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, 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 {
132 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_propagate_gc_state);
133 heap->propagate_gc_state_to_all_threads();
134 }
135 }
136
137 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
138 ShenandoahHeap* heap = ShenandoahHeap::heap();
139
140 // A full GC may be entered directly, or as an upgrade from a failed
141 // degenerated GC. In the latter case, self-forwarded objects may be
142 // present from the failed evacuation. Drain those marks before any phase
143 // (verify, update_roots, phase1_mark_heap) walks headers.
144 {
145 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_un_self_forward);
146 heap->un_self_forward_cset_regions();
147 }
148
149 if (heap->mode()->is_generational()) {
150 ShenandoahGenerationalFullGC::prepare();
151 }
152
153 if (ShenandoahVerify) {
154 heap->verifier()->verify_before_fullgc(_generation);
155 }
156
157 if (VerifyBeforeGC) {
158 Universe::verify();
159 }
160
161 // Degenerated GC may carry concurrent root flags when upgrading to
162 // full GC. We need to reset it before mutators resume.
163 heap->set_concurrent_strong_root_in_progress(false);
164 heap->set_concurrent_weak_root_in_progress(false);
165
166 heap->set_full_gc_in_progress(true);
167
168 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
169 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
170
171 {
172 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
173 heap->pre_full_gc_dump(_gc_timer);
174 }
175
176 {
177 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
178 // Full GC is supposed to recover from any GC state:
179
180 // a0. Remember if we have forwarded objects
181 bool has_forwarded_objects = heap->has_forwarded_objects();
182
183 // a1. Cancel evacuation, if in progress
184 if (heap->is_evacuation_in_progress()) {
185 heap->set_evacuation_in_progress(false);
186 }
187 assert(!heap->is_evacuation_in_progress(), "sanity");
188
189 // a2. Cancel update-refs, if in progress
190 if (heap->is_update_refs_in_progress()) {
191 heap->set_update_refs_in_progress(false);
192 }
193 assert(!heap->is_update_refs_in_progress(), "sanity");
194
195 // b. Cancel all concurrent marks, if in progress
196 if (heap->is_concurrent_mark_in_progress()) {
197 heap->cancel_concurrent_mark();
198 }
199 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
200
201 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
202 if (has_forwarded_objects) {
203 update_roots(true /*full_gc*/);
204 }
205
206 // d. Abandon reference discovery and clear all discovered references.
207 ShenandoahReferenceProcessor* rp = _generation->ref_processor();
208 rp->abandon_partial_discovery();
209
210 // e. Sync pinned region status from the CP marks
211 heap->sync_pinned_region_status();
212
213 if (heap->mode()->is_generational()) {
214 ShenandoahGenerationalFullGC::restore_top_before_promote(heap);
215 }
216
217 // The rest of prologue:
218 _preserved_marks->init(heap->workers()->active_workers());
219
220 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
221 }
222
223 if (UseTLAB) {
224 // Note: PLABs are also retired with GCLABs in generational mode.
225 heap->gclabs_retire(ResizeTLAB);
226 heap->tlabs_retire(ResizeTLAB);
227 }
228
229 OrderAccess::fence();
230
231 phase1_mark_heap();
232
233 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
234 // Coming out of Full GC, we would not have any forwarded objects.
235 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
236 heap->set_has_forwarded_objects(false);
237
238 heap->set_full_gc_move_in_progress(true);
239
240 // Setup workers for the rest
241 OrderAccess::fence();
242
243 // Initialize worker slices
244 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
245 for (uint i = 0; i < heap->max_workers(); i++) {
246 worker_slices[i] = new ShenandoahHeapRegionSet();
247 }
248
249 {
250 // The rest of code performs region moves, where region status is undefined
251 // until all phases run together.
252 ShenandoahHeapLocker lock(heap->lock());
253
254 phase2_calculate_target_addresses(worker_slices);
255
256 OrderAccess::fence();
257
258 phase3_update_references();
259
260 phase4_compact_objects(worker_slices);
261
262 phase5_epilog();
263 }
264 heap->start_idle_span();
265
266 // Resize metaspace
267 MetaspaceGC::compute_new_size();
268
269 // Free worker slices
270 for (uint i = 0; i < heap->max_workers(); i++) {
271 delete worker_slices[i];
272 }
273 FREE_C_HEAP_ARRAY(worker_slices);
274
275 heap->set_full_gc_move_in_progress(false);
276 heap->set_full_gc_in_progress(false);
277
278 DEBUG_ONLY(heap->assert_no_self_forwards());
279
280 if (ShenandoahVerify) {
281 heap->verifier()->verify_after_fullgc(_generation);
282 }
283
284 if (VerifyAfterGC) {
285 Universe::verify();
286 }
287
288 {
289 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
290 heap->post_full_gc_dump(_gc_timer);
291 }
292 }
293
294 void ShenandoahFullGC::phase1_mark_heap() {
295 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
296 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
297
298 ShenandoahHeap* heap = ShenandoahHeap::heap();
299
300 _generation->reset_mark_bitmap<true, true>();
301 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
302 assert(!_generation->is_mark_complete(), "sanity");
303
304 heap->set_unload_classes(_generation->heuristics()->can_unload_classes());
305
306 ShenandoahReferenceProcessor* rp = _generation->ref_processor();
307 // enable ("weak") refs discovery
308 rp->set_soft_reference_policy(true); // forcefully purge all soft references
309
310 ShenandoahSTWMark mark(_generation, true /*full_gc*/);
311 mark.mark();
312 heap->parallel_cleaning(_generation, true /* full_gc */);
313
314 if (ShenandoahHeap::heap()->mode()->is_generational()) {
315 ShenandoahGenerationalFullGC::log_live_in_old(heap);
316 }
317 }
318
319 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
320 private:
321 PreservedMarks* const _preserved_marks;
322 ShenandoahHeap* const _heap;
323 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
324 int _empty_regions_pos;
325 ShenandoahHeapRegion* _to_region;
326 ShenandoahHeapRegion* _from_region;
327 HeapWord* _compact_point;
328
329 public:
330 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
331 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
332 ShenandoahHeapRegion* to_region) :
333 _preserved_marks(preserved_marks),
334 _heap(ShenandoahHeap::heap()),
335 _empty_regions(empty_regions),
336 _empty_regions_pos(0),
337 _to_region(to_region),
338 _from_region(nullptr),
339 _compact_point(to_region->bottom()) {}
340
341 void set_from_region(ShenandoahHeapRegion* from_region) {
342 _from_region = from_region;
343 }
344
345 void finish() {
346 assert(_to_region != nullptr, "should not happen");
347 _to_region->set_new_top(_compact_point);
348 }
349
350 bool is_compact_same_region() {
351 return _from_region == _to_region;
352 }
353
354 int empty_regions_pos() {
355 return _empty_regions_pos;
356 }
357
358 void do_object(oop p) override {
359 shenandoah_assert_mark_complete(cast_from_oop<HeapWord*>(p));
360 assert(_from_region != nullptr, "must set before work");
361 assert(_heap->global_generation()->is_mark_complete(), "marking must be finished");
362 assert(_heap->marking_context()->is_marked(p), "must be marked");
363 assert(!_heap->marking_context()->allocated_after_mark_start(p), "must be truly marked");
364
365 size_t obj_size = p->size();
366 if (_compact_point + obj_size > _to_region->end()) {
367 finish();
368
369 // Object doesn't fit. Pick next empty region and start compacting there.
370 ShenandoahHeapRegion* new_to_region;
371 if (_empty_regions_pos < _empty_regions.length()) {
372 new_to_region = _empty_regions.at(_empty_regions_pos);
373 _empty_regions_pos++;
374 } else {
375 // Out of empty region? Compact within the same region.
376 new_to_region = _from_region;
377 }
378
379 assert(new_to_region != _to_region, "must not reuse same to-region");
380 assert(new_to_region != nullptr, "must not be null");
381 _to_region = new_to_region;
382 _compact_point = _to_region->bottom();
383 }
384
385 // Object fits into current region, record new location, if object does not move:
386 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
387 shenandoah_assert_not_forwarded(nullptr, p);
388 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
389 _preserved_marks->push_if_necessary(p, p->mark());
390 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
391 }
392 _compact_point += obj_size;
393 }
394 };
395
396 class ShenandoahPrepareForCompactionTask : public WorkerTask {
397 private:
398 PreservedMarksSet* const _preserved_marks;
399 ShenandoahHeap* const _heap;
400 ShenandoahHeapRegionSet** const _worker_slices;
401
402 public:
403 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
404 WorkerTask("Shenandoah Prepare For Compaction"),
405 _preserved_marks(preserved_marks),
406 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
407 }
408
409 static bool is_candidate_region(ShenandoahHeapRegion* r) {
410 // Empty region: get it into the slice to defragment the slice itself.
411 // We could have skipped this without violating correctness, but we really
412 // want to compact all live regions to the start of the heap, which sometimes
413 // means moving them into the fully empty regions.
414 if (r->is_empty()) return true;
415
416 // Can move the region, and this is not the humongous region. Humongous
417 // moves are special cased here, because their moves are handled separately.
418 return r->is_stw_move_allowed() && !r->is_humongous();
419 }
420
421 void work(uint worker_id) override;
422 private:
423 template<typename ClosureType>
424 void prepare_for_compaction(ClosureType& cl,
425 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
426 ShenandoahHeapRegionSetIterator& it,
427 ShenandoahHeapRegion* from_region);
428 };
429
430 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
431 ShenandoahParallelWorkerSession worker_session(worker_id);
432 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
433 ShenandoahHeapRegionSetIterator it(slice);
434 ShenandoahHeapRegion* from_region = it.next();
435 // No work?
436 if (from_region == nullptr) {
437 return;
438 }
439
440 // Sliding compaction. Walk all regions in the slice, and compact them.
441 // Remember empty regions and reuse them as needed.
442 ResourceMark rm;
443
444 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
445
446 if (_heap->mode()->is_generational()) {
447 ShenandoahPrepareForGenerationalCompactionObjectClosure cl(_preserved_marks->get(worker_id),
448 empty_regions, from_region, worker_id);
449 prepare_for_compaction(cl, empty_regions, it, from_region);
450 } else {
451 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
452 prepare_for_compaction(cl, empty_regions, it, from_region);
453 }
454 }
455
456 template<typename ClosureType>
457 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
458 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
459 ShenandoahHeapRegionSetIterator& it,
460 ShenandoahHeapRegion* from_region) {
461 while (from_region != nullptr) {
462 assert(is_candidate_region(from_region), "Sanity");
463 cl.set_from_region(from_region);
464 if (from_region->has_live()) {
465 _heap->marked_object_iterate(from_region, &cl);
466 }
467
468 // Compacted the region to somewhere else? From-region is empty then.
469 if (!cl.is_compact_same_region()) {
470 empty_regions.append(from_region);
471 }
472 from_region = it.next();
473 }
474 cl.finish();
475
476 // Mark all remaining regions as empty
477 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
478 ShenandoahHeapRegion* r = empty_regions.at(pos);
479 r->set_new_top(r->bottom());
480 }
481 }
482
483 void ShenandoahFullGC::calculate_target_humongous_objects() {
484 ShenandoahHeap* heap = ShenandoahHeap::heap();
485
486 // Compute the new addresses for humongous objects. We need to do this after addresses
487 // for regular objects are calculated, and we know what regions in heap suffix are
488 // available for humongous moves.
489 //
490 // Scan the heap backwards, because we are compacting humongous regions towards the end.
491 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
492 // humongous start there.
493 //
494 // The complication is potential non-movable regions during the scan. If such region is
495 // detected, then sliding restarts towards that non-movable region.
496
497 size_t to_begin = heap->num_regions();
498 size_t to_end = heap->num_regions();
499
500 log_debug(gc)("Full GC calculating target humongous objects from end %zu", to_end);
501 for (size_t c = heap->num_regions(); c > 0; c--) {
502 ShenandoahHeapRegion *r = heap->get_region(c - 1);
503 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
504 // To-region candidate: record this, and continue scan
505 to_begin = r->index();
506 continue;
507 }
508
509 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
510 // From-region candidate: movable humongous region
511 oop old_obj = cast_to_oop(r->bottom());
512 size_t words_size = old_obj->size();
513 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
514
515 size_t start = to_end - num_regions;
516
517 if (start >= to_begin && start != r->index()) {
518 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
519 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
520 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
521 to_end = start;
522 continue;
523 }
524 }
525
526 // Failed to fit. Scan starting from current region.
527 to_begin = r->index();
528 to_end = r->index();
529 }
530 }
531
532 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
533 public:
534 void heap_region_do(ShenandoahHeapRegion* r) override {
535 if (r->is_trash()) {
536 r->try_recycle_under_lock();
537 // No need to adjust_interval_for_recycled_old_region. That will be taken care of during freeset rebuild.
538 }
539 if (r->is_cset()) {
540 // Leave affiliation unchanged
541 r->make_regular_bypass();
542 }
543 if (r->is_empty_uncommitted()) {
544 r->make_committed_bypass();
545 }
546 assert (r->is_committed(), "only committed regions in heap now, see region %zu", r->index());
547
548 // Record current region occupancy: this communicates empty regions are free
549 // to the rest of Full GC code.
550 r->set_new_top(r->top());
551 }
552 };
553
554 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
555 private:
556 ShenandoahHeap* const _heap;
557 ShenandoahMarkingContext* const _ctx;
558
559 public:
560 ShenandoahTrashImmediateGarbageClosure() :
561 _heap(ShenandoahHeap::heap()),
562 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
563
564 void heap_region_do(ShenandoahHeapRegion* r) override {
565 if (r->is_humongous_start()) {
566 oop humongous_obj = cast_to_oop(r->bottom());
567 if (!_ctx->is_marked(humongous_obj)) {
568 assert(!r->has_live(), "Region %zu is not marked, should not have live", r->index());
569 _heap->trash_humongous_region_at(r);
570 } else {
571 assert(r->has_live(), "Region %zu should have live", r->index());
572 }
573 } else if (r->is_humongous_continuation()) {
574 // If we hit continuation, the non-live humongous starts should have been trashed already
575 assert(r->humongous_start_region()->has_live(), "Region %zu should have live", r->index());
576 } else if (r->is_regular()) {
577 if (!r->has_live()) {
578 r->make_trash_immediate();
579 }
580 }
581 }
582 };
583
584 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
585 ShenandoahHeap* heap = ShenandoahHeap::heap();
586
587 uint n_workers = heap->workers()->active_workers();
588 size_t n_regions = heap->num_regions();
589
590 // What we want to accomplish: have the dense prefix of data, while still balancing
591 // out the parallel work.
592 //
593 // Assuming the amount of work is driven by the live data that needs moving, we can slice
594 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
595 // thread takes all regions in its prefix subset, and then it takes some regions from
596 // the tail.
597 //
598 // Tail region selection becomes interesting.
599 //
600 // First, we want to distribute the regions fairly between the workers, and those regions
601 // might have different amount of live data. So, until we sure no workers need live data,
602 // we need to only take what the worker needs.
603 //
604 // Second, since we slide everything to the left in each slice, the most busy regions
605 // would be the ones on the left. Which means we want to have all workers have their after-tail
606 // regions as close to the left as possible.
607 //
608 // The easiest way to do this is to distribute after-tail regions in round-robin between
609 // workers that still need live data.
610 //
611 // Consider parallel workers A, B, C, then the target slice layout would be:
612 //
613 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
614 //
615 // (.....dense-prefix.....) (.....................tail...................)
616 // [all regions fully live] [left-most regions are fuller that right-most]
617 //
618
619 // Compute how much live data is there. This would approximate the size of dense prefix
620 // we target to create.
621 size_t total_live = 0;
622 for (size_t idx = 0; idx < n_regions; idx++) {
623 ShenandoahHeapRegion *r = heap->get_region(idx);
624 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
625 total_live += r->get_live_data_words();
626 }
627 }
628
629 // Estimate the size for the dense prefix. Note that we specifically count only the
630 // "full" regions, so there would be some non-full regions in the slice tail.
631 size_t live_per_worker = total_live / n_workers;
632 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
633 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
634 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
635 assert(prefix_regions_total <= n_regions, "Sanity");
636
637 // There might be non-candidate regions in the prefix. To compute where the tail actually
638 // ends up being, we need to account those as well.
639 size_t prefix_end = prefix_regions_total;
640 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
641 ShenandoahHeapRegion *r = heap->get_region(idx);
642 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
643 prefix_end++;
644 }
645 }
646 prefix_end = MIN2(prefix_end, n_regions);
647 assert(prefix_end <= n_regions, "Sanity");
648
649 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
650 // subset of dense prefix.
651 size_t prefix_idx = 0;
652
653 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
654
655 for (size_t wid = 0; wid < n_workers; wid++) {
656 ShenandoahHeapRegionSet* slice = worker_slices[wid];
657
658 live[wid] = 0;
659 size_t regs = 0;
660
661 // Add all prefix regions for this worker
662 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
663 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
664 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
665 slice->add_region(r);
666 live[wid] += r->get_live_data_words();
667 regs++;
668 }
669 prefix_idx++;
670 }
671 }
672
673 // Distribute the tail among workers in round-robin fashion.
674 size_t wid = n_workers - 1;
675
676 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
677 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
678 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
679 assert(wid < n_workers, "Sanity");
680
681 size_t live_region = r->get_live_data_words();
682
683 // Select next worker that still needs live data.
684 size_t old_wid = wid;
685 do {
686 wid++;
687 if (wid == n_workers) wid = 0;
688 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
689
690 if (old_wid == wid) {
691 // Circled back to the same worker? This means liveness data was
692 // miscalculated. Bump the live_per_worker limit so that
693 // everyone gets a piece of the leftover work.
694 live_per_worker += ShenandoahHeapRegion::region_size_words();
695 }
696
697 worker_slices[wid]->add_region(r);
698 live[wid] += live_region;
699 }
700 }
701
702 FREE_C_HEAP_ARRAY(live);
703
704 #ifdef ASSERT
705 ResourceBitMap map(n_regions);
706 for (size_t wid = 0; wid < n_workers; wid++) {
707 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
708 ShenandoahHeapRegion* r = it.next();
709 while (r != nullptr) {
710 size_t idx = r->index();
711 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: %zu", idx);
712 assert(!map.at(idx), "No region distributed twice: %zu", idx);
713 map.at_put(idx, true);
714 r = it.next();
715 }
716 }
717
718 for (size_t rid = 0; rid < n_regions; rid++) {
719 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
720 bool is_distributed = map.at(rid);
721 assert(is_distributed || !is_candidate, "All candidates are distributed: %zu", rid);
722 }
723 #endif
724 }
725
726 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
727 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
728 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
729
730 ShenandoahHeap* heap = ShenandoahHeap::heap();
731
732 // About to figure out which regions can be compacted, make sure pinning status
733 // had been updated in GC prologue.
734 heap->assert_pinned_region_status();
735
736 {
737 // Trash the immediately collectible regions before computing addresses
738 ShenandoahTrashImmediateGarbageClosure trash_immediate_garbage;
739 ShenandoahExcludeRegionClosure<FREE> cl(&trash_immediate_garbage);
740 heap->heap_region_iterate(&cl);
741
742 // Make sure regions are in good state: committed, active, clean.
743 // This is needed because we are potentially sliding the data through them.
744 ShenandoahEnsureHeapActiveClosure ecl;
745 heap->heap_region_iterate(&ecl);
746 }
747
748 // Compute the new addresses for regular objects
749 {
750 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
751
752 distribute_slices(worker_slices);
753
754 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
755 heap->workers()->run_task(&task);
756 }
757
758 // Compute the new addresses for humongous objects
759 {
760 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
761 calculate_target_humongous_objects();
762 }
763 }
764
765 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
766 private:
767 ShenandoahMarkingContext* const _ctx;
768
769 template <class T>
770 inline void do_oop_work(T* p) {
771 T o = RawAccess<>::oop_load(p);
772 if (!CompressedOops::is_null(o)) {
773 oop obj = CompressedOops::decode_not_null(o);
774 assert(_ctx->is_marked(obj), "must be marked");
775 if (FullGCForwarding::is_forwarded(obj)) {
776 oop forw = FullGCForwarding::forwardee(obj);
777 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
778 }
779 }
780 }
781
782 public:
783 ShenandoahAdjustPointersClosure() :
784 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
785
786 void do_oop(oop* p) { do_oop_work(p); }
787 void do_oop(narrowOop* p) { do_oop_work(p); }
788 void do_method(Method* m) {}
789 void do_nmethod(nmethod* nm) {}
790 };
791
792 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
793 private:
794 ShenandoahAdjustPointersClosure _cl;
795
796 public:
797 void do_object(oop p) override {
798 assert(ShenandoahHeap::heap()->global_generation()->is_mark_complete(), "marking must be complete");
799 assert(ShenandoahHeap::heap()->marking_context()->is_marked(p), "must be marked");
800 p->oop_iterate(&_cl);
801 }
802 };
803
804 class ShenandoahAdjustPointersTask : public WorkerTask {
805 private:
806 ShenandoahHeap* const _heap;
807 ShenandoahRegionIterator _regions;
808
809 public:
810 ShenandoahAdjustPointersTask() :
811 WorkerTask("Shenandoah Adjust Pointers"),
812 _heap(ShenandoahHeap::heap()) {
813 }
814
815 void work(uint worker_id) override {
816 ShenandoahParallelWorkerSession worker_session(worker_id);
817 ShenandoahAdjustPointersObjectClosure obj_cl;
818 ShenandoahHeapRegion* r = _regions.next();
819 while (r != nullptr) {
820 if (!r->is_humongous_continuation() && r->has_live()) {
821 _heap->marked_object_iterate(r, &obj_cl);
822 }
823 if (_heap->mode()->is_generational()) {
824 ShenandoahGenerationalFullGC::maybe_coalesce_and_fill_region(r);
825 }
826 r = _regions.next();
827 }
828 }
829 };
830
831 class ShenandoahAdjustRootPointersTask : public WorkerTask {
832 private:
833 ShenandoahRootAdjuster* _rp;
834 PreservedMarksSet* _preserved_marks;
835 public:
836 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
837 WorkerTask("Shenandoah Adjust Root Pointers"),
838 _rp(rp),
839 _preserved_marks(preserved_marks) {}
840
841 void work(uint worker_id) override {
842 ShenandoahParallelWorkerSession worker_session(worker_id);
843 ShenandoahAdjustPointersClosure cl;
844 _rp->roots_do(worker_id, &cl);
845 _preserved_marks->get(worker_id)->adjust_during_full_gc();
846 }
847 };
848
849 void ShenandoahFullGC::phase3_update_references() {
850 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
851 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
852
853 ShenandoahHeap* heap = ShenandoahHeap::heap();
854
855 WorkerThreads* workers = heap->workers();
856 uint nworkers = workers->active_workers();
857 {
858 #if COMPILER2_OR_JVMCI
859 DerivedPointerTable::clear();
860 #endif
861 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
862 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
863 workers->run_task(&task);
864 #if COMPILER2_OR_JVMCI
865 DerivedPointerTable::update_pointers();
866 #endif
867 }
868
869 ShenandoahAdjustPointersTask adjust_pointers_task;
870 workers->run_task(&adjust_pointers_task);
871 }
872
873 class ShenandoahCompactObjectsClosure : public ObjectClosure {
874 private:
875 uint const _worker_id;
876
877 public:
878 explicit ShenandoahCompactObjectsClosure(uint worker_id) :
879 _worker_id(worker_id) {}
880
881 void do_object(oop p) override {
882 assert(ShenandoahHeap::heap()->global_generation()->is_mark_complete(), "marking must be finished");
883 assert(ShenandoahHeap::heap()->marking_context()->is_marked(p), "must be marked");
884 size_t size = p->size();
885 if (FullGCForwarding::is_forwarded(p)) {
886 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
887 HeapWord* compact_to = cast_from_oop<HeapWord*>(FullGCForwarding::forwardee(p));
888 assert(compact_from != compact_to, "Forwarded object should move");
889 Copy::aligned_conjoint_words(compact_from, compact_to, size);
890 oop new_obj = cast_to_oop(compact_to);
891
892 // Restore the mark word before relativizing the stack chunk. The copy's
893 // mark word contains the full GC forwarding encoding, which would cause
894 // is_stackChunk() to read garbage (especially with compact headers).
895 new_obj->init_mark();
896 ContinuationGCSupport::relativize_stack_chunk(new_obj);
897 }
898 }
899 };
900
901 class ShenandoahCompactObjectsTask : public WorkerTask {
902 private:
903 ShenandoahHeap* const _heap;
904 ShenandoahHeapRegionSet** const _worker_slices;
905
906 public:
907 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
908 WorkerTask("Shenandoah Compact Objects"),
909 _heap(ShenandoahHeap::heap()),
910 _worker_slices(worker_slices) {
911 }
912
913 void work(uint worker_id) override {
914 ShenandoahParallelWorkerSession worker_session(worker_id);
915 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
916
917 ShenandoahCompactObjectsClosure cl(worker_id);
918 ShenandoahHeapRegion* r = slice.next();
919 while (r != nullptr) {
920 assert(!r->is_humongous(), "must not get humongous regions here");
921 if (r->has_live()) {
922 _heap->marked_object_iterate(r, &cl);
923 }
924 r->set_top(r->new_top());
925 r = slice.next();
926 }
927 }
928 };
929
930 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
931 private:
932 ShenandoahHeap* const _heap;
933 bool _is_generational;
934 size_t _young_regions, _young_usage, _young_humongous_waste;
935 size_t _old_regions, _old_usage, _old_humongous_waste;
936
937 public:
938 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()),
939 _is_generational(_heap->mode()->is_generational()),
940 _young_regions(0),
941 _young_usage(0),
942 _young_humongous_waste(0),
943 _old_regions(0),
944 _old_usage(0),
945 _old_humongous_waste(0)
946 {
947 _heap->free_set()->clear();
948 }
949
950 void heap_region_do(ShenandoahHeapRegion* r) override {
951 assert (!r->is_cset(), "cset regions should have been demoted already");
952
953 // Need to reset the complete-top-at-mark-start pointer here because
954 // the complete marking bitmap is no longer valid. This ensures
955 // size-based iteration in marked_object_iterate().
956 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
957 // pinned regions.
958 if (!r->is_pinned()) {
959 _heap->marking_context()->reset_top_at_mark_start(r);
960 }
961
962 size_t live = r->used();
963
964 // Make empty regions that have been allocated into regular
965 if (r->is_empty() && live > 0) {
966 if (!_is_generational) {
967 r->make_affiliated_maybe();
968 }
969 // else, generational mode compaction has already established affiliation.
970 r->make_regular_bypass();
971 if (ZapUnusedHeapArea) {
972 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
973 }
974 }
975
976 // Reclaim regular regions that became empty
977 if (r->is_regular() && live == 0) {
978 r->make_trash();
979 }
980
981 // Recycle all trash regions
982 if (r->is_trash()) {
983 live = 0;
984 r->try_recycle_under_lock();
985 // No need to adjust_interval_for_recycled_old_region. That will be taken care of during freeset rebuild.
986 } else {
987 if (r->is_old()) {
988 ShenandoahGenerationalFullGC::account_for_region(r, _old_regions, _old_usage, _old_humongous_waste);
989 } else if (r->is_young()) {
990 ShenandoahGenerationalFullGC::account_for_region(r, _young_regions, _young_usage, _young_humongous_waste);
991 }
992 }
993 r->set_live_data(live);
994 r->reset_alloc_metadata();
995 }
996 };
997
998 void ShenandoahFullGC::compact_humongous_objects() {
999 // Compact humongous regions, based on their fwdptr objects.
1000 //
1001 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1002 // humongous regions are already compacted, and do not require further moves, which alleviates
1003 // sliding costs. We may consider doing this in parallel in the future.
1004
1005 ShenandoahHeap* heap = ShenandoahHeap::heap();
1006
1007 for (size_t c = heap->num_regions(); c > 0; c--) {
1008 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1009 if (r->is_humongous_start()) {
1010 oop old_obj = cast_to_oop(r->bottom());
1011 if (!FullGCForwarding::is_forwarded(old_obj)) {
1012 // No need to move the object, it stays at the same slot
1013 continue;
1014 }
1015 size_t words_size = old_obj->size();
1016 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1017
1018 size_t old_start = r->index();
1019 size_t old_end = old_start + num_regions - 1;
1020 size_t new_start = heap->heap_region_index_containing(FullGCForwarding::forwardee(old_obj));
1021 size_t new_end = new_start + num_regions - 1;
1022 assert(old_start != new_start, "must be real move");
1023 assert(r->is_stw_move_allowed(), "Region %zu should be movable", r->index());
1024
1025 log_debug(gc)("Full GC compaction moves humongous object from region %zu to region %zu", old_start, new_start);
1026 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
1027 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
1028
1029 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1030 new_obj->init_mark();
1031
1032 {
1033 ShenandoahAffiliation original_affiliation = r->affiliation();
1034 for (size_t c = old_start; c <= old_end; c++) {
1035 ShenandoahHeapRegion* r = heap->get_region(c);
1036 // Leave humongous region affiliation unchanged.
1037 r->make_regular_bypass();
1038 r->set_top(r->bottom());
1039 }
1040
1041 for (size_t c = new_start; c <= new_end; c++) {
1042 ShenandoahHeapRegion* r = heap->get_region(c);
1043 if (c == new_start) {
1044 r->make_humongous_start_bypass(original_affiliation);
1045 } else {
1046 r->make_humongous_cont_bypass(original_affiliation);
1047 }
1048
1049 // Trailing region may be non-full, record the remainder there
1050 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1051 if ((c == new_end) && (remainder != 0)) {
1052 r->set_top(r->bottom() + remainder);
1053 } else {
1054 r->set_top(r->end());
1055 }
1056
1057 r->reset_alloc_metadata();
1058 }
1059 }
1060 }
1061 }
1062 }
1063
1064 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1065 // we need to remain able to walk pinned regions.
1066 // Since pinned region do not move and don't get compacted, we will get holes with
1067 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1068 // cannot be iterated over using oop->size()). The only way to safely iterate over those is using
1069 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1070 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1071 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1072 private:
1073 ShenandoahRegionIterator _regions;
1074
1075 public:
1076 ShenandoahMCResetCompleteBitmapTask() :
1077 WorkerTask("Shenandoah Reset Bitmap") {
1078 }
1079
1080 void work(uint worker_id) override {
1081 ShenandoahParallelWorkerSession worker_session(worker_id);
1082 ShenandoahHeapRegion* region = _regions.next();
1083 ShenandoahHeap* heap = ShenandoahHeap::heap();
1084 ShenandoahMarkingContext* const ctx = heap->marking_context();
1085 assert(heap->global_generation()->is_mark_complete(), "Marking must be complete");
1086 while (region != nullptr) {
1087 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1088 ctx->clear_bitmap(region);
1089 }
1090 region = _regions.next();
1091 }
1092 }
1093 };
1094
1095 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1096 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1097 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1098
1099 ShenandoahHeap* heap = ShenandoahHeap::heap();
1100
1101 // Compact regular objects first
1102 {
1103 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1104 ShenandoahCompactObjectsTask compact_task(worker_slices);
1105 heap->workers()->run_task(&compact_task);
1106 }
1107
1108 // Compact humongous objects after regular object moves
1109 {
1110 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1111 compact_humongous_objects();
1112 }
1113 }
1114
1115 void ShenandoahFullGC::phase5_epilog() {
1116 GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1117 ShenandoahHeap* heap = ShenandoahHeap::heap();
1118
1119 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1120 // and must ensure the bitmap is in sync.
1121 {
1122 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1123 ShenandoahMCResetCompleteBitmapTask task;
1124 heap->workers()->run_task(&task);
1125 }
1126
1127 // Bring regions in proper states after the collection, and set heap properties.
1128 {
1129 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1130 ShenandoahPostCompactClosure post_compact;
1131 heap->heap_region_iterate(&post_compact);
1132 heap->collection_set()->clear();
1133 {
1134 ShenandoahFreeSet* free_set = heap->free_set();
1135 size_t young_trashed_regions, old_trashed_regions, first_old, last_old, num_old;
1136 free_set->prepare_to_rebuild(young_trashed_regions, old_trashed_regions, first_old, last_old, num_old);
1137 // We also do not expand old generation size following Full GC because we have scrambled age populations and
1138 // no longer have objects separated by age into distinct regions.
1139 if (heap->mode()->is_generational()) {
1140 ShenandoahGenerationalFullGC::compute_balances();
1141 }
1142 heap->free_set()->finish_rebuild(young_trashed_regions, old_trashed_regions, num_old);
1143 }
1144
1145 // Set mark incomplete because the marking bitmaps have been reset except pinned regions.
1146 _generation->set_mark_incomplete();
1147
1148 heap->clear_cancelled_gc();
1149 }
1150
1151 _preserved_marks->restore(heap->workers());
1152 _preserved_marks->reclaim();
1153
1154 if (heap->mode()->is_generational()) {
1155 ShenandoahGenerationalFullGC::rebuild_remembered_set(heap);
1156 }
1157 }