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/gcTraceTime.inline.hpp"
30 #include "gc/shared/preservedMarks.inline.hpp"
31 #include "gc/shared/slidingForwarding.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/shenandoahConcurrentGC.hpp"
36 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
37 #include "gc/shenandoah/shenandoahFreeSet.hpp"
38 #include "gc/shenandoah/shenandoahFullGC.hpp"
39 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
40 #include "gc/shenandoah/shenandoahMark.inline.hpp"
41 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
42 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
43 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
44 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
45 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
46 #include "gc/shenandoah/shenandoahMetrics.hpp"
47 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp"
48 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
49 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
50 #include "gc/shenandoah/shenandoahSTWMark.hpp"
51 #include "gc/shenandoah/shenandoahUtils.hpp"
52 #include "gc/shenandoah/shenandoahVerifier.hpp"
53 #include "gc/shenandoah/shenandoahVMOperations.hpp"
54 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
55 #include "memory/metaspaceUtils.hpp"
56 #include "memory/universe.hpp"
57 #include "oops/compressedOops.inline.hpp"
58 #include "oops/oop.inline.hpp"
59 #include "runtime/javaThread.hpp"
60 #include "runtime/orderAccess.hpp"
61 #include "runtime/vmThread.hpp"
62 #include "utilities/copy.hpp"
63 #include "utilities/events.hpp"
64 #include "utilities/growableArray.hpp"
65
66 ShenandoahFullGC::ShenandoahFullGC() :
67 _gc_timer(ShenandoahHeap::heap()->gc_timer()),
68 _preserved_marks(new PreservedMarksSet(true)) {}
69
70 ShenandoahFullGC::~ShenandoahFullGC() {
71 delete _preserved_marks;
72 }
73
74 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
75 vmop_entry_full(cause);
76 // Always success
77 return true;
78 }
79
80 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
81 ShenandoahHeap* const heap = ShenandoahHeap::heap();
82 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
83 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
84
85 heap->try_inject_alloc_failure();
86 VM_ShenandoahFullGC op(cause, this);
87 VMThread::execute(&op);
88 }
89
90 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
91 static const char* msg = "Pause Full";
92 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
93 EventMark em("%s", msg);
94
95 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
96 ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
97 "full gc");
98
99 op_full(cause);
100 }
101
102 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
103 ShenandoahMetricsSnapshot metrics;
104 metrics.snap_before();
105
106 // Perform full GC
107 do_it(cause);
108
109 metrics.snap_after();
110
111 if (metrics.is_good_progress()) {
112 ShenandoahHeap::heap()->notify_gc_progress();
113 } else {
114 // Nothing to do. Tell the allocation path that we have failed to make
115 // progress, and it can finally fail.
116 ShenandoahHeap::heap()->notify_gc_no_progress();
117 }
118 }
119
120 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
121 ShenandoahHeap* heap = ShenandoahHeap::heap();
122
123 if (ShenandoahVerify) {
124 heap->verifier()->verify_before_fullgc();
125 }
126
127 if (VerifyBeforeGC) {
128 Universe::verify();
129 }
130
131 // Degenerated GC may carry concurrent root flags when upgrading to
132 // full GC. We need to reset it before mutators resume.
133 heap->set_concurrent_strong_root_in_progress(false);
134 heap->set_concurrent_weak_root_in_progress(false);
135
136 heap->set_full_gc_in_progress(true);
137
138 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
139 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
140
141 {
142 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
143 heap->pre_full_gc_dump(_gc_timer);
144 }
145
146 {
147 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
148 // Full GC is supposed to recover from any GC state:
149
150 // a0. Remember if we have forwarded objects
151 bool has_forwarded_objects = heap->has_forwarded_objects();
152
153 // a1. Cancel evacuation, if in progress
154 if (heap->is_evacuation_in_progress()) {
155 heap->set_evacuation_in_progress(false);
156 }
157 assert(!heap->is_evacuation_in_progress(), "sanity");
158
159 // a2. Cancel update-refs, if in progress
160 if (heap->is_update_refs_in_progress()) {
161 heap->set_update_refs_in_progress(false);
162 }
163 assert(!heap->is_update_refs_in_progress(), "sanity");
164
165 // b. Cancel concurrent mark, if in progress
166 if (heap->is_concurrent_mark_in_progress()) {
167 ShenandoahConcurrentGC::cancel();
168 heap->set_concurrent_mark_in_progress(false);
169 }
170 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
171
172 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
173 if (has_forwarded_objects) {
174 update_roots(true /*full_gc*/);
175 }
176
177 // d. Reset the bitmaps for new marking
178 heap->reset_mark_bitmap();
179 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
180 assert(!heap->marking_context()->is_complete(), "sanity");
181
182 // e. Abandon reference discovery and clear all discovered references.
183 ShenandoahReferenceProcessor* rp = heap->ref_processor();
184 rp->abandon_partial_discovery();
185
186 // f. Sync pinned region status from the CP marks
187 heap->sync_pinned_region_status();
188
189 // The rest of prologue:
190 _preserved_marks->init(heap->workers()->active_workers());
191
192 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
193 }
194
195 if (UseTLAB) {
196 heap->gclabs_retire(ResizeTLAB);
197 heap->tlabs_retire(ResizeTLAB);
198 }
199
200 OrderAccess::fence();
201
202 phase1_mark_heap();
203
204 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
205 // Coming out of Full GC, we would not have any forwarded objects.
206 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
207 heap->set_has_forwarded_objects(false);
208
209 heap->set_full_gc_move_in_progress(true);
210
211 // Setup workers for the rest
212 OrderAccess::fence();
213
214 // Initialize worker slices
215 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
216 for (uint i = 0; i < heap->max_workers(); i++) {
217 worker_slices[i] = new ShenandoahHeapRegionSet();
218 }
219
220 {
221 // The rest of code performs region moves, where region status is undefined
222 // until all phases run together.
223 ShenandoahHeapLocker lock(heap->lock());
224
225 SlidingForwarding::begin();
226
227 phase2_calculate_target_addresses(worker_slices);
228
229 OrderAccess::fence();
230
231 phase3_update_references();
232
233 phase4_compact_objects(worker_slices);
234 }
235
236 {
237 // Epilogue
238 _preserved_marks->restore(heap->workers());
239 _preserved_marks->reclaim();
240 SlidingForwarding::end();
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
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 ShenandoahPrepareForMarkClosure cl;
289 heap->heap_region_iterate(&cl);
290
291 heap->set_unload_classes(heap->heuristics()->can_unload_classes());
292
293 ShenandoahReferenceProcessor* rp = heap->ref_processor();
294 // enable ("weak") refs discovery
295 rp->set_soft_reference_policy(true); // forcefully purge all soft references
296
297 ShenandoahSTWMark mark(true /*full_gc*/);
298 mark.mark();
299 heap->parallel_cleaning(true /* full_gc */);
300 }
301
302 template <bool ALT_FWD>
303 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
304 private:
305 PreservedMarks* const _preserved_marks;
306 ShenandoahHeap* const _heap;
307 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
308 int _empty_regions_pos;
309 ShenandoahHeapRegion* _to_region;
310 ShenandoahHeapRegion* _from_region;
311 HeapWord* _compact_point;
312
313 public:
314 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
315 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
316 ShenandoahHeapRegion* to_region) :
317 _preserved_marks(preserved_marks),
318 _heap(ShenandoahHeap::heap()),
319 _empty_regions(empty_regions),
320 _empty_regions_pos(0),
321 _to_region(to_region),
322 _from_region(nullptr),
323 _compact_point(to_region->bottom()) {}
324
325 void set_from_region(ShenandoahHeapRegion* from_region) {
326 _from_region = from_region;
327 }
328
329 void finish_region() {
330 assert(_to_region != nullptr, "should not happen");
331 _to_region->set_new_top(_compact_point);
332 }
333
334 bool is_compact_same_region() {
335 return _from_region == _to_region;
336 }
337
338 int empty_regions_pos() {
339 return _empty_regions_pos;
340 }
341
342 void do_object(oop p) {
343 assert(_from_region != nullptr, "must set before work");
344 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
345 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
346
347 size_t obj_size = p->size();
348 if (_compact_point + obj_size > _to_region->end()) {
349 finish_region();
350
351 // Object doesn't fit. Pick next empty region and start compacting there.
352 ShenandoahHeapRegion* new_to_region;
353 if (_empty_regions_pos < _empty_regions.length()) {
354 new_to_region = _empty_regions.at(_empty_regions_pos);
355 _empty_regions_pos++;
356 } else {
357 // Out of empty region? Compact within the same region.
358 new_to_region = _from_region;
359 }
360
361 assert(new_to_region != _to_region, "must not reuse same to-region");
362 assert(new_to_region != nullptr, "must not be null");
363 _to_region = new_to_region;
364 _compact_point = _to_region->bottom();
365 }
366
367 // Object fits into current region, record new location:
368 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
369 shenandoah_assert_not_forwarded(nullptr, p);
370 _preserved_marks->push_if_necessary(p, p->mark());
371 SlidingForwarding::forward_to<ALT_FWD>(p, cast_to_oop(_compact_point));
372 _compact_point += obj_size;
373 }
374 };
375
376 class ShenandoahPrepareForCompactionTask : public WorkerTask {
377 private:
378 PreservedMarksSet* const _preserved_marks;
379 ShenandoahHeap* const _heap;
380 ShenandoahHeapRegionSet** const _worker_slices;
381
382 public:
383 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
384 WorkerTask("Shenandoah Prepare For Compaction"),
385 _preserved_marks(preserved_marks),
386 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
387 }
388
389 static bool is_candidate_region(ShenandoahHeapRegion* r) {
390 // Empty region: get it into the slice to defragment the slice itself.
391 // We could have skipped this without violating correctness, but we really
392 // want to compact all live regions to the start of the heap, which sometimes
393 // means moving them into the fully empty regions.
394 if (r->is_empty()) return true;
395
396 // Can move the region, and this is not the humongous region. Humongous
397 // moves are special cased here, because their moves are handled separately.
398 return r->is_stw_move_allowed() && !r->is_humongous();
399 }
400
401 void work(uint worker_id) {
402 if (UseAltGCForwarding) {
403 work_impl<true>(worker_id);
404 } else {
405 work_impl<false>(worker_id);
406 }
407 }
408
409 private:
410 template <bool ALT_FWD>
411 void work_impl(uint worker_id) {
412 ShenandoahParallelWorkerSession worker_session(worker_id);
413 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
414 ShenandoahHeapRegionSetIterator it(slice);
415 ShenandoahHeapRegion* from_region = it.next();
416 // No work?
417 if (from_region == nullptr) {
418 return;
419 }
420
421 // Sliding compaction. Walk all regions in the slice, and compact them.
422 // Remember empty regions and reuse them as needed.
423 ResourceMark rm;
424
425 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
426
427 ShenandoahPrepareForCompactionObjectClosure<ALT_FWD> cl(_preserved_marks->get(worker_id), empty_regions, from_region);
428
429 while (from_region != nullptr) {
430 assert(is_candidate_region(from_region), "Sanity");
431
432 cl.set_from_region(from_region);
433 if (from_region->has_live()) {
434 _heap->marked_object_iterate(from_region, &cl);
435 }
436
437 // Compacted the region to somewhere else? From-region is empty then.
438 if (!cl.is_compact_same_region()) {
439 empty_regions.append(from_region);
440 }
441 from_region = it.next();
442 }
443 cl.finish_region();
444
445 // Mark all remaining regions as empty
446 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
447 ShenandoahHeapRegion* r = empty_regions.at(pos);
448 r->set_new_top(r->bottom());
449 }
450 }
451 };
452
453 template <bool ALT_FWD>
454 void ShenandoahFullGC::calculate_target_humongous_objects_impl() {
455 ShenandoahHeap* heap = ShenandoahHeap::heap();
456
457 // Compute the new addresses for humongous objects. We need to do this after addresses
458 // for regular objects are calculated, and we know what regions in heap suffix are
459 // available for humongous moves.
460 //
461 // Scan the heap backwards, because we are compacting humongous regions towards the end.
462 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
463 // humongous start there.
464 //
465 // The complication is potential non-movable regions during the scan. If such region is
466 // detected, then sliding restarts towards that non-movable region.
467
468 size_t to_begin = heap->num_regions();
469 size_t to_end = heap->num_regions();
470
471 for (size_t c = heap->num_regions(); c > 0; c--) {
472 ShenandoahHeapRegion *r = heap->get_region(c - 1);
473 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
474 // To-region candidate: record this, and continue scan
475 to_begin = r->index();
476 continue;
477 }
478
479 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
480 // From-region candidate: movable humongous region
481 oop old_obj = cast_to_oop(r->bottom());
482 size_t words_size = old_obj->size();
483 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
484
485 size_t start = to_end - num_regions;
486
487 if (start >= to_begin && start != r->index()) {
488 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
489 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
490 SlidingForwarding::forward_to<ALT_FWD>(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
491 to_end = start;
492 continue;
493 }
494 }
495
496 // Failed to fit. Scan starting from current region.
497 to_begin = r->index();
498 to_end = r->index();
499 }
500 }
501
502 void ShenandoahFullGC::calculate_target_humongous_objects() {
503 if (UseAltGCForwarding) {
504 calculate_target_humongous_objects_impl<true>();
505 } else {
506 calculate_target_humongous_objects_impl<false>();
507 }
508 }
509
510 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
511 private:
512 ShenandoahHeap* const _heap;
513
514 public:
515 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
516 void heap_region_do(ShenandoahHeapRegion* r) {
517 if (r->is_trash()) {
518 r->recycle();
519 }
520 if (r->is_cset()) {
521 r->make_regular_bypass();
522 }
523 if (r->is_empty_uncommitted()) {
524 r->make_committed_bypass();
525 }
526 assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index());
527
528 // Record current region occupancy: this communicates empty regions are free
529 // to the rest of Full GC code.
530 r->set_new_top(r->top());
531 }
532 };
533
534 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
535 private:
536 ShenandoahHeap* const _heap;
537 ShenandoahMarkingContext* const _ctx;
538
539 public:
540 ShenandoahTrashImmediateGarbageClosure() :
541 _heap(ShenandoahHeap::heap()),
542 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
543
544 void heap_region_do(ShenandoahHeapRegion* r) {
545 if (r->is_humongous_start()) {
546 oop humongous_obj = cast_to_oop(r->bottom());
547 if (!_ctx->is_marked(humongous_obj)) {
548 assert(!r->has_live(),
549 "Region " SIZE_FORMAT " is not marked, should not have live", r->index());
550 _heap->trash_humongous_region_at(r);
551 } else {
552 assert(r->has_live(),
553 "Region " SIZE_FORMAT " should have live", r->index());
554 }
555 } else if (r->is_humongous_continuation()) {
556 // If we hit continuation, the non-live humongous starts should have been trashed already
557 assert(r->humongous_start_region()->has_live(),
558 "Region " SIZE_FORMAT " should have live", r->index());
559 } else if (r->is_regular()) {
560 if (!r->has_live()) {
561 r->make_trash_immediate();
562 }
563 }
564 }
565 };
566
567 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
568 ShenandoahHeap* heap = ShenandoahHeap::heap();
569
570 uint n_workers = heap->workers()->active_workers();
571 size_t n_regions = heap->num_regions();
572
573 // What we want to accomplish: have the dense prefix of data, while still balancing
574 // out the parallel work.
575 //
576 // Assuming the amount of work is driven by the live data that needs moving, we can slice
577 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
578 // thread takes all regions in its prefix subset, and then it takes some regions from
579 // the tail.
580 //
581 // Tail region selection becomes interesting.
582 //
583 // First, we want to distribute the regions fairly between the workers, and those regions
584 // might have different amount of live data. So, until we sure no workers need live data,
585 // we need to only take what the worker needs.
586 //
587 // Second, since we slide everything to the left in each slice, the most busy regions
588 // would be the ones on the left. Which means we want to have all workers have their after-tail
589 // regions as close to the left as possible.
590 //
591 // The easiest way to do this is to distribute after-tail regions in round-robin between
592 // workers that still need live data.
593 //
594 // Consider parallel workers A, B, C, then the target slice layout would be:
595 //
596 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
597 //
598 // (.....dense-prefix.....) (.....................tail...................)
599 // [all regions fully live] [left-most regions are fuller that right-most]
600 //
601
602 // Compute how much live data is there. This would approximate the size of dense prefix
603 // we target to create.
604 size_t total_live = 0;
605 for (size_t idx = 0; idx < n_regions; idx++) {
606 ShenandoahHeapRegion *r = heap->get_region(idx);
607 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
608 total_live += r->get_live_data_words();
609 }
610 }
611
612 // Estimate the size for the dense prefix. Note that we specifically count only the
613 // "full" regions, so there would be some non-full regions in the slice tail.
614 size_t live_per_worker = total_live / n_workers;
615 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
616 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
617 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
618 assert(prefix_regions_total <= n_regions, "Sanity");
619
620 // There might be non-candidate regions in the prefix. To compute where the tail actually
621 // ends up being, we need to account those as well.
622 size_t prefix_end = prefix_regions_total;
623 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
624 ShenandoahHeapRegion *r = heap->get_region(idx);
625 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
626 prefix_end++;
627 }
628 }
629 prefix_end = MIN2(prefix_end, n_regions);
630 assert(prefix_end <= n_regions, "Sanity");
631
632 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
633 // subset of dense prefix.
634 size_t prefix_idx = 0;
635
636 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
637
638 for (size_t wid = 0; wid < n_workers; wid++) {
639 ShenandoahHeapRegionSet* slice = worker_slices[wid];
640
641 live[wid] = 0;
642 size_t regs = 0;
643
644 // Add all prefix regions for this worker
645 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
646 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
647 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
648 slice->add_region(r);
649 live[wid] += r->get_live_data_words();
650 regs++;
651 }
652 prefix_idx++;
653 }
654 }
655
656 // Distribute the tail among workers in round-robin fashion.
657 size_t wid = n_workers - 1;
658
659 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
660 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
661 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
662 assert(wid < n_workers, "Sanity");
663
664 size_t live_region = r->get_live_data_words();
665
666 // Select next worker that still needs live data.
667 size_t old_wid = wid;
668 do {
669 wid++;
670 if (wid == n_workers) wid = 0;
671 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
672
673 if (old_wid == wid) {
674 // Circled back to the same worker? This means liveness data was
675 // miscalculated. Bump the live_per_worker limit so that
676 // everyone gets a piece of the leftover work.
677 live_per_worker += ShenandoahHeapRegion::region_size_words();
678 }
679
680 worker_slices[wid]->add_region(r);
681 live[wid] += live_region;
682 }
683 }
684
685 FREE_C_HEAP_ARRAY(size_t, live);
686
687 #ifdef ASSERT
688 ResourceBitMap map(n_regions);
689 for (size_t wid = 0; wid < n_workers; wid++) {
690 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
691 ShenandoahHeapRegion* r = it.next();
692 while (r != nullptr) {
693 size_t idx = r->index();
694 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx);
695 assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx);
696 map.at_put(idx, true);
697 r = it.next();
698 }
699 }
700
701 for (size_t rid = 0; rid < n_regions; rid++) {
702 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
703 bool is_distributed = map.at(rid);
704 assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid);
705 }
706 #endif
707 }
708
709 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
710 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
711 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
712
713 ShenandoahHeap* heap = ShenandoahHeap::heap();
714
715 // About to figure out which regions can be compacted, make sure pinning status
716 // had been updated in GC prologue.
717 heap->assert_pinned_region_status();
718
719 {
720 // Trash the immediately collectible regions before computing addresses
721 ShenandoahTrashImmediateGarbageClosure tigcl;
722 heap->heap_region_iterate(&tigcl);
723
724 // Make sure regions are in good state: committed, active, clean.
725 // This is needed because we are potentially sliding the data through them.
726 ShenandoahEnsureHeapActiveClosure ecl;
727 heap->heap_region_iterate(&ecl);
728 }
729
730 // Compute the new addresses for regular objects
731 {
732 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
733
734 distribute_slices(worker_slices);
735
736 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
737 heap->workers()->run_task(&task);
738 }
739
740 // Compute the new addresses for humongous objects
741 {
742 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
743 calculate_target_humongous_objects();
744 }
745 }
746
747 template <bool ALT_FWD>
748 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
749 private:
750 ShenandoahHeap* const _heap;
751 ShenandoahMarkingContext* const _ctx;
752
753 template <class T>
754 inline void do_oop_work(T* p) {
755 T o = RawAccess<>::oop_load(p);
756 if (!CompressedOops::is_null(o)) {
757 oop obj = CompressedOops::decode_not_null(o);
758 assert(_ctx->is_marked(obj), "must be marked");
759 if (SlidingForwarding::is_forwarded(obj)) {
760 oop forw = SlidingForwarding::forwardee<ALT_FWD>(obj);
761 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
762 }
763 }
764 }
765
766 public:
767 ShenandoahAdjustPointersClosure() :
768 _heap(ShenandoahHeap::heap()),
769 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
770
771 void do_oop(oop* p) { do_oop_work(p); }
772 void do_oop(narrowOop* p) { do_oop_work(p); }
773 void do_method(Method* m) {}
774 void do_nmethod(nmethod* nm) {}
775 };
776
777 template <bool ALT_FWD>
778 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
779 private:
780 ShenandoahHeap* const _heap;
781 ShenandoahAdjustPointersClosure<ALT_FWD> _cl;
782
783 public:
784 ShenandoahAdjustPointersObjectClosure() :
785 _heap(ShenandoahHeap::heap()) {
786 }
787 void do_object(oop p) {
788 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
789 p->oop_iterate(&_cl);
790 }
791 };
792
793 class ShenandoahAdjustPointersTask : public WorkerTask {
794 private:
795 ShenandoahHeap* const _heap;
796 ShenandoahRegionIterator _regions;
797
798 public:
799 ShenandoahAdjustPointersTask() :
800 WorkerTask("Shenandoah Adjust Pointers"),
801 _heap(ShenandoahHeap::heap()) {
802 }
803
804 private:
805 template <bool ALT_FWD>
806 void work_impl(uint worker_id) {
807 ShenandoahParallelWorkerSession worker_session(worker_id);
808 ShenandoahAdjustPointersObjectClosure<ALT_FWD> obj_cl;
809 ShenandoahHeapRegion* r = _regions.next();
810 while (r != nullptr) {
811 if (!r->is_humongous_continuation() && r->has_live()) {
812 _heap->marked_object_iterate(r, &obj_cl);
813 }
814 r = _regions.next();
815 }
816 }
817
818 public:
819 void work(uint worker_id) {
820 if (UseAltGCForwarding) {
821 work_impl<true>(worker_id);
822 } else {
823 work_impl<false>(worker_id);
824 }
825 }
826 };
827
828 class ShenandoahAdjustRootPointersTask : public WorkerTask {
829 private:
830 ShenandoahRootAdjuster* _rp;
831 PreservedMarksSet* _preserved_marks;
832
833 public:
834 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
835 WorkerTask("Shenandoah Adjust Root Pointers"),
836 _rp(rp),
837 _preserved_marks(preserved_marks) {}
838
839 private:
840 template <bool ALT_FWD>
841 void work_impl(uint worker_id) {
842 ShenandoahParallelWorkerSession worker_session(worker_id);
843 ShenandoahAdjustPointersClosure<ALT_FWD> cl;
844 _rp->roots_do(worker_id, &cl);
845 _preserved_marks->get(worker_id)->adjust_during_full_gc();
846 }
847
848 public:
849 void work(uint worker_id) {
850 if (UseAltGCForwarding) {
851 work_impl<true>(worker_id);
852 } else {
853 work_impl<false>(worker_id);
854 }
855 }
856 };
857
858 void ShenandoahFullGC::phase3_update_references() {
859 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
860 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
861
862 ShenandoahHeap* heap = ShenandoahHeap::heap();
863
864 WorkerThreads* workers = heap->workers();
865 uint nworkers = workers->active_workers();
866 {
867 #if COMPILER2_OR_JVMCI
868 DerivedPointerTable::clear();
869 #endif
870 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
871 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
872 workers->run_task(&task);
873 #if COMPILER2_OR_JVMCI
874 DerivedPointerTable::update_pointers();
875 #endif
876 }
877
878 ShenandoahAdjustPointersTask adjust_pointers_task;
879 workers->run_task(&adjust_pointers_task);
880 }
881
882 template <bool ALT_FWD>
883 class ShenandoahCompactObjectsClosure : public ObjectClosure {
884 private:
885 ShenandoahHeap* const _heap;
886 uint const _worker_id;
887
888 public:
889 ShenandoahCompactObjectsClosure(uint worker_id) :
890 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
891
892 void do_object(oop p) {
893 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
894 size_t size = p->size();
895 if (SlidingForwarding::is_forwarded(p)) {
896 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
897 HeapWord* compact_to = cast_from_oop<HeapWord*>(SlidingForwarding::forwardee<ALT_FWD>(p));
898 Copy::aligned_conjoint_words(compact_from, compact_to, size);
899 oop new_obj = cast_to_oop(compact_to);
900
901 ContinuationGCSupport::relativize_stack_chunk(new_obj);
902 new_obj->init_mark();
903 }
904 }
905 };
906
907 class ShenandoahCompactObjectsTask : public WorkerTask {
908 private:
909 ShenandoahHeap* const _heap;
910 ShenandoahHeapRegionSet** const _worker_slices;
911
912 public:
913 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
914 WorkerTask("Shenandoah Compact Objects"),
915 _heap(ShenandoahHeap::heap()),
916 _worker_slices(worker_slices) {
917 }
918
919 private:
920 template <bool ALT_FWD>
921 void work_impl(uint worker_id) {
922 ShenandoahParallelWorkerSession worker_session(worker_id);
923 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
924
925 ShenandoahCompactObjectsClosure<ALT_FWD> cl(worker_id);
926 ShenandoahHeapRegion* r = slice.next();
927 while (r != nullptr) {
928 assert(!r->is_humongous(), "must not get humongous regions here");
929 if (r->has_live()) {
930 _heap->marked_object_iterate(r, &cl);
931 }
932 r->set_top(r->new_top());
933 r = slice.next();
934 }
935 }
936
937 public:
938 void work(uint worker_id) {
939 if (UseAltGCForwarding) {
940 work_impl<true>(worker_id);
941 } else {
942 work_impl<false>(worker_id);
943 }
944 }
945 };
946
947 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
948 private:
949 ShenandoahHeap* const _heap;
950 size_t _live;
951
952 public:
953 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
954 _heap->free_set()->clear();
955 }
956
957 void heap_region_do(ShenandoahHeapRegion* r) {
958 assert (!r->is_cset(), "cset regions should have been demoted already");
959
960 // Need to reset the complete-top-at-mark-start pointer here because
961 // the complete marking bitmap is no longer valid. This ensures
962 // size-based iteration in marked_object_iterate().
963 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
964 // pinned regions.
965 if (!r->is_pinned()) {
966 _heap->complete_marking_context()->reset_top_at_mark_start(r);
967 }
968
969 size_t live = r->used();
970
971 // Make empty regions that have been allocated into regular
972 if (r->is_empty() && live > 0) {
973 r->make_regular_bypass();
974 if (ZapUnusedHeapArea) {
975 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
976 }
977 }
978
979 // Reclaim regular regions that became empty
980 if (r->is_regular() && live == 0) {
981 r->make_trash();
982 }
983
984 // Recycle all trash regions
985 if (r->is_trash()) {
986 live = 0;
987 r->recycle();
988 }
989
990 r->set_live_data(live);
991 r->reset_alloc_metadata();
992 _live += live;
993 }
994
995 size_t get_live() {
996 return _live;
997 }
998 };
999
1000 template <bool ALT_FWD>
1001 void ShenandoahFullGC::compact_humongous_objects_impl() {
1002 // Compact humongous regions, based on their fwdptr objects.
1003 //
1004 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1005 // humongous regions are already compacted, and do not require further moves, which alleviates
1006 // sliding costs. We may consider doing this in parallel in future.
1007
1008 ShenandoahHeap* heap = ShenandoahHeap::heap();
1009
1010 for (size_t c = heap->num_regions(); c > 0; c--) {
1011 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1012 if (r->is_humongous_start()) {
1013 oop old_obj = cast_to_oop(r->bottom());
1014 if (SlidingForwarding::is_not_forwarded(old_obj)) {
1015 // No need to move the object, it stays at the same slot
1016 continue;
1017 }
1018 size_t words_size = old_obj->size();
1019 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1020
1021 size_t old_start = r->index();
1022 size_t old_end = old_start + num_regions - 1;
1023 size_t new_start = heap->heap_region_index_containing(SlidingForwarding::forwardee<ALT_FWD>(old_obj));
1024 size_t new_end = new_start + num_regions - 1;
1025 assert(old_start != new_start, "must be real move");
1026 assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
1027
1028 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
1029 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
1030
1031 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1032 new_obj->init_mark();
1033
1034 {
1035 for (size_t c = old_start; c <= old_end; c++) {
1036 ShenandoahHeapRegion* r = heap->get_region(c);
1037 r->make_regular_bypass();
1038 r->set_top(r->bottom());
1039 }
1040
1041 for (size_t c = new_start; c <= new_end; c++) {
1042 ShenandoahHeapRegion* r = heap->get_region(c);
1043 if (c == new_start) {
1044 r->make_humongous_start_bypass();
1045 } else {
1046 r->make_humongous_cont_bypass();
1047 }
1048
1049 // Trailing region may be non-full, record the remainder there
1050 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1051 if ((c == new_end) && (remainder != 0)) {
1052 r->set_top(r->bottom() + remainder);
1053 } else {
1054 r->set_top(r->end());
1055 }
1056
1057 r->reset_alloc_metadata();
1058 }
1059 }
1060 }
1061 }
1062 }
1063
1064 void ShenandoahFullGC::compact_humongous_objects() {
1065 if (UseAltGCForwarding) {
1066 compact_humongous_objects_impl<true>();
1067 } else {
1068 compact_humongous_objects_impl<false>();
1069 }
1070 }
1071
1072 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1073 // we need to remain able to walk pinned regions.
1074 // Since pinned region do not move and don't get compacted, we will get holes with
1075 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1076 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1077 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1078 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1079 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1080 private:
1081 ShenandoahRegionIterator _regions;
1082
1083 public:
1084 ShenandoahMCResetCompleteBitmapTask() :
1085 WorkerTask("Shenandoah Reset Bitmap") {
1086 }
1087
1088 void work(uint worker_id) {
1089 ShenandoahParallelWorkerSession worker_session(worker_id);
1090 ShenandoahHeapRegion* region = _regions.next();
1091 ShenandoahHeap* heap = ShenandoahHeap::heap();
1092 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1093 while (region != nullptr) {
1094 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1095 ctx->clear_bitmap(region);
1096 }
1097 region = _regions.next();
1098 }
1099 }
1100 };
1101
1102 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1103 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1104 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1105
1106 ShenandoahHeap* heap = ShenandoahHeap::heap();
1107
1108 // Compact regular objects first
1109 {
1110 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1111 ShenandoahCompactObjectsTask compact_task(worker_slices);
1112 heap->workers()->run_task(&compact_task);
1113 }
1114
1115 // Compact humongous objects after regular object moves
1116 {
1117 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1118 compact_humongous_objects();
1119 }
1120
1121 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1122 // and must ensure the bitmap is in sync.
1123 {
1124 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1125 ShenandoahMCResetCompleteBitmapTask task;
1126 heap->workers()->run_task(&task);
1127 }
1128
1129 // Bring regions in proper states after the collection, and set heap properties.
1130 {
1131 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1132
1133 ShenandoahPostCompactClosure post_compact;
1134 heap->heap_region_iterate(&post_compact);
1135 heap->set_used(post_compact.get_live());
1136
1137 heap->collection_set()->clear();
1138 heap->free_set()->rebuild();
1139 }
1140
1141 heap->clear_cancelled_gc();
1142 }