1 /*
2 * Copyright (c) 2014, 2020, 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 "code/codeCache.hpp"
28 #include "gc/shared/gcTraceTime.inline.hpp"
29 #include "gc/shared/preservedMarks.inline.hpp"
30 #include "gc/shenandoah/shenandoahForwarding.inline.hpp"
31 #include "gc/shenandoah/shenandoahConcurrentMark.inline.hpp"
32 #include "gc/shenandoah/shenandoahConcurrentRoots.hpp"
33 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
34 #include "gc/shenandoah/shenandoahFreeSet.hpp"
35 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
36 #include "gc/shenandoah/shenandoahMarkCompact.hpp"
37 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
38 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
39 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
40 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
41 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
42 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
43 #include "gc/shenandoah/shenandoahTaskqueue.inline.hpp"
44 #include "gc/shenandoah/shenandoahUtils.hpp"
45 #include "gc/shenandoah/shenandoahVerifier.hpp"
46 #include "gc/shenandoah/shenandoahVMOperations.hpp"
47 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
48 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
49 #include "memory/metaspace.hpp"
50 #include "memory/universe.hpp"
51 #include "oops/compressedOops.inline.hpp"
52 #include "oops/oop.inline.hpp"
53 #include "runtime/biasedLocking.hpp"
54 #include "runtime/orderAccess.hpp"
55 #include "runtime/thread.hpp"
56 #include "utilities/copy.hpp"
57 #include "utilities/growableArray.hpp"
58 #include "gc/shared/workgroup.hpp"
59
ShenandoahMarkCompact()60 ShenandoahMarkCompact::ShenandoahMarkCompact() :
61 _gc_timer(NULL),
62 _preserved_marks(new PreservedMarksSet(true)) {}
63
initialize(GCTimer * gc_timer)64 void ShenandoahMarkCompact::initialize(GCTimer* gc_timer) {
65 _gc_timer = gc_timer;
66 }
67
do_it(GCCause::Cause gc_cause)68 void ShenandoahMarkCompact::do_it(GCCause::Cause gc_cause) {
69 ShenandoahHeap* heap = ShenandoahHeap::heap();
70
71 if (ShenandoahVerify) {
72 heap->verifier()->verify_before_fullgc();
73 }
74
75 if (VerifyBeforeGC) {
76 Universe::verify();
77 }
78
79 // Degenerated GC may carry concurrent root flags when upgrading to
80 // full GC. We need to reset it before mutators resume.
81 if (ShenandoahConcurrentRoots::can_do_concurrent_class_unloading()) {
82 heap->set_concurrent_strong_root_in_progress(false);
83 heap->set_concurrent_weak_root_in_progress(false);
84 }
85
86 heap->set_full_gc_in_progress(true);
87
88 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
89 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
90
91 {
92 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
93 heap->pre_full_gc_dump(_gc_timer);
94 }
95
96 {
97 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
98 // Full GC is supposed to recover from any GC state:
99
100 // a0. Remember if we have forwarded objects
101 bool has_forwarded_objects = heap->has_forwarded_objects();
102
103 // a1. Cancel evacuation, if in progress
104 if (heap->is_evacuation_in_progress()) {
105 heap->set_evacuation_in_progress(false);
106 }
107 assert(!heap->is_evacuation_in_progress(), "sanity");
108
109 // a2. Cancel update-refs, if in progress
110 if (heap->is_update_refs_in_progress()) {
111 heap->set_update_refs_in_progress(false);
112 }
113 assert(!heap->is_update_refs_in_progress(), "sanity");
114
115 // b. Cancel concurrent mark, if in progress
116 if (heap->is_concurrent_mark_in_progress()) {
117 heap->concurrent_mark()->cancel();
118 heap->set_concurrent_mark_in_progress(false);
119 }
120 assert(!heap->is_concurrent_mark_in_progress(), "sanity");
121
122 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
123 if (has_forwarded_objects) {
124 heap->concurrent_mark()->update_roots(ShenandoahPhaseTimings::full_gc_update_roots);
125 }
126
127 // d. Reset the bitmaps for new marking
128 heap->reset_mark_bitmap();
129 assert(heap->marking_context()->is_bitmap_clear(), "sanity");
130 assert(!heap->marking_context()->is_complete(), "sanity");
131
132 // e. Abandon reference discovery and clear all discovered references.
133 ShenandoahReferenceProcessor* rp = heap->ref_processor();
134 rp->abandon_partial_discovery();
135
136 // f. Set back forwarded objects bit back, in case some steps above dropped it.
137 heap->set_has_forwarded_objects(has_forwarded_objects);
138
139 // g. Sync pinned region status from the CP marks
140 heap->sync_pinned_region_status();
141
142 // The rest of prologue:
143 BiasedLocking::preserve_marks();
144 _preserved_marks->init(heap->workers()->active_workers());
145 }
146
147 if (UseTLAB) {
148 heap->gclabs_retire(ResizeTLAB);
149 heap->tlabs_retire(ResizeTLAB);
150 }
151
152 OrderAccess::fence();
153
154 phase1_mark_heap();
155
156 // Once marking is done, which may have fixed up forwarded objects, we can drop it.
157 // Coming out of Full GC, we would not have any forwarded objects.
158 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
159 heap->set_has_forwarded_objects(false);
160
161 heap->set_full_gc_move_in_progress(true);
162
163 // Setup workers for the rest
164 OrderAccess::fence();
165
166 // Initialize worker slices
167 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
168 for (uint i = 0; i < heap->max_workers(); i++) {
169 worker_slices[i] = new ShenandoahHeapRegionSet();
170 }
171
172 {
173 // The rest of code performs region moves, where region status is undefined
174 // until all phases run together.
175 ShenandoahHeapLocker lock(heap->lock());
176
177 phase2_calculate_target_addresses(worker_slices);
178
179 OrderAccess::fence();
180
181 phase3_update_references();
182
183 phase4_compact_objects(worker_slices);
184 }
185
186 {
187 // Epilogue
188 _preserved_marks->restore(heap->workers());
189 BiasedLocking::restore_marks();
190 _preserved_marks->reclaim();
191 }
192
193 // Resize metaspace
194 MetaspaceGC::compute_new_size();
195
196 // Free worker slices
197 for (uint i = 0; i < heap->max_workers(); i++) {
198 delete worker_slices[i];
199 }
200 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
201
202 heap->set_full_gc_move_in_progress(false);
203 heap->set_full_gc_in_progress(false);
204
205 if (ShenandoahVerify) {
206 heap->verifier()->verify_after_fullgc();
207 }
208
209 if (VerifyAfterGC) {
210 Universe::verify();
211 }
212
213 {
214 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
215 heap->post_full_gc_dump(_gc_timer);
216 }
217 }
218
219 class ShenandoahPrepareForMarkClosure: public ShenandoahHeapRegionClosure {
220 private:
221 ShenandoahMarkingContext* const _ctx;
222
223 public:
ShenandoahPrepareForMarkClosure()224 ShenandoahPrepareForMarkClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
225
heap_region_do(ShenandoahHeapRegion * r)226 void heap_region_do(ShenandoahHeapRegion *r) {
227 _ctx->capture_top_at_mark_start(r);
228 r->clear_live_data();
229 }
230 };
231
phase1_mark_heap()232 void ShenandoahMarkCompact::phase1_mark_heap() {
233 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
234 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
235
236 ShenandoahHeap* heap = ShenandoahHeap::heap();
237
238 ShenandoahPrepareForMarkClosure cl;
239 heap->heap_region_iterate(&cl);
240
241 ShenandoahConcurrentMark* cm = heap->concurrent_mark();
242
243 heap->set_unload_classes(heap->heuristics()->can_unload_classes());
244
245 ShenandoahReferenceProcessor* rp = heap->ref_processor();
246 // enable ("weak") refs discovery
247 rp->set_soft_reference_policy(true); // forcefully purge all soft references
248
249 cm->mark_roots(ShenandoahPhaseTimings::full_gc_scan_roots);
250 cm->finish_mark_from_roots(/* full_gc = */ true);
251 heap->mark_complete_marking_context();
252 heap->parallel_cleaning(true /* full_gc */);
253 }
254
255 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
256 private:
257 PreservedMarks* const _preserved_marks;
258 ShenandoahHeap* const _heap;
259 GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
260 int _empty_regions_pos;
261 ShenandoahHeapRegion* _to_region;
262 ShenandoahHeapRegion* _from_region;
263 HeapWord* _compact_point;
264
265 public:
ShenandoahPrepareForCompactionObjectClosure(PreservedMarks * preserved_marks,GrowableArray<ShenandoahHeapRegion * > & empty_regions,ShenandoahHeapRegion * to_region)266 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
267 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
268 ShenandoahHeapRegion* to_region) :
269 _preserved_marks(preserved_marks),
270 _heap(ShenandoahHeap::heap()),
271 _empty_regions(empty_regions),
272 _empty_regions_pos(0),
273 _to_region(to_region),
274 _from_region(NULL),
275 _compact_point(to_region->bottom()) {}
276
set_from_region(ShenandoahHeapRegion * from_region)277 void set_from_region(ShenandoahHeapRegion* from_region) {
278 _from_region = from_region;
279 }
280
finish_region()281 void finish_region() {
282 assert(_to_region != NULL, "should not happen");
283 _to_region->set_new_top(_compact_point);
284 }
285
is_compact_same_region()286 bool is_compact_same_region() {
287 return _from_region == _to_region;
288 }
289
empty_regions_pos()290 int empty_regions_pos() {
291 return _empty_regions_pos;
292 }
293
do_object(oop p)294 void do_object(oop p) {
295 assert(_from_region != NULL, "must set before work");
296 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
297 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
298
299 size_t obj_size = p->size();
300 if (_compact_point + obj_size > _to_region->end()) {
301 finish_region();
302
303 // Object doesn't fit. Pick next empty region and start compacting there.
304 ShenandoahHeapRegion* new_to_region;
305 if (_empty_regions_pos < _empty_regions.length()) {
306 new_to_region = _empty_regions.at(_empty_regions_pos);
307 _empty_regions_pos++;
308 } else {
309 // Out of empty region? Compact within the same region.
310 new_to_region = _from_region;
311 }
312
313 assert(new_to_region != _to_region, "must not reuse same to-region");
314 assert(new_to_region != NULL, "must not be NULL");
315 _to_region = new_to_region;
316 _compact_point = _to_region->bottom();
317 }
318
319 // Object fits into current region, record new location:
320 assert(_compact_point + obj_size <= _to_region->end(), "must fit");
321 shenandoah_assert_not_forwarded(NULL, p);
322 _preserved_marks->push_if_necessary(p, p->mark());
323 p->forward_to(oop(_compact_point));
324 _compact_point += obj_size;
325 }
326 };
327
328 class ShenandoahPrepareForCompactionTask : public AbstractGangTask {
329 private:
330 PreservedMarksSet* const _preserved_marks;
331 ShenandoahHeap* const _heap;
332 ShenandoahHeapRegionSet** const _worker_slices;
333
334 public:
ShenandoahPrepareForCompactionTask(PreservedMarksSet * preserved_marks,ShenandoahHeapRegionSet ** worker_slices)335 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
336 AbstractGangTask("Shenandoah Prepare For Compaction"),
337 _preserved_marks(preserved_marks),
338 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
339 }
340
is_candidate_region(ShenandoahHeapRegion * r)341 static bool is_candidate_region(ShenandoahHeapRegion* r) {
342 // Empty region: get it into the slice to defragment the slice itself.
343 // We could have skipped this without violating correctness, but we really
344 // want to compact all live regions to the start of the heap, which sometimes
345 // means moving them into the fully empty regions.
346 if (r->is_empty()) return true;
347
348 // Can move the region, and this is not the humongous region. Humongous
349 // moves are special cased here, because their moves are handled separately.
350 return r->is_stw_move_allowed() && !r->is_humongous();
351 }
352
work(uint worker_id)353 void work(uint worker_id) {
354 ShenandoahParallelWorkerSession worker_session(worker_id);
355 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
356 ShenandoahHeapRegionSetIterator it(slice);
357 ShenandoahHeapRegion* from_region = it.next();
358 // No work?
359 if (from_region == NULL) {
360 return;
361 }
362
363 // Sliding compaction. Walk all regions in the slice, and compact them.
364 // Remember empty regions and reuse them as needed.
365 ResourceMark rm;
366
367 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
368
369 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
370
371 while (from_region != NULL) {
372 assert(is_candidate_region(from_region), "Sanity");
373
374 cl.set_from_region(from_region);
375 if (from_region->has_live()) {
376 _heap->marked_object_iterate(from_region, &cl);
377 }
378
379 // Compacted the region to somewhere else? From-region is empty then.
380 if (!cl.is_compact_same_region()) {
381 empty_regions.append(from_region);
382 }
383 from_region = it.next();
384 }
385 cl.finish_region();
386
387 // Mark all remaining regions as empty
388 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
389 ShenandoahHeapRegion* r = empty_regions.at(pos);
390 r->set_new_top(r->bottom());
391 }
392 }
393 };
394
calculate_target_humongous_objects()395 void ShenandoahMarkCompact::calculate_target_humongous_objects() {
396 ShenandoahHeap* heap = ShenandoahHeap::heap();
397
398 // Compute the new addresses for humongous objects. We need to do this after addresses
399 // for regular objects are calculated, and we know what regions in heap suffix are
400 // available for humongous moves.
401 //
402 // Scan the heap backwards, because we are compacting humongous regions towards the end.
403 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
404 // humongous start there.
405 //
406 // The complication is potential non-movable regions during the scan. If such region is
407 // detected, then sliding restarts towards that non-movable region.
408
409 size_t to_begin = heap->num_regions();
410 size_t to_end = heap->num_regions();
411
412 for (size_t c = heap->num_regions(); c > 0; c--) {
413 ShenandoahHeapRegion *r = heap->get_region(c - 1);
414 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
415 // To-region candidate: record this, and continue scan
416 to_begin = r->index();
417 continue;
418 }
419
420 if (r->is_humongous_start() && r->is_stw_move_allowed()) {
421 // From-region candidate: movable humongous region
422 oop old_obj = oop(r->bottom());
423 size_t words_size = old_obj->size();
424 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
425
426 size_t start = to_end - num_regions;
427
428 if (start >= to_begin && start != r->index()) {
429 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
430 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
431 old_obj->forward_to(oop(heap->get_region(start)->bottom()));
432 to_end = start;
433 continue;
434 }
435 }
436
437 // Failed to fit. Scan starting from current region.
438 to_begin = r->index();
439 to_end = r->index();
440 }
441 }
442
443 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
444 private:
445 ShenandoahHeap* const _heap;
446
447 public:
ShenandoahEnsureHeapActiveClosure()448 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
heap_region_do(ShenandoahHeapRegion * r)449 void heap_region_do(ShenandoahHeapRegion* r) {
450 if (r->is_trash()) {
451 r->recycle();
452 }
453 if (r->is_cset()) {
454 r->make_regular_bypass();
455 }
456 if (r->is_empty_uncommitted()) {
457 r->make_committed_bypass();
458 }
459 assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index());
460
461 // Record current region occupancy: this communicates empty regions are free
462 // to the rest of Full GC code.
463 r->set_new_top(r->top());
464 }
465 };
466
467 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
468 private:
469 ShenandoahHeap* const _heap;
470 ShenandoahMarkingContext* const _ctx;
471
472 public:
ShenandoahTrashImmediateGarbageClosure()473 ShenandoahTrashImmediateGarbageClosure() :
474 _heap(ShenandoahHeap::heap()),
475 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
476
heap_region_do(ShenandoahHeapRegion * r)477 void heap_region_do(ShenandoahHeapRegion* r) {
478 if (r->is_humongous_start()) {
479 oop humongous_obj = oop(r->bottom());
480 if (!_ctx->is_marked(humongous_obj)) {
481 assert(!r->has_live(),
482 "Region " SIZE_FORMAT " is not marked, should not have live", r->index());
483 _heap->trash_humongous_region_at(r);
484 } else {
485 assert(r->has_live(),
486 "Region " SIZE_FORMAT " should have live", r->index());
487 }
488 } else if (r->is_humongous_continuation()) {
489 // If we hit continuation, the non-live humongous starts should have been trashed already
490 assert(r->humongous_start_region()->has_live(),
491 "Region " SIZE_FORMAT " should have live", r->index());
492 } else if (r->is_regular()) {
493 if (!r->has_live()) {
494 r->make_trash_immediate();
495 }
496 }
497 }
498 };
499
distribute_slices(ShenandoahHeapRegionSet ** worker_slices)500 void ShenandoahMarkCompact::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
501 ShenandoahHeap* heap = ShenandoahHeap::heap();
502
503 uint n_workers = heap->workers()->active_workers();
504 size_t n_regions = heap->num_regions();
505
506 // What we want to accomplish: have the dense prefix of data, while still balancing
507 // out the parallel work.
508 //
509 // Assuming the amount of work is driven by the live data that needs moving, we can slice
510 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
511 // thread takes all regions in its prefix subset, and then it takes some regions from
512 // the tail.
513 //
514 // Tail region selection becomes interesting.
515 //
516 // First, we want to distribute the regions fairly between the workers, and those regions
517 // might have different amount of live data. So, until we sure no workers need live data,
518 // we need to only take what the worker needs.
519 //
520 // Second, since we slide everything to the left in each slice, the most busy regions
521 // would be the ones on the left. Which means we want to have all workers have their after-tail
522 // regions as close to the left as possible.
523 //
524 // The easiest way to do this is to distribute after-tail regions in round-robin between
525 // workers that still need live data.
526 //
527 // Consider parallel workers A, B, C, then the target slice layout would be:
528 //
529 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
530 //
531 // (.....dense-prefix.....) (.....................tail...................)
532 // [all regions fully live] [left-most regions are fuller that right-most]
533 //
534
535 // Compute how much live data is there. This would approximate the size of dense prefix
536 // we target to create.
537 size_t total_live = 0;
538 for (size_t idx = 0; idx < n_regions; idx++) {
539 ShenandoahHeapRegion *r = heap->get_region(idx);
540 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
541 total_live += r->get_live_data_words();
542 }
543 }
544
545 // Estimate the size for the dense prefix. Note that we specifically count only the
546 // "full" regions, so there would be some non-full regions in the slice tail.
547 size_t live_per_worker = total_live / n_workers;
548 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
549 size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
550 prefix_regions_total = MIN2(prefix_regions_total, n_regions);
551 assert(prefix_regions_total <= n_regions, "Sanity");
552
553 // There might be non-candidate regions in the prefix. To compute where the tail actually
554 // ends up being, we need to account those as well.
555 size_t prefix_end = prefix_regions_total;
556 for (size_t idx = 0; idx < prefix_regions_total; idx++) {
557 ShenandoahHeapRegion *r = heap->get_region(idx);
558 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
559 prefix_end++;
560 }
561 }
562 prefix_end = MIN2(prefix_end, n_regions);
563 assert(prefix_end <= n_regions, "Sanity");
564
565 // Distribute prefix regions per worker: each thread definitely gets its own same-sized
566 // subset of dense prefix.
567 size_t prefix_idx = 0;
568
569 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
570
571 for (size_t wid = 0; wid < n_workers; wid++) {
572 ShenandoahHeapRegionSet* slice = worker_slices[wid];
573
574 live[wid] = 0;
575 size_t regs = 0;
576
577 // Add all prefix regions for this worker
578 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
579 ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
580 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
581 slice->add_region(r);
582 live[wid] += r->get_live_data_words();
583 regs++;
584 }
585 prefix_idx++;
586 }
587 }
588
589 // Distribute the tail among workers in round-robin fashion.
590 size_t wid = n_workers - 1;
591
592 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
593 ShenandoahHeapRegion *r = heap->get_region(tail_idx);
594 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
595 assert(wid < n_workers, "Sanity");
596
597 size_t live_region = r->get_live_data_words();
598
599 // Select next worker that still needs live data.
600 size_t old_wid = wid;
601 do {
602 wid++;
603 if (wid == n_workers) wid = 0;
604 } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
605
606 if (old_wid == wid) {
607 // Circled back to the same worker? This means liveness data was
608 // miscalculated. Bump the live_per_worker limit so that
609 // everyone gets a piece of the leftover work.
610 live_per_worker += ShenandoahHeapRegion::region_size_words();
611 }
612
613 worker_slices[wid]->add_region(r);
614 live[wid] += live_region;
615 }
616 }
617
618 FREE_C_HEAP_ARRAY(size_t, live);
619
620 #ifdef ASSERT
621 ResourceBitMap map(n_regions);
622 for (size_t wid = 0; wid < n_workers; wid++) {
623 ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
624 ShenandoahHeapRegion* r = it.next();
625 while (r != NULL) {
626 size_t idx = r->index();
627 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx);
628 assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx);
629 map.at_put(idx, true);
630 r = it.next();
631 }
632 }
633
634 for (size_t rid = 0; rid < n_regions; rid++) {
635 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
636 bool is_distributed = map.at(rid);
637 assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid);
638 }
639 #endif
640 }
641
phase2_calculate_target_addresses(ShenandoahHeapRegionSet ** worker_slices)642 void ShenandoahMarkCompact::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
643 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
644 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
645
646 ShenandoahHeap* heap = ShenandoahHeap::heap();
647
648 // About to figure out which regions can be compacted, make sure pinning status
649 // had been updated in GC prologue.
650 heap->assert_pinned_region_status();
651
652 {
653 // Trash the immediately collectible regions before computing addresses
654 ShenandoahTrashImmediateGarbageClosure tigcl;
655 heap->heap_region_iterate(&tigcl);
656
657 // Make sure regions are in good state: committed, active, clean.
658 // This is needed because we are potentially sliding the data through them.
659 ShenandoahEnsureHeapActiveClosure ecl;
660 heap->heap_region_iterate(&ecl);
661 }
662
663 // Compute the new addresses for regular objects
664 {
665 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
666
667 distribute_slices(worker_slices);
668
669 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
670 heap->workers()->run_task(&task);
671 }
672
673 // Compute the new addresses for humongous objects
674 {
675 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
676 calculate_target_humongous_objects();
677 }
678 }
679
680 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
681 private:
682 ShenandoahHeap* const _heap;
683 ShenandoahMarkingContext* const _ctx;
684
685 template <class T>
do_oop_work(T * p)686 inline void do_oop_work(T* p) {
687 T o = RawAccess<>::oop_load(p);
688 if (!CompressedOops::is_null(o)) {
689 oop obj = CompressedOops::decode_not_null(o);
690 assert(_ctx->is_marked(obj), "must be marked");
691 if (obj->is_forwarded()) {
692 oop forw = obj->forwardee();
693 RawAccess<IS_NOT_NULL>::oop_store(p, forw);
694 }
695 }
696 }
697
698 public:
ShenandoahAdjustPointersClosure()699 ShenandoahAdjustPointersClosure() :
700 _heap(ShenandoahHeap::heap()),
701 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
702
do_oop(oop * p)703 void do_oop(oop* p) { do_oop_work(p); }
do_oop(narrowOop * p)704 void do_oop(narrowOop* p) { do_oop_work(p); }
705 };
706
707 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
708 private:
709 ShenandoahHeap* const _heap;
710 ShenandoahAdjustPointersClosure _cl;
711
712 public:
ShenandoahAdjustPointersObjectClosure()713 ShenandoahAdjustPointersObjectClosure() :
714 _heap(ShenandoahHeap::heap()) {
715 }
do_object(oop p)716 void do_object(oop p) {
717 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
718 p->oop_iterate(&_cl);
719 }
720 };
721
722 class ShenandoahAdjustPointersTask : public AbstractGangTask {
723 private:
724 ShenandoahHeap* const _heap;
725 ShenandoahRegionIterator _regions;
726
727 public:
ShenandoahAdjustPointersTask()728 ShenandoahAdjustPointersTask() :
729 AbstractGangTask("Shenandoah Adjust Pointers"),
730 _heap(ShenandoahHeap::heap()) {
731 }
732
work(uint worker_id)733 void work(uint worker_id) {
734 ShenandoahParallelWorkerSession worker_session(worker_id);
735 ShenandoahAdjustPointersObjectClosure obj_cl;
736 ShenandoahHeapRegion* r = _regions.next();
737 while (r != NULL) {
738 if (!r->is_humongous_continuation() && r->has_live()) {
739 _heap->marked_object_iterate(r, &obj_cl);
740 }
741 r = _regions.next();
742 }
743 }
744 };
745
746 class ShenandoahAdjustRootPointersTask : public AbstractGangTask {
747 private:
748 ShenandoahRootAdjuster* _rp;
749 PreservedMarksSet* _preserved_marks;
750 public:
ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster * rp,PreservedMarksSet * preserved_marks)751 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
752 AbstractGangTask("Shenandoah Adjust Root Pointers"),
753 _rp(rp),
754 _preserved_marks(preserved_marks) {}
755
work(uint worker_id)756 void work(uint worker_id) {
757 ShenandoahParallelWorkerSession worker_session(worker_id);
758 ShenandoahAdjustPointersClosure cl;
759 _rp->roots_do(worker_id, &cl);
760 _preserved_marks->get(worker_id)->adjust_during_full_gc();
761 }
762 };
763
phase3_update_references()764 void ShenandoahMarkCompact::phase3_update_references() {
765 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
766 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
767
768 ShenandoahHeap* heap = ShenandoahHeap::heap();
769
770 WorkGang* workers = heap->workers();
771 uint nworkers = workers->active_workers();
772 {
773 #if COMPILER2_OR_JVMCI
774 DerivedPointerTable::clear();
775 #endif
776 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
777 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
778 workers->run_task(&task);
779 #if COMPILER2_OR_JVMCI
780 DerivedPointerTable::update_pointers();
781 #endif
782 }
783
784 ShenandoahAdjustPointersTask adjust_pointers_task;
785 workers->run_task(&adjust_pointers_task);
786 }
787
788 class ShenandoahCompactObjectsClosure : public ObjectClosure {
789 private:
790 ShenandoahHeap* const _heap;
791 uint const _worker_id;
792
793 public:
ShenandoahCompactObjectsClosure(uint worker_id)794 ShenandoahCompactObjectsClosure(uint worker_id) :
795 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
796
do_object(oop p)797 void do_object(oop p) {
798 assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
799 size_t size = (size_t)p->size();
800 if (p->is_forwarded()) {
801 HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
802 HeapWord* compact_to = cast_from_oop<HeapWord*>(p->forwardee());
803 Copy::aligned_conjoint_words(compact_from, compact_to, size);
804 oop new_obj = oop(compact_to);
805 new_obj->init_mark();
806 }
807 }
808 };
809
810 class ShenandoahCompactObjectsTask : public AbstractGangTask {
811 private:
812 ShenandoahHeap* const _heap;
813 ShenandoahHeapRegionSet** const _worker_slices;
814
815 public:
ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet ** worker_slices)816 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
817 AbstractGangTask("Shenandoah Compact Objects"),
818 _heap(ShenandoahHeap::heap()),
819 _worker_slices(worker_slices) {
820 }
821
work(uint worker_id)822 void work(uint worker_id) {
823 ShenandoahParallelWorkerSession worker_session(worker_id);
824 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
825
826 ShenandoahCompactObjectsClosure cl(worker_id);
827 ShenandoahHeapRegion* r = slice.next();
828 while (r != NULL) {
829 assert(!r->is_humongous(), "must not get humongous regions here");
830 if (r->has_live()) {
831 _heap->marked_object_iterate(r, &cl);
832 }
833 r->set_top(r->new_top());
834 r = slice.next();
835 }
836 }
837 };
838
839 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
840 private:
841 ShenandoahHeap* const _heap;
842 size_t _live;
843
844 public:
ShenandoahPostCompactClosure()845 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
846 _heap->free_set()->clear();
847 }
848
heap_region_do(ShenandoahHeapRegion * r)849 void heap_region_do(ShenandoahHeapRegion* r) {
850 assert (!r->is_cset(), "cset regions should have been demoted already");
851
852 // Need to reset the complete-top-at-mark-start pointer here because
853 // the complete marking bitmap is no longer valid. This ensures
854 // size-based iteration in marked_object_iterate().
855 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
856 // pinned regions.
857 if (!r->is_pinned()) {
858 _heap->complete_marking_context()->reset_top_at_mark_start(r);
859 }
860
861 size_t live = r->used();
862
863 // Make empty regions that have been allocated into regular
864 if (r->is_empty() && live > 0) {
865 r->make_regular_bypass();
866 }
867
868 // Reclaim regular regions that became empty
869 if (r->is_regular() && live == 0) {
870 r->make_trash();
871 }
872
873 // Recycle all trash regions
874 if (r->is_trash()) {
875 live = 0;
876 r->recycle();
877 }
878
879 r->set_live_data(live);
880 r->reset_alloc_metadata();
881 _live += live;
882 }
883
get_live()884 size_t get_live() {
885 return _live;
886 }
887 };
888
compact_humongous_objects()889 void ShenandoahMarkCompact::compact_humongous_objects() {
890 // Compact humongous regions, based on their fwdptr objects.
891 //
892 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
893 // humongous regions are already compacted, and do not require further moves, which alleviates
894 // sliding costs. We may consider doing this in parallel in future.
895
896 ShenandoahHeap* heap = ShenandoahHeap::heap();
897
898 for (size_t c = heap->num_regions(); c > 0; c--) {
899 ShenandoahHeapRegion* r = heap->get_region(c - 1);
900 if (r->is_humongous_start()) {
901 oop old_obj = oop(r->bottom());
902 if (!old_obj->is_forwarded()) {
903 // No need to move the object, it stays at the same slot
904 continue;
905 }
906 size_t words_size = old_obj->size();
907 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
908
909 size_t old_start = r->index();
910 size_t old_end = old_start + num_regions - 1;
911 size_t new_start = heap->heap_region_index_containing(old_obj->forwardee());
912 size_t new_end = new_start + num_regions - 1;
913 assert(old_start != new_start, "must be real move");
914 assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
915
916 Copy::aligned_conjoint_words(heap->get_region(old_start)->bottom(),
917 heap->get_region(new_start)->bottom(),
918 words_size);
919
920 oop new_obj = oop(heap->get_region(new_start)->bottom());
921 new_obj->init_mark();
922
923 {
924 for (size_t c = old_start; c <= old_end; c++) {
925 ShenandoahHeapRegion* r = heap->get_region(c);
926 r->make_regular_bypass();
927 r->set_top(r->bottom());
928 }
929
930 for (size_t c = new_start; c <= new_end; c++) {
931 ShenandoahHeapRegion* r = heap->get_region(c);
932 if (c == new_start) {
933 r->make_humongous_start_bypass();
934 } else {
935 r->make_humongous_cont_bypass();
936 }
937
938 // Trailing region may be non-full, record the remainder there
939 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
940 if ((c == new_end) && (remainder != 0)) {
941 r->set_top(r->bottom() + remainder);
942 } else {
943 r->set_top(r->end());
944 }
945
946 r->reset_alloc_metadata();
947 }
948 }
949 }
950 }
951 }
952
953 // This is slightly different to ShHeap::reset_next_mark_bitmap:
954 // we need to remain able to walk pinned regions.
955 // Since pinned region do not move and don't get compacted, we will get holes with
956 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
957 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
958 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
959 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
960 class ShenandoahMCResetCompleteBitmapTask : public AbstractGangTask {
961 private:
962 ShenandoahRegionIterator _regions;
963
964 public:
ShenandoahMCResetCompleteBitmapTask()965 ShenandoahMCResetCompleteBitmapTask() :
966 AbstractGangTask("Shenandoah Reset Bitmap") {
967 }
968
work(uint worker_id)969 void work(uint worker_id) {
970 ShenandoahParallelWorkerSession worker_session(worker_id);
971 ShenandoahHeapRegion* region = _regions.next();
972 ShenandoahHeap* heap = ShenandoahHeap::heap();
973 ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
974 while (region != NULL) {
975 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
976 ctx->clear_bitmap(region);
977 }
978 region = _regions.next();
979 }
980 }
981 };
982
phase4_compact_objects(ShenandoahHeapRegionSet ** worker_slices)983 void ShenandoahMarkCompact::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
984 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
985 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
986
987 ShenandoahHeap* heap = ShenandoahHeap::heap();
988
989 // Compact regular objects first
990 {
991 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
992 ShenandoahCompactObjectsTask compact_task(worker_slices);
993 heap->workers()->run_task(&compact_task);
994 }
995
996 // Compact humongous objects after regular object moves
997 {
998 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
999 compact_humongous_objects();
1000 }
1001
1002 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1003 // and must ensure the bitmap is in sync.
1004 {
1005 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1006 ShenandoahMCResetCompleteBitmapTask task;
1007 heap->workers()->run_task(&task);
1008 }
1009
1010 // Bring regions in proper states after the collection, and set heap properties.
1011 {
1012 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1013
1014 ShenandoahPostCompactClosure post_compact;
1015 heap->heap_region_iterate(&post_compact);
1016 heap->set_used(post_compact.get_live());
1017
1018 heap->collection_set()->clear();
1019 heap->free_set()->rebuild();
1020 }
1021
1022 heap->clear_cancelled_gc();
1023 }
1024