1 /*
2 * Copyright (c) 2014, 2021, Red Hat, Inc. All rights reserved.
3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
4 * Copyright (c) 2025, Oracle and/or its affiliates. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27
28 #include "compiler/oopMap.hpp"
29 #include "gc/shared/continuationGCSupport.hpp"
30 #include "gc/shared/fullGCForwarding.inline.hpp"
31 #include "gc/shared/gcTraceTime.inline.hpp"
32 #include "gc/shared/preservedMarks.inline.hpp"
33 #include "gc/shared/tlab_globals.hpp"
34 #include "gc/shared/workerThread.hpp"
35 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
36 #include "gc/shenandoah/shenandoahClosures.inline.hpp"
37 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
38 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
39 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
40 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
41 #include "gc/shenandoah/shenandoahFreeSet.hpp"
42 #include "gc/shenandoah/shenandoahFullGC.hpp"
43 #include "gc/shenandoah/shenandoahGenerationalFullGC.hpp"
44 #include "gc/shenandoah/shenandoahGlobalGeneration.hpp"
45 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
46 #include "gc/shenandoah/shenandoahMark.inline.hpp"
47 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
48 #include "gc/shenandoah/shenandoahHeapRegionClosures.hpp"
49 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
50 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
51 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
52 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
53 #include "gc/shenandoah/shenandoahMetrics.hpp"
54 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
55 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
56 #include "gc/shenandoah/shenandoahSTWMark.hpp"
57 #include "gc/shenandoah/shenandoahUtils.hpp"
58 #include "gc/shenandoah/shenandoahVerifier.hpp"
59 #include "gc/shenandoah/shenandoahVMOperations.hpp"
60 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
61 #include "memory/metaspaceUtils.hpp"
62 #include "memory/universe.hpp"
63 #include "oops/compressedOops.inline.hpp"
64 #include "oops/oop.inline.hpp"
65 #include "runtime/orderAccess.hpp"
66 #include "runtime/vmThread.hpp"
67 #include "utilities/copy.hpp"
68 #include "utilities/events.hpp"
69 #include "utilities/growableArray.hpp"
70
71 ShenandoahFullGC::ShenandoahFullGC() :
72 _gc_timer(ShenandoahHeap::heap()->gc_timer()),
73 _preserved_marks(new PreservedMarksSet(true)) {}
74
75 ShenandoahFullGC::~ShenandoahFullGC() {
76 delete _preserved_marks;
77 }
78
79 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
80 vmop_entry_full(cause);
81 // Always success
82 return true;
83 }
84
85 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
86 ShenandoahHeap* const heap = ShenandoahHeap::heap();
87 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
88 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
89
90 heap->try_inject_alloc_failure();
91 VM_ShenandoahFullGC op(cause, this);
92 VMThread::execute(&op);
93 }
94
95 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
96 static const char* msg = "Pause Full";
97 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
98 EventMark em("%s", msg);
99
100 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
101 ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
102 "full gc");
103
104 op_full(cause);
105 }
106
107 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
108 ShenandoahMetricsSnapshot metrics;
109 metrics.snap_before();
110
111 // Perform full GC
112 do_it(cause);
113
114 ShenandoahHeap* const heap = ShenandoahHeap::heap();
115
116 if (heap->mode()->is_generational()) {
117 ShenandoahGenerationalFullGC::handle_completion(heap);
118 }
119
120 metrics.snap_after();
121
122 if (metrics.is_good_progress(heap->global_generation())) {
123 heap->notify_gc_progress();
124 } else {
125 // Nothing to do. Tell the allocation path that we have failed to make
126 // progress, and it can finally fail.
127 heap->notify_gc_no_progress();
128 }
129
130 // Regardless if progress was made, we record that we completed a "successful" full GC.
131 heap->global_generation()->heuristics()->record_success_full();
132 heap->shenandoah_policy()->record_success_full();
133
134 {
135 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_propagate_gc_state);
136 heap->propagate_gc_state_to_all_threads();
137 }
138 }
139
140 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
141 ShenandoahHeap* heap = ShenandoahHeap::heap();
142
143 if (heap->mode()->is_generational()) {
144 ShenandoahGenerationalFullGC::prepare();
145 }
146
147 if (ShenandoahVerify) {
148 heap->verifier()->verify_before_fullgc();
149 }
150
151 if (VerifyBeforeGC) {
152 Universe::verify();
153 }
154
155 // Degenerated GC may carry concurrent root flags when upgrading to
156 // full GC. We need to reset it before mutators resume.
157 heap->set_concurrent_strong_root_in_progress(false);
158 heap->set_concurrent_weak_root_in_progress(false);
159
160 heap->set_full_gc_in_progress(true);
161
162 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
163 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
164
165 {
166 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
167 heap->pre_full_gc_dump(_gc_timer);
168 }
169
170 {
171 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
172 // Full GC is supposed to recover from any GC state:
173
174 // a0. Remember if we have forwarded objects
175 bool has_forwarded_objects = heap->has_forwarded_objects();
176
177 // a1. Cancel evacuation, if in progress
178 if (heap->is_evacuation_in_progress()) {
179 heap->set_evacuation_in_progress(false);
180 }
181 assert(!heap->is_evacuation_in_progress(), "sanity");
182
183 // a2. Cancel update-refs, if in progress
184 if (heap->is_update_refs_in_progress()) {
185 heap->set_update_refs_in_progress(false);
186 }
187 assert(!heap->is_update_refs_in_progress(), "sanity");
188
189 // b. Cancel all concurrent marks, if in progress
190 if (heap->is_concurrent_mark_in_progress()) {
191 heap->cancel_concurrent_mark();
192 }
193 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
194
195 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
196 if (has_forwarded_objects) {
197 update_roots(true /*full_gc*/);
198 }
199
200 // d. Abandon reference discovery and clear all discovered references.
201 ShenandoahReferenceProcessor* rp = heap->global_generation()->ref_processor();
202 rp->abandon_partial_discovery();
203
204 // e. Sync pinned region status from the CP marks
205 heap->sync_pinned_region_status();
206
207 if (heap->mode()->is_generational()) {
208 ShenandoahGenerationalFullGC::restore_top_before_promote(heap);
209 }
210
211 // The rest of prologue:
212 _preserved_marks->init(heap->workers()->active_workers());
213
214 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
215 }
216
217 if (UseTLAB) {
218 // Note: PLABs are also retired with GCLABs in generational mode.
219 heap->gclabs_retire(ResizeTLAB);
220 heap->tlabs_retire(ResizeTLAB);
221 }
222
223 OrderAccess::fence();
224
225 phase1_mark_heap();
226
227 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
228 // Coming out of Full GC, we would not have any forwarded objects.
229 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
230 heap->set_has_forwarded_objects(false);
231
232 heap->set_full_gc_move_in_progress(true);
233
234 // Setup workers for the rest
235 OrderAccess::fence();
236
237 // Initialize worker slices
238 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
239 for (uint i = 0; i < heap->max_workers(); i++) {
240 worker_slices[i] = new ShenandoahHeapRegionSet();
241 }
242
243 {
244 // The rest of code performs region moves, where region status is undefined
245 // until all phases run together.
246 ShenandoahHeapLocker lock(heap->lock());
247
248 phase2_calculate_target_addresses(worker_slices);
249
250 OrderAccess::fence();
251
252 phase3_update_references();
253
254 phase4_compact_objects(worker_slices);
255
256 phase5_epilog();
257 }
258
259 // Resize metaspace
260 MetaspaceGC::compute_new_size();
261
262 // Free worker slices
263 for (uint i = 0; i < heap->max_workers(); i++) {
264 delete worker_slices[i];
265 }
266 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
267
268 heap->set_full_gc_move_in_progress(false);
269 heap->set_full_gc_in_progress(false);
270
271 if (ShenandoahVerify) {
272 heap->verifier()->verify_after_fullgc();
273 }
274
275 if (VerifyAfterGC) {
276 Universe::verify();
277 }
278
279 {
280 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
281 heap->post_full_gc_dump(_gc_timer);
282 }
283 }
284
285 void ShenandoahFullGC::phase1_mark_heap() {
286 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
287 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
288
289 ShenandoahHeap* heap = ShenandoahHeap::heap();
290
291 heap->global_generation()->reset_mark_bitmap<true, true>();
292 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
293 assert(!heap->global_generation()->is_mark_complete(), "sanity");
294
295 heap->set_unload_classes(heap->global_generation()->heuristics()->can_unload_classes());
296
297 ShenandoahReferenceProcessor* rp = heap->global_generation()->ref_processor();
298 // enable ("weak") refs discovery
299 rp->set_soft_reference_policy(true); // forcefully purge all soft references
300
301 ShenandoahSTWMark mark(heap->global_generation(), true /*full_gc*/);
302 mark.mark();
303 heap->parallel_cleaning(true /* full_gc */);
304
305 if (ShenandoahHeap::heap()->mode()->is_generational()) {
306 ShenandoahGenerationalFullGC::log_live_in_old(heap);
307 }
308 }
309
310 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
311 private:
312 PreservedMarks* const _preserved_marks;
313 ShenandoahHeap* const _heap;
314 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
315 int _empty_regions_pos;
316 ShenandoahHeapRegion* _to_region;
317 ShenandoahHeapRegion* _from_region;
318 HeapWord* _compact_point;
319
320 public:
321 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
322 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
323 ShenandoahHeapRegion* to_region) :
324 _preserved_marks(preserved_marks),
325 _heap(ShenandoahHeap::heap()),
326 _empty_regions(empty_regions),
327 _empty_regions_pos(0),
328 _to_region(to_region),
329 _from_region(nullptr),
330 _compact_point(to_region->bottom()) {}
331
332 void set_from_region(ShenandoahHeapRegion* from_region) {
333 _from_region = from_region;
334 }
335
336 void finish() {
337 assert(_to_region != nullptr, "should not happen");
338 _to_region->set_new_top(_compact_point);
339 }
340
341 bool is_compact_same_region() {
342 return _from_region == _to_region;
343 }
344
345 int empty_regions_pos() {
346 return _empty_regions_pos;
347 }
348
349 void do_object(oop p) {
350 assert(_from_region != nullptr, "must set before work");
351 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
352 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
353
354 size_t obj_size = p->size();
355 if (_compact_point + obj_size > _to_region->end()) {
356 finish();
357
358 // Object doesn't fit. Pick next empty region and start compacting there.
359 ShenandoahHeapRegion* new_to_region;
360 if (_empty_regions_pos < _empty_regions.length()) {
361 new_to_region = _empty_regions.at(_empty_regions_pos);
362 _empty_regions_pos++;
363 } else {
364 // Out of empty region? Compact within the same region.
365 new_to_region = _from_region;
366 }
367
368 assert(new_to_region != _to_region, "must not reuse same to-region");
369 assert(new_to_region != nullptr, "must not be null");
370 _to_region = new_to_region;
371 _compact_point = _to_region->bottom();
372 }
373
374 // Object fits into current region, record new location, if object does not move:
375 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
376 shenandoah_assert_not_forwarded(nullptr, p);
377 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
378 _preserved_marks->push_if_necessary(p, p->mark());
379 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
380 }
381 _compact_point += obj_size;
382 }
383 };
384
385 class ShenandoahPrepareForCompactionTask : public WorkerTask {
386 private:
387 PreservedMarksSet* const _preserved_marks;
388 ShenandoahHeap* const _heap;
389 ShenandoahHeapRegionSet** const _worker_slices;
390
391 public:
392 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
393 WorkerTask("Shenandoah Prepare For Compaction"),
394 _preserved_marks(preserved_marks),
395 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
396 }
397
398 static bool is_candidate_region(ShenandoahHeapRegion* r) {
399 // Empty region: get it into the slice to defragment the slice itself.
400 // We could have skipped this without violating correctness, but we really
401 // want to compact all live regions to the start of the heap, which sometimes
402 // means moving them into the fully empty regions.
403 if (r->is_empty()) return true;
404
405 // Can move the region, and this is not the humongous region. Humongous
406 // moves are special cased here, because their moves are handled separately.
407 return r->is_stw_move_allowed() && !r->is_humongous();
408 }
409
410 void work(uint worker_id) override;
411 private:
412 template<typename ClosureType>
413 void prepare_for_compaction(ClosureType& cl,
414 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
415 ShenandoahHeapRegionSetIterator& it,
416 ShenandoahHeapRegion* from_region);
417 };
418
419 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
420 ShenandoahParallelWorkerSession worker_session(worker_id);
421 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
422 ShenandoahHeapRegionSetIterator it(slice);
423 ShenandoahHeapRegion* from_region = it.next();
424 // No work?
425 if (from_region == nullptr) {
426 return;
427 }
428
429 // Sliding compaction. Walk all regions in the slice, and compact them.
430 // Remember empty regions and reuse them as needed.
431 ResourceMark rm;
432
433 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
434
435 if (_heap->mode()->is_generational()) {
436 ShenandoahPrepareForGenerationalCompactionObjectClosure cl(_preserved_marks->get(worker_id),
437 empty_regions, from_region, worker_id);
438 prepare_for_compaction(cl, empty_regions, it, from_region);
439 } else {
440 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
441 prepare_for_compaction(cl, empty_regions, it, from_region);
442 }
443 }
444
445 template<typename ClosureType>
446 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
447 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
448 ShenandoahHeapRegionSetIterator& it,
449 ShenandoahHeapRegion* from_region) {
450 while (from_region != nullptr) {
451 assert(is_candidate_region(from_region), "Sanity");
452 cl.set_from_region(from_region);
453 if (from_region->has_live()) {
454 _heap->marked_object_iterate(from_region, &cl);
455 }
456
457 // Compacted the region to somewhere else? From-region is empty then.
458 if (!cl.is_compact_same_region()) {
459 empty_regions.append(from_region);
460 }
461 from_region = it.next();
462 }
463 cl.finish();
464
465 // Mark all remaining regions as empty
466 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
467 ShenandoahHeapRegion* r = empty_regions.at(pos);
468 r->set_new_top(r->bottom());
469 }
470 }
471
472 void ShenandoahFullGC::calculate_target_humongous_objects() {
473 ShenandoahHeap* heap = ShenandoahHeap::heap();
474
475 // Compute the new addresses for humongous objects. We need to do this after addresses
476 // for regular objects are calculated, and we know what regions in heap suffix are
477 // available for humongous moves.
478 //
479 // Scan the heap backwards, because we are compacting humongous regions towards the end.
480 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
481 // humongous start there.
482 //
483 // The complication is potential non-movable regions during the scan. If such region is
484 // detected, then sliding restarts towards that non-movable region.
485
486 size_t to_begin = heap->num_regions();
487 size_t to_end = heap->num_regions();
488
489 log_debug(gc)("Full GC calculating target humongous objects from end %zu", to_end);
490 for (size_t c = heap->num_regions(); c > 0; c--) {
491 ShenandoahHeapRegion *r = heap->get_region(c - 1);
492 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
493 // To-region candidate: record this, and continue scan
494 to_begin = r->index();
495 continue;
496 }
497
498 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
499 // From-region candidate: movable humongous region
500 oop old_obj = cast_to_oop(r->bottom());
501 size_t words_size = old_obj->size();
502 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
503
504 size_t start = to_end - num_regions;
505
506 if (start >= to_begin && start != r->index()) {
507 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
508 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
509 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
510 to_end = start;
511 continue;
512 }
513 }
514
515 // Failed to fit. Scan starting from current region.
516 to_begin = r->index();
517 to_end = r->index();
518 }
519 }
520
521 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
522 private:
523 ShenandoahHeap* const _heap;
524
525 public:
526 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
527 void heap_region_do(ShenandoahHeapRegion* r) {
528 if (r->is_trash()) {
529 r->try_recycle_under_lock();
530 }
531 if (r->is_cset()) {
532 // Leave affiliation unchanged
533 r->make_regular_bypass();
534 }
535 if (r->is_empty_uncommitted()) {
536 r->make_committed_bypass();
537 }
538 assert (r->is_committed(), "only committed regions in heap now, see region %zu", r->index());
539
540 // Record current region occupancy: this communicates empty regions are free
541 // to the rest of Full GC code.
542 r->set_new_top(r->top());
543 }
544 };
545
546 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
547 private:
548 ShenandoahHeap* const _heap;
549 ShenandoahMarkingContext* const _ctx;
550
551 public:
552 ShenandoahTrashImmediateGarbageClosure() :
553 _heap(ShenandoahHeap::heap()),
554 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
555
556 void heap_region_do(ShenandoahHeapRegion* r) override {
557 if (r->is_humongous_start()) {
558 oop humongous_obj = cast_to_oop(r->bottom());
559 if (!_ctx->is_marked(humongous_obj)) {
560 assert(!r->has_live(), "Region %zu is not marked, should not have live", r->index());
561 _heap->trash_humongous_region_at(r);
562 } else {
563 assert(r->has_live(), "Region %zu should have live", r->index());
564 }
565 } else if (r->is_humongous_continuation()) {
566 // If we hit continuation, the non-live humongous starts should have been trashed already
567 assert(r->humongous_start_region()->has_live(), "Region %zu should have live", r->index());
568 } else if (r->is_regular()) {
569 if (!r->has_live()) {
570 r->make_trash_immediate();
571 }
572 }
573 }
574 };
575
576 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
577 ShenandoahHeap* heap = ShenandoahHeap::heap();
578
579 uint n_workers = heap->workers()->active_workers();
580 size_t n_regions = heap->num_regions();
581
582 // What we want to accomplish: have the dense prefix of data, while still balancing
583 // out the parallel work.
584 //
585 // Assuming the amount of work is driven by the live data that needs moving, we can slice
586 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
587 // thread takes all regions in its prefix subset, and then it takes some regions from
588 // the tail.
589 //
590 // Tail region selection becomes interesting.
591 //
592 // First, we want to distribute the regions fairly between the workers, and those regions
593 // might have different amount of live data. So, until we sure no workers need live data,
594 // we need to only take what the worker needs.
595 //
596 // Second, since we slide everything to the left in each slice, the most busy regions
597 // would be the ones on the left. Which means we want to have all workers have their after-tail
598 // regions as close to the left as possible.
599 //
600 // The easiest way to do this is to distribute after-tail regions in round-robin between
601 // workers that still need live data.
602 //
603 // Consider parallel workers A, B, C, then the target slice layout would be:
604 //
605 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
606 //
607 // (.....dense-prefix.....) (.....................tail...................)
608 // [all regions fully live] [left-most regions are fuller that right-most]
609 //
610
611 // Compute how much live data is there. This would approximate the size of dense prefix
612 // we target to create.
613 size_t total_live = 0;
614 for (size_t idx = 0; idx < n_regions; idx++) {
615 ShenandoahHeapRegion *r = heap->get_region(idx);
616 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
617 total_live += r->get_live_data_words();
618 }
619 }
620
621 // Estimate the size for the dense prefix. Note that we specifically count only the
622 // "full" regions, so there would be some non-full regions in the slice tail.
623 size_t live_per_worker = total_live / n_workers;
624 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
625 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
626 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
627 assert(prefix_regions_total <= n_regions, "Sanity");
628
629 // There might be non-candidate regions in the prefix. To compute where the tail actually
630 // ends up being, we need to account those as well.
631 size_t prefix_end = prefix_regions_total;
632 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
633 ShenandoahHeapRegion *r = heap->get_region(idx);
634 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
635 prefix_end++;
636 }
637 }
638 prefix_end = MIN2(prefix_end, n_regions);
639 assert(prefix_end <= n_regions, "Sanity");
640
641 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
642 // subset of dense prefix.
643 size_t prefix_idx = 0;
644
645 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
646
647 for (size_t wid = 0; wid < n_workers; wid++) {
648 ShenandoahHeapRegionSet* slice = worker_slices[wid];
649
650 live[wid] = 0;
651 size_t regs = 0;
652
653 // Add all prefix regions for this worker
654 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
655 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
656 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
657 slice->add_region(r);
658 live[wid] += r->get_live_data_words();
659 regs++;
660 }
661 prefix_idx++;
662 }
663 }
664
665 // Distribute the tail among workers in round-robin fashion.
666 size_t wid = n_workers - 1;
667
668 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
669 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
670 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
671 assert(wid < n_workers, "Sanity");
672
673 size_t live_region = r->get_live_data_words();
674
675 // Select next worker that still needs live data.
676 size_t old_wid = wid;
677 do {
678 wid++;
679 if (wid == n_workers) wid = 0;
680 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
681
682 if (old_wid == wid) {
683 // Circled back to the same worker? This means liveness data was
684 // miscalculated. Bump the live_per_worker limit so that
685 // everyone gets a piece of the leftover work.
686 live_per_worker += ShenandoahHeapRegion::region_size_words();
687 }
688
689 worker_slices[wid]->add_region(r);
690 live[wid] += live_region;
691 }
692 }
693
694 FREE_C_HEAP_ARRAY(size_t, live);
695
696 #ifdef ASSERT
697 ResourceBitMap map(n_regions);
698 for (size_t wid = 0; wid < n_workers; wid++) {
699 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
700 ShenandoahHeapRegion* r = it.next();
701 while (r != nullptr) {
702 size_t idx = r->index();
703 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: %zu", idx);
704 assert(!map.at(idx), "No region distributed twice: %zu", idx);
705 map.at_put(idx, true);
706 r = it.next();
707 }
708 }
709
710 for (size_t rid = 0; rid < n_regions; rid++) {
711 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
712 bool is_distributed = map.at(rid);
713 assert(is_distributed || !is_candidate, "All candidates are distributed: %zu", rid);
714 }
715 #endif
716 }
717
718 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
719 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
720 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
721
722 ShenandoahHeap* heap = ShenandoahHeap::heap();
723
724 // About to figure out which regions can be compacted, make sure pinning status
725 // had been updated in GC prologue.
726 heap->assert_pinned_region_status();
727
728 {
729 // Trash the immediately collectible regions before computing addresses
730 ShenandoahTrashImmediateGarbageClosure trash_immediate_garbage;
731 ShenandoahExcludeRegionClosure<FREE> cl(&trash_immediate_garbage);
732 heap->heap_region_iterate(&cl);
733
734 // Make sure regions are in good state: committed, active, clean.
735 // This is needed because we are potentially sliding the data through them.
736 ShenandoahEnsureHeapActiveClosure ecl;
737 heap->heap_region_iterate(&ecl);
738 }
739
740 // Compute the new addresses for regular objects
741 {
742 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
743
744 distribute_slices(worker_slices);
745
746 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
747 heap->workers()->run_task(&task);
748 }
749
750 // Compute the new addresses for humongous objects
751 {
752 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
753 calculate_target_humongous_objects();
754 }
755 }
756
757 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
758 private:
759 ShenandoahHeap* const _heap;
760 ShenandoahMarkingContext* const _ctx;
761
762 template <class T>
763 inline void do_oop_work(T* p) {
764 T o = RawAccess<>::oop_load(p);
765 if (!CompressedOops::is_null(o)) {
766 oop obj = CompressedOops::decode_not_null(o);
767 assert(_ctx->is_marked(obj), "must be marked");
768 if (FullGCForwarding::is_forwarded(obj)) {
769 oop forw = FullGCForwarding::forwardee(obj);
770 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
771 }
772 }
773 }
774
775 public:
776 ShenandoahAdjustPointersClosure() :
777 _heap(ShenandoahHeap::heap()),
778 _ctx(ShenandoahHeap::heap()->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 ShenandoahHeap* const _heap;
789 ShenandoahAdjustPointersClosure _cl;
790
791 public:
792 ShenandoahAdjustPointersObjectClosure() :
793 _heap(ShenandoahHeap::heap()) {
794 }
795 void do_object(oop p) {
796 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
797 p->oop_iterate(&_cl);
798 }
799 };
800
801 class ShenandoahAdjustPointersTask : public WorkerTask {
802 private:
803 ShenandoahHeap* const _heap;
804 ShenandoahRegionIterator _regions;
805
806 public:
807 ShenandoahAdjustPointersTask() :
808 WorkerTask("Shenandoah Adjust Pointers"),
809 _heap(ShenandoahHeap::heap()) {
810 }
811
812 void work(uint worker_id) {
813 ShenandoahParallelWorkerSession worker_session(worker_id);
814 ShenandoahAdjustPointersObjectClosure obj_cl;
815 ShenandoahHeapRegion* r = _regions.next();
816 while (r != nullptr) {
817 if (!r->is_humongous_continuation() && r->has_live()) {
818 _heap->marked_object_iterate(r, &obj_cl);
819 }
820 if (_heap->mode()->is_generational()) {
821 ShenandoahGenerationalFullGC::maybe_coalesce_and_fill_region(r);
822 }
823 r = _regions.next();
824 }
825 }
826 };
827
828 class ShenandoahAdjustRootPointersTask : public WorkerTask {
829 private:
830 ShenandoahRootAdjuster* _rp;
831 PreservedMarksSet* _preserved_marks;
832 public:
833 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
834 WorkerTask("Shenandoah Adjust Root Pointers"),
835 _rp(rp),
836 _preserved_marks(preserved_marks) {}
837
838 void work(uint worker_id) {
839 ShenandoahParallelWorkerSession worker_session(worker_id);
840 ShenandoahAdjustPointersClosure cl;
841 _rp->roots_do(worker_id, &cl);
842 _preserved_marks->get(worker_id)->adjust_during_full_gc();
843 }
844 };
845
846 void ShenandoahFullGC::phase3_update_references() {
847 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
848 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
849
850 ShenandoahHeap* heap = ShenandoahHeap::heap();
851
852 WorkerThreads* workers = heap->workers();
853 uint nworkers = workers->active_workers();
854 {
855 #if COMPILER2_OR_JVMCI
856 DerivedPointerTable::clear();
857 #endif
858 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
859 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
860 workers->run_task(&task);
861 #if COMPILER2_OR_JVMCI
862 DerivedPointerTable::update_pointers();
863 #endif
864 }
865
866 ShenandoahAdjustPointersTask adjust_pointers_task;
867 workers->run_task(&adjust_pointers_task);
868 }
869
870 class ShenandoahCompactObjectsClosure : public ObjectClosure {
871 private:
872 ShenandoahHeap* const _heap;
873 uint const _worker_id;
874
875 public:
876 ShenandoahCompactObjectsClosure(uint worker_id) :
877 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
878
879 void do_object(oop p) {
880 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
881 size_t size = p->size();
882 if (FullGCForwarding::is_forwarded(p)) {
883 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
884 HeapWord* compact_to = cast_from_oop<HeapWord*>(FullGCForwarding::forwardee(p));
885 assert(compact_from != compact_to, "Forwarded object should move");
886 Copy::aligned_conjoint_words(compact_from, compact_to, size);
887 oop new_obj = cast_to_oop(compact_to);
888
889 ContinuationGCSupport::relativize_stack_chunk(new_obj);
890 new_obj->init_mark();
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) {
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) {
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->complete_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 } else {
980 if (r->is_old()) {
981 ShenandoahGenerationalFullGC::account_for_region(r, _old_regions, _old_usage, _old_humongous_waste);
982 } else if (r->is_young()) {
983 ShenandoahGenerationalFullGC::account_for_region(r, _young_regions, _young_usage, _young_humongous_waste);
984 }
985 }
986 r->set_live_data(live);
987 r->reset_alloc_metadata();
988 }
989
990 void update_generation_usage() {
991 if (_is_generational) {
992 _heap->old_generation()->establish_usage(_old_regions, _old_usage, _old_humongous_waste);
993 _heap->young_generation()->establish_usage(_young_regions, _young_usage, _young_humongous_waste);
994 } else {
995 assert(_old_regions == 0, "Old regions only expected in generational mode");
996 assert(_old_usage == 0, "Old usage only expected in generational mode");
997 assert(_old_humongous_waste == 0, "Old humongous waste only expected in generational mode");
998 }
999
1000 // In generational mode, global usage should be the sum of young and old. This is also true
1001 // for non-generational modes except that there are no old regions.
1002 _heap->global_generation()->establish_usage(_old_regions + _young_regions,
1003 _old_usage + _young_usage,
1004 _old_humongous_waste + _young_humongous_waste);
1005 }
1006 };
1007
1008 void ShenandoahFullGC::compact_humongous_objects() {
1009 // Compact humongous regions, based on their fwdptr objects.
1010 //
1011 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1012 // humongous regions are already compacted, and do not require further moves, which alleviates
1013 // sliding costs. We may consider doing this in parallel in the future.
1014
1015 ShenandoahHeap* heap = ShenandoahHeap::heap();
1016
1017 for (size_t c = heap->num_regions(); c > 0; c--) {
1018 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1019 if (r->is_humongous_start()) {
1020 oop old_obj = cast_to_oop(r->bottom());
1021 if (!FullGCForwarding::is_forwarded(old_obj)) {
1022 // No need to move the object, it stays at the same slot
1023 continue;
1024 }
1025 size_t words_size = old_obj->size();
1026 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1027
1028 size_t old_start = r->index();
1029 size_t old_end = old_start + num_regions - 1;
1030 size_t new_start = heap->heap_region_index_containing(FullGCForwarding::forwardee(old_obj));
1031 size_t new_end = new_start + num_regions - 1;
1032 assert(old_start != new_start, "must be real move");
1033 assert(r->is_stw_move_allowed(), "Region %zu should be movable", r->index());
1034
1035 log_debug(gc)("Full GC compaction moves humongous object from region %zu to region %zu", old_start, new_start);
1036 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
1037 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
1038
1039 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1040 new_obj->init_mark();
1041
1042 {
1043 ShenandoahAffiliation original_affiliation = r->affiliation();
1044 for (size_t c = old_start; c <= old_end; c++) {
1045 ShenandoahHeapRegion* r = heap->get_region(c);
1046 // Leave humongous region affiliation unchanged.
1047 r->make_regular_bypass();
1048 r->set_top(r->bottom());
1049 }
1050
1051 for (size_t c = new_start; c <= new_end; c++) {
1052 ShenandoahHeapRegion* r = heap->get_region(c);
1053 if (c == new_start) {
1054 r->make_humongous_start_bypass(original_affiliation);
1055 } else {
1056 r->make_humongous_cont_bypass(original_affiliation);
1057 }
1058
1059 // Trailing region may be non-full, record the remainder there
1060 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1061 if ((c == new_end) && (remainder != 0)) {
1062 r->set_top(r->bottom() + remainder);
1063 } else {
1064 r->set_top(r->end());
1065 }
1066
1067 r->reset_alloc_metadata();
1068 }
1069 }
1070 }
1071 }
1072 }
1073
1074 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1075 // we need to remain able to walk pinned regions.
1076 // Since pinned region do not move and don't get compacted, we will get holes with
1077 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1078 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1079 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1080 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1081 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1082 private:
1083 ShenandoahRegionIterator _regions;
1084
1085 public:
1086 ShenandoahMCResetCompleteBitmapTask() :
1087 WorkerTask("Shenandoah Reset Bitmap") {
1088 }
1089
1090 void work(uint worker_id) {
1091 ShenandoahParallelWorkerSession worker_session(worker_id);
1092 ShenandoahHeapRegion* region = _regions.next();
1093 ShenandoahHeap* heap = ShenandoahHeap::heap();
1094 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1095 while (region != nullptr) {
1096 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1097 ctx->clear_bitmap(region);
1098 }
1099 region = _regions.next();
1100 }
1101 }
1102 };
1103
1104 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1105 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1106 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1107
1108 ShenandoahHeap* heap = ShenandoahHeap::heap();
1109
1110 // Compact regular objects first
1111 {
1112 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1113 ShenandoahCompactObjectsTask compact_task(worker_slices);
1114 heap->workers()->run_task(&compact_task);
1115 }
1116
1117 // Compact humongous objects after regular object moves
1118 {
1119 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1120 compact_humongous_objects();
1121 }
1122 }
1123
1124 void ShenandoahFullGC::phase5_epilog() {
1125 GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1126 ShenandoahHeap* heap = ShenandoahHeap::heap();
1127
1128 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1129 // and must ensure the bitmap is in sync.
1130 {
1131 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1132 ShenandoahMCResetCompleteBitmapTask task;
1133 heap->workers()->run_task(&task);
1134 }
1135
1136 // Bring regions in proper states after the collection, and set heap properties.
1137 {
1138 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1139 ShenandoahPostCompactClosure post_compact;
1140 heap->heap_region_iterate(&post_compact);
1141 post_compact.update_generation_usage();
1142
1143 if (heap->mode()->is_generational()) {
1144 ShenandoahGenerationalFullGC::balance_generations_after_gc(heap);
1145 }
1146
1147 heap->collection_set()->clear();
1148 size_t young_cset_regions, old_cset_regions;
1149 size_t first_old, last_old, num_old;
1150 heap->free_set()->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old, last_old, num_old);
1151
1152 // We also do not expand old generation size following Full GC because we have scrambled age populations and
1153 // no longer have objects separated by age into distinct regions.
1154 if (heap->mode()->is_generational()) {
1155 ShenandoahGenerationalFullGC::compute_balances();
1156 }
1157
1158 heap->free_set()->finish_rebuild(young_cset_regions, old_cset_regions, num_old);
1159
1160 // Set mark incomplete because the marking bitmaps have been reset except pinned regions.
1161 heap->global_generation()->set_mark_incomplete();
1162
1163 heap->clear_cancelled_gc(true /* clear oom handler */);
1164 }
1165
1166 _preserved_marks->restore(heap->workers());
1167 _preserved_marks->reclaim();
1168
1169 // We defer generation resizing actions until after cset regions have been recycled. We do this even following an
1170 // abbreviated cycle.
1171 if (heap->mode()->is_generational()) {
1172 ShenandoahGenerationalFullGC::balance_generations_after_rebuilding_free_set();
1173 ShenandoahGenerationalFullGC::rebuild_remembered_set(heap);
1174 }
1175 }