1 /*
2 * Copyright (c) 2014, 2021, Red Hat, Inc. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26
27 #include "compiler/oopMap.hpp"
28 #include "gc/shared/continuationGCSupport.hpp"
29 #include "gc/shared/fullGCForwarding.inline.hpp"
30 #include "gc/shared/gcTraceTime.inline.hpp"
31 #include "gc/shared/preservedMarks.inline.hpp"
32 #include "gc/shared/tlab_globals.hpp"
33 #include "gc/shared/workerThread.hpp"
34 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
35 #include "gc/shenandoah/shenandoahClosures.inline.hpp"
36 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
37 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
38 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
39 #include "gc/shenandoah/shenandoahFreeSet.hpp"
40 #include "gc/shenandoah/shenandoahFullGC.hpp"
41 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
42 #include "gc/shenandoah/shenandoahMark.inline.hpp"
43 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
44 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
45 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
46 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
47 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
48 #include "gc/shenandoah/shenandoahMetrics.hpp"
49 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
50 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
51 #include "gc/shenandoah/shenandoahSTWMark.hpp"
52 #include "gc/shenandoah/shenandoahUtils.hpp"
53 #include "gc/shenandoah/shenandoahVerifier.hpp"
54 #include "gc/shenandoah/shenandoahVMOperations.hpp"
55 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
56 #include "memory/metaspaceUtils.hpp"
57 #include "memory/universe.hpp"
58 #include "oops/compressedOops.inline.hpp"
59 #include "oops/oop.inline.hpp"
60 #include "runtime/javaThread.hpp"
61 #include "runtime/orderAccess.hpp"
62 #include "runtime/vmThread.hpp"
63 #include "utilities/copy.hpp"
64 #include "utilities/events.hpp"
65 #include "utilities/growableArray.hpp"
66
67 ShenandoahFullGC::ShenandoahFullGC() :
68 _gc_timer(ShenandoahHeap::heap()->gc_timer()),
69 _preserved_marks(new PreservedMarksSet(true)) {}
70
71 ShenandoahFullGC::~ShenandoahFullGC() {
72 delete _preserved_marks;
73 }
74
75 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
76 vmop_entry_full(cause);
77 // Always success
78 return true;
79 }
80
81 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
82 ShenandoahHeap* const heap = ShenandoahHeap::heap();
83 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
84 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
85
86 heap->try_inject_alloc_failure();
87 VM_ShenandoahFullGC op(cause, this);
88 VMThread::execute(&op);
89 }
90
91 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
92 static const char* msg = "Pause Full";
93 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
94 EventMark em("%s", msg);
95
96 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
97 ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
98 "full gc");
99
100 op_full(cause);
101 }
102
103 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
104 ShenandoahMetricsSnapshot metrics;
105 metrics.snap_before();
106
107 // Perform full GC
108 do_it(cause);
109
110 ShenandoahHeap* const heap = ShenandoahHeap::heap();
111
112 metrics.snap_after();
113
114 if (metrics.is_good_progress()) {
115 heap->notify_gc_progress();
116 } else {
117 // Nothing to do. Tell the allocation path that we have failed to make
118 // progress, and it can finally fail.
119 heap->notify_gc_no_progress();
120 }
121
122 // Regardless if progress was made, we record that we completed a "successful" full GC.
123 heap->heuristics()->record_success_full();
124 heap->shenandoah_policy()->record_success_full();
125 }
126
127 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
128 ShenandoahHeap* heap = ShenandoahHeap::heap();
129
130 if (ShenandoahVerify) {
131 heap->verifier()->verify_before_fullgc();
132 }
133
134 if (VerifyBeforeGC) {
135 Universe::verify();
136 }
137
138 // Degenerated GC may carry concurrent root flags when upgrading to
139 // full GC. We need to reset it before mutators resume.
140 heap->set_concurrent_strong_root_in_progress(false);
141 heap->set_concurrent_weak_root_in_progress(false);
142
143 heap->set_full_gc_in_progress(true);
144
145 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
146 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
147
148 {
149 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
150 heap->pre_full_gc_dump(_gc_timer);
151 }
152
153 {
154 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
155 // Full GC is supposed to recover from any GC state:
156
157 // a0. Remember if we have forwarded objects
158 bool has_forwarded_objects = heap->has_forwarded_objects();
159
160 // a1. Cancel evacuation, if in progress
161 if (heap->is_evacuation_in_progress()) {
162 heap->set_evacuation_in_progress(false);
163 }
164 assert(!heap->is_evacuation_in_progress(), "sanity");
165
166 // a2. Cancel update-refs, if in progress
167 if (heap->is_update_refs_in_progress()) {
168 heap->set_update_refs_in_progress(false);
169 }
170 assert(!heap->is_update_refs_in_progress(), "sanity");
171
172 // b. Cancel concurrent mark, if in progress
173 if (heap->is_concurrent_mark_in_progress()) {
174 ShenandoahConcurrentGC::cancel();
175 heap->set_concurrent_mark_in_progress(false);
176 }
177 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
178
179 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
180 if (has_forwarded_objects) {
181 update_roots(true /*full_gc*/);
182 }
183
184 // d. Reset the bitmaps for new marking
185 heap->reset_mark_bitmap();
186 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
187 assert(!heap->marking_context()->is_complete(), "sanity");
188
189 // e. Abandon reference discovery and clear all discovered references.
190 ShenandoahReferenceProcessor* rp = heap->ref_processor();
191 rp->abandon_partial_discovery();
192
193 // f. Sync pinned region status from the CP marks
194 heap->sync_pinned_region_status();
195
196 // The rest of prologue:
197 _preserved_marks->init(heap->workers()->active_workers());
198
199 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
200 }
201
202 if (UseTLAB) {
203 heap->gclabs_retire(ResizeTLAB);
204 heap->tlabs_retire(ResizeTLAB);
205 }
206
207 OrderAccess::fence();
208
209 phase1_mark_heap();
210
211 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
212 // Coming out of Full GC, we would not have any forwarded objects.
213 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
214 heap->set_has_forwarded_objects(false);
215
216 heap->set_full_gc_move_in_progress(true);
217
218 // Setup workers for the rest
219 OrderAccess::fence();
220
221 // Initialize worker slices
222 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
223 for (uint i = 0; i < heap->max_workers(); i++) {
224 worker_slices[i] = new ShenandoahHeapRegionSet();
225 }
226
227 {
228 // The rest of code performs region moves, where region status is undefined
229 // until all phases run together.
230 ShenandoahHeapLocker lock(heap->lock());
231
232 phase2_calculate_target_addresses(worker_slices);
233
234 OrderAccess::fence();
235
236 phase3_update_references();
237
238 phase4_compact_objects(worker_slices);
239
240 phase5_epilog();
241 }
242
243 // Resize metaspace
244 MetaspaceGC::compute_new_size();
245
246 // Free worker slices
247 for (uint i = 0; i < heap->max_workers(); i++) {
248 delete worker_slices[i];
249 }
250 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
251
252 heap->set_full_gc_move_in_progress(false);
253 heap->set_full_gc_in_progress(false);
254
255 if (ShenandoahVerify) {
256 heap->verifier()->verify_after_fullgc();
257 }
258
259 if (VerifyAfterGC) {
260 Universe::verify();
261 }
262
263 {
264 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
265 heap->post_full_gc_dump(_gc_timer);
266 }
267 }
268
269 class ShenandoahPrepareForMarkClosure: public ShenandoahHeapRegionClosure {
270 private:
271 ShenandoahMarkingContext* const _ctx;
272
273 public:
274 ShenandoahPrepareForMarkClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
275
276 void heap_region_do(ShenandoahHeapRegion *r) {
277 _ctx->capture_top_at_mark_start(r);
278 r->clear_live_data();
279 }
280
281 bool is_thread_safe() { return true; }
282 };
283
284 void ShenandoahFullGC::phase1_mark_heap() {
285 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
286 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
287
288 ShenandoahHeap* heap = ShenandoahHeap::heap();
289
290 ShenandoahPrepareForMarkClosure cl;
291 heap->parallel_heap_region_iterate(&cl);
292
293 heap->set_unload_classes(heap->heuristics()->can_unload_classes());
294
295 ShenandoahReferenceProcessor* rp = heap->ref_processor();
296 // enable ("weak") refs discovery
297 rp->set_soft_reference_policy(true); // forcefully purge all soft references
298
299 ShenandoahSTWMark mark(true /*full_gc*/);
300 mark.mark();
301 heap->parallel_cleaning(true /* full_gc */);
302 }
303
304 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
305 private:
306 PreservedMarks* const _preserved_marks;
307 ShenandoahHeap* const _heap;
308 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
309 int _empty_regions_pos;
310 ShenandoahHeapRegion* _to_region;
311 ShenandoahHeapRegion* _from_region;
312 HeapWord* _compact_point;
313
314 public:
315 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
316 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
317 ShenandoahHeapRegion* to_region) :
318 _preserved_marks(preserved_marks),
319 _heap(ShenandoahHeap::heap()),
320 _empty_regions(empty_regions),
321 _empty_regions_pos(0),
322 _to_region(to_region),
323 _from_region(nullptr),
324 _compact_point(to_region->bottom()) {}
325
326 void set_from_region(ShenandoahHeapRegion* from_region) {
327 _from_region = from_region;
328 }
329
330 void finish() {
331 assert(_to_region != nullptr, "should not happen");
332 _to_region->set_new_top(_compact_point);
333 }
334
335 bool is_compact_same_region() {
336 return _from_region == _to_region;
337 }
338
339 int empty_regions_pos() {
340 return _empty_regions_pos;
341 }
342
343 void do_object(oop p) {
344 assert(_from_region != nullptr, "must set before work");
345 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
346 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
347
348 size_t obj_size = p->size();
349 if (_compact_point + obj_size > _to_region->end()) {
350 finish();
351
352 // Object doesn't fit. Pick next empty region and start compacting there.
353 ShenandoahHeapRegion* new_to_region;
354 if (_empty_regions_pos < _empty_regions.length()) {
355 new_to_region = _empty_regions.at(_empty_regions_pos);
356 _empty_regions_pos++;
357 } else {
358 // Out of empty region? Compact within the same region.
359 new_to_region = _from_region;
360 }
361
362 assert(new_to_region != _to_region, "must not reuse same to-region");
363 assert(new_to_region != nullptr, "must not be null");
364 _to_region = new_to_region;
365 _compact_point = _to_region->bottom();
366 }
367
368 // Object fits into current region, record new location, if object does not move:
369 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
370 shenandoah_assert_not_forwarded(nullptr, p);
371 if (_compact_point != cast_from_oop<HeapWord*>(p)) {
372 _preserved_marks->push_if_necessary(p, p->mark());
373 FullGCForwarding::forward_to(p, cast_to_oop(_compact_point));
374 }
375 _compact_point += obj_size;
376 }
377 };
378
379 class ShenandoahPrepareForCompactionTask : public WorkerTask {
380 private:
381 PreservedMarksSet* const _preserved_marks;
382 ShenandoahHeap* const _heap;
383 ShenandoahHeapRegionSet** const _worker_slices;
384
385 public:
386 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
387 WorkerTask("Shenandoah Prepare For Compaction"),
388 _preserved_marks(preserved_marks),
389 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
390 }
391
392 static bool is_candidate_region(ShenandoahHeapRegion* r) {
393 // Empty region: get it into the slice to defragment the slice itself.
394 // We could have skipped this without violating correctness, but we really
395 // want to compact all live regions to the start of the heap, which sometimes
396 // means moving them into the fully empty regions.
397 if (r->is_empty()) return true;
398
399 // Can move the region, and this is not the humongous region. Humongous
400 // moves are special cased here, because their moves are handled separately.
401 return r->is_stw_move_allowed() && !r->is_humongous();
402 }
403
404 void work(uint worker_id) override;
405 private:
406 template<typename ClosureType>
407 void prepare_for_compaction(ClosureType& cl,
408 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
409 ShenandoahHeapRegionSetIterator& it,
410 ShenandoahHeapRegion* from_region);
411 };
412
413 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
414 ShenandoahParallelWorkerSession worker_session(worker_id);
415 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
416 ShenandoahHeapRegionSetIterator it(slice);
417 ShenandoahHeapRegion* from_region = it.next();
418 // No work?
419 if (from_region == nullptr) {
420 return;
421 }
422
423 // Sliding compaction. Walk all regions in the slice, and compact them.
424 // Remember empty regions and reuse them as needed.
425 ResourceMark rm;
426
427 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
428
429 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
430 prepare_for_compaction(cl, empty_regions, it, from_region);
431 }
432
433 template<typename ClosureType>
434 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
435 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
436 ShenandoahHeapRegionSetIterator& it,
437 ShenandoahHeapRegion* from_region) {
438 while (from_region != nullptr) {
439 assert(is_candidate_region(from_region), "Sanity");
440 cl.set_from_region(from_region);
441 if (from_region->has_live()) {
442 _heap->marked_object_iterate(from_region, &cl);
443 }
444
445 // Compacted the region to somewhere else? From-region is empty then.
446 if (!cl.is_compact_same_region()) {
447 empty_regions.append(from_region);
448 }
449 from_region = it.next();
450 }
451 cl.finish();
452
453 // Mark all remaining regions as empty
454 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
455 ShenandoahHeapRegion* r = empty_regions.at(pos);
456 r->set_new_top(r->bottom());
457 }
458 }
459
460 void ShenandoahFullGC::calculate_target_humongous_objects() {
461 ShenandoahHeap* heap = ShenandoahHeap::heap();
462
463 // Compute the new addresses for humongous objects. We need to do this after addresses
464 // for regular objects are calculated, and we know what regions in heap suffix are
465 // available for humongous moves.
466 //
467 // Scan the heap backwards, because we are compacting humongous regions towards the end.
468 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
469 // humongous start there.
470 //
471 // The complication is potential non-movable regions during the scan. If such region is
472 // detected, then sliding restarts towards that non-movable region.
473
474 size_t to_begin = heap->num_regions();
475 size_t to_end = heap->num_regions();
476
477 for (size_t c = heap->num_regions(); c > 0; c--) {
478 ShenandoahHeapRegion *r = heap->get_region(c - 1);
479 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
480 // To-region candidate: record this, and continue scan
481 to_begin = r->index();
482 continue;
483 }
484
485 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
486 // From-region candidate: movable humongous region
487 oop old_obj = cast_to_oop(r->bottom());
488 size_t words_size = old_obj->size();
489 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
490
491 size_t start = to_end - num_regions;
492
493 if (start >= to_begin && start != r->index()) {
494 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
495 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
496 FullGCForwarding::forward_to(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
497 to_end = start;
498 continue;
499 }
500 }
501
502 // Failed to fit. Scan starting from current region.
503 to_begin = r->index();
504 to_end = r->index();
505 }
506 }
507
508 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
509 private:
510 ShenandoahHeap* const _heap;
511
512 public:
513 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
514 void heap_region_do(ShenandoahHeapRegion* r) {
515 if (r->is_trash()) {
516 r->recycle();
517 }
518 if (r->is_cset()) {
519 r->make_regular_bypass();
520 }
521 if (r->is_empty_uncommitted()) {
522 r->make_committed_bypass();
523 }
524 assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index());
525
526 // Record current region occupancy: this communicates empty regions are free
527 // to the rest of Full GC code.
528 r->set_new_top(r->top());
529 }
530 };
531
532 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
533 private:
534 ShenandoahHeap* const _heap;
535 ShenandoahMarkingContext* const _ctx;
536
537 public:
538 ShenandoahTrashImmediateGarbageClosure() :
539 _heap(ShenandoahHeap::heap()),
540 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
541
542 void heap_region_do(ShenandoahHeapRegion* r) {
543 if (r->is_humongous_start()) {
544 oop humongous_obj = cast_to_oop(r->bottom());
545 if (!_ctx->is_marked(humongous_obj)) {
546 assert(!r->has_live(),
547 "Region " SIZE_FORMAT " is not marked, should not have live", r->index());
548 _heap->trash_humongous_region_at(r);
549 } else {
550 assert(r->has_live(),
551 "Region " SIZE_FORMAT " should have live", r->index());
552 }
553 } else if (r->is_humongous_continuation()) {
554 // If we hit continuation, the non-live humongous starts should have been trashed already
555 assert(r->humongous_start_region()->has_live(),
556 "Region " SIZE_FORMAT " should have live", r->index());
557 } else if (r->is_regular()) {
558 if (!r->has_live()) {
559 r->make_trash_immediate();
560 }
561 }
562 }
563 };
564
565 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
566 ShenandoahHeap* heap = ShenandoahHeap::heap();
567
568 uint n_workers = heap->workers()->active_workers();
569 size_t n_regions = heap->num_regions();
570
571 // What we want to accomplish: have the dense prefix of data, while still balancing
572 // out the parallel work.
573 //
574 // Assuming the amount of work is driven by the live data that needs moving, we can slice
575 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
576 // thread takes all regions in its prefix subset, and then it takes some regions from
577 // the tail.
578 //
579 // Tail region selection becomes interesting.
580 //
581 // First, we want to distribute the regions fairly between the workers, and those regions
582 // might have different amount of live data. So, until we sure no workers need live data,
583 // we need to only take what the worker needs.
584 //
585 // Second, since we slide everything to the left in each slice, the most busy regions
586 // would be the ones on the left. Which means we want to have all workers have their after-tail
587 // regions as close to the left as possible.
588 //
589 // The easiest way to do this is to distribute after-tail regions in round-robin between
590 // workers that still need live data.
591 //
592 // Consider parallel workers A, B, C, then the target slice layout would be:
593 //
594 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
595 //
596 // (.....dense-prefix.....) (.....................tail...................)
597 // [all regions fully live] [left-most regions are fuller that right-most]
598 //
599
600 // Compute how much live data is there. This would approximate the size of dense prefix
601 // we target to create.
602 size_t total_live = 0;
603 for (size_t idx = 0; idx < n_regions; idx++) {
604 ShenandoahHeapRegion *r = heap->get_region(idx);
605 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
606 total_live += r->get_live_data_words();
607 }
608 }
609
610 // Estimate the size for the dense prefix. Note that we specifically count only the
611 // "full" regions, so there would be some non-full regions in the slice tail.
612 size_t live_per_worker = total_live / n_workers;
613 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
614 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
615 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
616 assert(prefix_regions_total <= n_regions, "Sanity");
617
618 // There might be non-candidate regions in the prefix. To compute where the tail actually
619 // ends up being, we need to account those as well.
620 size_t prefix_end = prefix_regions_total;
621 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
622 ShenandoahHeapRegion *r = heap->get_region(idx);
623 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
624 prefix_end++;
625 }
626 }
627 prefix_end = MIN2(prefix_end, n_regions);
628 assert(prefix_end <= n_regions, "Sanity");
629
630 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
631 // subset of dense prefix.
632 size_t prefix_idx = 0;
633
634 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
635
636 for (size_t wid = 0; wid < n_workers; wid++) {
637 ShenandoahHeapRegionSet* slice = worker_slices[wid];
638
639 live[wid] = 0;
640 size_t regs = 0;
641
642 // Add all prefix regions for this worker
643 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
644 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
645 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
646 slice->add_region(r);
647 live[wid] += r->get_live_data_words();
648 regs++;
649 }
650 prefix_idx++;
651 }
652 }
653
654 // Distribute the tail among workers in round-robin fashion.
655 size_t wid = n_workers - 1;
656
657 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
658 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
659 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
660 assert(wid < n_workers, "Sanity");
661
662 size_t live_region = r->get_live_data_words();
663
664 // Select next worker that still needs live data.
665 size_t old_wid = wid;
666 do {
667 wid++;
668 if (wid == n_workers) wid = 0;
669 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
670
671 if (old_wid == wid) {
672 // Circled back to the same worker? This means liveness data was
673 // miscalculated. Bump the live_per_worker limit so that
674 // everyone gets a piece of the leftover work.
675 live_per_worker += ShenandoahHeapRegion::region_size_words();
676 }
677
678 worker_slices[wid]->add_region(r);
679 live[wid] += live_region;
680 }
681 }
682
683 FREE_C_HEAP_ARRAY(size_t, live);
684
685 #ifdef ASSERT
686 ResourceBitMap map(n_regions);
687 for (size_t wid = 0; wid < n_workers; wid++) {
688 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
689 ShenandoahHeapRegion* r = it.next();
690 while (r != nullptr) {
691 size_t idx = r->index();
692 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx);
693 assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx);
694 map.at_put(idx, true);
695 r = it.next();
696 }
697 }
698
699 for (size_t rid = 0; rid < n_regions; rid++) {
700 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
701 bool is_distributed = map.at(rid);
702 assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid);
703 }
704 #endif
705 }
706
707 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
708 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
709 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
710
711 ShenandoahHeap* heap = ShenandoahHeap::heap();
712
713 // About to figure out which regions can be compacted, make sure pinning status
714 // had been updated in GC prologue.
715 heap->assert_pinned_region_status();
716
717 {
718 // Trash the immediately collectible regions before computing addresses
719 ShenandoahTrashImmediateGarbageClosure tigcl;
720 heap->heap_region_iterate(&tigcl);
721
722 // Make sure regions are in good state: committed, active, clean.
723 // This is needed because we are potentially sliding the data through them.
724 ShenandoahEnsureHeapActiveClosure ecl;
725 heap->heap_region_iterate(&ecl);
726 }
727
728 // Compute the new addresses for regular objects
729 {
730 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
731
732 distribute_slices(worker_slices);
733
734 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
735 heap->workers()->run_task(&task);
736 }
737
738 // Compute the new addresses for humongous objects
739 {
740 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
741 calculate_target_humongous_objects();
742 }
743 }
744
745 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
746 private:
747 ShenandoahHeap* const _heap;
748 ShenandoahMarkingContext* const _ctx;
749
750 template <class T>
751 inline void do_oop_work(T* p) {
752 T o = RawAccess<>::oop_load(p);
753 if (!CompressedOops::is_null(o)) {
754 oop obj = CompressedOops::decode_not_null(o);
755 assert(_ctx->is_marked(obj), "must be marked");
756 if (FullGCForwarding::is_forwarded(obj)) {
757 oop forw = FullGCForwarding::forwardee(obj);
758 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
759 }
760 }
761 }
762
763 public:
764 ShenandoahAdjustPointersClosure() :
765 _heap(ShenandoahHeap::heap()),
766 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
767
768 void do_oop(oop* p) { do_oop_work(p); }
769 void do_oop(narrowOop* p) { do_oop_work(p); }
770 void do_method(Method* m) {}
771 void do_nmethod(nmethod* nm) {}
772 };
773
774 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
775 private:
776 ShenandoahHeap* const _heap;
777 ShenandoahAdjustPointersClosure _cl;
778
779 public:
780 ShenandoahAdjustPointersObjectClosure() :
781 _heap(ShenandoahHeap::heap()) {
782 }
783 void do_object(oop p) {
784 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
785 p->oop_iterate(&_cl);
786 }
787 };
788
789 class ShenandoahAdjustPointersTask : public WorkerTask {
790 private:
791 ShenandoahHeap* const _heap;
792 ShenandoahRegionIterator _regions;
793
794 public:
795 ShenandoahAdjustPointersTask() :
796 WorkerTask("Shenandoah Adjust Pointers"),
797 _heap(ShenandoahHeap::heap()) {
798 }
799
800 void work(uint worker_id) {
801 ShenandoahParallelWorkerSession worker_session(worker_id);
802 ShenandoahAdjustPointersObjectClosure obj_cl;
803 ShenandoahHeapRegion* r = _regions.next();
804 while (r != nullptr) {
805 if (!r->is_humongous_continuation() && r->has_live()) {
806 _heap->marked_object_iterate(r, &obj_cl);
807 }
808 r = _regions.next();
809 }
810 }
811 };
812
813 class ShenandoahAdjustRootPointersTask : public WorkerTask {
814 private:
815 ShenandoahRootAdjuster* _rp;
816 PreservedMarksSet* _preserved_marks;
817 public:
818 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
819 WorkerTask("Shenandoah Adjust Root Pointers"),
820 _rp(rp),
821 _preserved_marks(preserved_marks) {}
822
823 void work(uint worker_id) {
824 ShenandoahParallelWorkerSession worker_session(worker_id);
825 ShenandoahAdjustPointersClosure cl;
826 _rp->roots_do(worker_id, &cl);
827 _preserved_marks->get(worker_id)->adjust_during_full_gc();
828 }
829 };
830
831 void ShenandoahFullGC::phase3_update_references() {
832 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
833 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
834
835 ShenandoahHeap* heap = ShenandoahHeap::heap();
836
837 WorkerThreads* workers = heap->workers();
838 uint nworkers = workers->active_workers();
839 {
840 #if COMPILER2_OR_JVMCI
841 DerivedPointerTable::clear();
842 #endif
843 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
844 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
845 workers->run_task(&task);
846 #if COMPILER2_OR_JVMCI
847 DerivedPointerTable::update_pointers();
848 #endif
849 }
850
851 ShenandoahAdjustPointersTask adjust_pointers_task;
852 workers->run_task(&adjust_pointers_task);
853 }
854
855 class ShenandoahCompactObjectsClosure : public ObjectClosure {
856 private:
857 ShenandoahHeap* const _heap;
858 uint const _worker_id;
859
860 public:
861 ShenandoahCompactObjectsClosure(uint worker_id) :
862 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
863
864 void do_object(oop p) {
865 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
866 size_t size = p->size();
867 if (FullGCForwarding::is_forwarded(p)) {
868 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
869 HeapWord* compact_to = cast_from_oop<HeapWord*>(FullGCForwarding::forwardee(p));
870 assert(compact_from != compact_to, "Forwarded object should move");
871 Copy::aligned_conjoint_words(compact_from, compact_to, size);
872 oop new_obj = cast_to_oop(compact_to);
873
874 ContinuationGCSupport::relativize_stack_chunk(new_obj);
875 new_obj->init_mark();
876 }
877 }
878 };
879
880 class ShenandoahCompactObjectsTask : public WorkerTask {
881 private:
882 ShenandoahHeap* const _heap;
883 ShenandoahHeapRegionSet** const _worker_slices;
884
885 public:
886 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
887 WorkerTask("Shenandoah Compact Objects"),
888 _heap(ShenandoahHeap::heap()),
889 _worker_slices(worker_slices) {
890 }
891
892 void work(uint worker_id) {
893 ShenandoahParallelWorkerSession worker_session(worker_id);
894 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
895
896 ShenandoahCompactObjectsClosure cl(worker_id);
897 ShenandoahHeapRegion* r = slice.next();
898 while (r != nullptr) {
899 assert(!r->is_humongous(), "must not get humongous regions here");
900 if (r->has_live()) {
901 _heap->marked_object_iterate(r, &cl);
902 }
903 r->set_top(r->new_top());
904 r = slice.next();
905 }
906 }
907 };
908
909 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
910 private:
911 ShenandoahHeap* const _heap;
912 size_t _live;
913
914 public:
915 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
916 }
917
918 void heap_region_do(ShenandoahHeapRegion* r) {
919 assert (!r->is_cset(), "cset regions should have been demoted already");
920
921 // Need to reset the complete-top-at-mark-start pointer here because
922 // the complete marking bitmap is no longer valid. This ensures
923 // size-based iteration in marked_object_iterate().
924 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
925 // pinned regions.
926 if (!r->is_pinned()) {
927 _heap->complete_marking_context()->reset_top_at_mark_start(r);
928 }
929
930 size_t live = r->used();
931
932 // Make empty regions that have been allocated into regular
933 if (r->is_empty() && live > 0) {
934 r->make_regular_bypass();
935 if (ZapUnusedHeapArea) {
936 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
937 }
938 }
939
940 // Reclaim regular regions that became empty
941 if (r->is_regular() && live == 0) {
942 r->make_trash();
943 }
944
945 // Recycle all trash regions
946 if (r->is_trash()) {
947 live = 0;
948 r->recycle();
949 }
950
951 r->set_live_data(live);
952 r->reset_alloc_metadata();
953 _live += live;
954 }
955
956 size_t get_live() {
957 return _live;
958 }
959 };
960
961 void ShenandoahFullGC::compact_humongous_objects() {
962 // Compact humongous regions, based on their fwdptr objects.
963 //
964 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
965 // humongous regions are already compacted, and do not require further moves, which alleviates
966 // sliding costs. We may consider doing this in parallel in the future.
967
968 ShenandoahHeap* heap = ShenandoahHeap::heap();
969
970 for (size_t c = heap->num_regions(); c > 0; c--) {
971 ShenandoahHeapRegion* r = heap->get_region(c - 1);
972 if (r->is_humongous_start()) {
973 oop old_obj = cast_to_oop(r->bottom());
974 if (!FullGCForwarding::is_forwarded(old_obj)) {
975 // No need to move the object, it stays at the same slot
976 continue;
977 }
978 size_t words_size = old_obj->size();
979 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
980
981 size_t old_start = r->index();
982 size_t old_end = old_start + num_regions - 1;
983 size_t new_start = heap->heap_region_index_containing(FullGCForwarding::forwardee(old_obj));
984 size_t new_end = new_start + num_regions - 1;
985 assert(old_start != new_start, "must be real move");
986 assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
987
988 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
989 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
990
991 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
992 new_obj->init_mark();
993
994 {
995 for (size_t c = old_start; c <= old_end; c++) {
996 ShenandoahHeapRegion* r = heap->get_region(c);
997 r->make_regular_bypass();
998 r->set_top(r->bottom());
999 }
1000
1001 for (size_t c = new_start; c <= new_end; c++) {
1002 ShenandoahHeapRegion* r = heap->get_region(c);
1003 if (c == new_start) {
1004 r->make_humongous_start_bypass();
1005 } else {
1006 r->make_humongous_cont_bypass();
1007 }
1008
1009 // Trailing region may be non-full, record the remainder there
1010 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1011 if ((c == new_end) && (remainder != 0)) {
1012 r->set_top(r->bottom() + remainder);
1013 } else {
1014 r->set_top(r->end());
1015 }
1016
1017 r->reset_alloc_metadata();
1018 }
1019 }
1020 }
1021 }
1022 }
1023
1024 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1025 // we need to remain able to walk pinned regions.
1026 // Since pinned region do not move and don't get compacted, we will get holes with
1027 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1028 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1029 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1030 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1031 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1032 private:
1033 ShenandoahRegionIterator _regions;
1034
1035 public:
1036 ShenandoahMCResetCompleteBitmapTask() :
1037 WorkerTask("Shenandoah Reset Bitmap") {
1038 }
1039
1040 void work(uint worker_id) {
1041 ShenandoahParallelWorkerSession worker_session(worker_id);
1042 ShenandoahHeapRegion* region = _regions.next();
1043 ShenandoahHeap* heap = ShenandoahHeap::heap();
1044 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1045 while (region != nullptr) {
1046 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1047 ctx->clear_bitmap(region);
1048 }
1049 region = _regions.next();
1050 }
1051 }
1052 };
1053
1054 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1055 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1056 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1057
1058 ShenandoahHeap* heap = ShenandoahHeap::heap();
1059
1060 // Compact regular objects first
1061 {
1062 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1063 ShenandoahCompactObjectsTask compact_task(worker_slices);
1064 heap->workers()->run_task(&compact_task);
1065 }
1066
1067 // Compact humongous objects after regular object moves
1068 {
1069 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1070 compact_humongous_objects();
1071 }
1072 }
1073
1074 void ShenandoahFullGC::phase5_epilog() {
1075 GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1076 ShenandoahHeap* heap = ShenandoahHeap::heap();
1077
1078 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1079 // and must ensure the bitmap is in sync.
1080 {
1081 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1082 ShenandoahMCResetCompleteBitmapTask task;
1083 heap->workers()->run_task(&task);
1084 }
1085
1086 // Bring regions in proper states after the collection, and set heap properties.
1087 {
1088 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1089 ShenandoahPostCompactClosure post_compact;
1090 heap->heap_region_iterate(&post_compact);
1091 heap->set_used(post_compact.get_live());
1092
1093 heap->collection_set()->clear();
1094 heap->free_set()->rebuild();
1095 heap->clear_cancelled_gc();
1096 }
1097
1098 _preserved_marks->restore(heap->workers());
1099 _preserved_marks->reclaim();
1100 }