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 _gc_timer(ShenandoahHeap::heap()->gc_timer()),
72 _preserved_marks(new PreservedMarksSet(true)) {}
73
74 ShenandoahFullGC::~ShenandoahFullGC() {
75 delete _preserved_marks;
76 }
77
78 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
79 vmop_entry_full(cause);
80 // Always success
81 return true;
82 }
83
84 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
85 ShenandoahHeap* const heap = ShenandoahHeap::heap();
86 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
87 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
88
89 heap->try_inject_alloc_failure();
90 VM_ShenandoahFullGC op(cause, this);
91 VMThread::execute(&op);
92 }
93
94 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
95 static const char* msg = "Pause Full";
96 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
97 EventMark em("%s", msg);
98
99 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
100 ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
101 "full gc");
102
103 op_full(cause);
104 }
105
106 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
107 ShenandoahMetricsSnapshot metrics;
108 metrics.snap_before();
109
110 // Perform full GC
111 do_it(cause);
112
113 ShenandoahHeap* const heap = ShenandoahHeap::heap();
114
115 if (heap->mode()->is_generational()) {
116 ShenandoahGenerationalFullGC::handle_completion(heap);
117 }
118
119 metrics.snap_after();
120
121 if (metrics.is_good_progress(heap->global_generation())) {
122 heap->notify_gc_progress();
123 } else {
124 // Nothing to do. Tell the allocation path that we have failed to make
125 // progress, and it can finally fail.
126 heap->notify_gc_no_progress();
127 }
128
129 // Regardless if progress was made, we record that we completed a "successful" full GC.
130 heap->global_generation()->heuristics()->record_success_full();
131 heap->shenandoah_policy()->record_success_full();
132
133 {
134 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_propagate_gc_state);
135 heap->propagate_gc_state_to_all_threads();
136 }
137 }
138
139 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
140 ShenandoahHeap* heap = ShenandoahHeap::heap();
141
142 if (heap->mode()->is_generational()) {
143 ShenandoahGenerationalFullGC::prepare();
144 }
145
146 if (ShenandoahVerify) {
147 heap->verifier()->verify_before_fullgc();
148 }
149
150 if (VerifyBeforeGC) {
151 Universe::verify();
152 }
153
154 // Degenerated GC may carry concurrent root flags when upgrading to
155 // full GC. We need to reset it before mutators resume.
156 heap->set_concurrent_strong_root_in_progress(false);
157 heap->set_concurrent_weak_root_in_progress(false);
158
159 heap->set_full_gc_in_progress(true);
160
161 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
162 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
163
164 {
165 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
166 heap->pre_full_gc_dump(_gc_timer);
167 }
168
169 {
170 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
171 // Full GC is supposed to recover from any GC state:
172
173 // a0. Remember if we have forwarded objects
174 bool has_forwarded_objects = heap->has_forwarded_objects();
175
176 // a1. Cancel evacuation, if in progress
177 if (heap->is_evacuation_in_progress()) {
178 heap->set_evacuation_in_progress(false);
179 }
180 assert(!heap->is_evacuation_in_progress(), "sanity");
181
182 // a2. Cancel update-refs, if in progress
183 if (heap->is_update_refs_in_progress()) {
184 heap->set_update_refs_in_progress(false);
185 }
186 assert(!heap->is_update_refs_in_progress(), "sanity");
187
188 // b. Cancel all concurrent marks, if in progress
189 if (heap->is_concurrent_mark_in_progress()) {
190 heap->cancel_concurrent_mark();
191 }
192 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
193
194 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
195 if (has_forwarded_objects) {
196 update_roots(true /*full_gc*/);
197 }
198
199 // d. Abandon reference discovery and clear all discovered references.
200 ShenandoahReferenceProcessor* rp = heap->global_generation()->ref_processor();
201 rp->abandon_partial_discovery();
202
203 // e. Sync pinned region status from the CP marks
204 heap->sync_pinned_region_status();
205
206 if (heap->mode()->is_generational()) {
207 ShenandoahGenerationalFullGC::restore_top_before_promote(heap);
208 }
209
210 // The rest of prologue:
211 _preserved_marks->init(heap->workers()->active_workers());
212
213 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
214 }
215
216 if (UseTLAB) {
217 // Note: PLABs are also retired with GCLABs in generational mode.
218 heap->gclabs_retire(ResizeTLAB);
219 heap->tlabs_retire(ResizeTLAB);
220 }
221
222 OrderAccess::fence();
223
224 phase1_mark_heap();
225
226 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
227 // Coming out of Full GC, we would not have any forwarded objects.
228 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
229 heap->set_has_forwarded_objects(false);
230
231 heap->set_full_gc_move_in_progress(true);
232
233 // Setup workers for the rest
234 OrderAccess::fence();
235
236 // Initialize worker slices
237 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
238 for (uint i = 0; i < heap->max_workers(); i++) {
239 worker_slices[i] = new ShenandoahHeapRegionSet();
240 }
241
242 {
243 // The rest of code performs region moves, where region status is undefined
244 // until all phases run together.
245 ShenandoahHeapLocker lock(heap->lock());
246
247 FullGCForwarding::begin();
248
249 phase2_calculate_target_addresses(worker_slices);
250
251 OrderAccess::fence();
252
253 phase3_update_references();
254
255 phase4_compact_objects(worker_slices);
256
257 phase5_epilog();
258
259 FullGCForwarding::end();
260 }
261
262 // Resize metaspace
263 MetaspaceGC::compute_new_size();
264
265 // Free worker slices
266 for (uint i = 0; i < heap->max_workers(); i++) {
267 delete worker_slices[i];
268 }
269 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
270
271 heap->set_full_gc_move_in_progress(false);
272 heap->set_full_gc_in_progress(false);
273
274 if (ShenandoahVerify) {
275 heap->verifier()->verify_after_fullgc();
276 }
277
278 if (VerifyAfterGC) {
279 Universe::verify();
280 }
281
282 {
283 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
284 heap->post_full_gc_dump(_gc_timer);
285 }
286 }
287
288 void ShenandoahFullGC::phase1_mark_heap() {
289 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
290 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
291
292 ShenandoahHeap* heap = ShenandoahHeap::heap();
293
294 heap->global_generation()->reset_mark_bitmap<true, true>();
295 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
296 assert(!heap->global_generation()->is_mark_complete(), "sanity");
297
298 heap->set_unload_classes(heap->global_generation()->heuristics()->can_unload_classes());
299
300 ShenandoahReferenceProcessor* rp = heap->global_generation()->ref_processor();
301 // enable ("weak") refs discovery
302 rp->set_soft_reference_policy(true); // forcefully purge all soft references
303
304 ShenandoahSTWMark mark(heap->global_generation(), true /*full_gc*/);
305 mark.mark();
306 heap->parallel_cleaning(true /* full_gc */);
307
308 if (ShenandoahHeap::heap()->mode()->is_generational()) {
309 ShenandoahGenerationalFullGC::log_live_in_old(heap);
310 }
311 }
312
313 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
314 private:
315 PreservedMarks* const _preserved_marks;
316 ShenandoahHeap* const _heap;
317 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
318 int _empty_regions_pos;
319 ShenandoahHeapRegion* _to_region;
320 ShenandoahHeapRegion* _from_region;
321 HeapWord* _compact_point;
322
323 public:
324 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
325 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
326 ShenandoahHeapRegion* to_region) :
327 _preserved_marks(preserved_marks),
328 _heap(ShenandoahHeap::heap()),
329 _empty_regions(empty_regions),
330 _empty_regions_pos(0),
331 _to_region(to_region),
332 _from_region(nullptr),
333 _compact_point(to_region->bottom()) {}
334
335 void set_from_region(ShenandoahHeapRegion* from_region) {
336 _from_region = from_region;
337 }
338
339 void finish() {
340 assert(_to_region != nullptr, "should not happen");
341 _to_region->set_new_top(_compact_point);
342 }
343
344 bool is_compact_same_region() {
345 return _from_region == _to_region;
346 }
347
348 int empty_regions_pos() {
349 return _empty_regions_pos;
350 }
351
352 void do_object(oop p) {
353 assert(_from_region != nullptr, "must set before work");
354 assert(_heap->gc_generation()->complete_marking_context()->is_marked(p), "must be marked");
355 assert(!_heap->gc_generation()->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
356
357 size_t old_size = p->size();
358 size_t new_size = p->copy_size(old_size, p->mark());
359 size_t obj_size = _compact_point == cast_from_oop<HeapWord*>(p) ? old_size : new_size;
360 if (_compact_point + obj_size > _to_region->end()) {
361 finish();
362
363 // Object doesn't fit. Pick next empty region and start compacting there.
364 ShenandoahHeapRegion* new_to_region;
365 if (_empty_regions_pos < _empty_regions.length()) {
366 new_to_region = _empty_regions.at(_empty_regions_pos);
367 _empty_regions_pos++;
368 } else {
369 // Out of empty region? Compact within the same region.
370 new_to_region = _from_region;
371 }
372
373 assert(new_to_region != _to_region, "must not reuse same to-region");
374 assert(new_to_region != nullptr, "must not be null");
375 _to_region = new_to_region;
376 _compact_point = _to_region->bottom();
377 obj_size = _compact_point == cast_from_oop<HeapWord*>(p) ? old_size : new_size;
378 }
379
380 // Object fits into current region, record new location, if object does not move:
381 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
382 shenandoah_assert_not_forwarded(nullptr, p);
383 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
384 _preserved_marks->push_if_necessary(p, p->mark());
385 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
386 }
387 _compact_point += obj_size;
388 }
389 };
390
391 class ShenandoahPrepareForCompactionTask : public WorkerTask {
392 private:
393 PreservedMarksSet* const _preserved_marks;
394 ShenandoahHeap* const _heap;
395 ShenandoahHeapRegionSet** const _worker_slices;
396
397 public:
398 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
399 WorkerTask("Shenandoah Prepare For Compaction"),
400 _preserved_marks(preserved_marks),
401 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
402 }
403
404 static bool is_candidate_region(ShenandoahHeapRegion* r) {
405 // Empty region: get it into the slice to defragment the slice itself.
406 // We could have skipped this without violating correctness, but we really
407 // want to compact all live regions to the start of the heap, which sometimes
408 // means moving them into the fully empty regions.
409 if (r->is_empty()) return true;
410
411 // Can move the region, and this is not the humongous region. Humongous
412 // moves are special cased here, because their moves are handled separately.
413 return r->is_stw_move_allowed() && !r->is_humongous();
414 }
415
416 void work(uint worker_id) override;
417 private:
418 template<typename ClosureType>
419 void prepare_for_compaction(ClosureType& cl,
420 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
421 ShenandoahHeapRegionSetIterator& it,
422 ShenandoahHeapRegion* from_region);
423 };
424
425 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
426 ShenandoahParallelWorkerSession worker_session(worker_id);
427 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
428 ShenandoahHeapRegionSetIterator it(slice);
429 ShenandoahHeapRegion* from_region = it.next();
430 // No work?
431 if (from_region == nullptr) {
432 return;
433 }
434
435 // Sliding compaction. Walk all regions in the slice, and compact them.
436 // Remember empty regions and reuse them as needed.
437 ResourceMark rm;
438
439 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
440
441 if (_heap->mode()->is_generational()) {
442 ShenandoahPrepareForGenerationalCompactionObjectClosure cl(_preserved_marks->get(worker_id),
443 empty_regions, from_region, worker_id);
444 prepare_for_compaction(cl, empty_regions, it, from_region);
445 } else {
446 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
447 prepare_for_compaction(cl, empty_regions, it, from_region);
448 }
449 }
450
451 template<typename ClosureType>
452 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
453 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
454 ShenandoahHeapRegionSetIterator& it,
455 ShenandoahHeapRegion* from_region) {
456 while (from_region != nullptr) {
457 assert(is_candidate_region(from_region), "Sanity");
458 cl.set_from_region(from_region);
459 if (from_region->has_live()) {
460 _heap->marked_object_iterate(from_region, &cl);
461 }
462
463 // Compacted the region to somewhere else? From-region is empty then.
464 if (!cl.is_compact_same_region()) {
465 empty_regions.append(from_region);
466 }
467 from_region = it.next();
468 }
469 cl.finish();
470
471 // Mark all remaining regions as empty
472 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
473 ShenandoahHeapRegion* r = empty_regions.at(pos);
474 r->set_new_top(r->bottom());
475 }
476 }
477
478 void ShenandoahFullGC::calculate_target_humongous_objects() {
479 ShenandoahHeap* heap = ShenandoahHeap::heap();
480
481 // Compute the new addresses for humongous objects. We need to do this after addresses
482 // for regular objects are calculated, and we know what regions in heap suffix are
483 // available for humongous moves.
484 //
485 // Scan the heap backwards, because we are compacting humongous regions towards the end.
486 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
487 // humongous start there.
488 //
489 // The complication is potential non-movable regions during the scan. If such region is
490 // detected, then sliding restarts towards that non-movable region.
491
492 size_t to_begin = heap->num_regions();
493 size_t to_end = heap->num_regions();
494
495 log_debug(gc)("Full GC calculating target humongous objects from end %zu", to_end);
496 for (size_t c = heap->num_regions(); c > 0; c--) {
497 ShenandoahHeapRegion *r = heap->get_region(c - 1);
498 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
499 // To-region candidate: record this, and continue scan
500 to_begin = r->index();
501 continue;
502 }
503
504 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
505 // From-region candidate: movable humongous region
506 oop old_obj = cast_to_oop(r->bottom());
507 size_t words_size = old_obj->size();
508 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
509
510 size_t start = to_end - num_regions;
511
512 if (start >= to_begin && start != r->index()) {
513 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
514 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
515 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
516 to_end = start;
517 continue;
518 }
519 }
520
521 // Failed to fit. Scan starting from current region.
522 to_begin = r->index();
523 to_end = r->index();
524 }
525 }
526
527 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
528 private:
529 ShenandoahHeap* const _heap;
530
531 public:
532 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
533 void heap_region_do(ShenandoahHeapRegion* r) {
534 if (r->is_trash()) {
535 r->try_recycle_under_lock();
536 }
537 if (r->is_cset()) {
538 // Leave affiliation unchanged
539 r->make_regular_bypass();
540 }
541 if (r->is_empty_uncommitted()) {
542 r->make_committed_bypass();
543 }
544 assert (r->is_committed(), "only committed regions in heap now, see region %zu", r->index());
545
546 // Record current region occupancy: this communicates empty regions are free
547 // to the rest of Full GC code.
548 r->set_new_top(r->top());
549 }
550 };
551
552 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
553 private:
554 ShenandoahHeap* const _heap;
555 ShenandoahMarkingContext* const _ctx;
556
557 public:
558 ShenandoahTrashImmediateGarbageClosure() :
559 _heap(ShenandoahHeap::heap()),
560 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
561
562 void heap_region_do(ShenandoahHeapRegion* r) override {
563 if (r->is_humongous_start()) {
564 oop humongous_obj = cast_to_oop(r->bottom());
565 if (!_ctx->is_marked(humongous_obj)) {
566 assert(!r->has_live(), "Region %zu is not marked, should not have live", r->index());
567 _heap->trash_humongous_region_at(r);
568 } else {
569 assert(r->has_live(), "Region %zu should have live", r->index());
570 }
571 } else if (r->is_humongous_continuation()) {
572 // If we hit continuation, the non-live humongous starts should have been trashed already
573 assert(r->humongous_start_region()->has_live(), "Region %zu should have live", r->index());
574 } else if (r->is_regular()) {
575 if (!r->has_live()) {
576 r->make_trash_immediate();
577 }
578 }
579 }
580 };
581
582 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
583 ShenandoahHeap* heap = ShenandoahHeap::heap();
584
585 uint n_workers = heap->workers()->active_workers();
586 size_t n_regions = heap->num_regions();
587
588 // What we want to accomplish: have the dense prefix of data, while still balancing
589 // out the parallel work.
590 //
591 // Assuming the amount of work is driven by the live data that needs moving, we can slice
592 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
593 // thread takes all regions in its prefix subset, and then it takes some regions from
594 // the tail.
595 //
596 // Tail region selection becomes interesting.
597 //
598 // First, we want to distribute the regions fairly between the workers, and those regions
599 // might have different amount of live data. So, until we sure no workers need live data,
600 // we need to only take what the worker needs.
601 //
602 // Second, since we slide everything to the left in each slice, the most busy regions
603 // would be the ones on the left. Which means we want to have all workers have their after-tail
604 // regions as close to the left as possible.
605 //
606 // The easiest way to do this is to distribute after-tail regions in round-robin between
607 // workers that still need live data.
608 //
609 // Consider parallel workers A, B, C, then the target slice layout would be:
610 //
611 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
612 //
613 // (.....dense-prefix.....) (.....................tail...................)
614 // [all regions fully live] [left-most regions are fuller that right-most]
615 //
616
617 // Compute how much live data is there. This would approximate the size of dense prefix
618 // we target to create.
619 size_t total_live = 0;
620 for (size_t idx = 0; idx < n_regions; idx++) {
621 ShenandoahHeapRegion *r = heap->get_region(idx);
622 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
623 total_live += r->get_live_data_words();
624 }
625 }
626
627 // Estimate the size for the dense prefix. Note that we specifically count only the
628 // "full" regions, so there would be some non-full regions in the slice tail.
629 size_t live_per_worker = total_live / n_workers;
630 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
631 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
632 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
633 assert(prefix_regions_total <= n_regions, "Sanity");
634
635 // There might be non-candidate regions in the prefix. To compute where the tail actually
636 // ends up being, we need to account those as well.
637 size_t prefix_end = prefix_regions_total;
638 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
639 ShenandoahHeapRegion *r = heap->get_region(idx);
640 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
641 prefix_end++;
642 }
643 }
644 prefix_end = MIN2(prefix_end, n_regions);
645 assert(prefix_end <= n_regions, "Sanity");
646
647 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
648 // subset of dense prefix.
649 size_t prefix_idx = 0;
650
651 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
652
653 for (size_t wid = 0; wid < n_workers; wid++) {
654 ShenandoahHeapRegionSet* slice = worker_slices[wid];
655
656 live[wid] = 0;
657 size_t regs = 0;
658
659 // Add all prefix regions for this worker
660 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
661 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
662 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
663 slice->add_region(r);
664 live[wid] += r->get_live_data_words();
665 regs++;
666 }
667 prefix_idx++;
668 }
669 }
670
671 // Distribute the tail among workers in round-robin fashion.
672 size_t wid = n_workers - 1;
673
674 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
675 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
676 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
677 assert(wid < n_workers, "Sanity");
678
679 size_t live_region = r->get_live_data_words();
680
681 // Select next worker that still needs live data.
682 size_t old_wid = wid;
683 do {
684 wid++;
685 if (wid == n_workers) wid = 0;
686 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
687
688 if (old_wid == wid) {
689 // Circled back to the same worker? This means liveness data was
690 // miscalculated. Bump the live_per_worker limit so that
691 // everyone gets a piece of the leftover work.
692 live_per_worker += ShenandoahHeapRegion::region_size_words();
693 }
694
695 worker_slices[wid]->add_region(r);
696 live[wid] += live_region;
697 }
698 }
699
700 FREE_C_HEAP_ARRAY(size_t, live);
701
702 #ifdef ASSERT
703 ResourceBitMap map(n_regions);
704 for (size_t wid = 0; wid < n_workers; wid++) {
705 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
706 ShenandoahHeapRegion* r = it.next();
707 while (r != nullptr) {
708 size_t idx = r->index();
709 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: %zu", idx);
710 assert(!map.at(idx), "No region distributed twice: %zu", idx);
711 map.at_put(idx, true);
712 r = it.next();
713 }
714 }
715
716 for (size_t rid = 0; rid < n_regions; rid++) {
717 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
718 bool is_distributed = map.at(rid);
719 assert(is_distributed || !is_candidate, "All candidates are distributed: %zu", rid);
720 }
721 #endif
722 }
723
724 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
725 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
726 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
727
728 ShenandoahHeap* heap = ShenandoahHeap::heap();
729
730 // About to figure out which regions can be compacted, make sure pinning status
731 // had been updated in GC prologue.
732 heap->assert_pinned_region_status();
733
734 {
735 // Trash the immediately collectible regions before computing addresses
736 ShenandoahTrashImmediateGarbageClosure trash_immediate_garbage;
737 ShenandoahExcludeRegionClosure<FREE> cl(&trash_immediate_garbage);
738 heap->heap_region_iterate(&cl);
739
740 // Make sure regions are in good state: committed, active, clean.
741 // This is needed because we are potentially sliding the data through them.
742 ShenandoahEnsureHeapActiveClosure ecl;
743 heap->heap_region_iterate(&ecl);
744 }
745
746 // Compute the new addresses for regular objects
747 {
748 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
749
750 distribute_slices(worker_slices);
751
752 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
753 heap->workers()->run_task(&task);
754 }
755
756 // Compute the new addresses for humongous objects
757 {
758 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
759 calculate_target_humongous_objects();
760 }
761 }
762
763 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
764 private:
765 ShenandoahHeap* const _heap;
766 ShenandoahMarkingContext* const _ctx;
767
768 template <class T>
769 inline void do_oop_work(T* p) {
770 T o = RawAccess<>::oop_load(p);
771 if (!CompressedOops::is_null(o)) {
772 oop obj = CompressedOops::decode_not_null(o);
773 assert(_ctx->is_marked(obj), "must be marked");
774 if (FullGCForwarding::is_forwarded(obj)) {
775 oop forw = FullGCForwarding::forwardee(obj);
776 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
777 }
778 }
779 }
780
781 public:
782 ShenandoahAdjustPointersClosure() :
783 _heap(ShenandoahHeap::heap()),
784 _ctx(ShenandoahHeap::heap()->gc_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 ShenandoahHeap* const _heap;
795 ShenandoahAdjustPointersClosure _cl;
796
797 public:
798 ShenandoahAdjustPointersObjectClosure() :
799 _heap(ShenandoahHeap::heap()) {
800 }
801 void do_object(oop p) {
802 assert(_heap->gc_generation()->complete_marking_context()->is_marked(p), "must be marked");
803 p->oop_iterate(&_cl);
804 }
805 };
806
807 class ShenandoahAdjustPointersTask : public WorkerTask {
808 private:
809 ShenandoahHeap* const _heap;
810 ShenandoahRegionIterator _regions;
811
812 public:
813 ShenandoahAdjustPointersTask() :
814 WorkerTask("Shenandoah Adjust Pointers"),
815 _heap(ShenandoahHeap::heap()) {
816 }
817
818 void work(uint worker_id) {
819 ShenandoahParallelWorkerSession worker_session(worker_id);
820 ShenandoahAdjustPointersObjectClosure obj_cl;
821 ShenandoahHeapRegion* r = _regions.next();
822 while (r != nullptr) {
823 if (!r->is_humongous_continuation() && r->has_live()) {
824 _heap->marked_object_iterate(r, &obj_cl);
825 }
826 if (_heap->mode()->is_generational()) {
827 ShenandoahGenerationalFullGC::maybe_coalesce_and_fill_region(r);
828 }
829 r = _regions.next();
830 }
831 }
832 };
833
834 class ShenandoahAdjustRootPointersTask : public WorkerTask {
835 private:
836 ShenandoahRootAdjuster* _rp;
837 PreservedMarksSet* _preserved_marks;
838 public:
839 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
840 WorkerTask("Shenandoah Adjust Root Pointers"),
841 _rp(rp),
842 _preserved_marks(preserved_marks) {}
843
844 void work(uint worker_id) {
845 ShenandoahParallelWorkerSession worker_session(worker_id);
846 ShenandoahAdjustPointersClosure cl;
847 _rp->roots_do(worker_id, &cl);
848 _preserved_marks->get(worker_id)->adjust_during_full_gc();
849 }
850 };
851
852 void ShenandoahFullGC::phase3_update_references() {
853 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
854 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
855
856 ShenandoahHeap* heap = ShenandoahHeap::heap();
857
858 WorkerThreads* workers = heap->workers();
859 uint nworkers = workers->active_workers();
860 {
861 #if COMPILER2_OR_JVMCI
862 DerivedPointerTable::clear();
863 #endif
864 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
865 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
866 workers->run_task(&task);
867 #if COMPILER2_OR_JVMCI
868 DerivedPointerTable::update_pointers();
869 #endif
870 }
871
872 ShenandoahAdjustPointersTask adjust_pointers_task;
873 workers->run_task(&adjust_pointers_task);
874 }
875
876 class ShenandoahCompactObjectsClosure : public ObjectClosure {
877 private:
878 ShenandoahHeap* const _heap;
879 uint const _worker_id;
880
881 public:
882 ShenandoahCompactObjectsClosure(uint worker_id) :
883 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
884
885 void do_object(oop p) {
886 assert(_heap->gc_generation()->complete_marking_context()->is_marked(p), "must be marked");
887 size_t size = p->size();
888 if (FullGCForwarding::is_forwarded(p)) {
889 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
890 HeapWord* compact_to = cast_from_oop<HeapWord*>(FullGCForwarding::forwardee(p));
891 assert(compact_from != compact_to, "Forwarded object should move");
892 Copy::aligned_conjoint_words(compact_from, compact_to, size);
893 oop new_obj = cast_to_oop(compact_to);
894
895 ContinuationGCSupport::relativize_stack_chunk(new_obj);
896 new_obj->init_mark();
897 new_obj->initialize_hash_if_necessary(p);
898 }
899 }
900 };
901
902 class ShenandoahCompactObjectsTask : public WorkerTask {
903 private:
904 ShenandoahHeap* const _heap;
905 ShenandoahHeapRegionSet** const _worker_slices;
906
907 public:
908 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
909 WorkerTask("Shenandoah Compact Objects"),
910 _heap(ShenandoahHeap::heap()),
911 _worker_slices(worker_slices) {
912 }
913
914 void work(uint worker_id) {
915 ShenandoahParallelWorkerSession worker_session(worker_id);
916 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
917
918 ShenandoahCompactObjectsClosure cl(worker_id);
919 ShenandoahHeapRegion* r = slice.next();
920 while (r != nullptr) {
921 assert(!r->is_humongous(), "must not get humongous regions here");
922 if (r->has_live()) {
923 _heap->marked_object_iterate(r, &cl);
924 }
925 r->set_top(r->new_top());
926 r = slice.next();
927 }
928 }
929 };
930
931 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
932 private:
933 ShenandoahHeap* const _heap;
934 bool _is_generational;
935 size_t _young_regions, _young_usage, _young_humongous_waste;
936 size_t _old_regions, _old_usage, _old_humongous_waste;
937
938 public:
939 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()),
940 _is_generational(_heap->mode()->is_generational()),
941 _young_regions(0),
942 _young_usage(0),
943 _young_humongous_waste(0),
944 _old_regions(0),
945 _old_usage(0),
946 _old_humongous_waste(0)
947 {
948 _heap->free_set()->clear();
949 }
950
951 void heap_region_do(ShenandoahHeapRegion* r) {
952 assert (!r->is_cset(), "cset regions should have been demoted already");
953
954 // Need to reset the complete-top-at-mark-start pointer here because
955 // the complete marking bitmap is no longer valid. This ensures
956 // size-based iteration in marked_object_iterate().
957 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
958 // pinned regions.
959 if (!r->is_pinned()) {
960 _heap->gc_generation()->complete_marking_context()->reset_top_at_mark_start(r);
961 }
962
963 size_t live = r->used();
964
965 // Make empty regions that have been allocated into regular
966 if (r->is_empty() && live > 0) {
967 if (!_is_generational) {
968 r->make_affiliated_maybe();
969 }
970 // else, generational mode compaction has already established affiliation.
971 r->make_regular_bypass();
972 if (ZapUnusedHeapArea) {
973 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
974 }
975 }
976
977 // Reclaim regular regions that became empty
978 if (r->is_regular() && live == 0) {
979 r->make_trash();
980 }
981
982 // Recycle all trash regions
983 if (r->is_trash()) {
984 live = 0;
985 r->try_recycle_under_lock();
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 void update_generation_usage() {
998 if (_is_generational) {
999 _heap->old_generation()->establish_usage(_old_regions, _old_usage, _old_humongous_waste);
1000 _heap->young_generation()->establish_usage(_young_regions, _young_usage, _young_humongous_waste);
1001 } else {
1002 assert(_old_regions == 0, "Old regions only expected in generational mode");
1003 assert(_old_usage == 0, "Old usage only expected in generational mode");
1004 assert(_old_humongous_waste == 0, "Old humongous waste only expected in generational mode");
1005 }
1006
1007 // In generational mode, global usage should be the sum of young and old. This is also true
1008 // for non-generational modes except that there are no old regions.
1009 _heap->global_generation()->establish_usage(_old_regions + _young_regions,
1010 _old_usage + _young_usage,
1011 _old_humongous_waste + _young_humongous_waste);
1012 }
1013 };
1014
1015 void ShenandoahFullGC::compact_humongous_objects() {
1016 // Compact humongous regions, based on their fwdptr objects.
1017 //
1018 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1019 // humongous regions are already compacted, and do not require further moves, which alleviates
1020 // sliding costs. We may consider doing this in parallel in the future.
1021
1022 ShenandoahHeap* heap = ShenandoahHeap::heap();
1023
1024 for (size_t c = heap->num_regions(); c > 0; c--) {
1025 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1026 if (r->is_humongous_start()) {
1027 oop old_obj = cast_to_oop(r->bottom());
1028 if (!FullGCForwarding::is_forwarded(old_obj)) {
1029 // No need to move the object, it stays at the same slot
1030 continue;
1031 }
1032 size_t words_size = old_obj->size();
1033 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1034
1035 size_t old_start = r->index();
1036 size_t old_end = old_start + num_regions - 1;
1037 size_t new_start = heap->heap_region_index_containing(FullGCForwarding::forwardee(old_obj));
1038 size_t new_end = new_start + num_regions - 1;
1039 assert(old_start != new_start, "must be real move");
1040 assert(r->is_stw_move_allowed(), "Region %zu should be movable", r->index());
1041
1042 log_debug(gc)("Full GC compaction moves humongous object from region %zu to region %zu", old_start, new_start);
1043 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
1044 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
1045
1046 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1047 new_obj->init_mark();
1048
1049 {
1050 ShenandoahAffiliation original_affiliation = r->affiliation();
1051 for (size_t c = old_start; c <= old_end; c++) {
1052 ShenandoahHeapRegion* r = heap->get_region(c);
1053 // Leave humongous region affiliation unchanged.
1054 r->make_regular_bypass();
1055 r->set_top(r->bottom());
1056 }
1057
1058 for (size_t c = new_start; c <= new_end; c++) {
1059 ShenandoahHeapRegion* r = heap->get_region(c);
1060 if (c == new_start) {
1061 r->make_humongous_start_bypass(original_affiliation);
1062 } else {
1063 r->make_humongous_cont_bypass(original_affiliation);
1064 }
1065
1066 // Trailing region may be non-full, record the remainder there
1067 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1068 if ((c == new_end) && (remainder != 0)) {
1069 r->set_top(r->bottom() + remainder);
1070 } else {
1071 r->set_top(r->end());
1072 }
1073
1074 r->reset_alloc_metadata();
1075 }
1076 }
1077 }
1078 }
1079 }
1080
1081 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1082 // we need to remain able to walk pinned regions.
1083 // Since pinned region do not move and don't get compacted, we will get holes with
1084 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1085 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1086 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1087 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1088 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1089 private:
1090 ShenandoahRegionIterator _regions;
1091
1092 public:
1093 ShenandoahMCResetCompleteBitmapTask() :
1094 WorkerTask("Shenandoah Reset Bitmap") {
1095 }
1096
1097 void work(uint worker_id) {
1098 ShenandoahParallelWorkerSession worker_session(worker_id);
1099 ShenandoahHeapRegion* region = _regions.next();
1100 ShenandoahHeap* heap = ShenandoahHeap::heap();
1101 ShenandoahMarkingContext* const ctx = heap->gc_generation()->complete_marking_context();
1102 while (region != nullptr) {
1103 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1104 ctx->clear_bitmap(region);
1105 }
1106 region = _regions.next();
1107 }
1108 }
1109 };
1110
1111 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1112 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1113 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1114
1115 ShenandoahHeap* heap = ShenandoahHeap::heap();
1116
1117 // Compact regular objects first
1118 {
1119 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1120 ShenandoahCompactObjectsTask compact_task(worker_slices);
1121 heap->workers()->run_task(&compact_task);
1122 }
1123
1124 // Compact humongous objects after regular object moves
1125 {
1126 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1127 compact_humongous_objects();
1128 }
1129 }
1130
1131 void ShenandoahFullGC::phase5_epilog() {
1132 GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1133 ShenandoahHeap* heap = ShenandoahHeap::heap();
1134
1135 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1136 // and must ensure the bitmap is in sync.
1137 {
1138 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1139 ShenandoahMCResetCompleteBitmapTask task;
1140 heap->workers()->run_task(&task);
1141 }
1142
1143 // Bring regions in proper states after the collection, and set heap properties.
1144 {
1145 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1146 ShenandoahPostCompactClosure post_compact;
1147 heap->heap_region_iterate(&post_compact);
1148 post_compact.update_generation_usage();
1149
1150 if (heap->mode()->is_generational()) {
1151 ShenandoahGenerationalFullGC::balance_generations_after_gc(heap);
1152 }
1153
1154 heap->collection_set()->clear();
1155 size_t young_cset_regions, old_cset_regions;
1156 size_t first_old, last_old, num_old;
1157 heap->free_set()->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old, last_old, num_old);
1158
1159 // We also do not expand old generation size following Full GC because we have scrambled age populations and
1160 // no longer have objects separated by age into distinct regions.
1161 if (heap->mode()->is_generational()) {
1162 ShenandoahGenerationalFullGC::compute_balances();
1163 }
1164
1165 heap->free_set()->finish_rebuild(young_cset_regions, old_cset_regions, num_old);
1166
1167 // Set mark incomplete because the marking bitmaps have been reset except pinned regions.
1168 heap->global_generation()->set_mark_incomplete();
1169
1170 heap->clear_cancelled_gc(true /* clear oom handler */);
1171 }
1172
1173 _preserved_marks->restore(heap->workers());
1174 _preserved_marks->reclaim();
1175
1176 // We defer generation resizing actions until after cset regions have been recycled. We do this even following an
1177 // abbreviated cycle.
1178 if (heap->mode()->is_generational()) {
1179 ShenandoahGenerationalFullGC::balance_generations_after_rebuilding_free_set();
1180 ShenandoahGenerationalFullGC::rebuild_remembered_set(heap);
1181 }
1182 }