FullGCAdjustTask.hpp"
30 #include "gc/g1/g1FullGCCompactTask.hpp"
31 #include "gc/g1/g1FullGCMarker.inline.hpp"
32 #include "gc/g1/g1FullGCMarkTask.hpp"
33 #include "gc/g1/g1FullGCPrepareTask.inline.hpp"
34 #include "gc/g1/g1FullGCResetMetadataTask.hpp"
35 #include "gc/g1/g1FullGCScope.hpp"
36 #include "gc/g1/g1OopClosures.hpp"
37 #include "gc/g1/g1Policy.hpp"
38 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
39 #include "gc/shared/gcTraceTime.inline.hpp"
40 #include "gc/shared/preservedMarks.inline.hpp"
41 #include "gc/shared/classUnloadingContext.hpp"
42 #include "gc/shared/referenceProcessor.hpp"
43 #include "gc/shared/verifyOption.hpp"
44 #include "gc/shared/weakProcessor.inline.hpp"
45 #include "gc/shared/workerPolicy.hpp"
46 #include "logging/log.hpp"
47 #include "runtime/handles.inline.hpp"
48 #include "utilities/debug.hpp"
49
50 static void clear_and_activate_derived_pointers() {
51 #if COMPILER2_OR_JVMCI
52 DerivedPointerTable::clear();
53 #endif
54 }
55
56 static void deactivate_derived_pointers() {
57 #if COMPILER2_OR_JVMCI
58 DerivedPointerTable::set_active(false);
59 #endif
60 }
61
62 static void update_derived_pointers() {
63 #if COMPILER2_OR_JVMCI
64 DerivedPointerTable::update_pointers();
65 #endif
66 }
67
68 G1CMBitMap* G1FullCollector::mark_bitmap() {
69 return _heap->concurrent_mark()->mark_bitmap();
70 }
71
72 ReferenceProcessor* G1FullCollector::reference_processor() {
73 return _heap->ref_processor_stw();
74 }
75
76 uint G1FullCollector::calc_active_workers() {
77 G1CollectedHeap* heap = G1CollectedHeap::heap();
78 uint max_worker_count = heap->workers()->max_workers();
79 // Only calculate number of workers if UseDynamicNumberOfGCThreads
80 // is enabled, otherwise use max.
81 if (!UseDynamicNumberOfGCThreads) {
82 return max_worker_count;
83 }
84
85 // Consider G1HeapWastePercent to decide max number of workers. Each worker
86 // will in average cause half a region waste.
87 uint max_wasted_regions_allowed = ((heap->num_regions() * G1HeapWastePercent) / 100);
88 uint waste_worker_count = MAX2((max_wasted_regions_allowed * 2) , 1u);
89 uint heap_waste_worker_limit = MIN2(waste_worker_count, max_worker_count);
90
91 // Also consider HeapSizePerGCThread by calling WorkerPolicy to calculate
92 // the number of workers.
93 uint current_active_workers = heap->workers()->active_workers();
94 uint active_worker_limit = WorkerPolicy::calc_active_workers(max_worker_count, current_active_workers, 0);
95
96 // Finally consider the amount of used regions.
97 uint used_worker_limit = heap->num_used_regions();
98 assert(used_worker_limit > 0, "Should never have zero used regions.");
99
100 // Update active workers to the lower of the limits.
101 uint worker_count = MIN3(heap_waste_worker_limit, active_worker_limit, used_worker_limit);
102 log_debug(gc, task)("Requesting %u active workers for full compaction (waste limited workers: %u, "
103 "adaptive workers: %u, used limited workers: %u)",
104 worker_count, heap_waste_worker_limit, active_worker_limit, used_worker_limit);
105 worker_count = heap->workers()->set_active_workers(worker_count);
106 log_info(gc, task)("Using %u workers of %u for full compaction", worker_count, max_worker_count);
107
108 return worker_count;
109 }
110
111 G1FullCollector::G1FullCollector(G1CollectedHeap* heap,
112 bool clear_soft_refs,
113 bool do_maximal_compaction,
114 G1FullGCTracer* tracer) :
115 _heap(heap),
116 _scope(heap->monitoring_support(), clear_soft_refs, do_maximal_compaction, tracer),
117 _num_workers(calc_active_workers()),
118 _has_compaction_targets(false),
119 _has_humongous(false),
120 _oop_queue_set(_num_workers),
121 _array_queue_set(_num_workers),
122 _preserved_marks_set(true),
123 _serial_compaction_point(this, nullptr),
124 _humongous_compaction_point(this, nullptr),
125 _is_alive(this, heap->concurrent_mark()->mark_bitmap()),
126 _is_alive_mutator(heap->ref_processor_stw(), &_is_alive),
127 _humongous_compaction_regions(8),
128 _always_subject_to_discovery(),
129 _is_subject_mutator(heap->ref_processor_stw(), &_always_subject_to_discovery),
130 _region_attr_table() {
131 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
132
133 _preserved_marks_set.init(_num_workers);
134 _markers = NEW_C_HEAP_ARRAY(G1FullGCMarker*, _num_workers, mtGC);
135 _compaction_points = NEW_C_HEAP_ARRAY(G1FullGCCompactionPoint*, _num_workers, mtGC);
136
137 _live_stats = NEW_C_HEAP_ARRAY(G1RegionMarkStats, _heap->max_regions(), mtGC);
138 _compaction_tops = NEW_C_HEAP_ARRAY(HeapWord*, _heap->max_regions(), mtGC);
139 for (uint j = 0; j < heap->max_regions(); j++) {
140 _live_stats[j].clear();
141 _compaction_tops[j] = nullptr;
142 }
143
144 for (uint i = 0; i < _num_workers; i++) {
145 _markers[i] = new G1FullGCMarker(this, i, _live_stats);
146 _compaction_points[i] = new G1FullGCCompactionPoint(this, _preserved_marks_set.get(i));
147 _oop_queue_set.register_queue(i, marker(i)->oop_stack());
148 _array_queue_set.register_queue(i, marker(i)->objarray_stack());
149 }
150 _serial_compaction_point.set_preserved_stack(_preserved_marks_set.get(0));
151 _humongous_compaction_point.set_preserved_stack(_preserved_marks_set.get(0));
152 _region_attr_table.initialize(heap->reserved(), G1HeapRegion::GrainBytes);
153 }
154
155 G1FullCollector::~G1FullCollector() {
156 for (uint i = 0; i < _num_workers; i++) {
157 delete _markers[i];
158 delete _compaction_points[i];
159 }
160
161 FREE_C_HEAP_ARRAY(G1FullGCMarker*, _markers);
162 FREE_C_HEAP_ARRAY(G1FullGCCompactionPoint*, _compaction_points);
163 FREE_C_HEAP_ARRAY(HeapWord*, _compaction_tops);
164 FREE_C_HEAP_ARRAY(G1RegionMarkStats, _live_stats);
165 }
166
167 class PrepareRegionsClosure : public G1HeapRegionClosure {
168 G1FullCollector* _collector;
169
170 public:
171 PrepareRegionsClosure(G1FullCollector* collector) : _collector(collector) { }
172
173 bool do_heap_region(G1HeapRegion* hr) {
174 hr->prepare_for_full_gc();
175 G1CollectedHeap::heap()->prepare_region_for_full_compaction(hr);
176 _collector->before_marking_update_attribute_table(hr);
177 return false;
178 }
179 };
180
181 void G1FullCollector::prepare_collection() {
182 _heap->policy()->record_full_collection_start();
183
184 // Verification needs the bitmap, so we should clear the bitmap only later.
185 bool in_concurrent_cycle = _heap->abort_concurrent_cycle();
186 _heap->verify_before_full_collection();
187 if (in_concurrent_cycle) {
188 GCTraceTime(Debug, gc) debug("Clear Bitmap");
189 _heap->concurrent_mark()->clear_bitmap(_heap->workers());
190 }
191
192 _heap->gc_prologue(true);
193 _heap->retire_tlabs();
194 _heap->flush_region_pin_cache();
195 _heap->prepare_heap_for_full_collection();
196
197 PrepareRegionsClosure cl(this);
198 _heap->heap_region_iterate(&cl);
199
200 reference_processor()->start_discovery(scope()->should_clear_soft_refs());
201
202 // Clear and activate derived pointer collection.
203 clear_and_activate_derived_pointers();
204 }
205
206 void G1FullCollector::collect() {
207 G1CollectedHeap::start_codecache_marking_cycle_if_inactive(false /* concurrent_mark_start */);
208
209 phase1_mark_live_objects();
210 verify_after_marking();
211
212 // Don't add any more derived pointers during later phases
213 deactivate_derived_pointers();
214
215 phase2_prepare_compaction();
216
217 if (has_compaction_targets()) {
218 phase3_adjust_pointers();
219
220 phase4_do_compaction();
221 } else {
222 // All regions have a high live ratio thus will not be compacted.
223 // The live ratio is only considered if do_maximal_compaction is false.
224 log_info(gc, phases) ("No Regions selected for compaction. Skipping Phase 3: Adjust pointers and Phase 4: Compact heap");
225 }
226
227 phase5_reset_metadata();
228
229 G1CollectedHeap::finish_codecache_marking_cycle();
230 }
231
232 void G1FullCollector::complete_collection() {
233 // Restore all marks.
234 restore_marks();
235
236 // When the pointers have been adjusted and moved, we can
237 // update the derived pointer table.
238 update_derived_pointers();
239
240 // Need completely cleared claim bits for the next concurrent marking or full gc.
241 ClassLoaderDataGraph::clear_claimed_marks();
242
243 // Prepare the bitmap for the next (potentially concurrent) marking.
244 _heap->concurrent_mark()->clear_bitmap(_heap->workers());
245
246 _heap->prepare_for_mutator_after_full_collection();
247
248 _heap->resize_all_tlabs();
249
250 _heap->young_regions_cardset()->clear();
251
252 _heap->policy()->record_full_collection_end();
253 _heap->gc_epilogue(true);
254
255 _heap->verify_after_full_collection();
256
257 _heap->print_heap_after_full_collection();
258 }
259
260 void G1FullCollector::before_marking_update_attribute_table(G1HeapRegion* hr) {
261 if (hr->is_free()) {
262 _region_attr_table.set_free(hr->hrm_index());
263 } else if (hr->is_humongous() || hr->has_pinned_objects()) {
264 // Humongous objects or pinned regions will never be moved in the "main"
265 // compaction phase, but non-pinned regions might afterwards in a special phase.
266 _region_attr_table.set_skip_compacting(hr->hrm_index());
267 } else {
268 // Everything else should be compacted.
269 _region_attr_table.set_compacting(hr->hrm_index());
270 }
271 }
272
273 class G1FullGCRefProcProxyTask : public RefProcProxyTask {
274 G1FullCollector& _collector;
275
276 public:
277 G1FullGCRefProcProxyTask(G1FullCollector &collector, uint max_workers)
278 : RefProcProxyTask("G1FullGCRefProcProxyTask", max_workers),
279 _collector(collector) {}
280
281 void work(uint worker_id) override {
282 assert(worker_id < _max_workers, "sanity");
283 G1IsAliveClosure is_alive(&_collector);
284 uint index = (_tm == RefProcThreadModel::Single) ? 0 : worker_id;
285 G1FullKeepAliveClosure keep_alive(_collector.marker(index));
286 BarrierEnqueueDiscoveredFieldClosure enqueue;
287 G1FollowStackClosure* complete_gc = _collector.marker(index)->stack_closure();
288 _rp_task->rp_work(worker_id, &is_alive, &keep_alive, &enqueue, complete_gc);
289 }
290 };
291
292 void G1FullCollector::phase1_mark_live_objects() {
293 // Recursively traverse all live objects and mark them.
294 GCTraceTime(Info, gc, phases) info("Phase 1: Mark live objects", scope()->timer());
295
296 {
297 // Do the actual marking.
298 G1FullGCMarkTask marking_task(this);
299 run_task(&marking_task);
300 }
301
302 {
303 uint old_active_mt_degree = reference_processor()->num_queues();
304 reference_processor()->set_active_mt_degree(workers());
305 GCTraceTime(Debug, gc, phases) debug("Phase 1: Reference Processing", scope()->timer());
306 // Process reference objects found during marking.
307 ReferenceProcessorPhaseTimes pt(scope()->timer(), reference_processor()->max_num_queues());
308 G1FullGCRefProcProxyTask task(*this, reference_processor()->max_num_queues());
309 const ReferenceProcessorStats& stats = reference_processor()->process_discovered_references(task, pt);
310 scope()->tracer()->report_gc_reference_stats(stats);
311 pt.print_all_references();
312 assert(marker(0)->oop_stack()->is_empty(), "Should be no oops on the stack");
313
314 reference_processor()->set_active_mt_degree(old_active_mt_degree);
315 }
316
317 {
318 GCTraceTime(Debug, gc, phases) debug("Phase 1: Flush Mark Stats Cache", scope()->timer());
319 for (uint i = 0; i < workers(); i++) {
320 marker(i)->flush_mark_stats_cache();
321 }
322 }
323
324 // Weak oops cleanup.
325 {
326 GCTraceTime(Debug, gc, phases) debug("Phase 1: Weak Processing", scope()->timer());
327 WeakProcessor::weak_oops_do(_heap->workers(), &_is_alive, &do_nothing_cl, 1);
328 }
329
330 // Class unloading and cleanup.
331 if (ClassUnloading) {
332 _heap->unload_classes_and_code("Phase 1: Class Unloading and Cleanup", &_is_alive, scope()->timer());
333 }
334
335 {
336 GCTraceTime(Debug, gc, phases) debug("Report Object Count", scope()->timer());
337 scope()->tracer()->report_object_count_after_gc(&_is_alive, _heap->workers());
338 }
339 #if TASKQUEUE_STATS
340 oop_queue_set()->print_and_reset_taskqueue_stats("Oop Queue");
341 array_queue_set()->print_and_reset_taskqueue_stats("ObjArrayOop Queue");
342 #endif
343 }
344
345 void G1FullCollector::phase2_prepare_compaction() {
346 GCTraceTime(Info, gc, phases) info("Phase 2: Prepare compaction", scope()->timer());
347
348 phase2a_determine_worklists();
349
350 if (!has_compaction_targets()) {
351 return;
352 }
353
354 bool has_free_compaction_targets = phase2b_forward_oops();
355
356 // Try to avoid OOM immediately after Full GC in case there are no free regions
357 // left after determining the result locations (i.e. this phase). Prepare to
358 // maximally compact the tail regions of the compaction queues serially.
359 if (scope()->do_maximal_compaction() || !has_free_compaction_targets) {
360 phase2c_prepare_serial_compaction();
361
362 if (scope()->do_maximal_compaction() &&
363 has_humongous() &&
364 serial_compaction_point()->has_regions()) {
365 phase2d_prepare_humongous_compaction();
366 }
367 }
368 }
369
370 void G1FullCollector::phase2a_determine_worklists() {
371 GCTraceTime(Debug, gc, phases) debug("Phase 2: Determine work lists", scope()->timer());
372
373 G1DetermineCompactionQueueClosure cl(this);
374 _heap->heap_region_iterate(&cl);
375 }
376
377 bool G1FullCollector::phase2b_forward_oops() {
378 GCTraceTime(Debug, gc, phases) debug("Phase 2: Prepare parallel compaction", scope()->timer());
379
380 G1FullGCPrepareTask task(this);
381 run_task(&task);
382
383 return task.has_free_compaction_targets();
384 }
385
386 uint G1FullCollector::truncate_parallel_cps() {
387 uint lowest_current = UINT_MAX;
388 for (uint i = 0; i < workers(); i++) {
389 G1FullGCCompactionPoint* cp = compaction_point(i);
390 if (cp->has_regions()) {
391 lowest_current = MIN2(lowest_current, cp->current_region()->hrm_index());
392 }
393 }
394
395 for (uint i = 0; i < workers(); i++) {
396 G1FullGCCompactionPoint* cp = compaction_point(i);
397 if (cp->has_regions()) {
398 cp->remove_at_or_above(lowest_current);
399 }
400 }
401 return lowest_current;
402 }
403
404 void G1FullCollector::phase2c_prepare_serial_compaction() {
405 GCTraceTime(Debug, gc, phases) debug("Phase 2: Prepare serial compaction", scope()->timer());
406 // At this point, we know that after parallel compaction there will be regions that
407 // are partially compacted into. Thus, the last compaction region of all
408 // compaction queues still have space in them. We try to re-compact these regions
409 // in serial to avoid a premature OOM when the mutator wants to allocate the first
410 // eden region after gc.
411
412 // For maximum compaction, we need to re-prepare all objects above the lowest
413 // region among the current regions for all thread compaction points. It may
414 // happen that due to the uneven distribution of objects to parallel threads, holes
415 // have been created as threads compact to different target regions between the
416 // lowest and the highest region in the tails of the compaction points.
417
418 uint start_serial = truncate_parallel_cps();
419 assert(start_serial < _heap->max_reserved_regions(), "Called on empty parallel compaction queues");
420
421 G1FullGCCompactionPoint* serial_cp = serial_compaction_point();
422 assert(!serial_cp->is_initialized(), "sanity!");
423
424 G1HeapRegion* start_hr = _heap->region_at(start_serial);
425 serial_cp->add(start_hr);
426 serial_cp->initialize(start_hr);
427
428 HeapWord* dense_prefix_top = compaction_top(start_hr);
429 G1SerialRePrepareClosure re_prepare(serial_cp, dense_prefix_top);
430
431 for (uint i = start_serial + 1; i < _heap->max_reserved_regions(); i++) {
432 if (is_compaction_target(i)) {
433 G1HeapRegion* current = _heap->region_at(i);
434 set_compaction_top(current, current->bottom());
435 serial_cp->add(current);
436 current->apply_to_marked_objects(mark_bitmap(), &re_prepare);
437 }
438 }
439 serial_cp->update();
440 }
441
442 void G1FullCollector::phase2d_prepare_humongous_compaction() {
443 GCTraceTime(Debug, gc, phases) debug("Phase 2: Prepare humongous compaction", scope()->timer());
444 G1FullGCCompactionPoint* serial_cp = serial_compaction_point();
445 assert(serial_cp->has_regions(), "Sanity!" );
446
447 uint last_serial_target = serial_cp->current_region()->hrm_index();
448 uint region_index = last_serial_target + 1;
449 uint max_reserved_regions = _heap->max_reserved_regions();
450
451 G1FullGCCompactionPoint* humongous_cp = humongous_compaction_point();
452
453 while (region_index < max_reserved_regions) {
454 G1HeapRegion* hr = _heap->region_at_or_null(region_index);
455
456 if (hr == nullptr) {
457 region_index++;
458 continue;
459 } else if (hr->is_starts_humongous()) {
460 size_t obj_size = cast_to_oop(hr->bottom())->size();
461 uint num_regions = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size);
462 // Even during last-ditch compaction we should not move pinned humongous objects.
463 if (!hr->has_pinned_objects()) {
464 humongous_cp->forward_humongous(hr);
465 }
466 region_index += num_regions; // Advance over all humongous regions.
467 continue;
468 } else if (is_compaction_target(region_index)) {
469 assert(!hr->has_pinned_objects(), "pinned regions should not be compaction targets");
470 // Add the region to the humongous compaction point.
471 humongous_cp->add(hr);
472 }
473 region_index++;
474 }
475 }
476
477 void G1FullCollector::phase3_adjust_pointers() {
478 // Adjust the pointers to reflect the new locations
479 GCTraceTime(Info, gc, phases) info("Phase 3: Adjust pointers", scope()->timer());
480
481 G1FullGCAdjustTask task(this);
482 run_task(&task);
483 }
484
485 void G1FullCollector::phase4_do_compaction() {
486 // Compact the heap using the compaction queues created in phase 2.
487 GCTraceTime(Info, gc, phases) info("Phase 4: Compact heap", scope()->timer());
488 G1FullGCCompactTask task(this);
489 run_task(&task);
490
491 // Serial compact to avoid OOM when very few free regions.
492 if (serial_compaction_point()->has_regions()) {
493 task.serial_compaction();
494 }
495
496 if (!_humongous_compaction_regions.is_empty()) {
497 assert(scope()->do_maximal_compaction(), "Only compact humongous during maximal compaction");
498 task.humongous_compaction();
499 }
500 }
501
502 void G1FullCollector::phase5_reset_metadata() {
503 // Clear region metadata that is invalid after GC for all regions.
504 GCTraceTime(Info, gc, phases) info("Phase 5: Reset Metadata", scope()->timer());
505 G1FullGCResetMetadataTask task(this);
506 run_task(&task);
507 }
508
509 void G1FullCollector::restore_marks() {
510 _preserved_marks_set.restore(_heap->workers());
511 _preserved_marks_set.reclaim();
512 }
513
514 void G1FullCollector::run_task(WorkerTask* task) {
515 _heap->workers()->run_task(task, _num_workers);
516 }
517
518 void G1FullCollector::verify_after_marking() {
519 if (!VerifyDuringGC || !_heap->verifier()->should_verify(G1HeapVerifier::G1VerifyFull)) {
520 // Only do verification if VerifyDuringGC and G1VerifyFull is set.
521 return;
522 }
523
524 #if COMPILER2_OR_JVMCI
525 DerivedPointerTableDeactivate dpt_deact;
526 #endif
527 _heap->prepare_for_verify();
528 // Note: we can verify only the heap here. When an object is
529 // marked, the previous value of the mark word (including
530 // identity hash values, ages, etc) is preserved, and the mark
531 // word is set to markWord::marked_value - effectively removing
532 // any hash values from the mark word. These hash values are
533 // used when verifying the dictionaries and so removing them
534 // from the mark word can make verification of the dictionaries
535 // fail. At the end of the GC, the original mark word values
536 // (including hash values) are restored to the appropriate
537 // objects.
538 GCTraceTime(Info, gc, verify) tm("Verifying During GC (full)");
539 _heap->verify(VerifyOption::G1UseFullMarking);
540 }