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