1 /*
  2  * Copyright (c) 2017, 2025, 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 "classfile/classLoaderDataGraph.hpp"
 26 #include "gc/g1/g1CollectedHeap.hpp"
 27 #include "gc/g1/g1FullCollector.inline.hpp"
 28 #include "gc/g1/g1FullGCAdjustTask.hpp"
 29 #include "gc/g1/g1FullGCCompactTask.hpp"
 30 #include "gc/g1/g1FullGCMarker.inline.hpp"
 31 #include "gc/g1/g1FullGCMarkTask.hpp"
 32 #include "gc/g1/g1FullGCPrepareTask.inline.hpp"
 33 #include "gc/g1/g1FullGCResetMetadataTask.hpp"
 34 #include "gc/g1/g1FullGCScope.hpp"
 35 #include "gc/g1/g1OopClosures.hpp"
 36 #include "gc/g1/g1Policy.hpp"
 37 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
 38 #include "gc/shared/classUnloadingContext.hpp"
 39 #include "gc/shared/gcTraceTime.inline.hpp"
 40 #include "gc/shared/preservedMarks.inline.hpp"
 41 #include "gc/shared/referenceProcessor.hpp"
 42 #include "gc/shared/verifyOption.hpp"
 43 #include "gc/shared/weakProcessor.inline.hpp"
 44 #include "gc/shared/workerPolicy.hpp"
 45 #include "logging/log.hpp"
 46 #include "runtime/handles.inline.hpp"
 47 #include "utilities/debug.hpp"
 48 
 49 static void clear_and_activate_derived_pointers() {
 50 #if COMPILER2_OR_JVMCI
 51   DerivedPointerTable::clear();
 52 #endif
 53 }
 54 
 55 static void deactivate_derived_pointers() {
 56 #if COMPILER2_OR_JVMCI
 57   DerivedPointerTable::set_active(false);
 58 #endif
 59 }
 60 
 61 static void update_derived_pointers() {
 62 #if COMPILER2_OR_JVMCI
 63   DerivedPointerTable::update_pointers();
 64 #endif
 65 }
 66 
 67 G1CMBitMap* G1FullCollector::mark_bitmap() {
 68   return _heap->concurrent_mark()->mark_bitmap();
 69 }
 70 
 71 ReferenceProcessor* G1FullCollector::reference_processor() {
 72   return _heap->ref_processor_stw();
 73 }
 74 
 75 uint G1FullCollector::calc_active_workers() {
 76   G1CollectedHeap* heap = G1CollectedHeap::heap();
 77   uint max_worker_count = heap->workers()->max_workers();
 78   // Only calculate number of workers if UseDynamicNumberOfGCThreads
 79   // is enabled, otherwise use max.
 80   if (!UseDynamicNumberOfGCThreads) {
 81     return max_worker_count;
 82   }
 83 
 84   // Consider G1HeapWastePercent to decide max number of workers. Each worker
 85   // will in average cause half a region waste.
 86   uint max_wasted_regions_allowed = ((heap->num_committed_regions() * G1HeapWastePercent) / 100);
 87   uint waste_worker_count = MAX2((max_wasted_regions_allowed * 2) , 1u);
 88   uint heap_waste_worker_limit = MIN2(waste_worker_count, max_worker_count);
 89 
 90   // Also consider HeapSizePerGCThread by calling WorkerPolicy to calculate
 91   // the number of workers.
 92   uint current_active_workers = heap->workers()->active_workers();
 93   uint active_worker_limit = WorkerPolicy::calc_active_workers(max_worker_count, current_active_workers, 0);
 94 
 95   // Finally consider the amount of used regions.
 96   uint used_worker_limit = heap->num_used_regions();
 97   assert(used_worker_limit > 0, "Should never have zero used regions.");
 98 
 99   // Update active workers to the lower of the limits.
100   uint worker_count = MIN3(heap_waste_worker_limit, active_worker_limit, used_worker_limit);
101   log_debug(gc, task)("Requesting %u active workers for full compaction (waste limited workers: %u, "
102                       "adaptive workers: %u, used limited workers: %u)",
103                       worker_count, heap_waste_worker_limit, active_worker_limit, used_worker_limit);
104   worker_count = heap->workers()->set_active_workers(worker_count);
105   log_info(gc, task)("Using %u workers of %u for full compaction", worker_count, max_worker_count);
106 
107   return worker_count;
108 }
109 
110 G1FullCollector::G1FullCollector(G1CollectedHeap* heap,
111                                  bool clear_soft_refs,
112                                  bool do_maximal_compaction,
113                                  G1FullGCTracer* tracer) :
114     _heap(heap),
115     _scope(heap->monitoring_support(), clear_soft_refs, do_maximal_compaction, tracer),
116     _num_workers(calc_active_workers()),
117     _has_compaction_targets(false),
118     _has_humongous(false),
119     _oop_queue_set(_num_workers),
120     _array_queue_set(_num_workers),
121     _preserved_marks_set(true),
122     _serial_compaction_point(this, nullptr),
123     _humongous_compaction_point(this, nullptr),
124     _is_alive(this, heap->concurrent_mark()->mark_bitmap()),
125     _is_alive_mutator(heap->ref_processor_stw(), &_is_alive),
126     _humongous_compaction_regions(8),
127     _always_subject_to_discovery(),
128     _is_subject_mutator(heap->ref_processor_stw(), &_always_subject_to_discovery),
129     _region_attr_table() {
130   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
131 
132   _preserved_marks_set.init(_num_workers);
133   _markers = NEW_C_HEAP_ARRAY(G1FullGCMarker*, _num_workers, mtGC);
134   _compaction_points = NEW_C_HEAP_ARRAY(G1FullGCCompactionPoint*, _num_workers, mtGC);
135 
136   _live_stats = NEW_C_HEAP_ARRAY(G1RegionMarkStats, _heap->max_num_regions(), mtGC);
137   _compaction_tops = NEW_C_HEAP_ARRAY(HeapWord*, _heap->max_num_regions(), mtGC);
138   for (uint j = 0; j < heap->max_num_regions(); j++) {
139     _live_stats[j].clear();
140     _compaction_tops[j] = nullptr;
141   }
142 
143   for (uint i = 0; i < _num_workers; i++) {
144     _markers[i] = new G1FullGCMarker(this, i, _live_stats);
145     _compaction_points[i] = new G1FullGCCompactionPoint(this, _preserved_marks_set.get(i));
146     _oop_queue_set.register_queue(i, marker(i)->oop_stack());
147     _array_queue_set.register_queue(i, marker(i)->objarray_stack());
148   }
149   _serial_compaction_point.set_preserved_stack(_preserved_marks_set.get(0));
150   _humongous_compaction_point.set_preserved_stack(_preserved_marks_set.get(0));
151   _region_attr_table.initialize(heap->reserved(), G1HeapRegion::GrainBytes);
152 }
153 
154 G1FullCollector::~G1FullCollector() {
155   for (uint i = 0; i < _num_workers; i++) {
156     delete _markers[i];
157     delete _compaction_points[i];
158   }
159 
160   FREE_C_HEAP_ARRAY(G1FullGCMarker*, _markers);
161   FREE_C_HEAP_ARRAY(G1FullGCCompactionPoint*, _compaction_points);
162   FREE_C_HEAP_ARRAY(HeapWord*, _compaction_tops);
163   FREE_C_HEAP_ARRAY(G1RegionMarkStats, _live_stats);
164 }
165 
166 class PrepareRegionsClosure : public G1HeapRegionClosure {
167   G1FullCollector* _collector;
168 
169 public:
170   PrepareRegionsClosure(G1FullCollector* collector) : _collector(collector) { }
171 
172   bool do_heap_region(G1HeapRegion* hr) {
173     hr->prepare_for_full_gc();
174     G1CollectedHeap::heap()->prepare_region_for_full_compaction(hr);
175     _collector->before_marking_update_attribute_table(hr);
176     return false;
177   }
178 };
179 
180 void G1FullCollector::prepare_collection() {
181   _heap->policy()->record_full_collection_start();
182 
183   // Verification needs the bitmap, so we should clear the bitmap only later.
184   bool in_concurrent_cycle = _heap->abort_concurrent_cycle();
185   _heap->verify_before_full_collection();
186   if (in_concurrent_cycle) {
187     GCTraceTime(Debug, gc) debug("Clear Bitmap");
188     _heap->concurrent_mark()->clear_bitmap(_heap->workers());
189   }
190 
191   _heap->gc_prologue(true);
192   _heap->retire_tlabs();
193   _heap->flush_region_pin_cache();
194   _heap->prepare_heap_for_full_collection();
195 
196   PrepareRegionsClosure cl(this);
197   _heap->heap_region_iterate(&cl);
198 
199   reference_processor()->start_discovery(scope()->should_clear_soft_refs());
200 
201   // Clear and activate derived pointer collection.
202   clear_and_activate_derived_pointers();
203 }
204 
205 void G1FullCollector::collect() {
206   G1CollectedHeap::start_codecache_marking_cycle_if_inactive(false /* concurrent_mark_start */);
207 
208   phase1_mark_live_objects();
209   verify_after_marking();
210 
211   // Don't add any more derived pointers during later phases
212   deactivate_derived_pointers();
213 
214   phase2_prepare_compaction();
215 
216   if (has_compaction_targets()) {
217     phase3_adjust_pointers();
218 
219     phase4_do_compaction();
220   } else {
221     // All regions have a high live ratio thus will not be compacted.
222     // The live ratio is only considered if do_maximal_compaction is false.
223     log_info(gc, phases) ("No Regions selected for compaction. Skipping Phase 3: Adjust pointers and Phase 4: Compact heap");
224   }
225 
226   phase5_reset_metadata();
227 }
228 
229 void G1FullCollector::complete_collection(size_t allocation_word_size) {
230   // Restore all marks.
231   restore_marks();
232 
233   // When the pointers have been adjusted and moved, we can
234   // update the derived pointer table.
235   update_derived_pointers();
236 
237   // Need completely cleared claim bits for the next concurrent marking or full gc.
238   ClassLoaderDataGraph::clear_claimed_marks();
239 
240   // Prepare the bitmap for the next (potentially concurrent) marking.
241   _heap->concurrent_mark()->clear_bitmap(_heap->workers());
242 
243   _heap->prepare_for_mutator_after_full_collection(allocation_word_size);
244 
245   _heap->resize_all_tlabs();
246 
247   _heap->policy()->record_full_collection_end(allocation_word_size);
248   _heap->gc_epilogue(true);
249 
250   _heap->verify_after_full_collection();
251 
252   _heap->print_heap_after_full_collection();
253 }
254 
255 void G1FullCollector::before_marking_update_attribute_table(G1HeapRegion* hr) {
256   if (hr->is_free()) {
257     _region_attr_table.set_free(hr->hrm_index());
258   } else if (hr->is_humongous() || hr->has_pinned_objects()) {
259     // Humongous objects or pinned regions will never be moved in the "main"
260     // compaction phase, but non-pinned regions might afterwards in a special phase.
261     _region_attr_table.set_skip_compacting(hr->hrm_index());
262   } else {
263     // Everything else should be compacted.
264     _region_attr_table.set_compacting(hr->hrm_index());
265   }
266 }
267 
268 class G1FullGCRefProcProxyTask : public RefProcProxyTask {
269   G1FullCollector& _collector;
270 
271 public:
272   G1FullGCRefProcProxyTask(G1FullCollector &collector, uint max_workers)
273     : RefProcProxyTask("G1FullGCRefProcProxyTask", max_workers),
274       _collector(collector) {}
275 
276   void work(uint worker_id) override {
277     assert(worker_id < _max_workers, "sanity");
278     G1IsAliveClosure is_alive(&_collector);
279     uint index = (_tm == RefProcThreadModel::Single) ? 0 : worker_id;
280     G1FullKeepAliveClosure keep_alive(_collector.marker(index));
281     BarrierEnqueueDiscoveredFieldClosure enqueue;
282     G1FollowStackClosure* complete_gc = _collector.marker(index)->stack_closure();
283     _rp_task->rp_work(worker_id, &is_alive, &keep_alive, &enqueue, complete_gc);
284   }
285 };
286 
287 void G1FullCollector::phase1_mark_live_objects() {
288   // Recursively traverse all live objects and mark them.
289   GCTraceTime(Info, gc, phases) info("Phase 1: Mark live objects", scope()->timer());
290 
291   {
292     // Do the actual marking.
293     G1FullGCMarkTask marking_task(this);
294     run_task(&marking_task);
295   }
296 
297   {
298     GCTraceTime(Debug, gc, phases) debug("Phase 1: Reference Processing", scope()->timer());
299     // Process reference objects found during marking.
300     ReferenceProcessorPhaseTimes pt(scope()->timer(), reference_processor()->max_num_queues());
301     G1FullGCRefProcProxyTask task(*this, reference_processor()->max_num_queues());
302     const ReferenceProcessorStats& stats = reference_processor()->process_discovered_references(task, _heap->workers(), pt);
303     scope()->tracer()->report_gc_reference_stats(stats);
304     pt.print_all_references();
305     assert(marker(0)->oop_stack()->is_empty(), "Should be no oops on the stack");
306   }
307 
308   {
309     GCTraceTime(Debug, gc, phases) debug("Phase 1: Flush Mark Stats Cache", scope()->timer());
310     for (uint i = 0; i < workers(); i++) {
311       marker(i)->flush_mark_stats_cache();
312     }
313   }
314 
315   // Weak oops cleanup.
316   {
317     GCTraceTime(Debug, gc, phases) debug("Phase 1: Weak Processing", scope()->timer());
318     WeakProcessor::weak_oops_do(_heap->workers(), &_is_alive, &do_nothing_cl, 1);
319   }
320 
321   // Class unloading and cleanup.
322   if (ClassUnloading) {
323     _heap->unload_classes_and_code("Phase 1: Class Unloading and Cleanup", &_is_alive, scope()->timer());
324   }
325 
326   {
327     GCTraceTime(Debug, gc, phases) debug("Report Object Count", scope()->timer());
328     scope()->tracer()->report_object_count_after_gc(&_is_alive, _heap->workers());
329   }
330 #if TASKQUEUE_STATS
331   oop_queue_set()->print_and_reset_taskqueue_stats("Oop Queue");
332   array_queue_set()->print_and_reset_taskqueue_stats("ObjArrayOop Queue");
333 #endif
334 }
335 
336 void G1FullCollector::phase2_prepare_compaction() {
337   GCTraceTime(Info, gc, phases) info("Phase 2: Prepare compaction", scope()->timer());
338 
339   phase2a_determine_worklists();
340 
341   if (!has_compaction_targets()) {
342     return;
343   }
344 
345   bool has_free_compaction_targets = phase2b_forward_oops();
346 
347   // Try to avoid OOM immediately after Full GC in case there are no free regions
348   // left after determining the result locations (i.e. this phase). Prepare to
349   // maximally compact the tail regions of the compaction queues serially.
350   if (scope()->do_maximal_compaction() || !has_free_compaction_targets) {
351     phase2c_prepare_serial_compaction();
352 
353     if (scope()->do_maximal_compaction() &&
354         has_humongous() &&
355         serial_compaction_point()->has_regions()) {
356       phase2d_prepare_humongous_compaction();
357     }
358   }
359 }
360 
361 void G1FullCollector::phase2a_determine_worklists() {
362   GCTraceTime(Debug, gc, phases) debug("Phase 2: Determine work lists", scope()->timer());
363 
364   G1DetermineCompactionQueueClosure cl(this);
365   _heap->heap_region_iterate(&cl);
366 }
367 
368 bool G1FullCollector::phase2b_forward_oops() {
369   GCTraceTime(Debug, gc, phases) debug("Phase 2: Prepare parallel compaction", scope()->timer());
370 
371   G1FullGCPrepareTask task(this);
372   run_task(&task);
373 
374   return task.has_free_compaction_targets();
375 }
376 
377 uint G1FullCollector::truncate_parallel_cps() {
378   uint lowest_current = UINT_MAX;
379   for (uint i = 0; i < workers(); i++) {
380     G1FullGCCompactionPoint* cp = compaction_point(i);
381     if (cp->has_regions()) {
382       lowest_current = MIN2(lowest_current, cp->current_region()->hrm_index());
383     }
384   }
385 
386   for (uint i = 0; i < workers(); i++) {
387     G1FullGCCompactionPoint* cp = compaction_point(i);
388     if (cp->has_regions()) {
389       cp->remove_at_or_above(lowest_current);
390     }
391   }
392   return lowest_current;
393 }
394 
395 void G1FullCollector::phase2c_prepare_serial_compaction() {
396   GCTraceTime(Debug, gc, phases) debug("Phase 2: Prepare serial compaction", scope()->timer());
397   // At this point, we know that after parallel compaction there will be regions that
398   // are partially compacted into. Thus, the last compaction region of all
399   // compaction queues still have space in them. We try to re-compact these regions
400   // in serial to avoid a premature OOM when the mutator wants to allocate the first
401   // eden region after gc.
402 
403   // For maximum compaction, we need to re-prepare all objects above the lowest
404   // region among the current regions for all thread compaction points. It may
405   // happen that due to the uneven distribution of objects to parallel threads, holes
406   // have been created as threads compact to different target regions between the
407   // lowest and the highest region in the tails of the compaction points.
408 
409   uint start_serial = truncate_parallel_cps();
410   assert(start_serial < _heap->max_num_regions(), "Called on empty parallel compaction queues");
411 
412   G1FullGCCompactionPoint* serial_cp = serial_compaction_point();
413   assert(!serial_cp->is_initialized(), "sanity!");
414 
415   G1HeapRegion* start_hr = _heap->region_at(start_serial);
416   serial_cp->add(start_hr);
417   serial_cp->initialize(start_hr);
418 
419   HeapWord* dense_prefix_top = compaction_top(start_hr);
420   G1SerialRePrepareClosure re_prepare(serial_cp, dense_prefix_top);
421 
422   for (uint i = start_serial + 1; i < _heap->max_num_regions(); i++) {
423     if (is_compaction_target(i)) {
424       G1HeapRegion* current = _heap->region_at(i);
425       set_compaction_top(current, current->bottom());
426       serial_cp->add(current);
427       current->apply_to_marked_objects(mark_bitmap(), &re_prepare);
428     }
429   }
430   serial_cp->update();
431 }
432 
433 void G1FullCollector::phase2d_prepare_humongous_compaction() {
434   GCTraceTime(Debug, gc, phases) debug("Phase 2: Prepare humongous compaction", scope()->timer());
435   G1FullGCCompactionPoint* serial_cp = serial_compaction_point();
436   assert(serial_cp->has_regions(), "Sanity!" );
437 
438   uint last_serial_target = serial_cp->current_region()->hrm_index();
439   uint region_index = last_serial_target + 1;
440   uint max_num_regions = _heap->max_num_regions();
441 
442   G1FullGCCompactionPoint* humongous_cp = humongous_compaction_point();
443 
444   while (region_index < max_num_regions) {
445     G1HeapRegion* hr = _heap->region_at_or_null(region_index);
446 
447     if (hr == nullptr) {
448       region_index++;
449       continue;
450     } else if (hr->is_starts_humongous()) {
451       size_t obj_size = cast_to_oop(hr->bottom())->size();
452       uint num_regions = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size);
453       // Even during last-ditch compaction we should not move pinned humongous objects.
454       if (!hr->has_pinned_objects()) {
455         humongous_cp->forward_humongous(hr);
456       }
457       region_index += num_regions; // Advance over all humongous regions.
458       continue;
459     } else if (is_compaction_target(region_index)) {
460       assert(!hr->has_pinned_objects(), "pinned regions should not be compaction targets");
461       // Add the region to the humongous compaction point.
462       humongous_cp->add(hr);
463     }
464     region_index++;
465   }
466 }
467 
468 void G1FullCollector::phase3_adjust_pointers() {
469   // Adjust the pointers to reflect the new locations
470   GCTraceTime(Info, gc, phases) info("Phase 3: Adjust pointers", scope()->timer());
471 
472   G1FullGCAdjustTask task(this);
473   run_task(&task);
474 }
475 
476 void G1FullCollector::phase4_do_compaction() {
477   // Compact the heap using the compaction queues created in phase 2.
478   GCTraceTime(Info, gc, phases) info("Phase 4: Compact heap", scope()->timer());
479   G1FullGCCompactTask task(this);
480   run_task(&task);
481 
482   // Serial compact to avoid OOM when very few free regions.
483   if (serial_compaction_point()->has_regions()) {
484     task.serial_compaction();
485   }
486 
487   if (!_humongous_compaction_regions.is_empty()) {
488     assert(scope()->do_maximal_compaction(), "Only compact humongous during maximal compaction");
489     task.humongous_compaction();
490   }
491 }
492 
493 void G1FullCollector::phase5_reset_metadata() {
494   // Clear region metadata that is invalid after GC for all regions.
495   GCTraceTime(Info, gc, phases) info("Phase 5: Reset Metadata", scope()->timer());
496   G1FullGCResetMetadataTask task(this);
497   run_task(&task);
498 }
499 
500 void G1FullCollector::restore_marks() {
501   _preserved_marks_set.restore(_heap->workers());
502   _preserved_marks_set.reclaim();
503 }
504 
505 void G1FullCollector::run_task(WorkerTask* task) {
506   _heap->workers()->run_task(task, _num_workers);
507 }
508 
509 void G1FullCollector::verify_after_marking() {
510   if (!VerifyDuringGC || !_heap->verifier()->should_verify(G1HeapVerifier::G1VerifyFull)) {
511     // Only do verification if VerifyDuringGC and G1VerifyFull is set.
512     return;
513   }
514 
515 #if COMPILER2_OR_JVMCI
516   DerivedPointerTableDeactivate dpt_deact;
517 #endif
518   _heap->prepare_for_verify();
519   // Note: we can verify only the heap here. When an object is
520   // marked, the previous value of the mark word (including
521   // identity hash values, ages, etc) is preserved, and the mark
522   // word is set to markWord::marked_value - effectively removing
523   // any hash values from the mark word. These hash values are
524   // used when verifying the dictionaries and so removing them
525   // from the mark word can make verification of the dictionaries
526   // fail. At the end of the GC, the original mark word values
527   // (including hash values) are restored to the appropriate
528   // objects.
529   GCTraceTime(Info, gc, verify) tm("Verifying During GC (full)");
530   _heap->verify(VerifyOption::G1UseFullMarking);
531 }