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