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/shenandoahStackWatermark.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 ShenandoahGC(ShenandoahHeap::heap()->global_generation()),
73 _gc_timer(ShenandoahHeap::heap()->gc_timer()),
74 _preserved_marks(new PreservedMarksSet(true)) {}
75
76 ShenandoahFullGC::~ShenandoahFullGC() {
77 delete _preserved_marks;
78 }
79
80 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
81 vmop_entry_full(cause);
82 // Always success
83 return true;
84 }
85
86 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
87 ShenandoahHeap* const heap = ShenandoahHeap::heap();
88 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
89 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
90
91 heap->try_inject_alloc_failure();
92 VM_ShenandoahFullGC op(cause, this);
93 VMThread::execute(&op);
94 }
95
96 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
97 static const char* msg = "Pause Full";
98 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
99 EventMark em("%s", msg);
100
101 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
102 ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
103 "full gc");
104
105 op_full(cause);
106 }
107
108 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
109 ShenandoahHeap* const heap = ShenandoahHeap::heap();
110
111 ShenandoahMetricsSnapshot metrics(heap->free_set());
112
113 // Perform full GC
114 do_it(cause);
115
116 if (heap->mode()->is_generational()) {
117 ShenandoahGenerationalFullGC::handle_completion(heap);
118 }
119
120 if (metrics.is_good_progress()) {
121 heap->notify_gc_progress();
122 } else {
123 // Nothing to do. Tell the allocation path that we have failed to make
124 // progress, and it can finally fail.
125 heap->notify_gc_no_progress();
126 }
127
128 // Regardless if progress was made, we record that we completed a "successful" full GC.
129 _generation->heuristics()->record_success_full();
130 heap->shenandoah_policy()->record_success_full();
131
132 // Leaving full GC, we need to flip barriers back to idle.
133 ShenandoahCodeRoots::arm_nmethods();
134 ShenandoahStackWatermark::change_epoch_id();
135
136 {
137 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_propagate_gc_state);
138 heap->propagate_gc_state_to_all_threads();
139 }
140 }
141
142 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
143 ShenandoahHeap* heap = ShenandoahHeap::heap();
144
145 if (heap->mode()->is_generational()) {
146 ShenandoahGenerationalFullGC::prepare();
147 }
148
149 if (ShenandoahVerify) {
150 heap->verifier()->verify_before_fullgc(_generation);
151 }
152
153 if (VerifyBeforeGC) {
154 Universe::verify();
155 }
156
157 // Degenerated GC may carry concurrent root flags when upgrading to
158 // full GC. We need to reset it before mutators resume.
159 heap->set_concurrent_strong_root_in_progress(false);
160 heap->set_concurrent_weak_root_in_progress(false);
161
162 heap->set_full_gc_in_progress(true);
163
164 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
165 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
166
167 {
168 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
169 heap->pre_full_gc_dump(_gc_timer);
170 }
171
172 {
173 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
174 // Full GC is supposed to recover from any GC state:
175
176 // a0. Remember if we have forwarded objects
177 bool has_forwarded_objects = heap->has_forwarded_objects();
178
179 // a1. Cancel evacuation, if in progress
180 if (heap->is_evacuation_in_progress()) {
181 heap->set_evacuation_in_progress(false);
182 }
183 assert(!heap->is_evacuation_in_progress(), "sanity");
184
185 // a2. Cancel update-refs, if in progress
186 if (heap->is_update_refs_in_progress()) {
187 heap->set_update_refs_in_progress(false);
188 }
189 assert(!heap->is_update_refs_in_progress(), "sanity");
190
191 // b. Cancel all concurrent marks, if in progress
192 if (heap->is_concurrent_mark_in_progress()) {
193 heap->cancel_concurrent_mark();
194 }
195 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
196
197 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
198 if (has_forwarded_objects) {
199 update_roots(true /*full_gc*/);
200 }
201
202 // d. Abandon reference discovery and clear all discovered references.
203 ShenandoahReferenceProcessor* rp = _generation->ref_processor();
204 rp->abandon_partial_discovery();
205
206 // e. Sync pinned region status from the CP marks
207 heap->sync_pinned_region_status();
208
209 if (heap->mode()->is_generational()) {
210 ShenandoahGenerationalFullGC::restore_top_before_promote(heap);
211 }
212
213 // The rest of prologue:
214 _preserved_marks->init(heap->workers()->active_workers());
215
216 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
217 }
218
219 if (UseTLAB) {
220 // Note: PLABs are also retired with GCLABs in generational mode.
221 heap->gclabs_retire(ResizeTLAB);
222 heap->tlabs_retire(ResizeTLAB);
223 }
224
225 OrderAccess::fence();
226
227 phase1_mark_heap();
228
229 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
230 // Coming out of Full GC, we would not have any forwarded objects.
231 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
232 heap->set_has_forwarded_objects(false);
233
234 heap->set_full_gc_move_in_progress(true);
235
236 // Setup workers for the rest
237 OrderAccess::fence();
238
239 // Initialize worker slices
240 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
241 for (uint i = 0; i < heap->max_workers(); i++) {
242 worker_slices[i] = new ShenandoahHeapRegionSet();
243 }
244
245 {
246 // The rest of code performs region moves, where region status is undefined
247 // until all phases run together.
248 ShenandoahHeapLocker lock(heap->lock());
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 heap->start_idle_span();
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(_generation);
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 _generation->reset_mark_bitmap<true, true>();
295 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
296 assert(!_generation->is_mark_complete(), "sanity");
297
298 heap->set_unload_classes(_generation->heuristics()->can_unload_classes());
299
300 ShenandoahReferenceProcessor* rp = _generation->ref_processor();
301 // enable ("weak") refs discovery
302 rp->set_soft_reference_policy(true); // forcefully purge all soft references
303
304 ShenandoahSTWMark mark(_generation, true /*full_gc*/);
305 mark.mark();
306 heap->parallel_cleaning(_generation, 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) override {
353 shenandoah_assert_mark_complete(cast_from_oop<HeapWord*>(p));
354 assert(_from_region != nullptr, "must set before work");
355 assert(_heap->global_generation()->is_mark_complete(), "marking must be finished");
356 assert(_heap->marking_context()->is_marked(p), "must be marked");
357 assert(!_heap->marking_context()->allocated_after_mark_start(p), "must be truly marked");
358
359 size_t obj_size = p->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 }
378
379 // Object fits into current region, record new location, if object does not move:
380 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
381 shenandoah_assert_not_forwarded(nullptr, p);
382 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
383 _preserved_marks->push_if_necessary(p, p->mark());
384 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
385 }
386 _compact_point += obj_size;
387 }
388 };
389
390 class ShenandoahPrepareForCompactionTask : public WorkerTask {
391 private:
392 PreservedMarksSet* const _preserved_marks;
393 ShenandoahHeap* const _heap;
394 ShenandoahHeapRegionSet** const _worker_slices;
395
396 public:
397 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
398 WorkerTask("Shenandoah Prepare For Compaction"),
399 _preserved_marks(preserved_marks),
400 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
401 }
402
403 static bool is_candidate_region(ShenandoahHeapRegion* r) {
404 // Empty region: get it into the slice to defragment the slice itself.
405 // We could have skipped this without violating correctness, but we really
406 // want to compact all live regions to the start of the heap, which sometimes
407 // means moving them into the fully empty regions.
408 if (r->is_empty()) return true;
409
410 // Can move the region, and this is not the humongous region. Humongous
411 // moves are special cased here, because their moves are handled separately.
412 return r->is_stw_move_allowed() && !r->is_humongous();
413 }
414
415 void work(uint worker_id) override;
416 private:
417 template<typename ClosureType>
418 void prepare_for_compaction(ClosureType& cl,
419 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
420 ShenandoahHeapRegionSetIterator& it,
421 ShenandoahHeapRegion* from_region);
422 };
423
424 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
425 ShenandoahParallelWorkerSession worker_session(worker_id);
426 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
427 ShenandoahHeapRegionSetIterator it(slice);
428 ShenandoahHeapRegion* from_region = it.next();
429 // No work?
430 if (from_region == nullptr) {
431 return;
432 }
433
434 // Sliding compaction. Walk all regions in the slice, and compact them.
435 // Remember empty regions and reuse them as needed.
436 ResourceMark rm;
437
438 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
439
440 if (_heap->mode()->is_generational()) {
441 ShenandoahPrepareForGenerationalCompactionObjectClosure cl(_preserved_marks->get(worker_id),
442 empty_regions, from_region, worker_id);
443 prepare_for_compaction(cl, empty_regions, it, from_region);
444 } else {
445 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
446 prepare_for_compaction(cl, empty_regions, it, from_region);
447 }
448 }
449
450 template<typename ClosureType>
451 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
452 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
453 ShenandoahHeapRegionSetIterator& it,
454 ShenandoahHeapRegion* from_region) {
455 while (from_region != nullptr) {
456 assert(is_candidate_region(from_region), "Sanity");
457 cl.set_from_region(from_region);
458 if (from_region->has_live()) {
459 _heap->marked_object_iterate(from_region, &cl);
460 }
461
462 // Compacted the region to somewhere else? From-region is empty then.
463 if (!cl.is_compact_same_region()) {
464 empty_regions.append(from_region);
465 }
466 from_region = it.next();
467 }
468 cl.finish();
469
470 // Mark all remaining regions as empty
471 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
472 ShenandoahHeapRegion* r = empty_regions.at(pos);
473 r->set_new_top(r->bottom());
474 }
475 }
476
477 void ShenandoahFullGC::calculate_target_humongous_objects() {
478 ShenandoahHeap* heap = ShenandoahHeap::heap();
479
480 // Compute the new addresses for humongous objects. We need to do this after addresses
481 // for regular objects are calculated, and we know what regions in heap suffix are
482 // available for humongous moves.
483 //
484 // Scan the heap backwards, because we are compacting humongous regions towards the end.
485 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
486 // humongous start there.
487 //
488 // The complication is potential non-movable regions during the scan. If such region is
489 // detected, then sliding restarts towards that non-movable region.
490
491 size_t to_begin = heap->num_regions();
492 size_t to_end = heap->num_regions();
493
494 log_debug(gc)("Full GC calculating target humongous objects from end %zu", to_end);
495 for (size_t c = heap->num_regions(); c > 0; c--) {
496 ShenandoahHeapRegion *r = heap->get_region(c - 1);
497 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
498 // To-region candidate: record this, and continue scan
499 to_begin = r->index();
500 continue;
501 }
502
503 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
504 // From-region candidate: movable humongous region
505 oop old_obj = cast_to_oop(r->bottom());
506 size_t words_size = old_obj->size();
507 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
508
509 size_t start = to_end - num_regions;
510
511 if (start >= to_begin && start != r->index()) {
512 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
513 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
514 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
515 to_end = start;
516 continue;
517 }
518 }
519
520 // Failed to fit. Scan starting from current region.
521 to_begin = r->index();
522 to_end = r->index();
523 }
524 }
525
526 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
527 public:
528 void heap_region_do(ShenandoahHeapRegion* r) override {
529 if (r->is_trash()) {
530 r->try_recycle_under_lock();
531 // No need to adjust_interval_for_recycled_old_region. That will be taken care of during freeset rebuild.
532 }
533 if (r->is_cset()) {
534 // Leave affiliation unchanged
535 r->make_regular_bypass();
536 }
537 if (r->is_empty_uncommitted()) {
538 r->make_committed_bypass();
539 }
540 assert (r->is_committed(), "only committed regions in heap now, see region %zu", r->index());
541
542 // Record current region occupancy: this communicates empty regions are free
543 // to the rest of Full GC code.
544 r->set_new_top(r->top());
545 }
546 };
547
548 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
549 private:
550 ShenandoahHeap* const _heap;
551 ShenandoahMarkingContext* const _ctx;
552
553 public:
554 ShenandoahTrashImmediateGarbageClosure() :
555 _heap(ShenandoahHeap::heap()),
556 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
557
558 void heap_region_do(ShenandoahHeapRegion* r) override {
559 if (r->is_humongous_start()) {
560 oop humongous_obj = cast_to_oop(r->bottom());
561 if (!_ctx->is_marked(humongous_obj)) {
562 assert(!r->has_live(), "Region %zu is not marked, should not have live", r->index());
563 _heap->trash_humongous_region_at(r);
564 } else {
565 assert(r->has_live(), "Region %zu should have live", r->index());
566 }
567 } else if (r->is_humongous_continuation()) {
568 // If we hit continuation, the non-live humongous starts should have been trashed already
569 assert(r->humongous_start_region()->has_live(), "Region %zu should have live", r->index());
570 } else if (r->is_regular()) {
571 if (!r->has_live()) {
572 r->make_trash_immediate();
573 }
574 }
575 }
576 };
577
578 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
579 ShenandoahHeap* heap = ShenandoahHeap::heap();
580
581 uint n_workers = heap->workers()->active_workers();
582 size_t n_regions = heap->num_regions();
583
584 // What we want to accomplish: have the dense prefix of data, while still balancing
585 // out the parallel work.
586 //
587 // Assuming the amount of work is driven by the live data that needs moving, we can slice
588 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
589 // thread takes all regions in its prefix subset, and then it takes some regions from
590 // the tail.
591 //
592 // Tail region selection becomes interesting.
593 //
594 // First, we want to distribute the regions fairly between the workers, and those regions
595 // might have different amount of live data. So, until we sure no workers need live data,
596 // we need to only take what the worker needs.
597 //
598 // Second, since we slide everything to the left in each slice, the most busy regions
599 // would be the ones on the left. Which means we want to have all workers have their after-tail
600 // regions as close to the left as possible.
601 //
602 // The easiest way to do this is to distribute after-tail regions in round-robin between
603 // workers that still need live data.
604 //
605 // Consider parallel workers A, B, C, then the target slice layout would be:
606 //
607 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
608 //
609 // (.....dense-prefix.....) (.....................tail...................)
610 // [all regions fully live] [left-most regions are fuller that right-most]
611 //
612
613 // Compute how much live data is there. This would approximate the size of dense prefix
614 // we target to create.
615 size_t total_live = 0;
616 for (size_t idx = 0; idx < n_regions; idx++) {
617 ShenandoahHeapRegion *r = heap->get_region(idx);
618 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
619 total_live += r->get_live_data_words();
620 }
621 }
622
623 // Estimate the size for the dense prefix. Note that we specifically count only the
624 // "full" regions, so there would be some non-full regions in the slice tail.
625 size_t live_per_worker = total_live / n_workers;
626 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
627 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
628 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
629 assert(prefix_regions_total <= n_regions, "Sanity");
630
631 // There might be non-candidate regions in the prefix. To compute where the tail actually
632 // ends up being, we need to account those as well.
633 size_t prefix_end = prefix_regions_total;
634 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
635 ShenandoahHeapRegion *r = heap->get_region(idx);
636 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
637 prefix_end++;
638 }
639 }
640 prefix_end = MIN2(prefix_end, n_regions);
641 assert(prefix_end <= n_regions, "Sanity");
642
643 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
644 // subset of dense prefix.
645 size_t prefix_idx = 0;
646
647 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
648
649 for (size_t wid = 0; wid < n_workers; wid++) {
650 ShenandoahHeapRegionSet* slice = worker_slices[wid];
651
652 live[wid] = 0;
653 size_t regs = 0;
654
655 // Add all prefix regions for this worker
656 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
657 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
658 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
659 slice->add_region(r);
660 live[wid] += r->get_live_data_words();
661 regs++;
662 }
663 prefix_idx++;
664 }
665 }
666
667 // Distribute the tail among workers in round-robin fashion.
668 size_t wid = n_workers - 1;
669
670 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
671 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
672 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
673 assert(wid < n_workers, "Sanity");
674
675 size_t live_region = r->get_live_data_words();
676
677 // Select next worker that still needs live data.
678 size_t old_wid = wid;
679 do {
680 wid++;
681 if (wid == n_workers) wid = 0;
682 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
683
684 if (old_wid == wid) {
685 // Circled back to the same worker? This means liveness data was
686 // miscalculated. Bump the live_per_worker limit so that
687 // everyone gets a piece of the leftover work.
688 live_per_worker += ShenandoahHeapRegion::region_size_words();
689 }
690
691 worker_slices[wid]->add_region(r);
692 live[wid] += live_region;
693 }
694 }
695
696 FREE_C_HEAP_ARRAY(size_t, live);
697
698 #ifdef ASSERT
699 ResourceBitMap map(n_regions);
700 for (size_t wid = 0; wid < n_workers; wid++) {
701 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
702 ShenandoahHeapRegion* r = it.next();
703 while (r != nullptr) {
704 size_t idx = r->index();
705 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: %zu", idx);
706 assert(!map.at(idx), "No region distributed twice: %zu", idx);
707 map.at_put(idx, true);
708 r = it.next();
709 }
710 }
711
712 for (size_t rid = 0; rid < n_regions; rid++) {
713 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
714 bool is_distributed = map.at(rid);
715 assert(is_distributed || !is_candidate, "All candidates are distributed: %zu", rid);
716 }
717 #endif
718 }
719
720 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
721 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
722 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
723
724 ShenandoahHeap* heap = ShenandoahHeap::heap();
725
726 // About to figure out which regions can be compacted, make sure pinning status
727 // had been updated in GC prologue.
728 heap->assert_pinned_region_status();
729
730 {
731 // Trash the immediately collectible regions before computing addresses
732 ShenandoahTrashImmediateGarbageClosure trash_immediate_garbage;
733 ShenandoahExcludeRegionClosure<FREE> cl(&trash_immediate_garbage);
734 heap->heap_region_iterate(&cl);
735
736 // Make sure regions are in good state: committed, active, clean.
737 // This is needed because we are potentially sliding the data through them.
738 ShenandoahEnsureHeapActiveClosure ecl;
739 heap->heap_region_iterate(&ecl);
740 }
741
742 // Compute the new addresses for regular objects
743 {
744 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
745
746 distribute_slices(worker_slices);
747
748 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
749 heap->workers()->run_task(&task);
750 }
751
752 // Compute the new addresses for humongous objects
753 {
754 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
755 calculate_target_humongous_objects();
756 }
757 }
758
759 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
760 private:
761 ShenandoahMarkingContext* const _ctx;
762
763 template <class T>
764 inline void do_oop_work(T* p) {
765 T o = RawAccess<>::oop_load(p);
766 if (!CompressedOops::is_null(o)) {
767 oop obj = CompressedOops::decode_not_null(o);
768 assert(_ctx->is_marked(obj), "must be marked");
769 if (FullGCForwarding::is_forwarded(obj)) {
770 oop forw = FullGCForwarding::forwardee(obj);
771 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
772 }
773 }
774 }
775
776 public:
777 ShenandoahAdjustPointersClosure() :
778 _ctx(ShenandoahHeap::heap()->global_generation()->complete_marking_context()) {}
779
780 void do_oop(oop* p) { do_oop_work(p); }
781 void do_oop(narrowOop* p) { do_oop_work(p); }
782 void do_method(Method* m) {}
783 void do_nmethod(nmethod* nm) {}
784 };
785
786 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
787 private:
788 ShenandoahAdjustPointersClosure _cl;
789
790 public:
791 void do_object(oop p) override {
792 assert(ShenandoahHeap::heap()->global_generation()->is_mark_complete(), "marking must be complete");
793 assert(ShenandoahHeap::heap()->marking_context()->is_marked(p), "must be marked");
794 p->oop_iterate(&_cl);
795 }
796 };
797
798 class ShenandoahAdjustPointersTask : public WorkerTask {
799 private:
800 ShenandoahHeap* const _heap;
801 ShenandoahRegionIterator _regions;
802
803 public:
804 ShenandoahAdjustPointersTask() :
805 WorkerTask("Shenandoah Adjust Pointers"),
806 _heap(ShenandoahHeap::heap()) {
807 }
808
809 void work(uint worker_id) override {
810 ShenandoahParallelWorkerSession worker_session(worker_id);
811 ShenandoahAdjustPointersObjectClosure obj_cl;
812 ShenandoahHeapRegion* r = _regions.next();
813 while (r != nullptr) {
814 if (!r->is_humongous_continuation() && r->has_live()) {
815 _heap->marked_object_iterate(r, &obj_cl);
816 }
817 if (_heap->mode()->is_generational()) {
818 ShenandoahGenerationalFullGC::maybe_coalesce_and_fill_region(r);
819 }
820 r = _regions.next();
821 }
822 }
823 };
824
825 class ShenandoahAdjustRootPointersTask : public WorkerTask {
826 private:
827 ShenandoahRootAdjuster* _rp;
828 PreservedMarksSet* _preserved_marks;
829 public:
830 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
831 WorkerTask("Shenandoah Adjust Root Pointers"),
832 _rp(rp),
833 _preserved_marks(preserved_marks) {}
834
835 void work(uint worker_id) override {
836 ShenandoahParallelWorkerSession worker_session(worker_id);
837 ShenandoahAdjustPointersClosure cl;
838 _rp->roots_do(worker_id, &cl);
839 _preserved_marks->get(worker_id)->adjust_during_full_gc();
840 }
841 };
842
843 void ShenandoahFullGC::phase3_update_references() {
844 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
845 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
846
847 ShenandoahHeap* heap = ShenandoahHeap::heap();
848
849 WorkerThreads* workers = heap->workers();
850 uint nworkers = workers->active_workers();
851 {
852 #if COMPILER2_OR_JVMCI
853 DerivedPointerTable::clear();
854 #endif
855 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
856 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
857 workers->run_task(&task);
858 #if COMPILER2_OR_JVMCI
859 DerivedPointerTable::update_pointers();
860 #endif
861 }
862
863 ShenandoahAdjustPointersTask adjust_pointers_task;
864 workers->run_task(&adjust_pointers_task);
865 }
866
867 class ShenandoahCompactObjectsClosure : public ObjectClosure {
868 private:
869 uint const _worker_id;
870
871 public:
872 explicit ShenandoahCompactObjectsClosure(uint worker_id) :
873 _worker_id(worker_id) {}
874
875 void do_object(oop p) override {
876 assert(ShenandoahHeap::heap()->global_generation()->is_mark_complete(), "marking must be finished");
877 assert(ShenandoahHeap::heap()->marking_context()->is_marked(p), "must be marked");
878 size_t size = p->size();
879 if (FullGCForwarding::is_forwarded(p)) {
880 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
881 HeapWord* compact_to = cast_from_oop<HeapWord*>(FullGCForwarding::forwardee(p));
882 assert(compact_from != compact_to, "Forwarded object should move");
883 Copy::aligned_conjoint_words(compact_from, compact_to, size);
884 oop new_obj = cast_to_oop(compact_to);
885
886 // Restore the mark word before relativizing the stack chunk. The copy's
887 // mark word contains the full GC forwarding encoding, which would cause
888 // is_stackChunk() to read garbage (especially with compact headers).
889 new_obj->init_mark();
890 ContinuationGCSupport::relativize_stack_chunk(new_obj);
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->init_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 }