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/gcTraceTime.inline.hpp"
29 #include "gc/shared/preservedMarks.inline.hpp"
30 #include "gc/shared/slidingForwarding.inline.hpp"
31 #include "gc/shared/tlab_globals.hpp"
32 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
33 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
34 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
35 #include "gc/shenandoah/shenandoahFreeSet.hpp"
36 #include "gc/shenandoah/shenandoahFullGC.hpp"
37 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
38 #include "gc/shenandoah/shenandoahMark.inline.hpp"
39 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
40 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
41 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
42 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
43 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
44 #include "gc/shenandoah/shenandoahMetrics.hpp"
45 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp"
46 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
47 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
48 #include "gc/shenandoah/shenandoahSTWMark.hpp"
49 #include "gc/shenandoah/shenandoahUtils.hpp"
50 #include "gc/shenandoah/shenandoahVerifier.hpp"
51 #include "gc/shenandoah/shenandoahVMOperations.hpp"
52 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
53 #include "memory/metaspaceUtils.hpp"
54 #include "memory/universe.hpp"
55 #include "oops/compressedOops.inline.hpp"
56 #include "oops/oop.inline.hpp"
57 #include "runtime/biasedLocking.hpp"
58 #include "runtime/orderAccess.hpp"
59 #include "runtime/thread.hpp"
60 #include "runtime/vmThread.hpp"
61 #include "utilities/copy.hpp"
62 #include "utilities/events.hpp"
63 #include "utilities/growableArray.hpp"
64 #include "gc/shared/workgroup.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 BiasedLocking::preserve_marks();
191 _preserved_marks->init(heap->workers()->active_workers());
192
193 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
194 }
195
196 if (UseTLAB) {
197 heap->gclabs_retire(ResizeTLAB);
198 heap->tlabs_retire(ResizeTLAB);
199 }
200
201 OrderAccess::fence();
202
203 phase1_mark_heap();
204
205 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
206 // Coming out of Full GC, we would not have any forwarded objects.
207 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
208 heap->set_has_forwarded_objects(false);
209
210 heap->set_full_gc_move_in_progress(true);
211
212 // Setup workers for the rest
213 OrderAccess::fence();
214
215 // Initialize worker slices
216 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
217 for (uint i = 0; i < heap->max_workers(); i++) {
218 worker_slices[i] = new ShenandoahHeapRegionSet();
219 }
220
221 {
222 // The rest of code performs region moves, where region status is undefined
223 // until all phases run together.
224 ShenandoahHeapLocker lock(heap->lock());
225
226 SlidingForwarding::begin();
227
228 phase2_calculate_target_addresses(worker_slices);
229
230 OrderAccess::fence();
231
232 phase3_update_references();
233
234 phase4_compact_objects(worker_slices);
235 }
236
237 {
238 // Epilogue
239 _preserved_marks->restore(heap->workers());
240 BiasedLocking::restore_marks();
241 _preserved_marks->reclaim();
242 SlidingForwarding::end();
243 }
244
245 // Resize metaspace
246 MetaspaceGC::compute_new_size();
247
248 // Free worker slices
249 for (uint i = 0; i < heap->max_workers(); i++) {
250 delete worker_slices[i];
251 }
252 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
253
254 heap->set_full_gc_move_in_progress(false);
255 heap->set_full_gc_in_progress(false);
256
257 if (ShenandoahVerify) {
258 heap->verifier()->verify_after_fullgc();
259 }
260
261 if (VerifyAfterGC) {
262 Universe::verify();
263 }
264
265 {
266 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
267 heap->post_full_gc_dump(_gc_timer);
268 }
269 }
270
271 class ShenandoahPrepareForMarkClosure: public ShenandoahHeapRegionClosure {
272 private:
273 ShenandoahMarkingContext* const _ctx;
274
275 public:
276 ShenandoahPrepareForMarkClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
277
278 void heap_region_do(ShenandoahHeapRegion *r) {
279 _ctx->capture_top_at_mark_start(r);
280 r->clear_live_data();
281 }
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->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 template <bool ALT_FWD>
305 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
306 private:
307 PreservedMarks* const _preserved_marks;
308 ShenandoahHeap* const _heap;
309 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
310 int _empty_regions_pos;
311 ShenandoahHeapRegion* _to_region;
312 ShenandoahHeapRegion* _from_region;
313 HeapWord* _compact_point;
314
315 public:
316 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
317 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
318 ShenandoahHeapRegion* to_region) :
319 _preserved_marks(preserved_marks),
320 _heap(ShenandoahHeap::heap()),
321 _empty_regions(empty_regions),
322 _empty_regions_pos(0),
323 _to_region(to_region),
324 _from_region(NULL),
325 _compact_point(to_region->bottom()) {}
326
327 void set_from_region(ShenandoahHeapRegion* from_region) {
328 _from_region = from_region;
329 }
330
331 void finish_region() {
332 assert(_to_region != NULL, "should not happen");
333 _to_region->set_new_top(_compact_point);
334 }
335
336 bool is_compact_same_region() {
337 return _from_region == _to_region;
338 }
339
340 int empty_regions_pos() {
341 return _empty_regions_pos;
342 }
343
344 void do_object(oop p) {
345 assert(_from_region != NULL, "must set before work");
346 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
347 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
348
349 size_t obj_size = p->size();
350 if (_compact_point + obj_size > _to_region->end()) {
351 finish_region();
352
353 // Object doesn't fit. Pick next empty region and start compacting there.
354 ShenandoahHeapRegion* new_to_region;
355 if (_empty_regions_pos < _empty_regions.length()) {
356 new_to_region = _empty_regions.at(_empty_regions_pos);
357 _empty_regions_pos++;
358 } else {
359 // Out of empty region? Compact within the same region.
360 new_to_region = _from_region;
361 }
362
363 assert(new_to_region != _to_region, "must not reuse same to-region");
364 assert(new_to_region != NULL, "must not be NULL");
365 _to_region = new_to_region;
366 _compact_point = _to_region->bottom();
367 }
368
369 // Object fits into current region, record new location:
370 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
371 shenandoah_assert_not_forwarded(NULL, p);
372 _preserved_marks->push_if_necessary(p, p->mark());
373 SlidingForwarding::forward_to<ALT_FWD>(p, cast_to_oop(_compact_point));
374 _compact_point += obj_size;
375 }
376 };
377
378 class ShenandoahPrepareForCompactionTask : public AbstractGangTask {
379 private:
380 PreservedMarksSet* const _preserved_marks;
381 ShenandoahHeap* const _heap;
382 ShenandoahHeapRegionSet** const _worker_slices;
383
384 public:
385 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
386 AbstractGangTask("Shenandoah Prepare For Compaction"),
387 _preserved_marks(preserved_marks),
388 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
389 }
390
391 static bool is_candidate_region(ShenandoahHeapRegion* r) {
392 // Empty region: get it into the slice to defragment the slice itself.
393 // We could have skipped this without violating correctness, but we really
394 // want to compact all live regions to the start of the heap, which sometimes
395 // means moving them into the fully empty regions.
396 if (r->is_empty()) return true;
397
398 // Can move the region, and this is not the humongous region. Humongous
399 // moves are special cased here, because their moves are handled separately.
400 return r->is_stw_move_allowed() && !r->is_humongous();
401 }
402
403 void work(uint worker_id) {
404 if (UseAltGCForwarding) {
405 work_impl<true>(worker_id);
406 } else {
407 work_impl<false>(worker_id);
408 }
409 }
410
411 private:
412 template <bool ALT_FWD>
413 void work_impl(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 == NULL) {
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<ALT_FWD> cl(_preserved_marks->get(worker_id), empty_regions, from_region);
430
431 while (from_region != NULL) {
432 assert(is_candidate_region(from_region), "Sanity");
433
434 cl.set_from_region(from_region);
435 if (from_region->has_live()) {
436 _heap->marked_object_iterate(from_region, &cl);
437 }
438
439 // Compacted the region to somewhere else? From-region is empty then.
440 if (!cl.is_compact_same_region()) {
441 empty_regions.append(from_region);
442 }
443 from_region = it.next();
444 }
445 cl.finish_region();
446
447 // Mark all remaining regions as empty
448 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
449 ShenandoahHeapRegion* r = empty_regions.at(pos);
450 r->set_new_top(r->bottom());
451 }
452 }
453 };
454
455 template <bool ALT_FWD>
456 void ShenandoahFullGC::calculate_target_humongous_objects_impl() {
457 ShenandoahHeap* heap = ShenandoahHeap::heap();
458
459 // Compute the new addresses for humongous objects. We need to do this after addresses
460 // for regular objects are calculated, and we know what regions in heap suffix are
461 // available for humongous moves.
462 //
463 // Scan the heap backwards, because we are compacting humongous regions towards the end.
464 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
465 // humongous start there.
466 //
467 // The complication is potential non-movable regions during the scan. If such region is
468 // detected, then sliding restarts towards that non-movable region.
469
470 size_t to_begin = heap->num_regions();
471 size_t to_end = heap->num_regions();
472
473 for (size_t c = heap->num_regions(); c > 0; c--) {
474 ShenandoahHeapRegion *r = heap->get_region(c - 1);
475 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
476 // To-region candidate: record this, and continue scan
477 to_begin = r->index();
478 continue;
479 }
480
481 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
482 // From-region candidate: movable humongous region
483 oop old_obj = cast_to_oop(r->bottom());
484 size_t words_size = old_obj->size();
485 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
486
487 size_t start = to_end - num_regions;
488
489 if (start >= to_begin && start != r->index()) {
490 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
491 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
492 SlidingForwarding::forward_to<ALT_FWD>(old_obj, cast_to_oop(heap->get_region(start)->bottom()));
493 to_end = start;
494 continue;
495 }
496 }
497
498 // Failed to fit. Scan starting from current region.
499 to_begin = r->index();
500 to_end = r->index();
501 }
502 }
503
504 void ShenandoahFullGC::calculate_target_humongous_objects() {
505 if (UseAltGCForwarding) {
506 calculate_target_humongous_objects_impl<true>();
507 } else {
508 calculate_target_humongous_objects_impl<false>();
509 }
510 }
511
512 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
513 private:
514 ShenandoahHeap* const _heap;
515
516 public:
517 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
518 void heap_region_do(ShenandoahHeapRegion* r) {
519 if (r->is_trash()) {
520 r->recycle();
521 }
522 if (r->is_cset()) {
523 r->make_regular_bypass();
524 }
525 if (r->is_empty_uncommitted()) {
526 r->make_committed_bypass();
527 }
528 assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index());
529
530 // Record current region occupancy: this communicates empty regions are free
531 // to the rest of Full GC code.
532 r->set_new_top(r->top());
533 }
534 };
535
536 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
537 private:
538 ShenandoahHeap* const _heap;
539 ShenandoahMarkingContext* const _ctx;
540
541 public:
542 ShenandoahTrashImmediateGarbageClosure() :
543 _heap(ShenandoahHeap::heap()),
544 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
545
546 void heap_region_do(ShenandoahHeapRegion* r) {
547 if (r->is_humongous_start()) {
548 oop humongous_obj = cast_to_oop(r->bottom());
549 if (!_ctx->is_marked(humongous_obj)) {
550 assert(!r->has_live(),
551 "Region " SIZE_FORMAT " is not marked, should not have live", r->index());
552 _heap->trash_humongous_region_at(r);
553 } else {
554 assert(r->has_live(),
555 "Region " SIZE_FORMAT " should have live", r->index());
556 }
557 } else if (r->is_humongous_continuation()) {
558 // If we hit continuation, the non-live humongous starts should have been trashed already
559 assert(r->humongous_start_region()->has_live(),
560 "Region " SIZE_FORMAT " should have live", r->index());
561 } else if (r->is_regular()) {
562 if (!r->has_live()) {
563 r->make_trash_immediate();
564 }
565 }
566 }
567 };
568
569 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
570 ShenandoahHeap* heap = ShenandoahHeap::heap();
571
572 uint n_workers = heap->workers()->active_workers();
573 size_t n_regions = heap->num_regions();
574
575 // What we want to accomplish: have the dense prefix of data, while still balancing
576 // out the parallel work.
577 //
578 // Assuming the amount of work is driven by the live data that needs moving, we can slice
579 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
580 // thread takes all regions in its prefix subset, and then it takes some regions from
581 // the tail.
582 //
583 // Tail region selection becomes interesting.
584 //
585 // First, we want to distribute the regions fairly between the workers, and those regions
586 // might have different amount of live data. So, until we sure no workers need live data,
587 // we need to only take what the worker needs.
588 //
589 // Second, since we slide everything to the left in each slice, the most busy regions
590 // would be the ones on the left. Which means we want to have all workers have their after-tail
591 // regions as close to the left as possible.
592 //
593 // The easiest way to do this is to distribute after-tail regions in round-robin between
594 // workers that still need live data.
595 //
596 // Consider parallel workers A, B, C, then the target slice layout would be:
597 //
598 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
599 //
600 // (.....dense-prefix.....) (.....................tail...................)
601 // [all regions fully live] [left-most regions are fuller that right-most]
602 //
603
604 // Compute how much live data is there. This would approximate the size of dense prefix
605 // we target to create.
606 size_t total_live = 0;
607 for (size_t idx = 0; idx < n_regions; idx++) {
608 ShenandoahHeapRegion *r = heap->get_region(idx);
609 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
610 total_live += r->get_live_data_words();
611 }
612 }
613
614 // Estimate the size for the dense prefix. Note that we specifically count only the
615 // "full" regions, so there would be some non-full regions in the slice tail.
616 size_t live_per_worker = total_live / n_workers;
617 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
618 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
619 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
620 assert(prefix_regions_total <= n_regions, "Sanity");
621
622 // There might be non-candidate regions in the prefix. To compute where the tail actually
623 // ends up being, we need to account those as well.
624 size_t prefix_end = prefix_regions_total;
625 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
626 ShenandoahHeapRegion *r = heap->get_region(idx);
627 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
628 prefix_end++;
629 }
630 }
631 prefix_end = MIN2(prefix_end, n_regions);
632 assert(prefix_end <= n_regions, "Sanity");
633
634 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
635 // subset of dense prefix.
636 size_t prefix_idx = 0;
637
638 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
639
640 for (size_t wid = 0; wid < n_workers; wid++) {
641 ShenandoahHeapRegionSet* slice = worker_slices[wid];
642
643 live[wid] = 0;
644 size_t regs = 0;
645
646 // Add all prefix regions for this worker
647 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
648 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
649 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
650 slice->add_region(r);
651 live[wid] += r->get_live_data_words();
652 regs++;
653 }
654 prefix_idx++;
655 }
656 }
657
658 // Distribute the tail among workers in round-robin fashion.
659 size_t wid = n_workers - 1;
660
661 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
662 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
663 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
664 assert(wid < n_workers, "Sanity");
665
666 size_t live_region = r->get_live_data_words();
667
668 // Select next worker that still needs live data.
669 size_t old_wid = wid;
670 do {
671 wid++;
672 if (wid == n_workers) wid = 0;
673 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
674
675 if (old_wid == wid) {
676 // Circled back to the same worker? This means liveness data was
677 // miscalculated. Bump the live_per_worker limit so that
678 // everyone gets a piece of the leftover work.
679 live_per_worker += ShenandoahHeapRegion::region_size_words();
680 }
681
682 worker_slices[wid]->add_region(r);
683 live[wid] += live_region;
684 }
685 }
686
687 FREE_C_HEAP_ARRAY(size_t, live);
688
689 #ifdef ASSERT
690 ResourceBitMap map(n_regions);
691 for (size_t wid = 0; wid < n_workers; wid++) {
692 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
693 ShenandoahHeapRegion* r = it.next();
694 while (r != NULL) {
695 size_t idx = r->index();
696 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx);
697 assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx);
698 map.at_put(idx, true);
699 r = it.next();
700 }
701 }
702
703 for (size_t rid = 0; rid < n_regions; rid++) {
704 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
705 bool is_distributed = map.at(rid);
706 assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid);
707 }
708 #endif
709 }
710
711 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
712 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
713 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
714
715 ShenandoahHeap* heap = ShenandoahHeap::heap();
716
717 // About to figure out which regions can be compacted, make sure pinning status
718 // had been updated in GC prologue.
719 heap->assert_pinned_region_status();
720
721 {
722 // Trash the immediately collectible regions before computing addresses
723 ShenandoahTrashImmediateGarbageClosure tigcl;
724 heap->heap_region_iterate(&tigcl);
725
726 // Make sure regions are in good state: committed, active, clean.
727 // This is needed because we are potentially sliding the data through them.
728 ShenandoahEnsureHeapActiveClosure ecl;
729 heap->heap_region_iterate(&ecl);
730 }
731
732 // Compute the new addresses for regular objects
733 {
734 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
735
736 distribute_slices(worker_slices);
737
738 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
739 heap->workers()->run_task(&task);
740 }
741
742 // Compute the new addresses for humongous objects
743 {
744 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
745 calculate_target_humongous_objects();
746 }
747 }
748
749 template <bool ALT_FWD>
750 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
751 private:
752 ShenandoahHeap* const _heap;
753 ShenandoahMarkingContext* const _ctx;
754
755 template <class T>
756 inline void do_oop_work(T* p) {
757 T o = RawAccess<>::oop_load(p);
758 if (!CompressedOops::is_null(o)) {
759 oop obj = CompressedOops::decode_not_null(o);
760 assert(_ctx->is_marked(obj), "must be marked");
761 if (SlidingForwarding::is_forwarded(obj)) {
762 oop forw = SlidingForwarding::forwardee<ALT_FWD>(obj);
763 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
764 }
765 }
766 }
767
768 public:
769 ShenandoahAdjustPointersClosure() :
770 _heap(ShenandoahHeap::heap()),
771 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
772
773 void do_oop(oop* p) { do_oop_work(p); }
774 void do_oop(narrowOop* p) { do_oop_work(p); }
775 };
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 AbstractGangTask {
794 private:
795 ShenandoahHeap* const _heap;
796 ShenandoahRegionIterator _regions;
797
798 public:
799 ShenandoahAdjustPointersTask() :
800 AbstractGangTask("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 != NULL) {
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 AbstractGangTask {
829 private:
830 ShenandoahRootAdjuster* _rp;
831 PreservedMarksSet* _preserved_marks;
832
833 public:
834 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
835 AbstractGangTask("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 WorkGang* 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 = (size_t)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 new_obj->init_mark();
901 }
902 }
903 };
904
905 class ShenandoahCompactObjectsTask : public AbstractGangTask {
906 private:
907 ShenandoahHeap* const _heap;
908 ShenandoahHeapRegionSet** const _worker_slices;
909
910 public:
911 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
912 AbstractGangTask("Shenandoah Compact Objects"),
913 _heap(ShenandoahHeap::heap()),
914 _worker_slices(worker_slices) {
915 }
916
917 private:
918 template <bool ALT_FWD>
919 void work_impl(uint worker_id) {
920 ShenandoahParallelWorkerSession worker_session(worker_id);
921 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
922
923 ShenandoahCompactObjectsClosure<ALT_FWD> cl(worker_id);
924 ShenandoahHeapRegion* r = slice.next();
925 while (r != NULL) {
926 assert(!r->is_humongous(), "must not get humongous regions here");
927 if (r->has_live()) {
928 _heap->marked_object_iterate(r, &cl);
929 }
930 r->set_top(r->new_top());
931 r = slice.next();
932 }
933 }
934
935 public:
936 void work(uint worker_id) {
937 if (UseAltGCForwarding) {
938 work_impl<true>(worker_id);
939 } else {
940 work_impl<false>(worker_id);
941 }
942 }
943 };
944
945 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
946 private:
947 ShenandoahHeap* const _heap;
948 size_t _live;
949
950 public:
951 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
952 _heap->free_set()->clear();
953 }
954
955 void heap_region_do(ShenandoahHeapRegion* r) {
956 assert (!r->is_cset(), "cset regions should have been demoted already");
957
958 // Need to reset the complete-top-at-mark-start pointer here because
959 // the complete marking bitmap is no longer valid. This ensures
960 // size-based iteration in marked_object_iterate().
961 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
962 // pinned regions.
963 if (!r->is_pinned()) {
964 _heap->complete_marking_context()->reset_top_at_mark_start(r);
965 }
966
967 size_t live = r->used();
968
969 // Make empty regions that have been allocated into regular
970 if (r->is_empty() && live > 0) {
971 r->make_regular_bypass();
972 }
973
974 // Reclaim regular regions that became empty
975 if (r->is_regular() && live == 0) {
976 r->make_trash();
977 }
978
979 // Recycle all trash regions
980 if (r->is_trash()) {
981 live = 0;
982 r->recycle();
983 }
984
985 r->set_live_data(live);
986 r->reset_alloc_metadata();
987 _live += live;
988 }
989
990 size_t get_live() {
991 return _live;
992 }
993 };
994
995 template <bool ALT_FWD>
996 void ShenandoahFullGC::compact_humongous_objects_impl() {
997 // Compact humongous regions, based on their fwdptr objects.
998 //
999 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
1000 // humongous regions are already compacted, and do not require further moves, which alleviates
1001 // sliding costs. We may consider doing this in parallel in future.
1002
1003 ShenandoahHeap* heap = ShenandoahHeap::heap();
1004
1005 for (size_t c = heap->num_regions(); c > 0; c--) {
1006 ShenandoahHeapRegion* r = heap->get_region(c - 1);
1007 if (r->is_humongous_start()) {
1008 oop old_obj = cast_to_oop(r->bottom());
1009 if (SlidingForwarding::is_not_forwarded(old_obj)) {
1010 // No need to move the object, it stays at the same slot
1011 continue;
1012 }
1013 size_t words_size = old_obj->size();
1014 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
1015
1016 size_t old_start = r->index();
1017 size_t old_end = old_start + num_regions - 1;
1018 size_t new_start = heap->heap_region_index_containing(SlidingForwarding::forwardee<ALT_FWD>(old_obj));
1019 size_t new_end = new_start + num_regions - 1;
1020 assert(old_start != new_start, "must be real move");
1021 assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
1022
1023 Copy::aligned_conjoint_words(heap->get_region(old_start)->bottom(),
1024 heap->get_region(new_start)->bottom(),
1025 words_size);
1026
1027 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
1028 new_obj->init_mark();
1029
1030 {
1031 for (size_t c = old_start; c <= old_end; c++) {
1032 ShenandoahHeapRegion* r = heap->get_region(c);
1033 r->make_regular_bypass();
1034 r->set_top(r->bottom());
1035 }
1036
1037 for (size_t c = new_start; c <= new_end; c++) {
1038 ShenandoahHeapRegion* r = heap->get_region(c);
1039 if (c == new_start) {
1040 r->make_humongous_start_bypass();
1041 } else {
1042 r->make_humongous_cont_bypass();
1043 }
1044
1045 // Trailing region may be non-full, record the remainder there
1046 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1047 if ((c == new_end) && (remainder != 0)) {
1048 r->set_top(r->bottom() + remainder);
1049 } else {
1050 r->set_top(r->end());
1051 }
1052
1053 r->reset_alloc_metadata();
1054 }
1055 }
1056 }
1057 }
1058 }
1059
1060 void ShenandoahFullGC::compact_humongous_objects() {
1061 if (UseAltGCForwarding) {
1062 compact_humongous_objects_impl<true>();
1063 } else {
1064 compact_humongous_objects_impl<false>();
1065 }
1066 }
1067
1068 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1069 // we need to remain able to walk pinned regions.
1070 // Since pinned region do not move and don't get compacted, we will get holes with
1071 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1072 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1073 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1074 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1075 class ShenandoahMCResetCompleteBitmapTask : public AbstractGangTask {
1076 private:
1077 ShenandoahRegionIterator _regions;
1078
1079 public:
1080 ShenandoahMCResetCompleteBitmapTask() :
1081 AbstractGangTask("Shenandoah Reset Bitmap") {
1082 }
1083
1084 void work(uint worker_id) {
1085 ShenandoahParallelWorkerSession worker_session(worker_id);
1086 ShenandoahHeapRegion* region = _regions.next();
1087 ShenandoahHeap* heap = ShenandoahHeap::heap();
1088 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1089 while (region != NULL) {
1090 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1091 ctx->clear_bitmap(region);
1092 }
1093 region = _regions.next();
1094 }
1095 }
1096 };
1097
1098 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1099 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1100 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1101
1102 ShenandoahHeap* heap = ShenandoahHeap::heap();
1103
1104 // Compact regular objects first
1105 {
1106 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1107 ShenandoahCompactObjectsTask compact_task(worker_slices);
1108 heap->workers()->run_task(&compact_task);
1109 }
1110
1111 // Compact humongous objects after regular object moves
1112 {
1113 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1114 compact_humongous_objects();
1115 }
1116
1117 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1118 // and must ensure the bitmap is in sync.
1119 {
1120 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1121 ShenandoahMCResetCompleteBitmapTask task;
1122 heap->workers()->run_task(&task);
1123 }
1124
1125 // Bring regions in proper states after the collection, and set heap properties.
1126 {
1127 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1128
1129 ShenandoahPostCompactClosure post_compact;
1130 heap->heap_region_iterate(&post_compact);
1131 heap->set_used(post_compact.get_live());
1132
1133 heap->collection_set()->clear();
1134 heap->free_set()->rebuild();
1135 }
1136
1137 heap->clear_cancelled_gc();
1138 }