1 /*
2 * Copyright (c) 2014, 2021, Red Hat, Inc. All rights reserved.
3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
4 * Copyright (c) 2025, 2026, Oracle and/or its affiliates. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27
28 #include "compiler/oopMap.hpp"
29 #include "gc/shared/continuationGCSupport.hpp"
30 #include "gc/shared/fullGCForwarding.inline.hpp"
31 #include "gc/shared/gcTraceTime.inline.hpp"
32 #include "gc/shared/preservedMarks.inline.hpp"
33 #include "gc/shared/tlab_globals.hpp"
34 #include "gc/shared/workerThread.hpp"
35 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
36 #include "gc/shenandoah/shenandoahClosures.inline.hpp"
37 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
38 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
39 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
40 #include "gc/shenandoah/shenandoahFreeSet.hpp"
41 #include "gc/shenandoah/shenandoahFullGC.hpp"
42 #include "gc/shenandoah/shenandoahGenerationalFullGC.hpp"
43 #include "gc/shenandoah/shenandoahGlobalGeneration.hpp"
44 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
45 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
46 #include "gc/shenandoah/shenandoahHeapRegionClosures.hpp"
47 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
48 #include "gc/shenandoah/shenandoahMark.inline.hpp"
49 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
50 #include "gc/shenandoah/shenandoahMetrics.hpp"
51 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
52 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
53 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
54 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
55 #include "gc/shenandoah/shenandoahSTWMark.hpp"
56 #include "gc/shenandoah/shenandoahUtils.hpp"
57 #include "gc/shenandoah/shenandoahVerifier.hpp"
58 #include "gc/shenandoah/shenandoahVMOperations.hpp"
59 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
60 #include "memory/metaspaceUtils.hpp"
61 #include "memory/universe.hpp"
62 #include "oops/compressedOops.inline.hpp"
63 #include "oops/oop.inline.hpp"
64 #include "runtime/orderAccess.hpp"
65 #include "runtime/vmThread.hpp"
66 #include "utilities/copy.hpp"
67 #include "utilities/events.hpp"
68 #include "utilities/growableArray.hpp"
69
70 ShenandoahFullGC::ShenandoahFullGC() :
71 ShenandoahGC(ShenandoahHeap::heap()->global_generation()),
72 _gc_timer(ShenandoahHeap::heap()->gc_timer()),
73 _preserved_marks(new PreservedMarksSet(true)) {}
74
75 ShenandoahFullGC::~ShenandoahFullGC() {
76 delete _preserved_marks;
77 }
78
79 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
80 vmop_entry_full(cause);
81 // Always success
82 return true;
83 }
84
85 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
86 ShenandoahHeap* const heap = ShenandoahHeap::heap();
87 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
88 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
89
90 heap->try_inject_alloc_failure();
91 VM_ShenandoahFullGC op(cause, this);
92 VMThread::execute(&op);
93 }
94
95 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
96 static const char* msg = "Pause Full";
97 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
98 EventMark em("%s", msg);
99
100 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
101 ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
102 "full gc");
103
104 op_full(cause);
105 }
106
107 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
108 ShenandoahHeap* const heap = ShenandoahHeap::heap();
109
110 ShenandoahMetricsSnapshot metrics(heap->free_set());
111
112 // Perform full GC
113 do_it(cause);
114
115 if (heap->mode()->is_generational()) {
116 ShenandoahGenerationalFullGC::handle_completion(heap);
117 }
118
119 if (metrics.is_good_progress()) {
120 heap->notify_gc_progress();
121 } else {
122 // Nothing to do. Tell the allocation path that we have failed to make
123 // progress, and it can finally fail.
124 heap->notify_gc_no_progress();
125 }
126
127 // Regardless if progress was made, we record that we completed a "successful" full GC.
128 _generation->heuristics()->record_success_full();
129 heap->shenandoah_policy()->record_success_full();
130
131 // Leaving full GC, we need to flip barriers back to idle.
132 ShenandoahCodeRoots::arm_nmethods();
133
134 {
135 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_propagate_gc_state);
136 heap->propagate_gc_state_to_all_threads();
137 }
138 }
139
140 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
141 ShenandoahHeap* heap = ShenandoahHeap::heap();
142 heap->release_injected_pins();
143
144 // A full GC may be entered directly, or as an upgrade from a failed
145 // degenerated GC. In the latter case, self-forwarded objects may be
146 // present from the failed evacuation. Drain those marks before any phase
147 // (verify, update_roots, phase1_mark_heap) walks headers.
148 {
149 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_un_self_forward);
150 heap->un_self_forward_cset_regions();
151 }
152
153 if (heap->mode()->is_generational()) {
154 ShenandoahGenerationalFullGC::prepare();
155 }
156
157 if (ShenandoahVerify) {
158 heap->verifier()->verify_before_fullgc(_generation);
159 }
160
161 if (VerifyBeforeGC) {
162 Universe::verify();
163 }
164
165 // Degenerated GC may carry concurrent root flags when upgrading to
166 // full GC. We need to reset it before mutators resume.
167 heap->set_concurrent_strong_root_in_progress(false);
168 heap->set_concurrent_weak_root_in_progress(false);
169
170 heap->set_full_gc_in_progress(true);
171
172 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
173 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
174
175 {
176 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
177 heap->pre_full_gc_dump(_gc_timer);
178 }
179
180 {
181 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
182 // Full GC is supposed to recover from any GC state:
183
184 // a0. Remember if we have forwarded objects
185 bool has_forwarded_objects = heap->has_forwarded_objects();
186
187 // a1. Cancel evacuation, if in progress
188 if (heap->is_evacuation_in_progress()) {
189 heap->set_evacuation_in_progress(false);
190 }
191 assert(!heap->is_evacuation_in_progress(), "sanity");
192
193 // a2. Cancel update-refs, if in progress
194 if (heap->is_update_refs_in_progress()) {
195 heap->set_update_refs_in_progress(false);
196 }
197 assert(!heap->is_update_refs_in_progress(), "sanity");
198
199 // b. Cancel all concurrent marks, if in progress
200 if (heap->is_concurrent_mark_in_progress()) {
201 heap->cancel_concurrent_mark();
202 }
203 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
204
205 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
206 if (has_forwarded_objects) {
207 update_roots(true /*full_gc*/);
208 }
209
210 // d. Abandon reference discovery and clear all discovered references.
211 ShenandoahReferenceProcessor* rp = _generation->ref_processor();
212 rp->abandon_partial_discovery();
213
214 // e. Sync pinned region status from the CP marks
215 heap->sync_pinned_region_status();
216
217 if (heap->mode()->is_generational()) {
218 ShenandoahGenerationalFullGC::restore_top_before_promote(heap);
219 }
220
221 // The rest of prologue:
222 _preserved_marks->init(heap->workers()->active_workers());
223
224 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
225 }
226
227 if (UseTLAB) {
228 // Note: PLABs are also retired with GCLABs in generational mode.
229 heap->gclabs_retire(ResizeTLAB);
230 heap->tlabs_retire(ResizeTLAB);
231 }
232
233 OrderAccess::fence();
234
235 phase1_mark_heap();
236
237 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
238 // Coming out of Full GC, we would not have any forwarded objects.
239 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
240 heap->set_has_forwarded_objects(false);
241
242 heap->set_full_gc_move_in_progress(true);
243
244 // Setup workers for the rest
245 OrderAccess::fence();
246
247 // Initialize worker slices
248 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
249 for (uint i = 0; i < heap->max_workers(); i++) {
250 worker_slices[i] = new ShenandoahHeapRegionSet();
251 }
252
253 {
254 // The rest of code performs region moves, where region status is undefined
255 // until all phases run together.
256 ShenandoahHeapLocker lock(heap->lock());
257
258 phase2_calculate_target_addresses(worker_slices);
259
260 OrderAccess::fence();
261
262 phase3_update_references();
263
264 phase4_compact_objects(worker_slices);
265
266 phase5_epilog();
267 }
268 heap->start_idle_span();
269
270 // Resize metaspace
271 MetaspaceGC::compute_new_size();
272
273 // Free worker slices
274 for (uint i = 0; i < heap->max_workers(); i++) {
275 delete worker_slices[i];
276 }
277 FREE_C_HEAP_ARRAY(worker_slices);
278
279 heap->set_full_gc_move_in_progress(false);
280 heap->set_full_gc_in_progress(false);
281
282 DEBUG_ONLY(heap->assert_no_self_forwards());
283
284 if (ShenandoahVerify) {
285 heap->verifier()->verify_after_fullgc(_generation);
286 }
287
288 if (VerifyAfterGC) {
289 Universe::verify();
290 }
291
292 {
293 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
294 heap->post_full_gc_dump(_gc_timer);
295 }
296 }
297
298 void ShenandoahFullGC::phase1_mark_heap() {
299 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
300 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
301
302 ShenandoahHeap* heap = ShenandoahHeap::heap();
303
304 _generation->reset_mark_bitmap<true, true>();
305 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
306 assert(!_generation->is_mark_complete(), "sanity");
307
308 heap->set_unload_classes(_generation->heuristics()->can_unload_classes());
309
310 ShenandoahReferenceProcessor* rp = _generation->ref_processor();
311 // enable ("weak") refs discovery
312 rp->set_soft_reference_policy(true); // forcefully purge all soft references
313
314 ShenandoahSTWMark mark(_generation, true /*full_gc*/);
315 mark.mark();
316 heap->parallel_cleaning(_generation, true /* full_gc */);
317
318 if (ShenandoahHeap::heap()->mode()->is_generational()) {
319 ShenandoahGenerationalFullGC::log_live_in_old(heap);
320 }
321 }
322
323 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
324 private:
325 PreservedMarks* const _preserved_marks;
326 ShenandoahHeap* const _heap;
327 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
328 int _empty_regions_pos;
329 ShenandoahHeapRegion* _to_region;
330 ShenandoahHeapRegion* _from_region;
331 HeapWord* _compact_point;
332
333 public:
334 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
335 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
336 ShenandoahHeapRegion* to_region) :
337 _preserved_marks(preserved_marks),
338 _heap(ShenandoahHeap::heap()),
339 _empty_regions(empty_regions),
340 _empty_regions_pos(0),
341 _to_region(to_region),
342 _from_region(nullptr),
343 _compact_point(to_region->bottom()) {}
344
345 void set_from_region(ShenandoahHeapRegion* from_region) {
346 _from_region = from_region;
347 }
348
349 void finish() {
350 assert(_to_region != nullptr, "should not happen");
351 _to_region->set_new_top(_compact_point);
352 }
353
354 bool is_compact_same_region() {
355 return _from_region == _to_region;
356 }
357
358 int empty_regions_pos() {
359 return _empty_regions_pos;
360 }
361
362 void do_object(oop p) override {
363 shenandoah_assert_mark_complete(cast_from_oop<HeapWord*>(p));
364 assert(_from_region != nullptr, "must set before work");
365 assert(_heap->global_generation()->is_mark_complete(), "marking must be finished");
366 assert(_heap->marking_context()->is_marked(p), "must be marked");
367 assert(!_heap->marking_context()->allocated_after_mark_start(p), "must be truly marked");
368
369 size_t obj_size = p->size();
370 if (_compact_point + obj_size > _to_region->end()) {
371 finish();
372
373 // Object doesn't fit. Pick next empty region and start compacting there.
374 ShenandoahHeapRegion* new_to_region;
375 if (_empty_regions_pos < _empty_regions.length()) {
376 new_to_region = _empty_regions.at(_empty_regions_pos);
377 _empty_regions_pos++;
378 } else {
379 // Out of empty region? Compact within the same region.
380 new_to_region = _from_region;
381 }
382
383 assert(new_to_region != _to_region, "must not reuse same to-region");
384 assert(new_to_region != nullptr, "must not be null");
385 _to_region = new_to_region;
386 _compact_point = _to_region->bottom();
387 }
388
389 // Object fits into current region, record new location, if object does not move:
390 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
391 shenandoah_assert_not_forwarded(nullptr, p);
392 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
393 _preserved_marks->push_if_necessary(p, p->mark());
394 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
395 }
396 _compact_point += obj_size;
397 }
398 };
399
400 class ShenandoahPrepareForCompactionTask : public WorkerTask {
401 private:
402 PreservedMarksSet* const _preserved_marks;
403 ShenandoahHeap* const _heap;
404 ShenandoahHeapRegionSet** const _worker_slices;
405
406 public:
407 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
408 WorkerTask("Shenandoah Prepare For Compaction"),
409 _preserved_marks(preserved_marks),
410 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
411 }
412
413 static bool is_candidate_region(ShenandoahHeapRegion* r) {
414 // Empty region: get it into the slice to defragment the slice itself.
415 // We could have skipped this without violating correctness, but we really
416 // want to compact all live regions to the start of the heap, which sometimes
417 // means moving them into the fully empty regions.
418 if (r->is_empty()) return true;
419
420 // Can move the region, and this is not the humongous region. Humongous
421 // moves are special cased here, because their moves are handled separately.
422 return r->is_stw_move_allowed() && !r->is_humongous();
423 }
424
425 void work(uint worker_id) override;
426 private:
427 template<typename ClosureType>
428 void prepare_for_compaction(ClosureType& cl,
429 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
430 ShenandoahHeapRegionSetIterator& it,
431 ShenandoahHeapRegion* from_region);
432 };
433
434 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
435 ShenandoahParallelWorkerSession worker_session(worker_id);
436 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
437 ShenandoahHeapRegionSetIterator it(slice);
438 ShenandoahHeapRegion* from_region = it.next();
439 // No work?
440 if (from_region == nullptr) {
441 return;
442 }
443
444 // Sliding compaction. Walk all regions in the slice, and compact them.
445 // Remember empty regions and reuse them as needed.
446 ResourceMark rm;
447
448 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
449
450 if (_heap->mode()->is_generational()) {
451 ShenandoahPrepareForGenerationalCompactionObjectClosure cl(_preserved_marks->get(worker_id),
452 empty_regions, from_region, worker_id);
453 prepare_for_compaction(cl, empty_regions, it, from_region);
454 } else {
455 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
456 prepare_for_compaction(cl, empty_regions, it, from_region);
457 }
458 }
459
460 template<typename ClosureType>
461 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
462 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
463 ShenandoahHeapRegionSetIterator& it,
464 ShenandoahHeapRegion* from_region) {
465 while (from_region != nullptr) {
466 assert(is_candidate_region(from_region), "Sanity");
467 cl.set_from_region(from_region);
468 if (from_region->has_live()) {
469 _heap->marked_object_iterate(from_region, &cl);
470 }
471
472 // Compacted the region to somewhere else? From-region is empty then.
473 if (!cl.is_compact_same_region()) {
474 empty_regions.append(from_region);
475 }
476 from_region = it.next();
477 }
478 cl.finish();
479
480 // Mark all remaining regions as empty
481 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
482 ShenandoahHeapRegion* r = empty_regions.at(pos);
483 r->set_new_top(r->bottom());
484 }
485 }
486
487 void ShenandoahFullGC::calculate_target_humongous_objects() {
488 ShenandoahHeap* heap = ShenandoahHeap::heap();
489
490 // Compute the new addresses for humongous objects. We need to do this after addresses
491 // for regular objects are calculated, and we know what regions in heap suffix are
492 // available for humongous moves.
493 //
494 // Scan the heap backwards, because we are compacting humongous regions towards the end.
495 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
496 // humongous start there.
497 //
498 // The complication is potential non-movable regions during the scan. If such region is
499 // detected, then sliding restarts towards that non-movable region.
500
501 size_t to_begin = heap->num_regions();
502 size_t to_end = heap->num_regions();
503
504 log_debug(gc)("Full GC calculating target humongous objects from end %zu", to_end);
505 for (size_t c = heap->num_regions(); c > 0; c--) {
506 ShenandoahHeapRegion *r = heap->get_region(c - 1);
507 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
508 // To-region candidate: record this, and continue scan
509 to_begin = r->index();
510 continue;
511 }
512
513 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
514 // From-region candidate: movable humongous region
515 oop old_obj = cast_to_oop(r->bottom());
516 size_t words_size = old_obj->size();
517 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
518
519 size_t start = to_end - num_regions;
520
521 if (start >= to_begin && start != r->index()) {
522 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
523 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
524 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
525 to_end = start;
526 continue;
527 }
528 }
529
530 // Failed to fit. Scan starting from current region.
531 to_begin = r->index();
532 to_end = r->index();
533 }
534 }
535
536 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
537 public:
538 void heap_region_do(ShenandoahHeapRegion* r) override {
539 if (r->is_trash()) {
540 r->try_recycle_under_lock();
541 // No need to adjust_interval_for_recycled_old_region. That will be taken care of during freeset rebuild.
542 }
543 if (r->is_cset()) {
544 // Leave affiliation unchanged
545 r->make_regular_bypass();
546 }
547 if (r->is_empty_uncommitted()) {
548 r->make_committed_bypass();
549 }
550 assert (r->is_committed(), "only committed regions in heap now, see region %zu", r->index());
551
552 // Record current region occupancy: this communicates empty regions are free
553 // to the rest of Full GC code.
554 r->set_new_top(r->top());
555 }
556 };
557
558 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
559 private:
560 ShenandoahHeap* const _heap;
561 ShenandoahMarkingContext* const _ctx;
562
563 public:
564 ShenandoahTrashImmediateGarbageClosure() :
565 _heap(ShenandoahHeap::heap()),
566 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
567
568 void heap_region_do(ShenandoahHeapRegion* r) override {
569 if (r->is_humongous_start()) {
570 oop humongous_obj = cast_to_oop(r->bottom());
571 if (!_ctx->is_marked(humongous_obj)) {
572 assert(!r->has_live(), "Region %zu is not marked, should not have live", r->index());
573 _heap->trash_humongous_region_at(r);
574 } else {
575 assert(r->has_live(), "Region %zu should have live", r->index());
576 }
577 } else if (r->is_humongous_continuation()) {
578 // If we hit continuation, the non-live humongous starts should have been trashed already
579 assert(r->humongous_start_region()->has_live(), "Region %zu should have live", r->index());
580 } else if (r->is_regular()) {
581 if (!r->has_live()) {
582 r->make_trash_immediate();
583 }
584 }
585 }
586 };
587
588 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
589 ShenandoahHeap* heap = ShenandoahHeap::heap();
590
591 uint n_workers = heap->workers()->active_workers();
592 size_t n_regions = heap->num_regions();
593
594 // What we want to accomplish: have the dense prefix of data, while still balancing
595 // out the parallel work.
596 //
597 // Assuming the amount of work is driven by the live data that needs moving, we can slice
598 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
599 // thread takes all regions in its prefix subset, and then it takes some regions from
600 // the tail.
601 //
602 // Tail region selection becomes interesting.
603 //
604 // First, we want to distribute the regions fairly between the workers, and those regions
605 // might have different amount of live data. So, until we sure no workers need live data,
606 // we need to only take what the worker needs.
607 //
608 // Second, since we slide everything to the left in each slice, the most busy regions
609 // would be the ones on the left. Which means we want to have all workers have their after-tail
610 // regions as close to the left as possible.
611 //
612 // The easiest way to do this is to distribute after-tail regions in round-robin between
613 // workers that still need live data.
614 //
615 // Consider parallel workers A, B, C, then the target slice layout would be:
616 //
617 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
618 //
619 // (.....dense-prefix.....) (.....................tail...................)
620 // [all regions fully live] [left-most regions are fuller that right-most]
621 //
622
623 // Compute how much live data is there. This would approximate the size of dense prefix
624 // we target to create.
625 size_t total_live = 0;
626 for (size_t idx = 0; idx < n_regions; idx++) {
627 ShenandoahHeapRegion *r = heap->get_region(idx);
628 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
629 total_live += r->get_live_data_words();
630 }
631 }
632
633 // Estimate the size for the dense prefix. Note that we specifically count only the
634 // "full" regions, so there would be some non-full regions in the slice tail.
635 size_t live_per_worker = total_live / n_workers;
636 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
637 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
638 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
639 assert(prefix_regions_total <= n_regions, "Sanity");
640
641 // There might be non-candidate regions in the prefix. To compute where the tail actually
642 // ends up being, we need to account those as well.
643 size_t prefix_end = prefix_regions_total;
644 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
645 ShenandoahHeapRegion *r = heap->get_region(idx);
646 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
647 prefix_end++;
648 }
649 }
650 prefix_end = MIN2(prefix_end, n_regions);
651 assert(prefix_end <= n_regions, "Sanity");
652
653 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
654 // subset of dense prefix.
655 size_t prefix_idx = 0;
656
657 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
658
659 for (size_t wid = 0; wid < n_workers; wid++) {
660 ShenandoahHeapRegionSet* slice = worker_slices[wid];
661
662 live[wid] = 0;
663 size_t regs = 0;
664
665 // Add all prefix regions for this worker
666 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
667 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
668 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
669 slice->add_region(r);
670 live[wid] += r->get_live_data_words();
671 regs++;
672 }
673 prefix_idx++;
674 }
675 }
676
677 // Distribute the tail among workers in round-robin fashion.
678 size_t wid = n_workers - 1;
679
680 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
681 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
682 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
683 assert(wid < n_workers, "Sanity");
684
685 size_t live_region = r->get_live_data_words();
686
687 // Select next worker that still needs live data.
688 size_t old_wid = wid;
689 do {
690 wid++;
691 if (wid == n_workers) wid = 0;
692 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
693
694 if (old_wid == wid) {
695 // Circled back to the same worker? This means liveness data was
696 // miscalculated. Bump the live_per_worker limit so that
697 // everyone gets a piece of the leftover work.
698 live_per_worker += ShenandoahHeapRegion::region_size_words();
699 }
700
701 worker_slices[wid]->add_region(r);
702 live[wid] += live_region;
703 }
704 }
705
706 FREE_C_HEAP_ARRAY(live);
707
708 #ifdef ASSERT
709 ResourceBitMap map(n_regions);
710 for (size_t wid = 0; wid < n_workers; wid++) {
711 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
712 ShenandoahHeapRegion* r = it.next();
713 while (r != nullptr) {
714 size_t idx = r->index();
715 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: %zu", idx);
716 assert(!map.at(idx), "No region distributed twice: %zu", idx);
717 map.at_put(idx, true);
718 r = it.next();
719 }
720 }
721
722 for (size_t rid = 0; rid < n_regions; rid++) {
723 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
724 bool is_distributed = map.at(rid);
725 assert(is_distributed || !is_candidate, "All candidates are distributed: %zu", rid);
726 }
727 #endif
728 }
729
730 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
731 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
732 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
733
734 ShenandoahHeap* heap = ShenandoahHeap::heap();
735
736 // About to figure out which regions can be compacted, make sure pinning status
737 // had been updated in GC prologue.
738 heap->assert_pinned_region_status();
739
740 {
741 // Trash the immediately collectible regions before computing addresses
742 ShenandoahTrashImmediateGarbageClosure trash_immediate_garbage;
743 ShenandoahExcludeRegionClosure<FREE> cl(&trash_immediate_garbage);
744 heap->heap_region_iterate(&cl);
745
746 // Make sure regions are in good state: committed, active, clean.
747 // This is needed because we are potentially sliding the data through them.
748 ShenandoahEnsureHeapActiveClosure ecl;
749 heap->heap_region_iterate(&ecl);
750 }
751
752 // Compute the new addresses for regular objects
753 {
754 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
755
756 distribute_slices(worker_slices);
757
758 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
759 heap->workers()->run_task(&task);
760 }
761
762 // Compute the new addresses for humongous objects
763 {
764 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
765 calculate_target_humongous_objects();
766 }
767 }
768
769 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
770 private:
771 ShenandoahMarkingContext* const _ctx;
772
773 template <class T>
774 inline void do_oop_work(T* p) {
775 T o = RawAccess<>::oop_load(p);
776 if (!CompressedOops::is_null(o)) {
777 oop obj = CompressedOops::decode_not_null(o);
778 assert(_ctx->is_marked(obj), "must be marked");
779 if (FullGCForwarding::is_forwarded(obj)) {
780 oop forw = FullGCForwarding::forwardee(obj);
781 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
782 }
783 }
784 }
785
786 public:
787 ShenandoahAdjustPointersClosure() :
788 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
789
790 void do_oop(oop* p) { do_oop_work(p); }
791 void do_oop(narrowOop* p) { do_oop_work(p); }
792 void do_method(Method* m) {}
793 void do_nmethod(nmethod* nm) {}
794 };
795
796 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
797 private:
798 ShenandoahAdjustPointersClosure _cl;
799
800 public:
801 void do_object(oop p) override {
802 assert(ShenandoahHeap::heap()->global_generation()->is_mark_complete(), "marking must be complete");
803 assert(ShenandoahHeap::heap()->marking_context()->is_marked(p), "must be marked");
804 p->oop_iterate(&_cl);
805 }
806 };
807
808 class ShenandoahAdjustPointersTask : public WorkerTask {
809 private:
810 ShenandoahHeap* const _heap;
811 ShenandoahRegionIterator _regions;
812
813 public:
814 ShenandoahAdjustPointersTask() :
815 WorkerTask("Shenandoah Adjust Pointers"),
816 _heap(ShenandoahHeap::heap()) {
817 }
818
819 void work(uint worker_id) override {
820 ShenandoahParallelWorkerSession worker_session(worker_id);
821 ShenandoahAdjustPointersObjectClosure obj_cl;
822 ShenandoahHeapRegion* r = _regions.next();
823 while (r != nullptr) {
824 if (!r->is_humongous_continuation() && r->has_live()) {
825 _heap->marked_object_iterate(r, &obj_cl);
826 }
827 if (_heap->mode()->is_generational()) {
828 ShenandoahGenerationalFullGC::maybe_coalesce_and_fill_region(r);
829 }
830 r = _regions.next();
831 }
832 }
833 };
834
835 class ShenandoahAdjustRootPointersTask : public WorkerTask {
836 private:
837 ShenandoahRootAdjuster* _rp;
838 PreservedMarksSet* _preserved_marks;
839 public:
840 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
841 WorkerTask("Shenandoah Adjust Root Pointers"),
842 _rp(rp),
843 _preserved_marks(preserved_marks) {}
844
845 void work(uint worker_id) override {
846 ShenandoahParallelWorkerSession worker_session(worker_id);
847 ShenandoahAdjustPointersClosure cl;
848 _rp->roots_do(worker_id, &cl);
849 _preserved_marks->get(worker_id)->adjust_during_full_gc();
850 }
851 };
852
853 void ShenandoahFullGC::phase3_update_references() {
854 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
855 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
856
857 ShenandoahHeap* heap = ShenandoahHeap::heap();
858
859 WorkerThreads* workers = heap->workers();
860 uint nworkers = workers->active_workers();
861 {
862 #ifdef COMPILER2
863 DerivedPointerTable::clear();
864 #endif // COMPILER2
865 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
866 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
867 workers->run_task(&task);
868 #ifdef COMPILER2
869 DerivedPointerTable::update_pointers();
870 #endif // COMPILER2
871 }
872
873 ShenandoahAdjustPointersTask adjust_pointers_task;
874 workers->run_task(&adjust_pointers_task);
875 }
876
877 class ShenandoahCompactObjectsClosure : public ObjectClosure {
878 private:
879 uint const _worker_id;
880
881 public:
882 explicit ShenandoahCompactObjectsClosure(uint worker_id) :
883 _worker_id(worker_id) {}
884
885 void do_object(oop p) override {
886 assert(ShenandoahHeap::heap()->global_generation()->is_mark_complete(), "marking must be finished");
887 assert(ShenandoahHeap::heap()->marking_context()->is_marked(p), "must be marked");
888 size_t size = p->size();
889 if (FullGCForwarding::is_forwarded(p)) {
890 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
891 HeapWord* compact_to = cast_from_oop<HeapWord*>(FullGCForwarding::forwardee(p));
892 assert(compact_from != compact_to, "Forwarded object should move");
893 Copy::aligned_conjoint_words(compact_from, compact_to, size);
894 oop new_obj = cast_to_oop(compact_to);
895
896 // Restore the mark word before relativizing the stack chunk. The copy's
897 // mark word contains the full GC forwarding encoding, which would cause
898 // is_stackChunk() to read garbage (especially with compact headers).
899 new_obj->init_mark();
900 ContinuationGCSupport::relativize_stack_chunk(new_obj);
901 }
902 }
903 };
904
905 class ShenandoahCompactObjectsTask : public WorkerTask {
906 private:
907 ShenandoahHeap* const _heap;
908 ShenandoahHeapRegionSet** const _worker_slices;
909
910 public:
911 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
912 WorkerTask("Shenandoah Compact Objects"),
913 _heap(ShenandoahHeap::heap()),
914 _worker_slices(worker_slices) {
915 }
916
917 void work(uint worker_id) override {
918 ShenandoahParallelWorkerSession worker_session(worker_id);
919 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
920
921 ShenandoahCompactObjectsClosure cl(worker_id);
922 ShenandoahHeapRegion* r = slice.next();
923 while (r != nullptr) {
924 assert(!r->is_humongous(), "must not get humongous regions here");
925 if (r->has_live()) {
926 _heap->marked_object_iterate(r, &cl);
927 }
928 r->set_top(r->new_top());
929 r = slice.next();
930 }
931 }
932 };
933
934 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
935 private:
936 ShenandoahHeap* const _heap;
937 bool _is_generational;
938 size_t _young_regions, _young_usage, _young_humongous_waste;
939 size_t _old_regions, _old_usage, _old_humongous_waste;
940
941 public:
942 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()),
943 _is_generational(_heap->mode()->is_generational()),
944 _young_regions(0),
945 _young_usage(0),
946 _young_humongous_waste(0),
947 _old_regions(0),
948 _old_usage(0),
949 _old_humongous_waste(0)
950 {
951 _heap->free_set()->clear();
952 }
953
954 void heap_region_do(ShenandoahHeapRegion* r) override {
955 assert (!r->is_cset(), "cset regions should have been demoted already");
956
957 // Need to reset the complete-top-at-mark-start pointer here because
958 // the complete marking bitmap is no longer valid. This ensures
959 // size-based iteration in marked_object_iterate().
960 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
961 // pinned regions.
962 if (!r->is_pinned()) {
963 _heap->marking_context()->reset_top_at_mark_start(r);
964 }
965
966 size_t live = r->used();
967
968 // Make empty regions that have been allocated into regular
969 if (r->is_empty() && live > 0) {
970 if (!_is_generational) {
971 r->make_affiliated_maybe();
972 }
973 // else, generational mode compaction has already established affiliation.
974 r->make_regular_bypass();
975 if (ZapUnusedHeapArea) {
976 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
977 }
978 }
979
980 // Reclaim regular regions that became empty
981 if (r->is_regular() && live == 0) {
982 r->make_trash();
983 }
984
985 // Recycle all trash regions
986 if (r->is_trash()) {
987 live = 0;
988 r->try_recycle_under_lock();
989 // No need to adjust_interval_for_recycled_old_region. That will be taken care of during freeset rebuild.
990 } else {
991 if (r->is_old()) {
992 ShenandoahGenerationalFullGC::account_for_region(r, _old_regions, _old_usage, _old_humongous_waste);
993 } else if (r->is_young()) {
994 ShenandoahGenerationalFullGC::account_for_region(r, _young_regions, _young_usage, _young_humongous_waste);
995 }
996 }
997 r->set_live_data(live);
998 r->reset_alloc_metadata();
999 }
1000 };
1001
1002 void ShenandoahFullGC::compact_humongous_objects() {
1003 // Compact humongous regions, based on their fwdptr objects.
1004 //
1005 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1006 // humongous regions are already compacted, and do not require further moves, which alleviates
1007 // sliding costs. We may consider doing this in parallel in the future.
1008
1009 ShenandoahHeap* heap = ShenandoahHeap::heap();
1010
1011 for (size_t c = heap->num_regions(); c > 0; c--) {
1012 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1013 if (r->is_humongous_start()) {
1014 oop old_obj = cast_to_oop(r->bottom());
1015 if (!FullGCForwarding::is_forwarded(old_obj)) {
1016 // No need to move the object, it stays at the same slot
1017 continue;
1018 }
1019 size_t words_size = old_obj->size();
1020 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1021
1022 size_t old_start = r->index();
1023 size_t old_end = old_start + num_regions - 1;
1024 size_t new_start = heap->heap_region_index_containing(FullGCForwarding::forwardee(old_obj));
1025 size_t new_end = new_start + num_regions - 1;
1026 assert(old_start != new_start, "must be real move");
1027 assert(r->is_stw_move_allowed(), "Region %zu should be movable", r->index());
1028
1029 log_debug(gc)("Full GC compaction moves humongous object from region %zu to region %zu", old_start, new_start);
1030 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
1031 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
1032
1033 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1034 new_obj->init_mark();
1035
1036 {
1037 ShenandoahAffiliation original_affiliation = r->affiliation();
1038 for (size_t c = old_start; c <= old_end; c++) {
1039 ShenandoahHeapRegion* r = heap->get_region(c);
1040 // Leave humongous region affiliation unchanged.
1041 r->make_regular_bypass();
1042 r->set_top(r->bottom());
1043 }
1044
1045 for (size_t c = new_start; c <= new_end; c++) {
1046 ShenandoahHeapRegion* r = heap->get_region(c);
1047 if (c == new_start) {
1048 r->make_humongous_start_bypass(original_affiliation);
1049 } else {
1050 r->make_humongous_cont_bypass(original_affiliation);
1051 }
1052
1053 // Trailing region may be non-full, record the remainder there
1054 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1055 if ((c == new_end) && (remainder != 0)) {
1056 r->set_top(r->bottom() + remainder);
1057 } else {
1058 r->set_top(r->end());
1059 }
1060
1061 r->reset_alloc_metadata();
1062 }
1063 }
1064 }
1065 }
1066 }
1067
1068 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1069 // we need to remain able to walk pinned regions.
1070 // Since pinned region do not move and don't get compacted, we will get holes with
1071 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1072 // cannot be iterated over using oop->size()). The only way to safely iterate over those is using
1073 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1074 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1075 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1076 private:
1077 ShenandoahRegionIterator _regions;
1078
1079 public:
1080 ShenandoahMCResetCompleteBitmapTask() :
1081 WorkerTask("Shenandoah Reset Bitmap") {
1082 }
1083
1084 void work(uint worker_id) override {
1085 ShenandoahParallelWorkerSession worker_session(worker_id);
1086 ShenandoahHeapRegion* region = _regions.next();
1087 ShenandoahHeap* heap = ShenandoahHeap::heap();
1088 ShenandoahMarkingContext* const ctx = heap->marking_context();
1089 assert(heap->global_generation()->is_mark_complete(), "Marking must be complete");
1090 while (region != nullptr) {
1091 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1092 ctx->clear_bitmap(region);
1093 }
1094 region = _regions.next();
1095 }
1096 }
1097 };
1098
1099 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1100 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1101 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1102
1103 ShenandoahHeap* heap = ShenandoahHeap::heap();
1104
1105 // Compact regular objects first
1106 {
1107 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1108 ShenandoahCompactObjectsTask compact_task(worker_slices);
1109 heap->workers()->run_task(&compact_task);
1110 }
1111
1112 // Compact humongous objects after regular object moves
1113 {
1114 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1115 compact_humongous_objects();
1116 }
1117 }
1118
1119 void ShenandoahFullGC::phase5_epilog() {
1120 GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1121 ShenandoahHeap* heap = ShenandoahHeap::heap();
1122
1123 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1124 // and must ensure the bitmap is in sync.
1125 {
1126 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1127 ShenandoahMCResetCompleteBitmapTask task;
1128 heap->workers()->run_task(&task);
1129 }
1130
1131 // Bring regions in proper states after the collection, and set heap properties.
1132 {
1133 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1134 ShenandoahPostCompactClosure post_compact;
1135 heap->heap_region_iterate(&post_compact);
1136 heap->collection_set()->clear();
1137 {
1138 ShenandoahFreeSet* free_set = heap->free_set();
1139 size_t young_trashed_regions, old_trashed_regions, first_old, last_old, num_old;
1140 free_set->prepare_to_rebuild(young_trashed_regions, old_trashed_regions, first_old, last_old, num_old);
1141 // We also do not expand old generation size following Full GC because we have scrambled age populations and
1142 // no longer have objects separated by age into distinct regions.
1143 if (heap->mode()->is_generational()) {
1144 ShenandoahGenerationalFullGC::compute_balances();
1145 }
1146 heap->free_set()->finish_rebuild(young_trashed_regions, old_trashed_regions, num_old);
1147 }
1148
1149 // Set mark incomplete because the marking bitmaps have been reset except pinned regions.
1150 _generation->set_mark_incomplete();
1151
1152 heap->clear_cancelled_gc();
1153 }
1154
1155 _preserved_marks->restore(heap->workers());
1156 _preserved_marks->reclaim();
1157
1158 if (heap->mode()->is_generational()) {
1159 ShenandoahGenerationalFullGC::rebuild_remembered_set(heap);
1160 }
1161 }