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/shenandoahCollectorPolicy.hpp"
38 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
39 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
40 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
41 #include "gc/shenandoah/shenandoahFreeSet.hpp"
42 #include "gc/shenandoah/shenandoahFullGC.hpp"
43 #include "gc/shenandoah/shenandoahGenerationalFullGC.hpp"
44 #include "gc/shenandoah/shenandoahGlobalGeneration.hpp"
45 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
46 #include "gc/shenandoah/shenandoahMark.inline.hpp"
47 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
48 #include "gc/shenandoah/shenandoahHeapRegionClosures.hpp"
49 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
50 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
51 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
52 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
53 #include "gc/shenandoah/shenandoahMetrics.hpp"
54 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
55 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
56 #include "gc/shenandoah/shenandoahSTWMark.hpp"
57 #include "gc/shenandoah/shenandoahUtils.hpp"
58 #include "gc/shenandoah/shenandoahVerifier.hpp"
59 #include "gc/shenandoah/shenandoahVMOperations.hpp"
60 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
61 #include "memory/metaspaceUtils.hpp"
62 #include "memory/universe.hpp"
63 #include "oops/compressedOops.inline.hpp"
64 #include "oops/oop.inline.hpp"
65 #include "runtime/orderAccess.hpp"
66 #include "runtime/vmThread.hpp"
67 #include "utilities/copy.hpp"
68 #include "utilities/events.hpp"
69 #include "utilities/growableArray.hpp"
70
71 ShenandoahFullGC::ShenandoahFullGC() :
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 ShenandoahMetricsSnapshot metrics;
109 metrics.snap_before();
110
111 // Perform full GC
112 do_it(cause);
113
114 ShenandoahHeap* const heap = ShenandoahHeap::heap();
115
116 if (heap->mode()->is_generational()) {
117 ShenandoahGenerationalFullGC::handle_completion(heap);
118 }
119
120 metrics.snap_after();
121
122 if (metrics.is_good_progress(heap->global_generation())) {
123 heap->notify_gc_progress();
124 } else {
125 // Nothing to do. Tell the allocation path that we have failed to make
126 // progress, and it can finally fail.
127 heap->notify_gc_no_progress();
128 }
129
130 // Regardless if progress was made, we record that we completed a "successful" full GC.
131 heap->global_generation()->heuristics()->record_success_full();
132 heap->shenandoah_policy()->record_success_full();
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
143 if (heap->mode()->is_generational()) {
144 ShenandoahGenerationalFullGC::prepare();
145 }
146
147 if (ShenandoahVerify) {
148 heap->verifier()->verify_before_fullgc();
149 }
150
151 if (VerifyBeforeGC) {
152 Universe::verify();
153 }
154
155 // Degenerated GC may carry concurrent root flags when upgrading to
156 // full GC. We need to reset it before mutators resume.
157 heap->set_concurrent_strong_root_in_progress(false);
158 heap->set_concurrent_weak_root_in_progress(false);
159
160 heap->set_full_gc_in_progress(true);
161
162 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
163 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
164
165 {
166 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
167 heap->pre_full_gc_dump(_gc_timer);
168 }
169
170 {
171 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
172 // Full GC is supposed to recover from any GC state:
173
174 // a0. Remember if we have forwarded objects
175 bool has_forwarded_objects = heap->has_forwarded_objects();
176
177 // a1. Cancel evacuation, if in progress
178 if (heap->is_evacuation_in_progress()) {
179 heap->set_evacuation_in_progress(false);
180 }
181 assert(!heap->is_evacuation_in_progress(), "sanity");
182
183 // a2. Cancel update-refs, if in progress
184 if (heap->is_update_refs_in_progress()) {
185 heap->set_update_refs_in_progress(false);
186 }
187 assert(!heap->is_update_refs_in_progress(), "sanity");
188
189 // b. Cancel all concurrent marks, if in progress
190 if (heap->is_concurrent_mark_in_progress()) {
191 heap->cancel_concurrent_mark();
192 }
193 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
194
195 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
196 if (has_forwarded_objects) {
197 update_roots(true /*full_gc*/);
198 }
199
200 // d. Abandon reference discovery and clear all discovered references.
201 ShenandoahReferenceProcessor* rp = heap->global_generation()->ref_processor();
202 rp->abandon_partial_discovery();
203
204 // e. Sync pinned region status from the CP marks
205 heap->sync_pinned_region_status();
206
207 if (heap->mode()->is_generational()) {
208 ShenandoahGenerationalFullGC::restore_top_before_promote(heap);
209 }
210
211 // The rest of prologue:
212 _preserved_marks->init(heap->workers()->active_workers());
213
214 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
215 }
216
217 if (UseTLAB) {
218 // Note: PLABs are also retired with GCLABs in generational mode.
219 heap->gclabs_retire(ResizeTLAB);
220 heap->tlabs_retire(ResizeTLAB);
221 }
222
223 OrderAccess::fence();
224
225 phase1_mark_heap();
226
227 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
228 // Coming out of Full GC, we would not have any forwarded objects.
229 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
230 heap->set_has_forwarded_objects(false);
231
232 heap->set_full_gc_move_in_progress(true);
233
234 // Setup workers for the rest
235 OrderAccess::fence();
236
237 // Initialize worker slices
238 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
239 for (uint i = 0; i < heap->max_workers(); i++) {
240 worker_slices[i] = new ShenandoahHeapRegionSet();
241 }
242
243 {
244 // The rest of code performs region moves, where region status is undefined
245 // until all phases run together.
246 ShenandoahHeapLocker lock(heap->lock());
247
248 FullGCForwarding::begin();
249
250 phase2_calculate_target_addresses(worker_slices);
251
252 OrderAccess::fence();
253
254 phase3_update_references();
255
256 phase4_compact_objects(worker_slices);
257
258 phase5_epilog();
259
260 FullGCForwarding::end();
261 }
262
263 // Resize metaspace
264 MetaspaceGC::compute_new_size();
265
266 // Free worker slices
267 for (uint i = 0; i < heap->max_workers(); i++) {
268 delete worker_slices[i];
269 }
270 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
271
272 heap->set_full_gc_move_in_progress(false);
273 heap->set_full_gc_in_progress(false);
274
275 if (ShenandoahVerify) {
276 heap->verifier()->verify_after_fullgc();
277 }
278
279 if (VerifyAfterGC) {
280 Universe::verify();
281 }
282
283 {
284 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
285 heap->post_full_gc_dump(_gc_timer);
286 }
287 }
288
289 void ShenandoahFullGC::phase1_mark_heap() {
290 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
291 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
292
293 ShenandoahHeap* heap = ShenandoahHeap::heap();
294
295 heap->global_generation()->reset_mark_bitmap<true, true>();
296 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
297 assert(!heap->global_generation()->is_mark_complete(), "sanity");
298
299 heap->set_unload_classes(heap->global_generation()->heuristics()->can_unload_classes());
300
301 ShenandoahReferenceProcessor* rp = heap->global_generation()->ref_processor();
302 // enable ("weak") refs discovery
303 rp->set_soft_reference_policy(true); // forcefully purge all soft references
304
305 ShenandoahSTWMark mark(heap->global_generation(), true /*full_gc*/);
306 mark.mark();
307 heap->parallel_cleaning(true /* full_gc */);
308
309 if (ShenandoahHeap::heap()->mode()->is_generational()) {
310 ShenandoahGenerationalFullGC::log_live_in_old(heap);
311 }
312 }
313
314 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
315 private:
316 PreservedMarks* const _preserved_marks;
317 ShenandoahHeap* const _heap;
318 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
319 int _empty_regions_pos;
320 ShenandoahHeapRegion* _to_region;
321 ShenandoahHeapRegion* _from_region;
322 HeapWord* _compact_point;
323
324 public:
325 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
326 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
327 ShenandoahHeapRegion* to_region) :
328 _preserved_marks(preserved_marks),
329 _heap(ShenandoahHeap::heap()),
330 _empty_regions(empty_regions),
331 _empty_regions_pos(0),
332 _to_region(to_region),
333 _from_region(nullptr),
334 _compact_point(to_region->bottom()) {}
335
336 void set_from_region(ShenandoahHeapRegion* from_region) {
337 _from_region = from_region;
338 }
339
340 void finish() {
341 assert(_to_region != nullptr, "should not happen");
342 _to_region->set_new_top(_compact_point);
343 }
344
345 bool is_compact_same_region() {
346 return _from_region == _to_region;
347 }
348
349 int empty_regions_pos() {
350 return _empty_regions_pos;
351 }
352
353 void do_object(oop p) {
354 assert(_from_region != nullptr, "must set before work");
355 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
356 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
357
358 size_t old_size = p->size();
359 size_t new_size = p->copy_size(old_size, p->mark());
360 size_t obj_size = _compact_point == cast_from_oop<HeapWord*>(p) ? old_size : new_size;
361 if (_compact_point + obj_size > _to_region->end()) {
362 finish();
363
364 // Object doesn't fit. Pick next empty region and start compacting there.
365 ShenandoahHeapRegion* new_to_region;
366 if (_empty_regions_pos < _empty_regions.length()) {
367 new_to_region = _empty_regions.at(_empty_regions_pos);
368 _empty_regions_pos++;
369 } else {
370 // Out of empty region? Compact within the same region.
371 new_to_region = _from_region;
372 }
373
374 assert(new_to_region != _to_region, "must not reuse same to-region");
375 assert(new_to_region != nullptr, "must not be null");
376 _to_region = new_to_region;
377 _compact_point = _to_region->bottom();
378 obj_size = _compact_point == cast_from_oop<HeapWord*>(p) ? old_size : new_size;
379 }
380
381 // Object fits into current region, record new location, if object does not move:
382 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
383 shenandoah_assert_not_forwarded(nullptr, p);
384 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
385 _preserved_marks->push_if_necessary(p, p->mark());
386 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
387 }
388 _compact_point += obj_size;
389 }
390 };
391
392 class ShenandoahPrepareForCompactionTask : public WorkerTask {
393 private:
394 PreservedMarksSet* const _preserved_marks;
395 ShenandoahHeap* const _heap;
396 ShenandoahHeapRegionSet** const _worker_slices;
397
398 public:
399 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
400 WorkerTask("Shenandoah Prepare For Compaction"),
401 _preserved_marks(preserved_marks),
402 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
403 }
404
405 static bool is_candidate_region(ShenandoahHeapRegion* r) {
406 // Empty region: get it into the slice to defragment the slice itself.
407 // We could have skipped this without violating correctness, but we really
408 // want to compact all live regions to the start of the heap, which sometimes
409 // means moving them into the fully empty regions.
410 if (r->is_empty()) return true;
411
412 // Can move the region, and this is not the humongous region. Humongous
413 // moves are special cased here, because their moves are handled separately.
414 return r->is_stw_move_allowed() && !r->is_humongous();
415 }
416
417 void work(uint worker_id) override;
418 private:
419 template<typename ClosureType>
420 void prepare_for_compaction(ClosureType& cl,
421 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
422 ShenandoahHeapRegionSetIterator& it,
423 ShenandoahHeapRegion* from_region);
424 };
425
426 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
427 ShenandoahParallelWorkerSession worker_session(worker_id);
428 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
429 ShenandoahHeapRegionSetIterator it(slice);
430 ShenandoahHeapRegion* from_region = it.next();
431 // No work?
432 if (from_region == nullptr) {
433 return;
434 }
435
436 // Sliding compaction. Walk all regions in the slice, and compact them.
437 // Remember empty regions and reuse them as needed.
438 ResourceMark rm;
439
440 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
441
442 if (_heap->mode()->is_generational()) {
443 ShenandoahPrepareForGenerationalCompactionObjectClosure cl(_preserved_marks->get(worker_id),
444 empty_regions, from_region, worker_id);
445 prepare_for_compaction(cl, empty_regions, it, from_region);
446 } else {
447 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
448 prepare_for_compaction(cl, empty_regions, it, from_region);
449 }
450 }
451
452 template<typename ClosureType>
453 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
454 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
455 ShenandoahHeapRegionSetIterator& it,
456 ShenandoahHeapRegion* from_region) {
457 while (from_region != nullptr) {
458 assert(is_candidate_region(from_region), "Sanity");
459 cl.set_from_region(from_region);
460 if (from_region->has_live()) {
461 _heap->marked_object_iterate(from_region, &cl);
462 }
463
464 // Compacted the region to somewhere else? From-region is empty then.
465 if (!cl.is_compact_same_region()) {
466 empty_regions.append(from_region);
467 }
468 from_region = it.next();
469 }
470 cl.finish();
471
472 // Mark all remaining regions as empty
473 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
474 ShenandoahHeapRegion* r = empty_regions.at(pos);
475 r->set_new_top(r->bottom());
476 }
477 }
478
479 void ShenandoahFullGC::calculate_target_humongous_objects() {
480 ShenandoahHeap* heap = ShenandoahHeap::heap();
481
482 // Compute the new addresses for humongous objects. We need to do this after addresses
483 // for regular objects are calculated, and we know what regions in heap suffix are
484 // available for humongous moves.
485 //
486 // Scan the heap backwards, because we are compacting humongous regions towards the end.
487 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
488 // humongous start there.
489 //
490 // The complication is potential non-movable regions during the scan. If such region is
491 // detected, then sliding restarts towards that non-movable region.
492
493 size_t to_begin = heap->num_regions();
494 size_t to_end = heap->num_regions();
495
496 log_debug(gc)("Full GC calculating target humongous objects from end %zu", to_end);
497 for (size_t c = heap->num_regions(); c > 0; c--) {
498 ShenandoahHeapRegion *r = heap->get_region(c - 1);
499 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
500 // To-region candidate: record this, and continue scan
501 to_begin = r->index();
502 continue;
503 }
504
505 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
506 // From-region candidate: movable humongous region
507 oop old_obj = cast_to_oop(r->bottom());
508 size_t words_size = old_obj->size();
509 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
510
511 size_t start = to_end - num_regions;
512
513 if (start >= to_begin && start != r->index()) {
514 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
515 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
516 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
517 to_end = start;
518 continue;
519 }
520 }
521
522 // Failed to fit. Scan starting from current region.
523 to_begin = r->index();
524 to_end = r->index();
525 }
526 }
527
528 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
529 private:
530 ShenandoahHeap* const _heap;
531
532 public:
533 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
534 void heap_region_do(ShenandoahHeapRegion* r) {
535 if (r->is_trash()) {
536 r->try_recycle_under_lock();
537 }
538 if (r->is_cset()) {
539 // Leave affiliation unchanged
540 r->make_regular_bypass();
541 }
542 if (r->is_empty_uncommitted()) {
543 r->make_committed_bypass();
544 }
545 assert (r->is_committed(), "only committed regions in heap now, see region %zu", r->index());
546
547 // Record current region occupancy: this communicates empty regions are free
548 // to the rest of Full GC code.
549 r->set_new_top(r->top());
550 }
551 };
552
553 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
554 private:
555 ShenandoahHeap* const _heap;
556 ShenandoahMarkingContext* const _ctx;
557
558 public:
559 ShenandoahTrashImmediateGarbageClosure() :
560 _heap(ShenandoahHeap::heap()),
561 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
562
563 void heap_region_do(ShenandoahHeapRegion* r) override {
564 if (r->is_humongous_start()) {
565 oop humongous_obj = cast_to_oop(r->bottom());
566 if (!_ctx->is_marked(humongous_obj)) {
567 assert(!r->has_live(), "Region %zu is not marked, should not have live", r->index());
568 _heap->trash_humongous_region_at(r);
569 } else {
570 assert(r->has_live(), "Region %zu should have live", r->index());
571 }
572 } else if (r->is_humongous_continuation()) {
573 // If we hit continuation, the non-live humongous starts should have been trashed already
574 assert(r->humongous_start_region()->has_live(), "Region %zu should have live", r->index());
575 } else if (r->is_regular()) {
576 if (!r->has_live()) {
577 r->make_trash_immediate();
578 }
579 }
580 }
581 };
582
583 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
584 ShenandoahHeap* heap = ShenandoahHeap::heap();
585
586 uint n_workers = heap->workers()->active_workers();
587 size_t n_regions = heap->num_regions();
588
589 // What we want to accomplish: have the dense prefix of data, while still balancing
590 // out the parallel work.
591 //
592 // Assuming the amount of work is driven by the live data that needs moving, we can slice
593 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
594 // thread takes all regions in its prefix subset, and then it takes some regions from
595 // the tail.
596 //
597 // Tail region selection becomes interesting.
598 //
599 // First, we want to distribute the regions fairly between the workers, and those regions
600 // might have different amount of live data. So, until we sure no workers need live data,
601 // we need to only take what the worker needs.
602 //
603 // Second, since we slide everything to the left in each slice, the most busy regions
604 // would be the ones on the left. Which means we want to have all workers have their after-tail
605 // regions as close to the left as possible.
606 //
607 // The easiest way to do this is to distribute after-tail regions in round-robin between
608 // workers that still need live data.
609 //
610 // Consider parallel workers A, B, C, then the target slice layout would be:
611 //
612 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
613 //
614 // (.....dense-prefix.....) (.....................tail...................)
615 // [all regions fully live] [left-most regions are fuller that right-most]
616 //
617
618 // Compute how much live data is there. This would approximate the size of dense prefix
619 // we target to create.
620 size_t total_live = 0;
621 for (size_t idx = 0; idx < n_regions; idx++) {
622 ShenandoahHeapRegion *r = heap->get_region(idx);
623 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
624 total_live += r->get_live_data_words();
625 }
626 }
627
628 // Estimate the size for the dense prefix. Note that we specifically count only the
629 // "full" regions, so there would be some non-full regions in the slice tail.
630 size_t live_per_worker = total_live / n_workers;
631 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
632 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
633 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
634 assert(prefix_regions_total <= n_regions, "Sanity");
635
636 // There might be non-candidate regions in the prefix. To compute where the tail actually
637 // ends up being, we need to account those as well.
638 size_t prefix_end = prefix_regions_total;
639 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
640 ShenandoahHeapRegion *r = heap->get_region(idx);
641 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
642 prefix_end++;
643 }
644 }
645 prefix_end = MIN2(prefix_end, n_regions);
646 assert(prefix_end <= n_regions, "Sanity");
647
648 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
649 // subset of dense prefix.
650 size_t prefix_idx = 0;
651
652 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
653
654 for (size_t wid = 0; wid < n_workers; wid++) {
655 ShenandoahHeapRegionSet* slice = worker_slices[wid];
656
657 live[wid] = 0;
658 size_t regs = 0;
659
660 // Add all prefix regions for this worker
661 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
662 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
663 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
664 slice->add_region(r);
665 live[wid] += r->get_live_data_words();
666 regs++;
667 }
668 prefix_idx++;
669 }
670 }
671
672 // Distribute the tail among workers in round-robin fashion.
673 size_t wid = n_workers - 1;
674
675 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
676 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
677 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
678 assert(wid < n_workers, "Sanity");
679
680 size_t live_region = r->get_live_data_words();
681
682 // Select next worker that still needs live data.
683 size_t old_wid = wid;
684 do {
685 wid++;
686 if (wid == n_workers) wid = 0;
687 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
688
689 if (old_wid == wid) {
690 // Circled back to the same worker? This means liveness data was
691 // miscalculated. Bump the live_per_worker limit so that
692 // everyone gets a piece of the leftover work.
693 live_per_worker += ShenandoahHeapRegion::region_size_words();
694 }
695
696 worker_slices[wid]->add_region(r);
697 live[wid] += live_region;
698 }
699 }
700
701 FREE_C_HEAP_ARRAY(size_t, live);
702
703 #ifdef ASSERT
704 ResourceBitMap map(n_regions);
705 for (size_t wid = 0; wid < n_workers; wid++) {
706 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
707 ShenandoahHeapRegion* r = it.next();
708 while (r != nullptr) {
709 size_t idx = r->index();
710 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: %zu", idx);
711 assert(!map.at(idx), "No region distributed twice: %zu", idx);
712 map.at_put(idx, true);
713 r = it.next();
714 }
715 }
716
717 for (size_t rid = 0; rid < n_regions; rid++) {
718 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
719 bool is_distributed = map.at(rid);
720 assert(is_distributed || !is_candidate, "All candidates are distributed: %zu", rid);
721 }
722 #endif
723 }
724
725 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
726 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
727 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
728
729 ShenandoahHeap* heap = ShenandoahHeap::heap();
730
731 // About to figure out which regions can be compacted, make sure pinning status
732 // had been updated in GC prologue.
733 heap->assert_pinned_region_status();
734
735 {
736 // Trash the immediately collectible regions before computing addresses
737 ShenandoahTrashImmediateGarbageClosure trash_immediate_garbage;
738 ShenandoahExcludeRegionClosure<FREE> cl(&trash_immediate_garbage);
739 heap->heap_region_iterate(&cl);
740
741 // Make sure regions are in good state: committed, active, clean.
742 // This is needed because we are potentially sliding the data through them.
743 ShenandoahEnsureHeapActiveClosure ecl;
744 heap->heap_region_iterate(&ecl);
745 }
746
747 // Compute the new addresses for regular objects
748 {
749 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
750
751 distribute_slices(worker_slices);
752
753 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
754 heap->workers()->run_task(&task);
755 }
756
757 // Compute the new addresses for humongous objects
758 {
759 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
760 calculate_target_humongous_objects();
761 }
762 }
763
764 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
765 private:
766 ShenandoahHeap* const _heap;
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 _heap(ShenandoahHeap::heap()),
785 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
786
787 void do_oop(oop* p) { do_oop_work(p); }
788 void do_oop(narrowOop* p) { do_oop_work(p); }
789 void do_method(Method* m) {}
790 void do_nmethod(nmethod* nm) {}
791 };
792
793 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
794 private:
795 ShenandoahHeap* const _heap;
796 ShenandoahAdjustPointersClosure _cl;
797
798 public:
799 ShenandoahAdjustPointersObjectClosure() :
800 _heap(ShenandoahHeap::heap()) {
801 }
802 void do_object(oop p) {
803 assert(_heap->complete_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) {
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) {
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 #if COMPILER2_OR_JVMCI
863 DerivedPointerTable::clear();
864 #endif
865 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
866 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
867 workers->run_task(&task);
868 #if COMPILER2_OR_JVMCI
869 DerivedPointerTable::update_pointers();
870 #endif
871 }
872
873 ShenandoahAdjustPointersTask adjust_pointers_task;
874 workers->run_task(&adjust_pointers_task);
875 }
876
877 class ShenandoahCompactObjectsClosure : public ObjectClosure {
878 private:
879 ShenandoahHeap* const _heap;
880 uint const _worker_id;
881
882 public:
883 ShenandoahCompactObjectsClosure(uint worker_id) :
884 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
885
886 void do_object(oop p) {
887 assert(_heap->complete_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 ContinuationGCSupport::relativize_stack_chunk(new_obj);
897 new_obj->init_mark();
898 new_obj->initialize_hash_if_necessary(p);
899 }
900 }
901 };
902
903 class ShenandoahCompactObjectsTask : public WorkerTask {
904 private:
905 ShenandoahHeap* const _heap;
906 ShenandoahHeapRegionSet** const _worker_slices;
907
908 public:
909 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
910 WorkerTask("Shenandoah Compact Objects"),
911 _heap(ShenandoahHeap::heap()),
912 _worker_slices(worker_slices) {
913 }
914
915 void work(uint worker_id) {
916 ShenandoahParallelWorkerSession worker_session(worker_id);
917 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
918
919 ShenandoahCompactObjectsClosure cl(worker_id);
920 ShenandoahHeapRegion* r = slice.next();
921 while (r != nullptr) {
922 assert(!r->is_humongous(), "must not get humongous regions here");
923 if (r->has_live()) {
924 _heap->marked_object_iterate(r, &cl);
925 }
926 r->set_top(r->new_top());
927 r = slice.next();
928 }
929 }
930 };
931
932 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
933 private:
934 ShenandoahHeap* const _heap;
935 bool _is_generational;
936 size_t _young_regions, _young_usage, _young_humongous_waste;
937 size_t _old_regions, _old_usage, _old_humongous_waste;
938
939 public:
940 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()),
941 _is_generational(_heap->mode()->is_generational()),
942 _young_regions(0),
943 _young_usage(0),
944 _young_humongous_waste(0),
945 _old_regions(0),
946 _old_usage(0),
947 _old_humongous_waste(0)
948 {
949 _heap->free_set()->clear();
950 }
951
952 void heap_region_do(ShenandoahHeapRegion* r) {
953 assert (!r->is_cset(), "cset regions should have been demoted already");
954
955 // Need to reset the complete-top-at-mark-start pointer here because
956 // the complete marking bitmap is no longer valid. This ensures
957 // size-based iteration in marked_object_iterate().
958 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
959 // pinned regions.
960 if (!r->is_pinned()) {
961 _heap->complete_marking_context()->reset_top_at_mark_start(r);
962 }
963
964 size_t live = r->used();
965
966 // Make empty regions that have been allocated into regular
967 if (r->is_empty() && live > 0) {
968 if (!_is_generational) {
969 r->make_affiliated_maybe();
970 }
971 // else, generational mode compaction has already established affiliation.
972 r->make_regular_bypass();
973 if (ZapUnusedHeapArea) {
974 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
975 }
976 }
977
978 // Reclaim regular regions that became empty
979 if (r->is_regular() && live == 0) {
980 r->make_trash();
981 }
982
983 // Recycle all trash regions
984 if (r->is_trash()) {
985 live = 0;
986 r->try_recycle_under_lock();
987 } else {
988 if (r->is_old()) {
989 ShenandoahGenerationalFullGC::account_for_region(r, _old_regions, _old_usage, _old_humongous_waste);
990 } else if (r->is_young()) {
991 ShenandoahGenerationalFullGC::account_for_region(r, _young_regions, _young_usage, _young_humongous_waste);
992 }
993 }
994 r->set_live_data(live);
995 r->reset_alloc_metadata();
996 }
997
998 void update_generation_usage() {
999 if (_is_generational) {
1000 _heap->old_generation()->establish_usage(_old_regions, _old_usage, _old_humongous_waste);
1001 _heap->young_generation()->establish_usage(_young_regions, _young_usage, _young_humongous_waste);
1002 } else {
1003 assert(_old_regions == 0, "Old regions only expected in generational mode");
1004 assert(_old_usage == 0, "Old usage only expected in generational mode");
1005 assert(_old_humongous_waste == 0, "Old humongous waste only expected in generational mode");
1006 }
1007
1008 // In generational mode, global usage should be the sum of young and old. This is also true
1009 // for non-generational modes except that there are no old regions.
1010 _heap->global_generation()->establish_usage(_old_regions + _young_regions,
1011 _old_usage + _young_usage,
1012 _old_humongous_waste + _young_humongous_waste);
1013 }
1014 };
1015
1016 void ShenandoahFullGC::compact_humongous_objects() {
1017 // Compact humongous regions, based on their fwdptr objects.
1018 //
1019 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1020 // humongous regions are already compacted, and do not require further moves, which alleviates
1021 // sliding costs. We may consider doing this in parallel in the future.
1022
1023 ShenandoahHeap* heap = ShenandoahHeap::heap();
1024
1025 for (size_t c = heap->num_regions(); c > 0; c--) {
1026 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1027 if (r->is_humongous_start()) {
1028 oop old_obj = cast_to_oop(r->bottom());
1029 if (!FullGCForwarding::is_forwarded(old_obj)) {
1030 // No need to move the object, it stays at the same slot
1031 continue;
1032 }
1033 size_t words_size = old_obj->size();
1034 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1035
1036 size_t old_start = r->index();
1037 size_t old_end = old_start + num_regions - 1;
1038 size_t new_start = heap->heap_region_index_containing(FullGCForwarding::forwardee(old_obj));
1039 size_t new_end = new_start + num_regions - 1;
1040 assert(old_start != new_start, "must be real move");
1041 assert(r->is_stw_move_allowed(), "Region %zu should be movable", r->index());
1042
1043 log_debug(gc)("Full GC compaction moves humongous object from region %zu to region %zu", old_start, new_start);
1044 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
1045 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
1046
1047 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1048 new_obj->init_mark();
1049
1050 {
1051 ShenandoahAffiliation original_affiliation = r->affiliation();
1052 for (size_t c = old_start; c <= old_end; c++) {
1053 ShenandoahHeapRegion* r = heap->get_region(c);
1054 // Leave humongous region affiliation unchanged.
1055 r->make_regular_bypass();
1056 r->set_top(r->bottom());
1057 }
1058
1059 for (size_t c = new_start; c <= new_end; c++) {
1060 ShenandoahHeapRegion* r = heap->get_region(c);
1061 if (c == new_start) {
1062 r->make_humongous_start_bypass(original_affiliation);
1063 } else {
1064 r->make_humongous_cont_bypass(original_affiliation);
1065 }
1066
1067 // Trailing region may be non-full, record the remainder there
1068 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1069 if ((c == new_end) && (remainder != 0)) {
1070 r->set_top(r->bottom() + remainder);
1071 } else {
1072 r->set_top(r->end());
1073 }
1074
1075 r->reset_alloc_metadata();
1076 }
1077 }
1078 }
1079 }
1080 }
1081
1082 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1083 // we need to remain able to walk pinned regions.
1084 // Since pinned region do not move and don't get compacted, we will get holes with
1085 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1086 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1087 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1088 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1089 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1090 private:
1091 ShenandoahRegionIterator _regions;
1092
1093 public:
1094 ShenandoahMCResetCompleteBitmapTask() :
1095 WorkerTask("Shenandoah Reset Bitmap") {
1096 }
1097
1098 void work(uint worker_id) {
1099 ShenandoahParallelWorkerSession worker_session(worker_id);
1100 ShenandoahHeapRegion* region = _regions.next();
1101 ShenandoahHeap* heap = ShenandoahHeap::heap();
1102 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1103 while (region != nullptr) {
1104 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1105 ctx->clear_bitmap(region);
1106 }
1107 region = _regions.next();
1108 }
1109 }
1110 };
1111
1112 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1113 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1114 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1115
1116 ShenandoahHeap* heap = ShenandoahHeap::heap();
1117
1118 // Compact regular objects first
1119 {
1120 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1121 ShenandoahCompactObjectsTask compact_task(worker_slices);
1122 heap->workers()->run_task(&compact_task);
1123 }
1124
1125 // Compact humongous objects after regular object moves
1126 {
1127 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1128 compact_humongous_objects();
1129 }
1130 }
1131
1132 void ShenandoahFullGC::phase5_epilog() {
1133 GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1134 ShenandoahHeap* heap = ShenandoahHeap::heap();
1135
1136 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1137 // and must ensure the bitmap is in sync.
1138 {
1139 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1140 ShenandoahMCResetCompleteBitmapTask task;
1141 heap->workers()->run_task(&task);
1142 }
1143
1144 // Bring regions in proper states after the collection, and set heap properties.
1145 {
1146 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1147 ShenandoahPostCompactClosure post_compact;
1148 heap->heap_region_iterate(&post_compact);
1149 post_compact.update_generation_usage();
1150
1151 if (heap->mode()->is_generational()) {
1152 ShenandoahGenerationalFullGC::balance_generations_after_gc(heap);
1153 }
1154
1155 heap->collection_set()->clear();
1156 size_t young_cset_regions, old_cset_regions;
1157 size_t first_old, last_old, num_old;
1158 heap->free_set()->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old, last_old, num_old);
1159
1160 // We also do not expand old generation size following Full GC because we have scrambled age populations and
1161 // no longer have objects separated by age into distinct regions.
1162 if (heap->mode()->is_generational()) {
1163 ShenandoahGenerationalFullGC::compute_balances();
1164 }
1165
1166 heap->free_set()->finish_rebuild(young_cset_regions, old_cset_regions, num_old);
1167
1168 // Set mark incomplete because the marking bitmaps have been reset except pinned regions.
1169 heap->global_generation()->set_mark_incomplete();
1170
1171 heap->clear_cancelled_gc(true /* clear oom handler */);
1172 }
1173
1174 _preserved_marks->restore(heap->workers());
1175 _preserved_marks->reclaim();
1176
1177 // We defer generation resizing actions until after cset regions have been recycled. We do this even following an
1178 // abbreviated cycle.
1179 if (heap->mode()->is_generational()) {
1180 ShenandoahGenerationalFullGC::balance_generations_after_rebuilding_free_set();
1181 ShenandoahGenerationalFullGC::rebuild_remembered_set(heap);
1182 }
1183 }