1 /*
2 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. 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 "classfile/classLoaderDataGraph.hpp"
26 #include "classfile/javaClasses.inline.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/parallel/objectStartArray.inline.hpp"
34 #include "gc/parallel/parallelArguments.hpp"
35 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
36 #include "gc/parallel/parMarkBitMap.inline.hpp"
37 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
38 #include "gc/parallel/psCompactionManager.inline.hpp"
39 #include "gc/parallel/psOldGen.hpp"
40 #include "gc/parallel/psParallelCompact.inline.hpp"
41 #include "gc/parallel/psPromotionManager.inline.hpp"
42 #include "gc/parallel/psRootType.hpp"
43 #include "gc/parallel/psScavenge.hpp"
44 #include "gc/parallel/psStringDedup.hpp"
45 #include "gc/parallel/psYoungGen.hpp"
46 #include "gc/shared/classUnloadingContext.hpp"
47 #include "gc/shared/fullGCForwarding.inline.hpp"
48 #include "gc/shared/gcCause.hpp"
49 #include "gc/shared/gcHeapSummary.hpp"
50 #include "gc/shared/gcId.hpp"
51 #include "gc/shared/gcLocker.hpp"
52 #include "gc/shared/gcTimer.hpp"
53 #include "gc/shared/gcTrace.hpp"
54 #include "gc/shared/gcTraceTime.inline.hpp"
55 #include "gc/shared/gcVMOperations.hpp"
56 #include "gc/shared/isGCActiveMark.hpp"
57 #include "gc/shared/oopStorage.inline.hpp"
58 #include "gc/shared/oopStorageSet.inline.hpp"
59 #include "gc/shared/oopStorageSetParState.inline.hpp"
60 #include "gc/shared/parallelCleaning.hpp"
61 #include "gc/shared/preservedMarks.inline.hpp"
62 #include "gc/shared/referencePolicy.hpp"
63 #include "gc/shared/referenceProcessor.hpp"
64 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
65 #include "gc/shared/spaceDecorator.hpp"
66 #include "gc/shared/taskTerminator.hpp"
67 #include "gc/shared/weakProcessor.inline.hpp"
68 #include "gc/shared/workerPolicy.hpp"
69 #include "gc/shared/workerThread.hpp"
70 #include "gc/shared/workerUtils.hpp"
71 #include "logging/log.hpp"
72 #include "memory/iterator.inline.hpp"
73 #include "memory/memoryReserver.hpp"
74 #include "memory/metaspaceUtils.hpp"
75 #include "memory/resourceArea.hpp"
76 #include "memory/universe.hpp"
77 #include "nmt/memTracker.hpp"
78 #include "oops/access.inline.hpp"
79 #include "oops/flatArrayKlass.inline.hpp"
80 #include "oops/instanceClassLoaderKlass.inline.hpp"
81 #include "oops/instanceKlass.inline.hpp"
82 #include "oops/instanceMirrorKlass.inline.hpp"
83 #include "oops/methodData.hpp"
84 #include "oops/objArrayKlass.inline.hpp"
85 #include "oops/oop.inline.hpp"
86 #include "runtime/atomicAccess.hpp"
87 #include "runtime/handles.inline.hpp"
88 #include "runtime/java.hpp"
89 #include "runtime/safepoint.hpp"
90 #include "runtime/threads.hpp"
91 #include "runtime/vmThread.hpp"
92 #include "services/memoryService.hpp"
93 #include "utilities/align.hpp"
94 #include "utilities/debug.hpp"
95 #include "utilities/events.hpp"
96 #include "utilities/formatBuffer.hpp"
97 #include "utilities/macros.hpp"
98 #include "utilities/stack.inline.hpp"
99 #if INCLUDE_JVMCI
100 #include "jvmci/jvmci.hpp"
101 #endif
102
103 #include <math.h>
104
105 // All sizes are in HeapWords.
106 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
107 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
108 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
109 const size_t ParallelCompactData::RegionSizeBytes =
110 RegionSize << LogHeapWordSize;
111 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
112 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
113 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
114
115 const ParallelCompactData::RegionData::region_sz_t
116 ParallelCompactData::RegionData::dc_shift = 27;
117
118 const ParallelCompactData::RegionData::region_sz_t
119 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
120
121 const ParallelCompactData::RegionData::region_sz_t
122 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
123
124 const ParallelCompactData::RegionData::region_sz_t
125 ParallelCompactData::RegionData::los_mask = ~dc_mask;
126
127 const ParallelCompactData::RegionData::region_sz_t
128 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
129
130 const ParallelCompactData::RegionData::region_sz_t
131 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
132
133 bool ParallelCompactData::RegionData::is_clear() {
134 return (_destination == nullptr) &&
135 (_source_region == 0) &&
136 (_partial_obj_addr == nullptr) &&
137 (_partial_obj_size == 0) &&
138 (_dc_and_los == 0) &&
139 (_shadow_state == 0);
140 }
141
142 #ifdef ASSERT
143 void ParallelCompactData::RegionData::verify_clear() {
144 assert(_destination == nullptr, "inv");
145 assert(_source_region == 0, "inv");
146 assert(_partial_obj_addr == nullptr, "inv");
147 assert(_partial_obj_size == 0, "inv");
148 assert(_dc_and_los == 0, "inv");
149 assert(_shadow_state == 0, "inv");
150 }
151 #endif
152
153 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
154
155 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
156 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
157
158 void SplitInfo::record(size_t split_region_idx, HeapWord* split_point, size_t preceding_live_words) {
159 assert(split_region_idx != 0, "precondition");
160
161 // Obj denoted by split_point will be deferred to the next space.
162 assert(split_point != nullptr, "precondition");
163
164 const ParallelCompactData& sd = PSParallelCompact::summary_data();
165
166 PSParallelCompact::RegionData* split_region_ptr = sd.region(split_region_idx);
167 assert(preceding_live_words < split_region_ptr->data_size(), "inv");
168
169 HeapWord* preceding_destination = split_region_ptr->destination();
170 assert(preceding_destination != nullptr, "inv");
171
172 // How many regions does the preceding part occupy
173 uint preceding_destination_count;
174 if (preceding_live_words == 0) {
175 preceding_destination_count = 0;
176 } else {
177 // -1 so that the ending address doesn't fall on the region-boundary
178 if (sd.region_align_down(preceding_destination) ==
179 sd.region_align_down(preceding_destination + preceding_live_words - 1)) {
180 preceding_destination_count = 1;
181 } else {
182 preceding_destination_count = 2;
183 }
184 }
185
186 _split_region_idx = split_region_idx;
187 _split_point = split_point;
188 _preceding_live_words = preceding_live_words;
189 _preceding_destination = preceding_destination;
190 _preceding_destination_count = preceding_destination_count;
191 }
192
193 void SplitInfo::clear()
194 {
195 _split_region_idx = 0;
196 _split_point = nullptr;
197 _preceding_live_words = 0;
198 _preceding_destination = nullptr;
199 _preceding_destination_count = 0;
200 assert(!is_valid(), "sanity");
201 }
202
203 #ifdef ASSERT
204 void SplitInfo::verify_clear()
205 {
206 assert(_split_region_idx == 0, "not clear");
207 assert(_split_point == nullptr, "not clear");
208 assert(_preceding_live_words == 0, "not clear");
209 assert(_preceding_destination == nullptr, "not clear");
210 assert(_preceding_destination_count == 0, "not clear");
211 }
212 #endif // #ifdef ASSERT
213
214
215 void PSParallelCompact::print_on(outputStream* st) {
216 _mark_bitmap.print_on(st);
217 }
218
219 ParallelCompactData::ParallelCompactData() :
220 _heap_start(nullptr),
221 DEBUG_ONLY(_heap_end(nullptr) COMMA)
222 _region_vspace(nullptr),
223 _reserved_byte_size(0),
224 _region_data(nullptr),
225 _region_count(0) {}
226
227 bool ParallelCompactData::initialize(MemRegion reserved_heap)
228 {
229 _heap_start = reserved_heap.start();
230 const size_t heap_size = reserved_heap.word_size();
231 DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
232
233 assert(region_align_down(_heap_start) == _heap_start,
234 "region start not aligned");
235
236 return initialize_region_data(heap_size);
237 }
238
239 PSVirtualSpace*
240 ParallelCompactData::create_vspace(size_t count, size_t element_size)
241 {
242 const size_t raw_bytes = count * element_size;
243 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
244 const size_t granularity = os::vm_allocation_granularity();
245 const size_t rs_align = MAX2(page_sz, granularity);
246
247 _reserved_byte_size = align_up(raw_bytes, rs_align);
248
249 ReservedSpace rs = MemoryReserver::reserve(_reserved_byte_size,
250 rs_align,
251 page_sz,
252 mtGC);
253
254 if (!rs.is_reserved()) {
255 // Failed to reserve memory.
256 return nullptr;
257 }
258
259 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
260 rs.size(), page_sz);
261
262 MemTracker::record_virtual_memory_tag(rs, mtGC);
263
264 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
265
266 if (!vspace->expand_by(_reserved_byte_size)) {
267 // Failed to commit memory.
268
269 delete vspace;
270
271 // Release memory reserved in the space.
272 MemoryReserver::release(rs);
273
274 return nullptr;
275 }
276
277 return vspace;
278 }
279
280 bool ParallelCompactData::initialize_region_data(size_t heap_size)
281 {
282 assert(is_aligned(heap_size, RegionSize), "precondition");
283
284 const size_t count = heap_size >> Log2RegionSize;
285 _region_vspace = create_vspace(count, sizeof(RegionData));
286 if (_region_vspace != nullptr) {
287 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
288 _region_count = count;
289 return true;
290 }
291 return false;
292 }
293
294 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
295 assert(beg_region <= _region_count, "beg_region out of range");
296 assert(end_region <= _region_count, "end_region out of range");
297
298 const size_t region_cnt = end_region - beg_region;
299 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
300 }
301
302 // The total live words on src_region would overflow the target space, so find
303 // the overflowing object and record the split point. The invariant is that an
304 // obj should not cross space boundary.
305 HeapWord* ParallelCompactData::summarize_split_space(size_t src_region,
306 SplitInfo& split_info,
307 HeapWord* const destination,
308 HeapWord* const target_end,
309 HeapWord** target_next) {
310 assert(destination <= target_end, "sanity");
311 assert(destination + _region_data[src_region].data_size() > target_end,
312 "region should not fit into target space");
313 assert(is_region_aligned(target_end), "sanity");
314
315 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
316
317 if (destination + partial_obj_size > target_end) {
318 assert(partial_obj_size > 0, "inv");
319 // The overflowing obj is from a previous region.
320 //
321 // source-regions:
322 //
323 // ***************
324 // | A|AA |
325 // ***************
326 // ^
327 // | split-point
328 //
329 // dest-region:
330 //
331 // ********
332 // |~~~~A |
333 // ********
334 // ^^
335 // || target-space-end
336 // |
337 // | destination
338 //
339 // AAA would overflow target-space.
340 //
341 HeapWord* overflowing_obj = _region_data[src_region].partial_obj_addr();
342 size_t split_region = addr_to_region_idx(overflowing_obj);
343
344 // The number of live words before the overflowing object on this split region
345 size_t preceding_live_words;
346 if (is_region_aligned(overflowing_obj)) {
347 preceding_live_words = 0;
348 } else {
349 // Words accounted by the overflowing object on the split region
350 size_t overflowing_size = pointer_delta(region_align_up(overflowing_obj), overflowing_obj);
351 preceding_live_words = region(split_region)->data_size() - overflowing_size;
352 }
353
354 split_info.record(split_region, overflowing_obj, preceding_live_words);
355
356 // The [overflowing_obj, src_region_start) part has been accounted for, so
357 // must move back the new_top, now that this overflowing obj is deferred.
358 HeapWord* new_top = destination - pointer_delta(region_to_addr(src_region), overflowing_obj);
359
360 // If the overflowing obj was relocated to its original destination,
361 // those destination regions would have their source_region set. Now that
362 // this overflowing obj is relocated somewhere else, reset the
363 // source_region.
364 {
365 size_t range_start = addr_to_region_idx(region_align_up(new_top));
366 size_t range_end = addr_to_region_idx(region_align_up(destination));
367 for (size_t i = range_start; i < range_end; ++i) {
368 region(i)->set_source_region(0);
369 }
370 }
371
372 // Update new top of target space
373 *target_next = new_top;
374
375 return overflowing_obj;
376 }
377
378 // Obj-iteration to locate the overflowing obj
379 HeapWord* region_start = region_to_addr(src_region);
380 HeapWord* region_end = region_start + RegionSize;
381 HeapWord* cur_addr = region_start + partial_obj_size;
382 size_t live_words = partial_obj_size;
383
384 while (true) {
385 assert(cur_addr < region_end, "inv");
386 cur_addr = PSParallelCompact::mark_bitmap()->find_obj_beg(cur_addr, region_end);
387 // There must be an overflowing obj in this region
388 assert(cur_addr < region_end, "inv");
389
390 oop obj = cast_to_oop(cur_addr);
391 size_t obj_size = obj->size();
392 if (destination + live_words + obj_size > target_end) {
393 // Found the overflowing obj
394 split_info.record(src_region, cur_addr, live_words);
395 *target_next = destination + live_words;
396 return cur_addr;
397 }
398
399 live_words += obj_size;
400 cur_addr += obj_size;
401 }
402 }
403
404 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
405 HeapWord** full_region_prefix_end) {
406 size_t cur_region = addr_to_region_idx(space->bottom());
407 const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
408 size_t live_words = 0;
409 if (full_region_prefix_end == nullptr) {
410 for (/* empty */; cur_region < end_region; ++cur_region) {
411 live_words += _region_data[cur_region].data_size();
412 }
413 } else {
414 bool first_set = false;
415 for (/* empty */; cur_region < end_region; ++cur_region) {
416 size_t live_words_in_region = _region_data[cur_region].data_size();
417 if (!first_set && live_words_in_region < RegionSize) {
418 *full_region_prefix_end = region_to_addr(cur_region);
419 first_set = true;
420 }
421 live_words += live_words_in_region;
422 }
423 if (!first_set) {
424 // All regions are full of live objs.
425 assert(is_region_aligned(space->top()), "inv");
426 *full_region_prefix_end = space->top();
427 }
428 assert(*full_region_prefix_end != nullptr, "postcondition");
429 assert(is_region_aligned(*full_region_prefix_end), "inv");
430 assert(*full_region_prefix_end >= space->bottom(), "in-range");
431 assert(*full_region_prefix_end <= space->top(), "in-range");
432 }
433 return live_words;
434 }
435
436 bool ParallelCompactData::summarize(SplitInfo& split_info,
437 HeapWord* source_beg, HeapWord* source_end,
438 HeapWord** source_next,
439 HeapWord* target_beg, HeapWord* target_end,
440 HeapWord** target_next)
441 {
442 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
443 log_develop_trace(gc, compaction)(
444 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
445 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
446 p2i(source_beg), p2i(source_end), p2i(source_next_val),
447 p2i(target_beg), p2i(target_end), p2i(*target_next));
448
449 size_t cur_region = addr_to_region_idx(source_beg);
450 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
451
452 HeapWord *dest_addr = target_beg;
453 for (/* empty */; cur_region < end_region; cur_region++) {
454 size_t words = _region_data[cur_region].data_size();
455
456 // Skip empty ones
457 if (words == 0) {
458 continue;
459 }
460
461 if (split_info.is_split(cur_region)) {
462 assert(words > split_info.preceding_live_words(), "inv");
463 words -= split_info.preceding_live_words();
464 }
465
466 _region_data[cur_region].set_destination(dest_addr);
467
468 // If cur_region does not fit entirely into the target space, find a point
469 // at which the source space can be 'split' so that part is copied to the
470 // target space and the rest is copied elsewhere.
471 if (dest_addr + words > target_end) {
472 assert(source_next != nullptr, "source_next is null when splitting");
473 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
474 target_end, target_next);
475 return false;
476 }
477
478 uint destination_count = split_info.is_split(cur_region)
479 ? split_info.preceding_destination_count()
480 : 0;
481
482 HeapWord* const last_addr = dest_addr + words - 1;
483 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
484 const size_t dest_region_2 = addr_to_region_idx(last_addr);
485
486 // Initially assume that the destination regions will be the same and
487 // adjust the value below if necessary. Under this assumption, if
488 // cur_region == dest_region_2, then cur_region will be compacted
489 // completely into itself.
490 destination_count += cur_region == dest_region_2 ? 0 : 1;
491 if (dest_region_1 != dest_region_2) {
492 // Destination regions differ; adjust destination_count.
493 destination_count += 1;
494 // Data from cur_region will be copied to the start of dest_region_2.
495 _region_data[dest_region_2].set_source_region(cur_region);
496 } else if (is_region_aligned(dest_addr)) {
497 // Data from cur_region will be copied to the start of the destination
498 // region.
499 _region_data[dest_region_1].set_source_region(cur_region);
500 }
501
502 _region_data[cur_region].set_destination_count(destination_count);
503 dest_addr += words;
504 }
505
506 *target_next = dest_addr;
507 return true;
508 }
509
510 #ifdef ASSERT
511 void ParallelCompactData::verify_clear() {
512 for (uint cur_idx = 0; cur_idx < region_count(); ++cur_idx) {
513 if (!region(cur_idx)->is_clear()) {
514 log_warning(gc)("Uncleared Region: %u", cur_idx);
515 region(cur_idx)->verify_clear();
516 }
517 }
518 }
519 #endif // #ifdef ASSERT
520
521 STWGCTimer PSParallelCompact::_gc_timer;
522 ParallelOldTracer PSParallelCompact::_gc_tracer;
523 elapsedTimer PSParallelCompact::_accumulated_time;
524 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
525 CollectorCounters* PSParallelCompact::_counters = nullptr;
526 ParMarkBitMap PSParallelCompact::_mark_bitmap;
527 ParallelCompactData PSParallelCompact::_summary_data;
528
529 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
530
531 class PCAdjustPointerClosure: public BasicOopIterateClosure {
532 template <typename T>
533 void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
534
535 public:
536 virtual void do_oop(oop* p) { do_oop_work(p); }
537 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
538
539 virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
540 };
541
542 static PCAdjustPointerClosure pc_adjust_pointer_closure;
543
544 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
545
546 void PSParallelCompact::post_initialize() {
547 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
548 _span_based_discoverer.set_span(heap->reserved_region());
549 _ref_processor =
550 new ReferenceProcessor(&_span_based_discoverer,
551 ParallelGCThreads, // mt processing degree
552 ParallelGCThreads, // mt discovery degree
553 false, // concurrent_discovery
554 &_is_alive_closure); // non-header is alive closure
555
556 _counters = new CollectorCounters("Parallel full collection pauses", 1);
557
558 // Initialize static fields in ParCompactionManager.
559 ParCompactionManager::initialize(mark_bitmap());
560 }
561
562 bool PSParallelCompact::initialize_aux_data() {
563 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
564 MemRegion mr = heap->reserved_region();
565 assert(mr.byte_size() != 0, "heap should be reserved");
566
567 initialize_space_info();
568
569 if (!_mark_bitmap.initialize(mr)) {
570 vm_shutdown_during_initialization(
571 err_msg("Unable to allocate %zuKB bitmaps for parallel "
572 "garbage collection for the requested %zuKB heap.",
573 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
574 return false;
575 }
576
577 if (!_summary_data.initialize(mr)) {
578 vm_shutdown_during_initialization(
579 err_msg("Unable to allocate %zuKB card tables for parallel "
580 "garbage collection for the requested %zuKB heap.",
581 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
582 return false;
583 }
584
585 return true;
586 }
587
588 void PSParallelCompact::initialize_space_info()
589 {
590 memset(&_space_info, 0, sizeof(_space_info));
591
592 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
593 PSYoungGen* young_gen = heap->young_gen();
594
595 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
596 _space_info[eden_space_id].set_space(young_gen->eden_space());
597 _space_info[from_space_id].set_space(young_gen->from_space());
598 _space_info[to_space_id].set_space(young_gen->to_space());
599
600 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
601 }
602
603 void
604 PSParallelCompact::clear_data_covering_space(SpaceId id)
605 {
606 // At this point, top is the value before GC, new_top() is the value that will
607 // be set at the end of GC. The marking bitmap is cleared to top; nothing
608 // should be marked above top. The summary data is cleared to the larger of
609 // top & new_top.
610 MutableSpace* const space = _space_info[id].space();
611 HeapWord* const bot = space->bottom();
612 HeapWord* const top = space->top();
613 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
614
615 _mark_bitmap.clear_range(bot, top);
616
617 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
618 const size_t end_region =
619 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
620 _summary_data.clear_range(beg_region, end_region);
621
622 // Clear the data used to 'split' regions.
623 SplitInfo& split_info = _space_info[id].split_info();
624 if (split_info.is_valid()) {
625 split_info.clear();
626 }
627 DEBUG_ONLY(split_info.verify_clear();)
628 }
629
630 void PSParallelCompact::pre_compact()
631 {
632 // Update the from & to space pointers in space_info, since they are swapped
633 // at each young gen gc. Do the update unconditionally (even though a
634 // promotion failure does not swap spaces) because an unknown number of young
635 // collections will have swapped the spaces an unknown number of times.
636 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
637 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
638 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
639 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
640
641 heap->increment_total_collections(true);
642
643 CodeCache::on_gc_marking_cycle_start();
644
645 heap->print_before_gc();
646 heap->trace_heap_before_gc(&_gc_tracer);
647
648 // Fill in TLABs
649 heap->ensure_parsability(true); // retire TLABs
650
651 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
652 Universe::verify("Before GC");
653 }
654
655 DEBUG_ONLY(mark_bitmap()->verify_clear();)
656 DEBUG_ONLY(summary_data().verify_clear();)
657 }
658
659 void PSParallelCompact::post_compact()
660 {
661 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
662 ParCompactionManager::remove_all_shadow_regions();
663
664 CodeCache::on_gc_marking_cycle_finish();
665 CodeCache::arm_all_nmethods();
666
667 // Need to clear claim bits for the next full-gc (marking and adjust-pointers).
668 ClassLoaderDataGraph::clear_claimed_marks();
669
670 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
671 // Clear the marking bitmap, summary data and split info.
672 clear_data_covering_space(SpaceId(id));
673 {
674 MutableSpace* space = _space_info[id].space();
675 HeapWord* top = space->top();
676 HeapWord* new_top = _space_info[id].new_top();
677 if (ZapUnusedHeapArea && new_top < top) {
678 space->mangle_region(MemRegion(new_top, top));
679 }
680 // Update top(). Must be done after clearing the bitmap and summary data.
681 space->set_top(new_top);
682 }
683 }
684
685 #ifdef ASSERT
686 {
687 mark_bitmap()->verify_clear();
688 summary_data().verify_clear();
689 }
690 #endif
691
692 ParCompactionManager::flush_all_string_dedup_requests();
693
694 MutableSpace* const eden_space = _space_info[eden_space_id].space();
695 MutableSpace* const from_space = _space_info[from_space_id].space();
696 MutableSpace* const to_space = _space_info[to_space_id].space();
697
698 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
699 bool eden_empty = eden_space->is_empty();
700
701 // Update heap occupancy information which is used as input to the soft ref
702 // clearing policy at the next gc.
703 Universe::heap()->update_capacity_and_used_at_gc();
704
705 bool young_gen_empty = eden_empty && from_space->is_empty() &&
706 to_space->is_empty();
707
708 PSCardTable* ct = heap->card_table();
709 MemRegion old_mr = heap->old_gen()->committed();
710 if (young_gen_empty) {
711 ct->clear_MemRegion(old_mr);
712 } else {
713 ct->dirty_MemRegion(old_mr);
714 }
715
716 heap->prune_scavengable_nmethods();
717
718 #if COMPILER2_OR_JVMCI
719 DerivedPointerTable::update_pointers();
720 #endif
721
722 // Signal that we have completed a visit to all live objects.
723 Universe::heap()->record_whole_heap_examined_timestamp();
724 }
725
726 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
727 HeapWord* full_region_prefix_end) {
728 const size_t region_size = ParallelCompactData::RegionSize;
729 const ParallelCompactData& sd = summary_data();
730
731 // Iteration starts with the region *after* the full-region-prefix-end.
732 const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
733 // If final region is not full, iteration stops before that region,
734 // because fill_dense_prefix_end assumes that prefix_end <= top.
735 const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
736 assert(start_region <= end_region, "inv");
737
738 size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
739 const RegionData* cur_region = start_region;
740 for (/* empty */; cur_region < end_region; ++cur_region) {
741 assert(region_size >= cur_region->data_size(), "inv");
742 size_t dead_size = region_size - cur_region->data_size();
743 if (max_waste < dead_size) {
744 break;
745 }
746 max_waste -= dead_size;
747 }
748
749 HeapWord* const prefix_end = sd.region_to_addr(cur_region);
750 assert(sd.is_region_aligned(prefix_end), "postcondition");
751 assert(prefix_end >= full_region_prefix_end, "in-range");
752 assert(prefix_end <= old_space->top(), "in-range");
753 return prefix_end;
754 }
755
756 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
757 // Comparing two sizes to decide if filling is required:
758 //
759 // The size of the filler (min-obj-size) is 2 heap words with the default
760 // MinObjAlignment, since both markword and klass take 1 heap word.
761 // With +UseCompactObjectHeaders, the minimum filler size is only one word,
762 // because the Klass* gets encoded in the mark-word.
763 //
764 // The size of the gap (if any) right before dense-prefix-end is
765 // MinObjAlignment.
766 //
767 // Need to fill in the gap only if it's smaller than min-obj-size, and the
768 // filler obj will extend to next region.
769
770 if (MinObjAlignment >= checked_cast<int>(CollectedHeap::min_fill_size())) {
771 return;
772 }
773
774 assert(!UseCompactObjectHeaders, "Compact headers can allocate small objects");
775 assert(CollectedHeap::min_fill_size() == 2, "inv");
776 HeapWord* const dense_prefix_end = dense_prefix(id);
777 assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
778 assert(dense_prefix_end <= space(id)->top(), "precondition");
779 if (dense_prefix_end == space(id)->top()) {
780 // Must not have single-word gap right before prefix-end/top.
781 return;
782 }
783 RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
784
785 if (region_after_dense_prefix->partial_obj_size() != 0 ||
786 _mark_bitmap.is_marked(dense_prefix_end)) {
787 // The region after the dense prefix starts with live bytes.
788 return;
789 }
790
791 HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
792 if (block_start == dense_prefix_end - 1) {
793 assert(!_mark_bitmap.is_marked(block_start), "inv");
794 // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
795 // The filler object will extend into region_after_dense_prefix.
796 const size_t obj_len = 2; // min-fill-size
797 HeapWord* const obj_beg = dense_prefix_end - 1;
798 CollectedHeap::fill_with_object(obj_beg, obj_len);
799 _mark_bitmap.mark_obj(obj_beg);
800 _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
801 region_after_dense_prefix->set_partial_obj_size(1);
802 region_after_dense_prefix->set_partial_obj_addr(obj_beg);
803 assert(start_array(id) != nullptr, "sanity");
804 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
805 }
806 }
807
808 bool PSParallelCompact::check_maximum_compaction(bool should_do_max_compaction,
809 size_t total_live_words,
810 MutableSpace* const old_space,
811 HeapWord* full_region_prefix_end) {
812
813 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
814
815 // Check System.GC
816 bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
817 && GCCause::is_user_requested_gc(heap->gc_cause());
818
819 // Check if all live objs are too much for old-gen.
820 const bool is_old_gen_too_full = (total_live_words >= old_space->capacity_in_words());
821
822 // If all regions in old-gen are full
823 const bool is_region_full =
824 full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
825
826 return should_do_max_compaction
827 || is_max_on_system_gc
828 || is_old_gen_too_full
829 || is_region_full;
830 }
831
832 void PSParallelCompact::summary_phase(bool should_do_max_compaction)
833 {
834 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
835
836 MutableSpace* const old_space = _space_info[old_space_id].space();
837 {
838 size_t total_live_words = 0;
839 HeapWord* full_region_prefix_end = nullptr;
840 {
841 // old-gen
842 size_t live_words = _summary_data.live_words_in_space(old_space,
843 &full_region_prefix_end);
844 total_live_words += live_words;
845 }
846 // young-gen
847 for (uint i = eden_space_id; i < last_space_id; ++i) {
848 const MutableSpace* space = _space_info[i].space();
849 size_t live_words = _summary_data.live_words_in_space(space);
850 total_live_words += live_words;
851 _space_info[i].set_new_top(space->bottom() + live_words);
852 _space_info[i].set_dense_prefix(space->bottom());
853 }
854
855 should_do_max_compaction = check_maximum_compaction(should_do_max_compaction,
856 total_live_words,
857 old_space,
858 full_region_prefix_end);
859 {
860 GCTraceTime(Info, gc, phases) tm("Summary Phase: expand", &_gc_timer);
861 // Try to expand old-gen in order to fit all live objs and waste.
862 size_t target_capacity_bytes = total_live_words * HeapWordSize
863 + old_space->capacity_in_bytes() * (MarkSweepDeadRatio / 100);
864 ParallelScavengeHeap::heap()->old_gen()->try_expand_till_size(target_capacity_bytes);
865 }
866
867 HeapWord* dense_prefix_end = should_do_max_compaction
868 ? full_region_prefix_end
869 : compute_dense_prefix_for_old_space(old_space,
870 full_region_prefix_end);
871 SpaceId id = old_space_id;
872 _space_info[id].set_dense_prefix(dense_prefix_end);
873
874 if (dense_prefix_end != old_space->bottom()) {
875 fill_dense_prefix_end(id);
876 }
877
878 // Compacting objs in [dense_prefix_end, old_space->top())
879 _summary_data.summarize(_space_info[id].split_info(),
880 dense_prefix_end, old_space->top(), nullptr,
881 dense_prefix_end, old_space->end(),
882 _space_info[id].new_top_addr());
883 }
884
885 // Summarize the remaining spaces in the young gen. The initial target space
886 // is the old gen. If a space does not fit entirely into the target, then the
887 // remainder is compacted into the space itself and that space becomes the new
888 // target.
889 SpaceId dst_space_id = old_space_id;
890 HeapWord* dst_space_end = old_space->end();
891 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
892 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
893 const MutableSpace* space = _space_info[id].space();
894 const size_t live = pointer_delta(_space_info[id].new_top(),
895 space->bottom());
896 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
897
898 if (live > 0 && live <= available) {
899 // All the live data will fit.
900 bool done = _summary_data.summarize(_space_info[id].split_info(),
901 space->bottom(), space->top(),
902 nullptr,
903 *new_top_addr, dst_space_end,
904 new_top_addr);
905 assert(done, "space must fit into old gen");
906
907 // Reset the new_top value for the space.
908 _space_info[id].set_new_top(space->bottom());
909 } else if (live > 0) {
910 // Attempt to fit part of the source space into the target space.
911 HeapWord* next_src_addr = nullptr;
912 bool done = _summary_data.summarize(_space_info[id].split_info(),
913 space->bottom(), space->top(),
914 &next_src_addr,
915 *new_top_addr, dst_space_end,
916 new_top_addr);
917 assert(!done, "space should not fit into old gen");
918 assert(next_src_addr != nullptr, "sanity");
919
920 // The source space becomes the new target, so the remainder is compacted
921 // within the space itself.
922 dst_space_id = SpaceId(id);
923 dst_space_end = space->end();
924 new_top_addr = _space_info[id].new_top_addr();
925 done = _summary_data.summarize(_space_info[id].split_info(),
926 next_src_addr, space->top(),
927 nullptr,
928 space->bottom(), dst_space_end,
929 new_top_addr);
930 assert(done, "space must fit when compacted into itself");
931 assert(*new_top_addr <= space->top(), "usage should not grow");
932 }
933 }
934 }
935
936 bool PSParallelCompact::invoke(bool clear_all_soft_refs, bool should_do_max_compaction) {
937 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
938 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
939 "should be in vm thread");
940 assert(ref_processor() != nullptr, "Sanity");
941
942 SvcGCMarker sgcm(SvcGCMarker::FULL);
943 IsSTWGCActiveMark mark;
944
945 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
946
947 GCIdMark gc_id_mark;
948 _gc_timer.register_gc_start();
949 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
950
951 GCCause::Cause gc_cause = heap->gc_cause();
952 PSOldGen* old_gen = heap->old_gen();
953 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
954
955 // Make sure data structures are sane, make the heap parsable, and do other
956 // miscellaneous bookkeeping.
957 pre_compact();
958
959 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
960
961 {
962 const uint active_workers =
963 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
964 ParallelScavengeHeap::heap()->workers().active_workers(),
965 Threads::number_of_non_daemon_threads());
966 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
967
968 GCTraceCPUTime tcpu(&_gc_tracer);
969 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
970
971 heap->pre_full_gc_dump(&_gc_timer);
972
973 TraceCollectorStats tcs(counters());
974 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
975
976 if (log_is_enabled(Debug, gc, heap, exit)) {
977 accumulated_time()->start();
978 }
979
980 // Let the size policy know we're starting
981 size_policy->major_collection_begin();
982
983 #if COMPILER2_OR_JVMCI
984 DerivedPointerTable::clear();
985 #endif
986
987 ref_processor()->start_discovery(clear_all_soft_refs);
988
989 marking_phase(&_gc_tracer);
990
991 summary_phase(should_do_max_compaction);
992
993 #if COMPILER2_OR_JVMCI
994 assert(DerivedPointerTable::is_active(), "Sanity");
995 DerivedPointerTable::set_active(false);
996 #endif
997
998 forward_to_new_addr();
999
1000 adjust_pointers();
1001
1002 compact();
1003
1004 ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1005
1006 ParCompactionManager::verify_all_region_stack_empty();
1007
1008 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1009 // done before resizing.
1010 post_compact();
1011
1012 size_policy->major_collection_end();
1013
1014 size_policy->sample_old_gen_used_bytes(MAX2(pre_gc_values.old_gen_used(), old_gen->used_in_bytes()));
1015
1016 if (UseAdaptiveSizePolicy) {
1017 heap->resize_after_full_gc();
1018 }
1019
1020 heap->resize_all_tlabs();
1021
1022 // Resize the metaspace capacity after a collection
1023 MetaspaceGC::compute_new_size();
1024
1025 if (log_is_enabled(Debug, gc, heap, exit)) {
1026 accumulated_time()->stop();
1027 }
1028
1029 heap->print_heap_change(pre_gc_values);
1030
1031 // Track memory usage and detect low memory
1032 MemoryService::track_memory_usage();
1033 heap->update_counters();
1034
1035 heap->post_full_gc_dump(&_gc_timer);
1036
1037 size_policy->record_gc_pause_end_instant();
1038 }
1039
1040 heap->gc_epilogue(true);
1041
1042 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1043 Universe::verify("After GC");
1044 }
1045
1046 heap->print_after_gc();
1047 heap->trace_heap_after_gc(&_gc_tracer);
1048
1049 _gc_timer.register_gc_end();
1050
1051 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1052 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1053
1054 return true;
1055 }
1056
1057 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1058 ParCompactionManager* _cm;
1059
1060 public:
1061 PCAddThreadRootsMarkingTaskClosure(ParCompactionManager* cm) : _cm(cm) { }
1062 void do_thread(Thread* thread) {
1063 ResourceMark rm;
1064
1065 MarkingNMethodClosure mark_and_push_in_blobs(&_cm->_mark_and_push_closure);
1066
1067 thread->oops_do(&_cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1068
1069 // Do the real work
1070 _cm->follow_marking_stacks();
1071 }
1072 };
1073
1074 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1075 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1076
1077 ParCompactionManager* cm =
1078 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1079
1080 do {
1081 ScannerTask task;
1082 if (ParCompactionManager::steal(worker_id, task)) {
1083 cm->follow_contents(task, true);
1084 }
1085 cm->follow_marking_stacks();
1086 } while (!terminator.offer_termination());
1087 }
1088
1089 class MarkFromRootsTask : public WorkerTask {
1090 NMethodMarkingScope _nmethod_marking_scope;
1091 ThreadsClaimTokenScope _threads_claim_token_scope;
1092 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1093 TaskTerminator _terminator;
1094 uint _active_workers;
1095
1096 public:
1097 MarkFromRootsTask(uint active_workers) :
1098 WorkerTask("MarkFromRootsTask"),
1099 _nmethod_marking_scope(),
1100 _threads_claim_token_scope(),
1101 _terminator(active_workers, ParCompactionManager::marking_stacks()),
1102 _active_workers(active_workers) {}
1103
1104 virtual void work(uint worker_id) {
1105 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1106 cm->create_marking_stats_cache();
1107 {
1108 CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1109 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1110
1111 // Do the real work
1112 cm->follow_marking_stacks();
1113 }
1114
1115 {
1116 PCAddThreadRootsMarkingTaskClosure closure(cm);
1117 Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1118 }
1119
1120 // Mark from OopStorages
1121 {
1122 _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1123 // Do the real work
1124 cm->follow_marking_stacks();
1125 }
1126
1127 if (_active_workers > 1) {
1128 steal_marking_work(_terminator, worker_id);
1129 }
1130 }
1131 };
1132
1133 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1134 TaskTerminator _terminator;
1135
1136 public:
1137 ParallelCompactRefProcProxyTask(uint max_workers)
1138 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1139 _terminator(_max_workers, ParCompactionManager::marking_stacks()) {}
1140
1141 void work(uint worker_id) override {
1142 assert(worker_id < _max_workers, "sanity");
1143 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1144 BarrierEnqueueDiscoveredFieldClosure enqueue;
1145 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1146 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1147 }
1148
1149 void prepare_run_task_hook() override {
1150 _terminator.reset_for_reuse(_queue_count);
1151 }
1152 };
1153
1154 static void flush_marking_stats_cache(const uint num_workers) {
1155 for (uint i = 0; i < num_workers; ++i) {
1156 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1157 cm->flush_and_destroy_marking_stats_cache();
1158 }
1159 }
1160
1161 class PSParallelCleaningTask : public WorkerTask {
1162 bool _unloading_occurred;
1163 CodeCacheUnloadingTask _code_cache_task;
1164 // Prune dead klasses from subklass/sibling/implementor lists.
1165 KlassCleaningTask _klass_cleaning_task;
1166
1167 public:
1168 PSParallelCleaningTask(bool unloading_occurred) :
1169 WorkerTask("PS Parallel Cleaning"),
1170 _unloading_occurred(unloading_occurred),
1171 _code_cache_task(unloading_occurred),
1172 _klass_cleaning_task() {}
1173
1174 void work(uint worker_id) {
1175 #if INCLUDE_JVMCI
1176 if (EnableJVMCI && worker_id == 0) {
1177 // Serial work; only first worker.
1178 // Clean JVMCI metadata handles.
1179 JVMCI::do_unloading(_unloading_occurred);
1180 }
1181 #endif
1182
1183 // Do first pass of code cache cleaning.
1184 _code_cache_task.work(worker_id);
1185
1186 // Clean all klasses that were not unloaded.
1187 // The weak metadata in klass doesn't need to be
1188 // processed if there was no unloading.
1189 if (_unloading_occurred) {
1190 _klass_cleaning_task.work();
1191 }
1192 }
1193 };
1194
1195 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1196 // Recursively traverse all live objects and mark them
1197 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1198
1199 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1200
1201 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1202 {
1203 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1204
1205 MarkFromRootsTask task(active_gc_threads);
1206 ParallelScavengeHeap::heap()->workers().run_task(&task);
1207 }
1208
1209 // Process reference objects found during marking
1210 {
1211 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1212
1213 ReferenceProcessorStats stats;
1214 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1215
1216 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1217 stats = ref_processor()->process_discovered_references(task, &ParallelScavengeHeap::heap()->workers(), pt);
1218
1219 gc_tracer->report_gc_reference_stats(stats);
1220 pt.print_all_references();
1221 }
1222
1223 {
1224 GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1225
1226 flush_marking_stats_cache(active_gc_threads);
1227 }
1228
1229 // This is the point where the entire marking should have completed.
1230 ParCompactionManager::verify_all_marking_stack_empty();
1231
1232 {
1233 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1234 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1235 is_alive_closure(),
1236 &do_nothing_cl,
1237 1);
1238 }
1239
1240 {
1241 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1242
1243 ClassUnloadingContext ctx(active_gc_threads /* num_nmethod_unlink_workers */,
1244 false /* unregister_nmethods_during_purge */,
1245 false /* lock_nmethod_free_separately */);
1246
1247 {
1248 CodeCache::UnlinkingScope scope(is_alive_closure());
1249
1250 // Follow system dictionary roots and unload classes.
1251 bool unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1252
1253 PSParallelCleaningTask task{unloading_occurred};
1254 ParallelScavengeHeap::heap()->workers().run_task(&task);
1255 }
1256
1257 {
1258 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1259 // Release unloaded nmethod's memory.
1260 ctx.purge_nmethods();
1261 }
1262 {
1263 GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1264 ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1265 }
1266 {
1267 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1268 ctx.free_nmethods();
1269 }
1270 {
1271 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1272 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
1273 ClassLoaderDataGraph::purge(true /* at_safepoint */);
1274 DEBUG_ONLY(MetaspaceUtils::verify();)
1275 }
1276 }
1277
1278 {
1279 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
1280 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers());
1281 }
1282 #if TASKQUEUE_STATS
1283 ParCompactionManager::print_and_reset_taskqueue_stats();
1284 #endif
1285 }
1286
1287 template<typename Func>
1288 void PSParallelCompact::adjust_in_space_helper(SpaceId id, volatile uint* claim_counter, Func&& on_stripe) {
1289 MutableSpace* sp = PSParallelCompact::space(id);
1290 HeapWord* const bottom = sp->bottom();
1291 HeapWord* const top = sp->top();
1292 if (bottom == top) {
1293 return;
1294 }
1295
1296 const uint num_regions_per_stripe = 2;
1297 const size_t region_size = ParallelCompactData::RegionSize;
1298 const size_t stripe_size = num_regions_per_stripe * region_size;
1299
1300 while (true) {
1301 uint counter = AtomicAccess::fetch_then_add(claim_counter, num_regions_per_stripe);
1302 HeapWord* cur_stripe = bottom + counter * region_size;
1303 if (cur_stripe >= top) {
1304 break;
1305 }
1306 HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1307 on_stripe(cur_stripe, stripe_end);
1308 }
1309 }
1310
1311 void PSParallelCompact::adjust_in_old_space(volatile uint* claim_counter) {
1312 // Regions in old-space shouldn't be split.
1313 assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1314
1315 auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1316 assert(mark_bitmap()->is_marked(obj_start), "inv");
1317 oop obj = cast_to_oop(obj_start);
1318 return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1319 };
1320
1321 adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1322 assert(_summary_data.is_region_aligned(stripe_start), "inv");
1323 RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1324 HeapWord* obj_start;
1325 if (cur_region->partial_obj_size() != 0) {
1326 obj_start = cur_region->partial_obj_addr();
1327 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1328 } else {
1329 obj_start = stripe_start;
1330 }
1331
1332 while (obj_start < stripe_end) {
1333 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1334 if (obj_start >= stripe_end) {
1335 break;
1336 }
1337 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1338 }
1339 });
1340 }
1341
1342 void PSParallelCompact::adjust_in_young_space(SpaceId id, volatile uint* claim_counter) {
1343 adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1344 HeapWord* obj_start = stripe_start;
1345 while (obj_start < stripe_end) {
1346 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1347 if (obj_start >= stripe_end) {
1348 break;
1349 }
1350 oop obj = cast_to_oop(obj_start);
1351 obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1352 }
1353 });
1354 }
1355
1356 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, volatile uint* claim_counters) {
1357 auto start_time = Ticks::now();
1358 adjust_in_old_space(&claim_counters[0]);
1359 for (uint id = eden_space_id; id < last_space_id; ++id) {
1360 adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1361 }
1362 log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1363 }
1364
1365 class PSAdjustTask final : public WorkerTask {
1366 ThreadsClaimTokenScope _threads_claim_token_scope;
1367 WeakProcessor::Task _weak_proc_task;
1368 OopStorageSetStrongParState<false, false> _oop_storage_iter;
1369 uint _nworkers;
1370 volatile bool _code_cache_claimed;
1371 volatile uint _claim_counters[PSParallelCompact::last_space_id] = {};
1372
1373 bool try_claim_code_cache_task() {
1374 return AtomicAccess::load(&_code_cache_claimed) == false
1375 && AtomicAccess::cmpxchg(&_code_cache_claimed, false, true) == false;
1376 }
1377
1378 public:
1379 PSAdjustTask(uint nworkers) :
1380 WorkerTask("PSAdjust task"),
1381 _threads_claim_token_scope(),
1382 _weak_proc_task(nworkers),
1383 _oop_storage_iter(),
1384 _nworkers(nworkers),
1385 _code_cache_claimed(false) {
1386
1387 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1388 }
1389
1390 void work(uint worker_id) {
1391 {
1392 // Pointers in heap.
1393 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1394 cm->preserved_marks()->adjust_during_full_gc();
1395
1396 PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1397 }
1398
1399 {
1400 // All (strong and weak) CLDs.
1401 CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1402 ClassLoaderDataGraph::cld_do(&cld_closure);
1403 }
1404
1405 {
1406 // Threads stack frames. No need to visit on-stack nmethods, because all
1407 // nmethods are visited in one go via CodeCache::nmethods_do.
1408 ResourceMark rm;
1409 Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1410 if (try_claim_code_cache_task()) {
1411 NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1412 CodeCache::nmethods_do(&adjust_code);
1413 }
1414 }
1415
1416 {
1417 // VM internal strong and weak roots.
1418 _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1419 AlwaysTrueClosure always_alive;
1420 _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1421 }
1422 }
1423 };
1424
1425 void PSParallelCompact::adjust_pointers() {
1426 // Adjust the pointers to reflect the new locations
1427 GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1428 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1429 PSAdjustTask task(nworkers);
1430 ParallelScavengeHeap::heap()->workers().run_task(&task);
1431 }
1432
1433 // Split [start, end) evenly for a number of workers and return the
1434 // range for worker_id.
1435 static void split_regions_for_worker(size_t start, size_t end,
1436 uint worker_id, uint num_workers,
1437 size_t* worker_start, size_t* worker_end) {
1438 assert(start < end, "precondition");
1439 assert(num_workers > 0, "precondition");
1440 assert(worker_id < num_workers, "precondition");
1441
1442 size_t num_regions = end - start;
1443 size_t num_regions_per_worker = num_regions / num_workers;
1444 size_t remainder = num_regions % num_workers;
1445 // The first few workers will get one extra.
1446 *worker_start = start + worker_id * num_regions_per_worker
1447 + MIN2(checked_cast<size_t>(worker_id), remainder);
1448 *worker_end = *worker_start + num_regions_per_worker
1449 + (worker_id < remainder ? 1 : 0);
1450 }
1451
1452 void PSParallelCompact::forward_to_new_addr() {
1453 GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1454 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1455
1456 struct ForwardTask final : public WorkerTask {
1457 uint _num_workers;
1458
1459 explicit ForwardTask(uint num_workers) :
1460 WorkerTask("PSForward task"),
1461 _num_workers(num_workers) {}
1462
1463 static void forward_objs_in_range(ParCompactionManager* cm,
1464 HeapWord* start,
1465 HeapWord* end,
1466 HeapWord* destination) {
1467 HeapWord* cur_addr = start;
1468 HeapWord* new_addr = destination;
1469
1470 while (cur_addr < end) {
1471 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, end);
1472 if (cur_addr >= end) {
1473 return;
1474 }
1475 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1476 oop obj = cast_to_oop(cur_addr);
1477 if (new_addr != cur_addr) {
1478 cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1479 FullGCForwarding::forward_to(obj, cast_to_oop(new_addr));
1480 }
1481 size_t obj_size = obj->size();
1482 new_addr += obj_size;
1483 cur_addr += obj_size;
1484 }
1485 }
1486
1487 void work(uint worker_id) override {
1488 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1489 for (uint id = old_space_id; id < last_space_id; ++id) {
1490 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1491 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1492 HeapWord* top = sp->top();
1493
1494 if (dense_prefix_addr == top) {
1495 // Empty space
1496 continue;
1497 }
1498
1499 const SplitInfo& split_info = _space_info[SpaceId(id)].split_info();
1500 size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1501 size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1502 size_t start_region;
1503 size_t end_region;
1504 split_regions_for_worker(dense_prefix_region, top_region,
1505 worker_id, _num_workers,
1506 &start_region, &end_region);
1507 for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1508 RegionData* region_ptr = _summary_data.region(cur_region);
1509 size_t partial_obj_size = region_ptr->partial_obj_size();
1510
1511 if (partial_obj_size == ParallelCompactData::RegionSize) {
1512 // No obj-start
1513 continue;
1514 }
1515
1516 HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1517 HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1518
1519 if (split_info.is_split(cur_region)) {
1520 // Part 1: will be relocated to space-1
1521 HeapWord* preceding_destination = split_info.preceding_destination();
1522 HeapWord* split_point = split_info.split_point();
1523 forward_objs_in_range(cm, region_start + partial_obj_size, split_point, preceding_destination + partial_obj_size);
1524
1525 // Part 2: will be relocated to space-2
1526 HeapWord* destination = region_ptr->destination();
1527 forward_objs_in_range(cm, split_point, region_end, destination);
1528 } else {
1529 HeapWord* destination = region_ptr->destination();
1530 forward_objs_in_range(cm, region_start + partial_obj_size, region_end, destination + partial_obj_size);
1531 }
1532 }
1533 }
1534 }
1535 } task(nworkers);
1536
1537 ParallelScavengeHeap::heap()->workers().run_task(&task);
1538 DEBUG_ONLY(verify_forward();)
1539 }
1540
1541 #ifdef ASSERT
1542 void PSParallelCompact::verify_forward() {
1543 HeapWord* const old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1544 // The destination addr for the first live obj after dense-prefix.
1545 HeapWord* bump_ptr = old_dense_prefix_addr
1546 + _summary_data.addr_to_region_ptr(old_dense_prefix_addr)->partial_obj_size();
1547 SpaceId bump_ptr_space = old_space_id;
1548
1549 for (uint id = old_space_id; id < last_space_id; ++id) {
1550 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1551 // Only verify objs after dense-prefix, because those before dense-prefix are not moved (forwarded).
1552 HeapWord* cur_addr = dense_prefix(SpaceId(id));
1553 HeapWord* top = sp->top();
1554
1555 while (cur_addr < top) {
1556 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1557 if (cur_addr >= top) {
1558 break;
1559 }
1560 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1561 assert(bump_ptr <= _space_info[bump_ptr_space].new_top(), "inv");
1562 // Move to the space containing cur_addr
1563 if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1564 bump_ptr = space(space_id(cur_addr))->bottom();
1565 bump_ptr_space = space_id(bump_ptr);
1566 }
1567 oop obj = cast_to_oop(cur_addr);
1568 if (cur_addr == bump_ptr) {
1569 assert(!FullGCForwarding::is_forwarded(obj), "inv");
1570 } else {
1571 assert(FullGCForwarding::forwardee(obj) == cast_to_oop(bump_ptr), "inv");
1572 }
1573 bump_ptr += obj->size();
1574 cur_addr += obj->size();
1575 }
1576 }
1577 }
1578 #endif
1579
1580 // Helper class to print 8 region numbers per line and then print the total at the end.
1581 class FillableRegionLogger : public StackObj {
1582 private:
1583 Log(gc, compaction) log;
1584 static const int LineLength = 8;
1585 size_t _regions[LineLength];
1586 int _next_index;
1587 bool _enabled;
1588 size_t _total_regions;
1589 public:
1590 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1591 ~FillableRegionLogger() {
1592 log.trace("%zu initially fillable regions", _total_regions);
1593 }
1594
1595 void print_line() {
1596 if (!_enabled || _next_index == 0) {
1597 return;
1598 }
1599 FormatBuffer<> line("Fillable: ");
1600 for (int i = 0; i < _next_index; i++) {
1601 line.append(" %7zu", _regions[i]);
1602 }
1603 log.trace("%s", line.buffer());
1604 _next_index = 0;
1605 }
1606
1607 void handle(size_t region) {
1608 if (!_enabled) {
1609 return;
1610 }
1611 _regions[_next_index++] = region;
1612 if (_next_index == LineLength) {
1613 print_line();
1614 }
1615 _total_regions++;
1616 }
1617 };
1618
1619 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1620 {
1621 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1622
1623 // Find the threads that are active
1624 uint worker_id = 0;
1625
1626 // Find all regions that are available (can be filled immediately) and
1627 // distribute them to the thread stacks. The iteration is done in reverse
1628 // order (high to low) so the regions will be removed in ascending order.
1629
1630 const ParallelCompactData& sd = PSParallelCompact::summary_data();
1631
1632 // id + 1 is used to test termination so unsigned can
1633 // be used with an old_space_id == 0.
1634 FillableRegionLogger region_logger;
1635 for (unsigned int id = last_space_id - 1; id + 1 > old_space_id; --id) {
1636 SpaceInfo* const space_info = _space_info + id;
1637 HeapWord* const new_top = space_info->new_top();
1638
1639 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1640 const size_t end_region =
1641 sd.addr_to_region_idx(sd.region_align_up(new_top));
1642
1643 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1644 if (sd.region(cur)->claim_unsafe()) {
1645 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1646 bool result = sd.region(cur)->mark_normal();
1647 assert(result, "Must succeed at this point.");
1648 cm->region_stack()->push(cur);
1649 region_logger.handle(cur);
1650 // Assign regions to tasks in round-robin fashion.
1651 if (++worker_id == parallel_gc_threads) {
1652 worker_id = 0;
1653 }
1654 }
1655 }
1656 region_logger.print_line();
1657 }
1658 }
1659
1660 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1661 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1662
1663 ParCompactionManager* cm =
1664 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1665
1666 // Drain the stacks that have been preloaded with regions
1667 // that are ready to fill.
1668
1669 cm->drain_region_stacks();
1670
1671 guarantee(cm->region_stack()->is_empty(), "Not empty");
1672
1673 size_t region_index = 0;
1674
1675 while (true) {
1676 if (ParCompactionManager::steal(worker_id, region_index)) {
1677 PSParallelCompact::fill_and_update_region(cm, region_index);
1678 cm->drain_region_stacks();
1679 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1680 // Fill and update an unavailable region with the help of a shadow region
1681 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1682 cm->drain_region_stacks();
1683 } else {
1684 if (terminator->offer_termination()) {
1685 break;
1686 }
1687 // Go around again.
1688 }
1689 }
1690 }
1691
1692 class FillDensePrefixAndCompactionTask: public WorkerTask {
1693 TaskTerminator _terminator;
1694
1695 public:
1696 FillDensePrefixAndCompactionTask(uint active_workers) :
1697 WorkerTask("FillDensePrefixAndCompactionTask"),
1698 _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1699 }
1700
1701 virtual void work(uint worker_id) {
1702 if (worker_id == 0) {
1703 auto start = Ticks::now();
1704 PSParallelCompact::fill_dead_objs_in_dense_prefix();
1705 log_trace(gc, phases)("Fill dense prefix by worker 0: %.3f ms", (Ticks::now() - start).seconds() * 1000);
1706 }
1707 compaction_with_stealing_work(&_terminator, worker_id);
1708 }
1709 };
1710
1711 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1712 #ifdef ASSERT
1713 {
1714 assert(start < end, "precondition");
1715 assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1716 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1717 if (start != bottom) {
1718 // The preceding live obj.
1719 HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1720 HeapWord* obj_end = obj_start + cast_to_oop(obj_start)->size();
1721 assert(obj_end == start, "precondition");
1722 }
1723 }
1724 #endif
1725
1726 CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1727 HeapWord* addr = start;
1728 do {
1729 size_t size = cast_to_oop(addr)->size();
1730 start_array(old_space_id)->update_for_block(addr, addr + size);
1731 addr += size;
1732 } while (addr < end);
1733 }
1734
1735 void PSParallelCompact::fill_dead_objs_in_dense_prefix() {
1736 ParMarkBitMap* bitmap = mark_bitmap();
1737
1738 HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1739 HeapWord* const prefix_end = dense_prefix(old_space_id);
1740
1741 const size_t region_size = ParallelCompactData::RegionSize;
1742
1743 // Fill dead space in [start_addr, end_addr)
1744 HeapWord* const start_addr = bottom;
1745 HeapWord* const end_addr = prefix_end;
1746
1747 for (HeapWord* cur_addr = start_addr; cur_addr < end_addr; /* empty */) {
1748 RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1749 if (cur_region_ptr->data_size() == region_size) {
1750 // Full; no dead space. Next region.
1751 if (_summary_data.is_region_aligned(cur_addr)) {
1752 cur_addr += region_size;
1753 } else {
1754 cur_addr = _summary_data.region_align_up(cur_addr);
1755 }
1756 continue;
1757 }
1758
1759 // Fill dead space inside cur_region.
1760 if (_summary_data.is_region_aligned(cur_addr)) {
1761 cur_addr += cur_region_ptr->partial_obj_size();
1762 }
1763
1764 HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1765 assert(region_end_addr <= end_addr, "inv");
1766 while (cur_addr < region_end_addr) {
1767 // Use end_addr to allow filler-obj to cross region boundary.
1768 HeapWord* live_start = bitmap->find_obj_beg(cur_addr, end_addr);
1769 if (cur_addr != live_start) {
1770 // Found dead space [cur_addr, live_start).
1771 fill_range_in_dense_prefix(cur_addr, live_start);
1772 }
1773 if (live_start >= region_end_addr) {
1774 cur_addr = live_start;
1775 break;
1776 }
1777 assert(bitmap->is_marked(live_start), "inv");
1778 cur_addr = live_start + cast_to_oop(live_start)->size();
1779 }
1780 }
1781 }
1782
1783 void PSParallelCompact::compact() {
1784 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1785
1786 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1787
1788 initialize_shadow_regions(active_gc_threads);
1789 prepare_region_draining_tasks(active_gc_threads);
1790
1791 {
1792 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1793
1794 FillDensePrefixAndCompactionTask task(active_gc_threads);
1795 ParallelScavengeHeap::heap()->workers().run_task(&task);
1796
1797 #ifdef ASSERT
1798 verify_filler_in_dense_prefix();
1799
1800 // Verify that all regions have been processed.
1801 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1802 verify_complete(SpaceId(id));
1803 }
1804 #endif
1805 }
1806 }
1807
1808 #ifdef ASSERT
1809 void PSParallelCompact::verify_filler_in_dense_prefix() {
1810 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1811 HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1812
1813 const size_t region_size = ParallelCompactData::RegionSize;
1814
1815 for (HeapWord* cur_addr = bottom; cur_addr < dense_prefix_end; /* empty */) {
1816 RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1817 if (cur_region_ptr->data_size() == region_size) {
1818 // Full; no dead space. Next region.
1819 if (_summary_data.is_region_aligned(cur_addr)) {
1820 cur_addr += region_size;
1821 } else {
1822 cur_addr = _summary_data.region_align_up(cur_addr);
1823 }
1824 continue;
1825 }
1826
1827 // This region contains filler objs.
1828 if (_summary_data.is_region_aligned(cur_addr)) {
1829 cur_addr += cur_region_ptr->partial_obj_size();
1830 }
1831
1832 HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1833 assert(region_end_addr <= dense_prefix_end, "inv");
1834
1835 while (cur_addr < region_end_addr) {
1836 oop obj = cast_to_oop(cur_addr);
1837 oopDesc::verify(obj);
1838 if (!mark_bitmap()->is_marked(cur_addr)) {
1839 Klass* k = cast_to_oop(cur_addr)->klass();
1840 assert(k == Universe::fillerArrayKlass() || k == vmClasses::FillerObject_klass(), "inv");
1841 }
1842 cur_addr += obj->size();
1843 }
1844 }
1845 }
1846
1847 void PSParallelCompact::verify_complete(SpaceId space_id) {
1848 // All Regions served as compaction targets, from dense_prefix() to
1849 // new_top(), should be marked as filled and all Regions between new_top()
1850 // and top() should be available (i.e., should have been emptied).
1851 ParallelCompactData& sd = summary_data();
1852 SpaceInfo si = _space_info[space_id];
1853 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1854 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1855 const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1856 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1857 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1858
1859 size_t cur_region;
1860 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1861 const RegionData* const c = sd.region(cur_region);
1862 assert(c->completed(), "region %zu not filled: destination_count=%u",
1863 cur_region, c->destination_count());
1864 }
1865
1866 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1867 const RegionData* const c = sd.region(cur_region);
1868 assert(c->available(), "region %zu not empty: destination_count=%u",
1869 cur_region, c->destination_count());
1870 }
1871 }
1872 #endif // #ifdef ASSERT
1873
1874 // Return the SpaceId for the space containing addr. If addr is not in the
1875 // heap, last_space_id is returned. In debug mode it expects the address to be
1876 // in the heap and asserts such.
1877 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1878 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1879
1880 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1881 if (_space_info[id].space()->contains(addr)) {
1882 return SpaceId(id);
1883 }
1884 }
1885
1886 assert(false, "no space contains the addr");
1887 return last_space_id;
1888 }
1889
1890 // Skip over count live words starting from beg, and return the address of the
1891 // next live word. Callers must also ensure that there are enough live words in
1892 // the range [beg, end) to skip.
1893 HeapWord* PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1894 {
1895 ParMarkBitMap* m = mark_bitmap();
1896 HeapWord* cur_addr = beg;
1897 while (true) {
1898 cur_addr = m->find_obj_beg(cur_addr, end);
1899 assert(cur_addr < end, "inv");
1900 size_t obj_size = cast_to_oop(cur_addr)->size();
1901 // Strictly greater-than
1902 if (obj_size > count) {
1903 return cur_addr + count;
1904 }
1905 count -= obj_size;
1906 cur_addr += obj_size;
1907 }
1908 }
1909
1910 // On starting to fill a destination region (dest-region), we need to know the
1911 // location of the word that will be at the start of the dest-region after
1912 // compaction. A dest-region can have one or more source regions, but only the
1913 // first source-region contains this location. This location is retrieved by
1914 // calling `first_src_addr` on a dest-region.
1915 // Conversely, a source-region has a dest-region which holds the destination of
1916 // the first live word on this source-region, based on which the destination
1917 // for the rest of live words can be derived.
1918 //
1919 // Note:
1920 // There is some complication due to space-boundary-fragmentation (an obj can't
1921 // cross space-boundary) -- a source-region may be split and behave like two
1922 // distinct regions with their own dest-region, as depicted below.
1923 //
1924 // source-region: region-n
1925 //
1926 // **********************
1927 // | A|A~~~~B|B |
1928 // **********************
1929 // n-1 n n+1
1930 //
1931 // AA, BB denote two live objs. ~~~~ denotes unknown number of live objs.
1932 //
1933 // Assuming the dest-region for region-n is the final region before
1934 // old-space-end and its first-live-word is the middle of AA, the heap content
1935 // will look like the following after compaction:
1936 //
1937 // ************** *************
1938 // A|A~~~~ | |BB |
1939 // ************** *************
1940 // ^ ^
1941 // | old-space-end | eden-space-start
1942 //
1943 // Therefore, in this example, region-n will have two dest-regions:
1944 // 1. the final region in old-space
1945 // 2. the first region in eden-space.
1946 // To handle this special case, we introduce the concept of split-region, whose
1947 // contents are relocated to two spaces. `SplitInfo` captures all necessary
1948 // info about the split, the first part, spliting-point, and the second part.
1949 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1950 SpaceId src_space_id,
1951 size_t src_region_idx)
1952 {
1953 const size_t RegionSize = ParallelCompactData::RegionSize;
1954 const ParallelCompactData& sd = summary_data();
1955 assert(sd.is_region_aligned(dest_addr), "precondition");
1956
1957 const RegionData* const src_region_ptr = sd.region(src_region_idx);
1958 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1959
1960 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1961 HeapWord* const src_region_destination = src_region_ptr->destination();
1962
1963 HeapWord* const region_start = sd.region_to_addr(src_region_idx);
1964 HeapWord* const region_end = sd.region_to_addr(src_region_idx) + RegionSize;
1965
1966 // Identify the actual destination for the first live words on this region,
1967 // taking split-region into account.
1968 HeapWord* region_start_destination;
1969 const SplitInfo& split_info = _space_info[src_space_id].split_info();
1970 if (split_info.is_split(src_region_idx)) {
1971 // The second part of this split region; use the recorded split point.
1972 if (dest_addr == src_region_destination) {
1973 return split_info.split_point();
1974 }
1975 region_start_destination = split_info.preceding_destination();
1976 } else {
1977 region_start_destination = src_region_destination;
1978 }
1979
1980 // Calculate the offset to be skipped
1981 size_t words_to_skip = pointer_delta(dest_addr, region_start_destination);
1982
1983 HeapWord* result;
1984 if (partial_obj_size > words_to_skip) {
1985 result = region_start + words_to_skip;
1986 } else {
1987 words_to_skip -= partial_obj_size;
1988 result = skip_live_words(region_start + partial_obj_size, region_end, words_to_skip);
1989 }
1990
1991 if (split_info.is_split(src_region_idx)) {
1992 assert(result < split_info.split_point(), "postcondition");
1993 } else {
1994 assert(result < region_end, "postcondition");
1995 }
1996
1997 return result;
1998 }
1999
2000 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2001 SpaceId src_space_id,
2002 size_t beg_region,
2003 HeapWord* end_addr)
2004 {
2005 ParallelCompactData& sd = summary_data();
2006
2007 #ifdef ASSERT
2008 MutableSpace* const src_space = _space_info[src_space_id].space();
2009 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2010 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2011 "src_space_id does not match beg_addr");
2012 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2013 "src_space_id does not match end_addr");
2014 #endif // #ifdef ASSERT
2015
2016 RegionData* const beg = sd.region(beg_region);
2017 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2018
2019 // Regions up to new_top() are enqueued if they become available.
2020 HeapWord* const new_top = _space_info[src_space_id].new_top();
2021 RegionData* const enqueue_end =
2022 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2023
2024 for (RegionData* cur = beg; cur < end; ++cur) {
2025 assert(cur->data_size() > 0, "region must have live data");
2026 cur->decrement_destination_count();
2027 if (cur < enqueue_end && cur->available() && cur->claim()) {
2028 if (cur->mark_normal()) {
2029 cm->push_region(sd.region(cur));
2030 } else if (cur->mark_copied()) {
2031 // Try to copy the content of the shadow region back to its corresponding
2032 // heap region if the shadow region is filled. Otherwise, the GC thread
2033 // fills the shadow region will copy the data back (see
2034 // MoveAndUpdateShadowClosure::complete_region).
2035 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2036 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2037 cur->set_completed();
2038 }
2039 }
2040 }
2041 }
2042
2043 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2044 SpaceId& src_space_id,
2045 HeapWord*& src_space_top,
2046 HeapWord* end_addr)
2047 {
2048 ParallelCompactData& sd = PSParallelCompact::summary_data();
2049
2050 size_t src_region_idx = 0;
2051
2052 // Skip empty regions (if any) up to the top of the space.
2053 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2054 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2055 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2056 const RegionData* const top_region_ptr = sd.addr_to_region_ptr(top_aligned_up);
2057
2058 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2059 ++src_region_ptr;
2060 }
2061
2062 if (src_region_ptr < top_region_ptr) {
2063 // Found the first non-empty region in the same space.
2064 src_region_idx = sd.region(src_region_ptr);
2065 closure.set_source(sd.region_to_addr(src_region_idx));
2066 return src_region_idx;
2067 }
2068
2069 // Switch to a new source space and find the first non-empty region.
2070 uint space_id = src_space_id + 1;
2071 assert(space_id < last_space_id, "not enough spaces");
2072
2073 for (/* empty */; space_id < last_space_id; ++space_id) {
2074 HeapWord* bottom = _space_info[space_id].space()->bottom();
2075 HeapWord* top = _space_info[space_id].space()->top();
2076 // Skip empty space
2077 if (bottom == top) {
2078 continue;
2079 }
2080
2081 // Identify the first region that contains live words in this space
2082 size_t cur_region = sd.addr_to_region_idx(bottom);
2083 size_t end_region = sd.addr_to_region_idx(sd.region_align_up(top));
2084
2085 for (/* empty */ ; cur_region < end_region; ++cur_region) {
2086 RegionData* cur = sd.region(cur_region);
2087 if (cur->live_obj_size() > 0) {
2088 HeapWord* region_start_addr = sd.region_to_addr(cur_region);
2089
2090 src_space_id = SpaceId(space_id);
2091 src_space_top = top;
2092 closure.set_source(region_start_addr);
2093 return cur_region;
2094 }
2095 }
2096 }
2097
2098 ShouldNotReachHere();
2099 }
2100
2101 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2102 ParallelCompactData& sd = summary_data();
2103 assert(sd.is_region_aligned(region_start_addr), "precondition");
2104
2105 // Use per-region partial_obj_size to locate the end of the obj, that extends
2106 // to region_start_addr.
2107 size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2108 size_t end_region_idx = sd.region_count();
2109 size_t accumulated_size = 0;
2110 for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2111 size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2112 accumulated_size += cur_partial_obj_size;
2113 if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2114 break;
2115 }
2116 }
2117 return region_start_addr + accumulated_size;
2118 }
2119
2120 // Use region_idx as the destination region, and evacuate all live objs on its
2121 // source regions to this destination region.
2122 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2123 {
2124 ParMarkBitMap* const bitmap = mark_bitmap();
2125 ParallelCompactData& sd = summary_data();
2126 RegionData* const region_ptr = sd.region(region_idx);
2127
2128 // Get the source region and related info.
2129 size_t src_region_idx = region_ptr->source_region();
2130 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2131 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2132 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2133
2134 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2135
2136 // Adjust src_region_idx to prepare for decrementing destination counts (the
2137 // destination count is not decremented when a region is copied to itself).
2138 if (src_region_idx == region_idx) {
2139 src_region_idx += 1;
2140 }
2141
2142 // source-region:
2143 //
2144 // **********
2145 // | ~~~ |
2146 // **********
2147 // ^
2148 // |-- closure.source() / first_src_addr
2149 //
2150 //
2151 // ~~~ : live words
2152 //
2153 // destination-region:
2154 //
2155 // **********
2156 // | |
2157 // **********
2158 // ^
2159 // |-- region-start
2160 if (bitmap->is_unmarked(closure.source())) {
2161 // An object overflows the previous destination region, so this
2162 // destination region should copy the remainder of the object or as much as
2163 // will fit.
2164 HeapWord* const old_src_addr = closure.source();
2165 {
2166 HeapWord* region_start = sd.region_align_down(closure.source());
2167 HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2168 HeapWord* obj_end;
2169 if (obj_start != closure.source()) {
2170 assert(bitmap->is_marked(obj_start), "inv");
2171 // Found the actual obj-start, try to find the obj-end using either
2172 // size() if this obj is completely contained in the current region.
2173 HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2174 HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2175 ? nullptr
2176 : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2177 // This obj extends to next region iff partial_obj_addr of the *next*
2178 // region is the same as obj-start.
2179 if (partial_obj_start == obj_start) {
2180 // This obj extends to next region.
2181 obj_end = partial_obj_end(next_region_start);
2182 } else {
2183 // Completely contained in this region; safe to use size().
2184 obj_end = obj_start + cast_to_oop(obj_start)->size();
2185 }
2186 } else {
2187 // This obj extends to current region.
2188 obj_end = partial_obj_end(region_start);
2189 }
2190 size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2191 closure.copy_partial_obj(partial_obj_size);
2192 }
2193
2194 if (closure.is_full()) {
2195 decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2196 closure.complete_region(dest_addr, region_ptr);
2197 return;
2198 }
2199
2200 // Finished copying without using up the current destination-region
2201 HeapWord* const end_addr = sd.region_align_down(closure.source());
2202 if (sd.region_align_down(old_src_addr) != end_addr) {
2203 assert(sd.region_align_up(old_src_addr) == end_addr, "only one region");
2204 // The partial object was copied from more than one source region.
2205 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2206
2207 // Move to the next source region, possibly switching spaces as well. All
2208 // args except end_addr may be modified.
2209 src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2210 }
2211 }
2212
2213 // Handle the rest obj-by-obj, where we know obj-start.
2214 do {
2215 HeapWord* cur_addr = closure.source();
2216 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2217 src_space_top);
2218 // To handle the case where the final obj in source region extends to next region.
2219 HeapWord* final_obj_start = (end_addr == src_space_top)
2220 ? nullptr
2221 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2222 // Apply closure on objs inside [cur_addr, end_addr)
2223 do {
2224 cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2225 if (cur_addr == end_addr) {
2226 break;
2227 }
2228 size_t obj_size;
2229 if (final_obj_start == cur_addr) {
2230 obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2231 } else {
2232 // This obj doesn't extend into next region; size() is safe to use.
2233 obj_size = cast_to_oop(cur_addr)->size();
2234 }
2235 closure.do_addr(cur_addr, obj_size);
2236 cur_addr += obj_size;
2237 } while (cur_addr < end_addr && !closure.is_full());
2238
2239 if (closure.is_full()) {
2240 decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2241 closure.complete_region(dest_addr, region_ptr);
2242 return;
2243 }
2244
2245 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2246
2247 // Move to the next source region, possibly switching spaces as well. All
2248 // args except end_addr may be modified.
2249 src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2250 } while (true);
2251 }
2252
2253 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2254 {
2255 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2256 fill_region(cm, cl, region_idx);
2257 }
2258
2259 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2260 {
2261 // Get a shadow region first
2262 ParallelCompactData& sd = summary_data();
2263 RegionData* const region_ptr = sd.region(region_idx);
2264 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2265 // The InvalidShadow return value indicates the corresponding heap region is available,
2266 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2267 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2268 if (shadow_region == ParCompactionManager::InvalidShadow) {
2269 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2270 region_ptr->shadow_to_normal();
2271 return fill_region(cm, cl, region_idx);
2272 } else {
2273 MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2274 return fill_region(cm, cl, region_idx);
2275 }
2276 }
2277
2278 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2279 {
2280 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2281 }
2282
2283 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
2284 {
2285 size_t next = cm->next_shadow_region();
2286 ParallelCompactData& sd = summary_data();
2287 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2288 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2289
2290 while (next < old_new_top) {
2291 if (sd.region(next)->mark_shadow()) {
2292 region_idx = next;
2293 return true;
2294 }
2295 next = cm->move_next_shadow_region_by(active_gc_threads);
2296 }
2297
2298 return false;
2299 }
2300
2301 // The shadow region is an optimization to address region dependencies in full GC. The basic
2302 // idea is making more regions available by temporally storing their live objects in empty
2303 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2304 // GC threads need not wait destination regions to be available before processing sources.
2305 //
2306 // A typical workflow would be:
2307 // After draining its own stack and failing to steal from others, a GC worker would pick an
2308 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2309 // the shadow region by copying live objects from source regions of the unavailable one. Once
2310 // the unavailable region becomes available, the data in the shadow region will be copied back.
2311 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2312 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2313 {
2314 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2315
2316 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2317 SpaceInfo* const space_info = _space_info + id;
2318 MutableSpace* const space = space_info->space();
2319
2320 const size_t beg_region =
2321 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2322 const size_t end_region =
2323 sd.addr_to_region_idx(sd.region_align_down(space->end()));
2324
2325 for (size_t cur = beg_region; cur < end_region; ++cur) {
2326 ParCompactionManager::push_shadow_region(cur);
2327 }
2328 }
2329
2330 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2331 for (uint i = 0; i < parallel_gc_threads; i++) {
2332 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2333 cm->set_next_shadow_region(beg_region + i);
2334 }
2335 }
2336
2337 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2338 {
2339 size_t words = MIN2(partial_obj_size, words_remaining());
2340
2341 // This test is necessary; if omitted, the pointer updates to a partial object
2342 // that crosses the dense prefix boundary could be overwritten.
2343 if (source() != copy_destination()) {
2344 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2345 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2346 }
2347 update_state(words);
2348 }
2349
2350 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2351 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2352 region_ptr->set_completed();
2353 }
2354
2355 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2356 assert(destination() != nullptr, "sanity");
2357 _source = addr;
2358
2359 // The start_array must be updated even if the object is not moving.
2360 if (_start_array != nullptr) {
2361 _start_array->update_for_block(destination(), destination() + words);
2362 }
2363
2364 // Avoid overflow
2365 words = MIN2(words, words_remaining());
2366 assert(words > 0, "inv");
2367
2368 if (copy_destination() != source()) {
2369 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2370 assert(source() != destination(), "inv");
2371 assert(FullGCForwarding::is_forwarded(cast_to_oop(source())), "inv");
2372 assert(FullGCForwarding::forwardee(cast_to_oop(source())) == cast_to_oop(destination()), "inv");
2373 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2374 cast_to_oop(copy_destination())->reinit_mark();
2375 }
2376
2377 update_state(words);
2378 }
2379
2380 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2381 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2382 // Record the shadow region index
2383 region_ptr->set_shadow_region(_shadow);
2384 // Mark the shadow region as filled to indicate the data is ready to be
2385 // copied back
2386 region_ptr->mark_filled();
2387 // Try to copy the content of the shadow region back to its corresponding
2388 // heap region if available; the GC thread that decreases the destination
2389 // count to zero will do the copying otherwise (see
2390 // PSParallelCompact::decrement_destination_counts).
2391 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2392 region_ptr->set_completed();
2393 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2394 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2395 }
2396 }