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