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/tlab_globals.hpp"
31 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
32 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
33 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
34 #include "gc/shenandoah/shenandoahFreeSet.hpp"
35 #include "gc/shenandoah/shenandoahFullGC.hpp"
36 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
37 #include "gc/shenandoah/shenandoahMark.inline.hpp"
38 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
39 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
40 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
41 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
42 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
43 #include "gc/shenandoah/shenandoahMetrics.hpp"
44 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp"
45 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
46 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
47 #include "gc/shenandoah/shenandoahSTWMark.hpp"
48 #include "gc/shenandoah/shenandoahUtils.hpp"
49 #include "gc/shenandoah/shenandoahVerifier.hpp"
50 #include "gc/shenandoah/shenandoahVMOperations.hpp"
51 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
52 #include "memory/metaspaceUtils.hpp"
53 #include "memory/universe.hpp"
54 #include "oops/compressedOops.inline.hpp"
55 #include "oops/oop.inline.hpp"
56 #include "runtime/biasedLocking.hpp"
57 #include "runtime/orderAccess.hpp"
58 #include "runtime/thread.hpp"
59 #include "runtime/vmThread.hpp"
60 #include "utilities/copy.hpp"
61 #include "utilities/events.hpp"
62 #include "utilities/growableArray.hpp"
63 #include "gc/shared/workgroup.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 BiasedLocking::preserve_marks();
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 phase2_calculate_target_addresses(worker_slices);
226
227 OrderAccess::fence();
228
229 phase3_update_references();
230
231 phase4_compact_objects(worker_slices);
232 }
233
234 {
235 // Epilogue
236 _preserved_marks->restore(heap->workers());
237 BiasedLocking::restore_marks();
238 _preserved_marks->reclaim();
239 }
240
241 // Resize metaspace
242 MetaspaceGC::compute_new_size();
243
244 // Free worker slices
245 for (uint i = 0; i < heap->max_workers(); i++) {
246 delete worker_slices[i];
247 }
248 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
249
250 heap->set_full_gc_move_in_progress(false);
251 heap->set_full_gc_in_progress(false);
252
253 if (ShenandoahVerify) {
254 heap->verifier()->verify_after_fullgc();
255 }
256
257 if (VerifyAfterGC) {
258 Universe::verify();
259 }
260
261 {
262 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
263 heap->post_full_gc_dump(_gc_timer);
264 }
265 }
266
267 class ShenandoahPrepareForMarkClosure: public ShenandoahHeapRegionClosure {
268 private:
269 ShenandoahMarkingContext* const _ctx;
270
271 public:
272 ShenandoahPrepareForMarkClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
273
274 void heap_region_do(ShenandoahHeapRegion *r) {
275 _ctx->capture_top_at_mark_start(r);
276 r->clear_live_data();
277 }
278 };
279
280 void ShenandoahFullGC::phase1_mark_heap() {
281 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
282 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
283
284 ShenandoahHeap* heap = ShenandoahHeap::heap();
285
286 ShenandoahPrepareForMarkClosure cl;
287 heap->heap_region_iterate(&cl);
288
289 heap->set_unload_classes(heap->heuristics()->can_unload_classes());
290
291 ShenandoahReferenceProcessor* rp = heap->ref_processor();
292 // enable ("weak") refs discovery
293 rp->set_soft_reference_policy(true); // forcefully purge all soft references
294
295 ShenandoahSTWMark mark(true /*full_gc*/);
296 mark.mark();
297 heap->parallel_cleaning(true /* full_gc */);
298 }
299
300 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
301 private:
302 PreservedMarks* const _preserved_marks;
303 ShenandoahHeap* const _heap;
304 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
305 int _empty_regions_pos;
306 ShenandoahHeapRegion* _to_region;
307 ShenandoahHeapRegion* _from_region;
308 HeapWord* _compact_point;
309
310 public:
311 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
312 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
313 ShenandoahHeapRegion* to_region) :
314 _preserved_marks(preserved_marks),
315 _heap(ShenandoahHeap::heap()),
316 _empty_regions(empty_regions),
317 _empty_regions_pos(0),
318 _to_region(to_region),
319 _from_region(NULL),
320 _compact_point(to_region->bottom()) {}
321
322 void set_from_region(ShenandoahHeapRegion* from_region) {
323 _from_region = from_region;
324 }
325
326 void finish_region() {
327 assert(_to_region != NULL, "should not happen");
328 _to_region->set_new_top(_compact_point);
329 }
330
331 bool is_compact_same_region() {
332 return _from_region == _to_region;
333 }
334
335 int empty_regions_pos() {
336 return _empty_regions_pos;
337 }
338
339 void do_object(oop p) {
340 assert(_from_region != NULL, "must set before work");
341 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
342 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
343
344 size_t obj_size = p->size();
345 if (_compact_point + obj_size > _to_region->end()) {
346 finish_region();
347
348 // Object doesn't fit. Pick next empty region and start compacting there.
349 ShenandoahHeapRegion* new_to_region;
350 if (_empty_regions_pos < _empty_regions.length()) {
351 new_to_region = _empty_regions.at(_empty_regions_pos);
352 _empty_regions_pos++;
353 } else {
354 // Out of empty region? Compact within the same region.
355 new_to_region = _from_region;
356 }
357
358 assert(new_to_region != _to_region, "must not reuse same to-region");
359 assert(new_to_region != NULL, "must not be NULL");
360 _to_region = new_to_region;
361 _compact_point = _to_region->bottom();
362 }
363
364 // Object fits into current region, record new location:
365 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
366 shenandoah_assert_not_forwarded(NULL, p);
367 _preserved_marks->push_if_necessary(p, p->mark());
368 p->forward_to(cast_to_oop(_compact_point));
369 _compact_point += obj_size;
370 }
371 };
372
373 class ShenandoahPrepareForCompactionTask : public AbstractGangTask {
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 AbstractGangTask("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 == NULL) {
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 != NULL) {
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 != NULL) {
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 };
751
752 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
753 private:
754 ShenandoahHeap* const _heap;
755 ShenandoahAdjustPointersClosure _cl;
756
757 public:
758 ShenandoahAdjustPointersObjectClosure() :
759 _heap(ShenandoahHeap::heap()) {
760 }
761 void do_object(oop p) {
762 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
763 p->oop_iterate(&_cl);
764 }
765 };
766
767 class ShenandoahAdjustPointersTask : public AbstractGangTask {
768 private:
769 ShenandoahHeap* const _heap;
770 ShenandoahRegionIterator _regions;
771
772 public:
773 ShenandoahAdjustPointersTask() :
774 AbstractGangTask("Shenandoah Adjust Pointers"),
775 _heap(ShenandoahHeap::heap()) {
776 }
777
778 void work(uint worker_id) {
779 ShenandoahParallelWorkerSession worker_session(worker_id);
780 ShenandoahAdjustPointersObjectClosure obj_cl;
781 ShenandoahHeapRegion* r = _regions.next();
782 while (r != NULL) {
783 if (!r->is_humongous_continuation() && r->has_live()) {
784 _heap->marked_object_iterate(r, &obj_cl);
785 }
786 r = _regions.next();
787 }
788 }
789 };
790
791 class ShenandoahAdjustRootPointersTask : public AbstractGangTask {
792 private:
793 ShenandoahRootAdjuster* _rp;
794 PreservedMarksSet* _preserved_marks;
795 public:
796 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
797 AbstractGangTask("Shenandoah Adjust Root Pointers"),
798 _rp(rp),
799 _preserved_marks(preserved_marks) {}
800
801 void work(uint worker_id) {
802 ShenandoahParallelWorkerSession worker_session(worker_id);
803 ShenandoahAdjustPointersClosure cl;
804 _rp->roots_do(worker_id, &cl);
805 _preserved_marks->get(worker_id)->adjust_during_full_gc();
806 }
807 };
808
809 void ShenandoahFullGC::phase3_update_references() {
810 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
811 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
812
813 ShenandoahHeap* heap = ShenandoahHeap::heap();
814
815 WorkGang* workers = heap->workers();
816 uint nworkers = workers->active_workers();
817 {
818 #if COMPILER2_OR_JVMCI
819 DerivedPointerTable::clear();
820 #endif
821 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
822 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
823 workers->run_task(&task);
824 #if COMPILER2_OR_JVMCI
825 DerivedPointerTable::update_pointers();
826 #endif
827 }
828
829 ShenandoahAdjustPointersTask adjust_pointers_task;
830 workers->run_task(&adjust_pointers_task);
831 }
832
833 class ShenandoahCompactObjectsClosure : public ObjectClosure {
834 private:
835 ShenandoahHeap* const _heap;
836 uint const _worker_id;
837
838 public:
839 ShenandoahCompactObjectsClosure(uint worker_id) :
840 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
841
842 void do_object(oop p) {
843 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
844 size_t size = (size_t)p->size();
845 if (p->is_forwarded()) {
846 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
847 HeapWord* compact_to = cast_from_oop<HeapWord*>(p->forwardee());
848 Copy::aligned_conjoint_words(compact_from, compact_to, size);
849 oop new_obj = cast_to_oop(compact_to);
850 new_obj->init_mark();
851 }
852 }
853 };
854
855 class ShenandoahCompactObjectsTask : public AbstractGangTask {
856 private:
857 ShenandoahHeap* const _heap;
858 ShenandoahHeapRegionSet** const _worker_slices;
859
860 public:
861 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
862 AbstractGangTask("Shenandoah Compact Objects"),
863 _heap(ShenandoahHeap::heap()),
864 _worker_slices(worker_slices) {
865 }
866
867 void work(uint worker_id) {
868 ShenandoahParallelWorkerSession worker_session(worker_id);
869 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
870
871 ShenandoahCompactObjectsClosure cl(worker_id);
872 ShenandoahHeapRegion* r = slice.next();
873 while (r != NULL) {
874 assert(!r->is_humongous(), "must not get humongous regions here");
875 if (r->has_live()) {
876 _heap->marked_object_iterate(r, &cl);
877 }
878 r->set_top(r->new_top());
879 r = slice.next();
880 }
881 }
882 };
883
884 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
885 private:
886 ShenandoahHeap* const _heap;
887 size_t _live;
888
889 public:
890 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
891 _heap->free_set()->clear();
892 }
893
894 void heap_region_do(ShenandoahHeapRegion* r) {
895 assert (!r->is_cset(), "cset regions should have been demoted already");
896
897 // Need to reset the complete-top-at-mark-start pointer here because
898 // the complete marking bitmap is no longer valid. This ensures
899 // size-based iteration in marked_object_iterate().
900 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
901 // pinned regions.
902 if (!r->is_pinned()) {
903 _heap->complete_marking_context()->reset_top_at_mark_start(r);
904 }
905
906 size_t live = r->used();
907
908 // Make empty regions that have been allocated into regular
909 if (r->is_empty() && live > 0) {
910 r->make_regular_bypass();
911 if (ZapUnusedHeapArea) {
912 SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
913 }
914 }
915
916 // Reclaim regular regions that became empty
917 if (r->is_regular() && live == 0) {
918 r->make_trash();
919 }
920
921 // Recycle all trash regions
922 if (r->is_trash()) {
923 live = 0;
924 r->recycle();
925 }
926
927 r->set_live_data(live);
928 r->reset_alloc_metadata();
929 _live += live;
930 }
931
932 size_t get_live() {
933 return _live;
934 }
935 };
936
937 void ShenandoahFullGC::compact_humongous_objects() {
938 // Compact humongous regions, based on their fwdptr objects.
939 //
940 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
941 // humongous regions are already compacted, and do not require further moves, which alleviates
942 // sliding costs. We may consider doing this in parallel in future.
943
944 ShenandoahHeap* heap = ShenandoahHeap::heap();
945
946 for (size_t c = heap->num_regions(); c > 0; c--) {
947 ShenandoahHeapRegion* r = heap->get_region(c - 1);
948 if (r->is_humongous_start()) {
949 oop old_obj = cast_to_oop(r->bottom());
950 if (!old_obj->is_forwarded()) {
951 // No need to move the object, it stays at the same slot
952 continue;
953 }
954 size_t words_size = old_obj->size();
955 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
956
957 size_t old_start = r->index();
958 size_t old_end = old_start + num_regions - 1;
959 size_t new_start = heap->heap_region_index_containing(old_obj->forwardee());
960 size_t new_end = new_start + num_regions - 1;
961 assert(old_start != new_start, "must be real move");
962 assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
963
964 Copy::aligned_conjoint_words(heap->get_region(old_start)->bottom(),
965 heap->get_region(new_start)->bottom(),
966 words_size);
967
968 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
969 new_obj->init_mark();
970
971 {
972 for (size_t c = old_start; c <= old_end; c++) {
973 ShenandoahHeapRegion* r = heap->get_region(c);
974 r->make_regular_bypass();
975 r->set_top(r->bottom());
976 }
977
978 for (size_t c = new_start; c <= new_end; c++) {
979 ShenandoahHeapRegion* r = heap->get_region(c);
980 if (c == new_start) {
981 r->make_humongous_start_bypass();
982 } else {
983 r->make_humongous_cont_bypass();
984 }
985
986 // Trailing region may be non-full, record the remainder there
987 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
988 if ((c == new_end) && (remainder != 0)) {
989 r->set_top(r->bottom() + remainder);
990 } else {
991 r->set_top(r->end());
992 }
993
994 r->reset_alloc_metadata();
995 }
996 }
997 }
998 }
999 }
1000
1001 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1002 // we need to remain able to walk pinned regions.
1003 // Since pinned region do not move and don't get compacted, we will get holes with
1004 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1005 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1006 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1007 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1008 class ShenandoahMCResetCompleteBitmapTask : public AbstractGangTask {
1009 private:
1010 ShenandoahRegionIterator _regions;
1011
1012 public:
1013 ShenandoahMCResetCompleteBitmapTask() :
1014 AbstractGangTask("Shenandoah Reset Bitmap") {
1015 }
1016
1017 void work(uint worker_id) {
1018 ShenandoahParallelWorkerSession worker_session(worker_id);
1019 ShenandoahHeapRegion* region = _regions.next();
1020 ShenandoahHeap* heap = ShenandoahHeap::heap();
1021 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1022 while (region != NULL) {
1023 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1024 ctx->clear_bitmap(region);
1025 }
1026 region = _regions.next();
1027 }
1028 }
1029 };
1030
1031 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1032 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1033 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1034
1035 ShenandoahHeap* heap = ShenandoahHeap::heap();
1036
1037 // Compact regular objects first
1038 {
1039 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1040 ShenandoahCompactObjectsTask compact_task(worker_slices);
1041 heap->workers()->run_task(&compact_task);
1042 }
1043
1044 // Compact humongous objects after regular object moves
1045 {
1046 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1047 compact_humongous_objects();
1048 }
1049
1050 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1051 // and must ensure the bitmap is in sync.
1052 {
1053 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1054 ShenandoahMCResetCompleteBitmapTask task;
1055 heap->workers()->run_task(&task);
1056 }
1057
1058 // Bring regions in proper states after the collection, and set heap properties.
1059 {
1060 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1061
1062 ShenandoahPostCompactClosure post_compact;
1063 heap->heap_region_iterate(&post_compact);
1064 heap->set_used(post_compact.get_live());
1065
1066 heap->collection_set()->clear();
1067 heap->free_set()->rebuild();
1068 }
1069
1070 heap->clear_cancelled_gc();
1071 }