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 "compiler/oopMap.hpp"
32 #include "gc/parallel/objectStartArray.inline.hpp"
33 #include "gc/parallel/parallelArguments.hpp"
34 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
35 #include "gc/parallel/parMarkBitMap.inline.hpp"
36 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
37 #include "gc/parallel/psCompactionManager.inline.hpp"
38 #include "gc/parallel/psOldGen.hpp"
39 #include "gc/parallel/psParallelCompact.inline.hpp"
40 #include "gc/parallel/psPromotionManager.inline.hpp"
41 #include "gc/parallel/psRootType.hpp"
42 #include "gc/parallel/psScavenge.hpp"
43 #include "gc/parallel/psStringDedup.hpp"
44 #include "gc/parallel/psYoungGen.hpp"
45 #include "gc/shared/classUnloadingContext.hpp"
46 #include "gc/shared/fullGCForwarding.inline.hpp"
47 #include "gc/shared/gcCause.hpp"
48 #include "gc/shared/gcHeapSummary.hpp"
49 #include "gc/shared/gcId.hpp"
50 #include "gc/shared/gcLocker.hpp"
51 #include "gc/shared/gcTimer.hpp"
52 #include "gc/shared/gcTrace.hpp"
53 #include "gc/shared/gcTraceTime.inline.hpp"
54 #include "gc/shared/gcVMOperations.hpp"
55 #include "gc/shared/isGCActiveMark.hpp"
56 #include "gc/shared/oopStorage.inline.hpp"
57 #include "gc/shared/oopStorageSet.inline.hpp"
58 #include "gc/shared/oopStorageSetParState.inline.hpp"
59 #include "gc/shared/preservedMarks.inline.hpp"
60 #include "gc/shared/referencePolicy.hpp"
61 #include "gc/shared/referenceProcessor.hpp"
62 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
63 #include "gc/shared/spaceDecorator.hpp"
64 #include "gc/shared/strongRootsScope.hpp"
65 #include "gc/shared/taskTerminator.hpp"
66 #include "gc/shared/weakProcessor.inline.hpp"
67 #include "gc/shared/workerPolicy.hpp"
68 #include "gc/shared/workerThread.hpp"
69 #include "gc/shared/workerUtils.hpp"
70 #include "logging/log.hpp"
71 #include "memory/iterator.inline.hpp"
72 #include "memory/memoryReserver.hpp"
73 #include "memory/metaspaceUtils.hpp"
74 #include "memory/resourceArea.hpp"
75 #include "memory/universe.hpp"
76 #include "nmt/memTracker.hpp"
77 #include "oops/access.inline.hpp"
78 #include "oops/flatArrayKlass.inline.hpp"
79 #include "oops/instanceClassLoaderKlass.inline.hpp"
80 #include "oops/instanceKlass.inline.hpp"
81 #include "oops/instanceMirrorKlass.inline.hpp"
82 #include "oops/methodData.hpp"
83 #include "oops/objArrayKlass.inline.hpp"
84 #include "oops/oop.inline.hpp"
85 #include "runtime/atomic.hpp"
86 #include "runtime/handles.inline.hpp"
87 #include "runtime/java.hpp"
88 #include "runtime/safepoint.hpp"
89 #include "runtime/threads.hpp"
90 #include "runtime/vmThread.hpp"
91 #include "services/memoryService.hpp"
92 #include "utilities/align.hpp"
93 #include "utilities/debug.hpp"
94 #include "utilities/events.hpp"
95 #include "utilities/formatBuffer.hpp"
96 #include "utilities/macros.hpp"
97 #include "utilities/stack.inline.hpp"
98 #if INCLUDE_JVMCI
99 #include "jvmci/jvmci.hpp"
100 #endif
101
102 #include <math.h>
103
104 // All sizes are in HeapWords.
105 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
106 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
107 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
108 const size_t ParallelCompactData::RegionSizeBytes =
109 RegionSize << LogHeapWordSize;
110 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
111 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
112 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
113
114 const ParallelCompactData::RegionData::region_sz_t
115 ParallelCompactData::RegionData::dc_shift = 27;
116
117 const ParallelCompactData::RegionData::region_sz_t
118 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
119
120 const ParallelCompactData::RegionData::region_sz_t
121 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
122
123 const ParallelCompactData::RegionData::region_sz_t
124 ParallelCompactData::RegionData::los_mask = ~dc_mask;
125
126 const ParallelCompactData::RegionData::region_sz_t
127 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
128
129 const ParallelCompactData::RegionData::region_sz_t
130 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
131
132 bool ParallelCompactData::RegionData::is_clear() {
133 return (_destination == nullptr) &&
134 (_source_region == 0) &&
135 (_partial_obj_addr == nullptr) &&
136 (_partial_obj_size == 0) &&
137 (_dc_and_los == 0) &&
138 (_shadow_state == 0);
139 }
140
141 #ifdef ASSERT
142 void ParallelCompactData::RegionData::verify_clear() {
143 assert(_destination == nullptr, "inv");
144 assert(_source_region == 0, "inv");
145 assert(_partial_obj_addr == nullptr, "inv");
146 assert(_partial_obj_size == 0, "inv");
147 assert(_dc_and_los == 0, "inv");
148 assert(_shadow_state == 0, "inv");
149 }
150 #endif
151
152 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
153
154 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
155 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
156
157 void SplitInfo::record(size_t split_region_idx, HeapWord* split_point, size_t preceding_live_words) {
158 assert(split_region_idx != 0, "precondition");
159
160 // Obj denoted by split_point will be deferred to the next space.
161 assert(split_point != nullptr, "precondition");
162
163 const ParallelCompactData& sd = PSParallelCompact::summary_data();
164
165 PSParallelCompact::RegionData* split_region_ptr = sd.region(split_region_idx);
166 assert(preceding_live_words < split_region_ptr->data_size(), "inv");
167
168 HeapWord* preceding_destination = split_region_ptr->destination();
169 assert(preceding_destination != nullptr, "inv");
170
171 // How many regions does the preceding part occupy
172 uint preceding_destination_count;
173 if (preceding_live_words == 0) {
174 preceding_destination_count = 0;
175 } else {
176 // -1 so that the ending address doesn't fall on the region-boundary
177 if (sd.region_align_down(preceding_destination) ==
178 sd.region_align_down(preceding_destination + preceding_live_words - 1)) {
179 preceding_destination_count = 1;
180 } else {
181 preceding_destination_count = 2;
182 }
183 }
184
185 _split_region_idx = split_region_idx;
186 _split_point = split_point;
187 _preceding_live_words = preceding_live_words;
188 _preceding_destination = preceding_destination;
189 _preceding_destination_count = preceding_destination_count;
190 }
191
192 void SplitInfo::clear()
193 {
194 _split_region_idx = 0;
195 _split_point = nullptr;
196 _preceding_live_words = 0;
197 _preceding_destination = nullptr;
198 _preceding_destination_count = 0;
199 assert(!is_valid(), "sanity");
200 }
201
202 #ifdef ASSERT
203 void SplitInfo::verify_clear()
204 {
205 assert(_split_region_idx == 0, "not clear");
206 assert(_split_point == nullptr, "not clear");
207 assert(_preceding_live_words == 0, "not clear");
208 assert(_preceding_destination == nullptr, "not clear");
209 assert(_preceding_destination_count == 0, "not clear");
210 }
211 #endif // #ifdef ASSERT
212
213
214 void PSParallelCompact::print_on(outputStream* st) {
215 _mark_bitmap.print_on(st);
216 }
217
218 ParallelCompactData::ParallelCompactData() :
219 _heap_start(nullptr),
220 DEBUG_ONLY(_heap_end(nullptr) COMMA)
221 _region_vspace(nullptr),
222 _reserved_byte_size(0),
223 _region_data(nullptr),
224 _region_count(0) {}
225
226 bool ParallelCompactData::initialize(MemRegion reserved_heap)
227 {
228 _heap_start = reserved_heap.start();
229 const size_t heap_size = reserved_heap.word_size();
230 DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
231
232 assert(region_align_down(_heap_start) == _heap_start,
233 "region start not aligned");
234
235 return initialize_region_data(heap_size);
236 }
237
238 PSVirtualSpace*
239 ParallelCompactData::create_vspace(size_t count, size_t element_size)
240 {
241 const size_t raw_bytes = count * element_size;
242 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
243 const size_t granularity = os::vm_allocation_granularity();
244 const size_t rs_align = MAX2(page_sz, granularity);
245
246 _reserved_byte_size = align_up(raw_bytes, rs_align);
247
248 ReservedSpace rs = MemoryReserver::reserve(_reserved_byte_size,
249 rs_align,
250 page_sz,
251 mtGC);
252
253 if (!rs.is_reserved()) {
254 // Failed to reserve memory.
255 return nullptr;
256 }
257
258 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
259 rs.size(), page_sz);
260
261 MemTracker::record_virtual_memory_tag(rs, mtGC);
262
263 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
264
265 if (!vspace->expand_by(_reserved_byte_size)) {
266 // Failed to commit memory.
267
268 delete vspace;
269
270 // Release memory reserved in the space.
271 MemoryReserver::release(rs);
272
273 return nullptr;
274 }
275
276 return vspace;
277 }
278
279 bool ParallelCompactData::initialize_region_data(size_t heap_size)
280 {
281 assert(is_aligned(heap_size, RegionSize), "precondition");
282
283 const size_t count = heap_size >> Log2RegionSize;
284 _region_vspace = create_vspace(count, sizeof(RegionData));
285 if (_region_vspace != nullptr) {
286 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
287 _region_count = count;
288 return true;
289 }
290 return false;
291 }
292
293 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
294 assert(beg_region <= _region_count, "beg_region out of range");
295 assert(end_region <= _region_count, "end_region out of range");
296
297 const size_t region_cnt = end_region - beg_region;
298 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
299 }
300
301 void
302 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
303 {
304 assert(is_region_aligned(beg), "not RegionSize aligned");
305 assert(is_region_aligned(end), "not RegionSize aligned");
306
307 size_t cur_region = addr_to_region_idx(beg);
308 const size_t end_region = addr_to_region_idx(end);
309 HeapWord* addr = beg;
310 while (cur_region < end_region) {
311 _region_data[cur_region].set_destination(addr);
312 _region_data[cur_region].set_destination_count(0);
313 _region_data[cur_region].set_source_region(cur_region);
314
315 // Update live_obj_size so the region appears completely full.
316 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
317 _region_data[cur_region].set_live_obj_size(live_size);
318
319 ++cur_region;
320 addr += RegionSize;
321 }
322 }
323
324 // The total live words on src_region would overflow the target space, so find
325 // the overflowing object and record the split point. The invariant is that an
326 // obj should not cross space boundary.
327 HeapWord* ParallelCompactData::summarize_split_space(size_t src_region,
328 SplitInfo& split_info,
329 HeapWord* const destination,
330 HeapWord* const target_end,
331 HeapWord** target_next) {
332 assert(destination <= target_end, "sanity");
333 assert(destination + _region_data[src_region].data_size() > target_end,
334 "region should not fit into target space");
335 assert(is_region_aligned(target_end), "sanity");
336
337 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
338
339 if (destination + partial_obj_size > target_end) {
340 assert(partial_obj_size > 0, "inv");
341 // The overflowing obj is from a previous region.
342 //
343 // source-regions:
344 //
345 // ***************
346 // | A|AA |
347 // ***************
348 // ^
349 // | split-point
350 //
351 // dest-region:
352 //
353 // ********
354 // |~~~~A |
355 // ********
356 // ^^
357 // || target-space-end
358 // |
359 // | destination
360 //
361 // AAA would overflow target-space.
362 //
363 HeapWord* overflowing_obj = _region_data[src_region].partial_obj_addr();
364 size_t split_region = addr_to_region_idx(overflowing_obj);
365
366 // The number of live words before the overflowing object on this split region
367 size_t preceding_live_words;
368 if (is_region_aligned(overflowing_obj)) {
369 preceding_live_words = 0;
370 } else {
371 // Words accounted by the overflowing object on the split region
372 size_t overflowing_size = pointer_delta(region_align_up(overflowing_obj), overflowing_obj);
373 preceding_live_words = region(split_region)->data_size() - overflowing_size;
374 }
375
376 split_info.record(split_region, overflowing_obj, preceding_live_words);
377
378 // The [overflowing_obj, src_region_start) part has been accounted for, so
379 // must move back the new_top, now that this overflowing obj is deferred.
380 HeapWord* new_top = destination - pointer_delta(region_to_addr(src_region), overflowing_obj);
381
382 // If the overflowing obj was relocated to its original destination,
383 // those destination regions would have their source_region set. Now that
384 // this overflowing obj is relocated somewhere else, reset the
385 // source_region.
386 {
387 size_t range_start = addr_to_region_idx(region_align_up(new_top));
388 size_t range_end = addr_to_region_idx(region_align_up(destination));
389 for (size_t i = range_start; i < range_end; ++i) {
390 region(i)->set_source_region(0);
391 }
392 }
393
394 // Update new top of target space
395 *target_next = new_top;
396
397 return overflowing_obj;
398 }
399
400 // Obj-iteration to locate the overflowing obj
401 HeapWord* region_start = region_to_addr(src_region);
402 HeapWord* region_end = region_start + RegionSize;
403 HeapWord* cur_addr = region_start + partial_obj_size;
404 size_t live_words = partial_obj_size;
405
406 while (true) {
407 assert(cur_addr < region_end, "inv");
408 cur_addr = PSParallelCompact::mark_bitmap()->find_obj_beg(cur_addr, region_end);
409 // There must be an overflowing obj in this region
410 assert(cur_addr < region_end, "inv");
411
412 oop obj = cast_to_oop(cur_addr);
413 size_t obj_size = obj->size();
414 if (destination + live_words + obj_size > target_end) {
415 // Found the overflowing obj
416 split_info.record(src_region, cur_addr, live_words);
417 *target_next = destination + live_words;
418 return cur_addr;
419 }
420
421 live_words += obj_size;
422 cur_addr += obj_size;
423 }
424 }
425
426 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
427 HeapWord** full_region_prefix_end) {
428 size_t cur_region = addr_to_region_idx(space->bottom());
429 const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
430 size_t live_words = 0;
431 if (full_region_prefix_end == nullptr) {
432 for (/* empty */; cur_region < end_region; ++cur_region) {
433 live_words += _region_data[cur_region].data_size();
434 }
435 } else {
436 bool first_set = false;
437 for (/* empty */; cur_region < end_region; ++cur_region) {
438 size_t live_words_in_region = _region_data[cur_region].data_size();
439 if (!first_set && live_words_in_region < RegionSize) {
440 *full_region_prefix_end = region_to_addr(cur_region);
441 first_set = true;
442 }
443 live_words += live_words_in_region;
444 }
445 if (!first_set) {
446 // All regions are full of live objs.
447 assert(is_region_aligned(space->top()), "inv");
448 *full_region_prefix_end = space->top();
449 }
450 assert(*full_region_prefix_end != nullptr, "postcondition");
451 assert(is_region_aligned(*full_region_prefix_end), "inv");
452 assert(*full_region_prefix_end >= space->bottom(), "in-range");
453 assert(*full_region_prefix_end <= space->top(), "in-range");
454 }
455 return live_words;
456 }
457
458 bool ParallelCompactData::summarize(SplitInfo& split_info,
459 HeapWord* source_beg, HeapWord* source_end,
460 HeapWord** source_next,
461 HeapWord* target_beg, HeapWord* target_end,
462 HeapWord** target_next)
463 {
464 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
465 log_develop_trace(gc, compaction)(
466 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
467 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
468 p2i(source_beg), p2i(source_end), p2i(source_next_val),
469 p2i(target_beg), p2i(target_end), p2i(*target_next));
470
471 size_t cur_region = addr_to_region_idx(source_beg);
472 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
473
474 HeapWord *dest_addr = target_beg;
475 for (/* empty */; cur_region < end_region; cur_region++) {
476 size_t words = _region_data[cur_region].data_size();
477
478 // Skip empty ones
479 if (words == 0) {
480 continue;
481 }
482
483 if (split_info.is_split(cur_region)) {
484 assert(words > split_info.preceding_live_words(), "inv");
485 words -= split_info.preceding_live_words();
486 }
487
488 _region_data[cur_region].set_destination(dest_addr);
489
490 // If cur_region does not fit entirely into the target space, find a point
491 // at which the source space can be 'split' so that part is copied to the
492 // target space and the rest is copied elsewhere.
493 if (dest_addr + words > target_end) {
494 assert(source_next != nullptr, "source_next is null when splitting");
495 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
496 target_end, target_next);
497 return false;
498 }
499
500 uint destination_count = split_info.is_split(cur_region)
501 ? split_info.preceding_destination_count()
502 : 0;
503
504 HeapWord* const last_addr = dest_addr + words - 1;
505 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
506 const size_t dest_region_2 = addr_to_region_idx(last_addr);
507
508 // Initially assume that the destination regions will be the same and
509 // adjust the value below if necessary. Under this assumption, if
510 // cur_region == dest_region_2, then cur_region will be compacted
511 // completely into itself.
512 destination_count += cur_region == dest_region_2 ? 0 : 1;
513 if (dest_region_1 != dest_region_2) {
514 // Destination regions differ; adjust destination_count.
515 destination_count += 1;
516 // Data from cur_region will be copied to the start of dest_region_2.
517 _region_data[dest_region_2].set_source_region(cur_region);
518 } else if (is_region_aligned(dest_addr)) {
519 // Data from cur_region will be copied to the start of the destination
520 // region.
521 _region_data[dest_region_1].set_source_region(cur_region);
522 }
523
524 _region_data[cur_region].set_destination_count(destination_count);
525 dest_addr += words;
526 }
527
528 *target_next = dest_addr;
529 return true;
530 }
531
532 #ifdef ASSERT
533 void ParallelCompactData::verify_clear() {
534 for (uint cur_idx = 0; cur_idx < region_count(); ++cur_idx) {
535 if (!region(cur_idx)->is_clear()) {
536 log_warning(gc)("Uncleared Region: %u", cur_idx);
537 region(cur_idx)->verify_clear();
538 }
539 }
540 }
541 #endif // #ifdef ASSERT
542
543 STWGCTimer PSParallelCompact::_gc_timer;
544 ParallelOldTracer PSParallelCompact::_gc_tracer;
545 elapsedTimer PSParallelCompact::_accumulated_time;
546 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
547 CollectorCounters* PSParallelCompact::_counters = nullptr;
548 ParMarkBitMap PSParallelCompact::_mark_bitmap;
549 ParallelCompactData PSParallelCompact::_summary_data;
550
551 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
552
553 class PCAdjustPointerClosure: public BasicOopIterateClosure {
554 template <typename T>
555 void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
556
557 public:
558 virtual void do_oop(oop* p) { do_oop_work(p); }
559 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
560
561 virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
562 };
563
564 static PCAdjustPointerClosure pc_adjust_pointer_closure;
565
566 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
567
568 void PSParallelCompact::post_initialize() {
569 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
570 _span_based_discoverer.set_span(heap->reserved_region());
571 _ref_processor =
572 new ReferenceProcessor(&_span_based_discoverer,
573 ParallelGCThreads, // mt processing degree
574 ParallelGCThreads, // mt discovery degree
575 false, // concurrent_discovery
576 &_is_alive_closure); // non-header is alive closure
577
578 _counters = new CollectorCounters("Parallel full collection pauses", 1);
579
580 // Initialize static fields in ParCompactionManager.
581 ParCompactionManager::initialize(mark_bitmap());
582 }
583
584 bool PSParallelCompact::initialize_aux_data() {
585 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
586 MemRegion mr = heap->reserved_region();
587 assert(mr.byte_size() != 0, "heap should be reserved");
588
589 initialize_space_info();
590
591 if (!_mark_bitmap.initialize(mr)) {
592 vm_shutdown_during_initialization(
593 err_msg("Unable to allocate %zuKB bitmaps for parallel "
594 "garbage collection for the requested %zuKB heap.",
595 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
596 return false;
597 }
598
599 if (!_summary_data.initialize(mr)) {
600 vm_shutdown_during_initialization(
601 err_msg("Unable to allocate %zuKB card tables for parallel "
602 "garbage collection for the requested %zuKB heap.",
603 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
604 return false;
605 }
606
607 return true;
608 }
609
610 void PSParallelCompact::initialize_space_info()
611 {
612 memset(&_space_info, 0, sizeof(_space_info));
613
614 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
615 PSYoungGen* young_gen = heap->young_gen();
616
617 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
618 _space_info[eden_space_id].set_space(young_gen->eden_space());
619 _space_info[from_space_id].set_space(young_gen->from_space());
620 _space_info[to_space_id].set_space(young_gen->to_space());
621
622 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
623 }
624
625 void
626 PSParallelCompact::clear_data_covering_space(SpaceId id)
627 {
628 // At this point, top is the value before GC, new_top() is the value that will
629 // be set at the end of GC. The marking bitmap is cleared to top; nothing
630 // should be marked above top. The summary data is cleared to the larger of
631 // top & new_top.
632 MutableSpace* const space = _space_info[id].space();
633 HeapWord* const bot = space->bottom();
634 HeapWord* const top = space->top();
635 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
636
637 _mark_bitmap.clear_range(bot, top);
638
639 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
640 const size_t end_region =
641 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
642 _summary_data.clear_range(beg_region, end_region);
643
644 // Clear the data used to 'split' regions.
645 SplitInfo& split_info = _space_info[id].split_info();
646 if (split_info.is_valid()) {
647 split_info.clear();
648 }
649 DEBUG_ONLY(split_info.verify_clear();)
650 }
651
652 void PSParallelCompact::pre_compact()
653 {
654 // Update the from & to space pointers in space_info, since they are swapped
655 // at each young gen gc. Do the update unconditionally (even though a
656 // promotion failure does not swap spaces) because an unknown number of young
657 // collections will have swapped the spaces an unknown number of times.
658 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
659 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
660 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
661 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
662
663 // Increment the invocation count
664 heap->increment_total_collections(true);
665
666 CodeCache::on_gc_marking_cycle_start();
667
668 heap->print_heap_before_gc();
669 heap->trace_heap_before_gc(&_gc_tracer);
670
671 // Fill in TLABs
672 heap->ensure_parsability(true); // retire TLABs
673
674 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
675 Universe::verify("Before GC");
676 }
677
678 DEBUG_ONLY(mark_bitmap()->verify_clear();)
679 DEBUG_ONLY(summary_data().verify_clear();)
680 }
681
682 void PSParallelCompact::post_compact()
683 {
684 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
685 ParCompactionManager::remove_all_shadow_regions();
686
687 CodeCache::on_gc_marking_cycle_finish();
688 CodeCache::arm_all_nmethods();
689
690 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
691 // Clear the marking bitmap, summary data and split info.
692 clear_data_covering_space(SpaceId(id));
693 {
694 MutableSpace* space = _space_info[id].space();
695 HeapWord* top = space->top();
696 HeapWord* new_top = _space_info[id].new_top();
697 if (ZapUnusedHeapArea && new_top < top) {
698 space->mangle_region(MemRegion(new_top, top));
699 }
700 // Update top(). Must be done after clearing the bitmap and summary data.
701 space->set_top(new_top);
702 }
703 }
704
705 #ifdef ASSERT
706 {
707 mark_bitmap()->verify_clear();
708 summary_data().verify_clear();
709 }
710 #endif
711
712 ParCompactionManager::flush_all_string_dedup_requests();
713
714 MutableSpace* const eden_space = _space_info[eden_space_id].space();
715 MutableSpace* const from_space = _space_info[from_space_id].space();
716 MutableSpace* const to_space = _space_info[to_space_id].space();
717
718 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
719 bool eden_empty = eden_space->is_empty();
720
721 // Update heap occupancy information which is used as input to the soft ref
722 // clearing policy at the next gc.
723 Universe::heap()->update_capacity_and_used_at_gc();
724
725 bool young_gen_empty = eden_empty && from_space->is_empty() &&
726 to_space->is_empty();
727
728 PSCardTable* ct = heap->card_table();
729 MemRegion old_mr = heap->old_gen()->committed();
730 if (young_gen_empty) {
731 ct->clear_MemRegion(old_mr);
732 } else {
733 ct->dirty_MemRegion(old_mr);
734 }
735
736 {
737 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
738 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
739 ClassLoaderDataGraph::purge(true /* at_safepoint */);
740 DEBUG_ONLY(MetaspaceUtils::verify();)
741 }
742
743 // Need to clear claim bits for the next mark.
744 ClassLoaderDataGraph::clear_claimed_marks();
745
746 heap->prune_scavengable_nmethods();
747
748 #if COMPILER2_OR_JVMCI
749 DerivedPointerTable::update_pointers();
750 #endif
751
752 // Signal that we have completed a visit to all live objects.
753 Universe::heap()->record_whole_heap_examined_timestamp();
754 }
755
756 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
757 HeapWord* full_region_prefix_end) {
758 const size_t region_size = ParallelCompactData::RegionSize;
759 const ParallelCompactData& sd = summary_data();
760
761 // Iteration starts with the region *after* the full-region-prefix-end.
762 const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
763 // If final region is not full, iteration stops before that region,
764 // because fill_dense_prefix_end assumes that prefix_end <= top.
765 const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
766 assert(start_region <= end_region, "inv");
767
768 size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
769 const RegionData* cur_region = start_region;
770 for (/* empty */; cur_region < end_region; ++cur_region) {
771 assert(region_size >= cur_region->data_size(), "inv");
772 size_t dead_size = region_size - cur_region->data_size();
773 if (max_waste < dead_size) {
774 break;
775 }
776 max_waste -= dead_size;
777 }
778
779 HeapWord* const prefix_end = sd.region_to_addr(cur_region);
780 assert(sd.is_region_aligned(prefix_end), "postcondition");
781 assert(prefix_end >= full_region_prefix_end, "in-range");
782 assert(prefix_end <= old_space->top(), "in-range");
783 return prefix_end;
784 }
785
786 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
787 // Comparing two sizes to decide if filling is required:
788 //
789 // The size of the filler (min-obj-size) is 2 heap words with the default
790 // MinObjAlignment, since both markword and klass take 1 heap word.
791 // With +UseCompactObjectHeaders, the minimum filler size is only one word,
792 // because the Klass* gets encoded in the mark-word.
793 //
794 // The size of the gap (if any) right before dense-prefix-end is
795 // MinObjAlignment.
796 //
797 // Need to fill in the gap only if it's smaller than min-obj-size, and the
798 // filler obj will extend to next region.
799
800 if (MinObjAlignment >= checked_cast<int>(CollectedHeap::min_fill_size())) {
801 return;
802 }
803
804 assert(!UseCompactObjectHeaders, "Compact headers can allocate small objects");
805 assert(CollectedHeap::min_fill_size() == 2, "inv");
806 HeapWord* const dense_prefix_end = dense_prefix(id);
807 assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
808 assert(dense_prefix_end <= space(id)->top(), "precondition");
809 if (dense_prefix_end == space(id)->top()) {
810 // Must not have single-word gap right before prefix-end/top.
811 return;
812 }
813 RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
814
815 if (region_after_dense_prefix->partial_obj_size() != 0 ||
816 _mark_bitmap.is_marked(dense_prefix_end)) {
817 // The region after the dense prefix starts with live bytes.
818 return;
819 }
820
821 HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
822 if (block_start == dense_prefix_end - 1) {
823 assert(!_mark_bitmap.is_marked(block_start), "inv");
824 // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
825 // The filler object will extend into region_after_dense_prefix.
826 const size_t obj_len = 2; // min-fill-size
827 HeapWord* const obj_beg = dense_prefix_end - 1;
828 CollectedHeap::fill_with_object(obj_beg, obj_len);
829 _mark_bitmap.mark_obj(obj_beg);
830 _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
831 region_after_dense_prefix->set_partial_obj_size(1);
832 region_after_dense_prefix->set_partial_obj_addr(obj_beg);
833 assert(start_array(id) != nullptr, "sanity");
834 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
835 }
836 }
837
838 bool PSParallelCompact::check_maximum_compaction(size_t total_live_words,
839 MutableSpace* const old_space,
840 HeapWord* full_region_prefix_end) {
841
842 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
843
844 // Check System.GC
845 bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
846 && GCCause::is_user_requested_gc(heap->gc_cause());
847
848 // Check if all live objs are larger than old-gen.
849 const bool is_old_gen_overflowing = (total_live_words > old_space->capacity_in_words());
850
851 // JVM flags
852 const uint total_invocations = heap->total_full_collections();
853 assert(total_invocations >= _maximum_compaction_gc_num, "sanity");
854 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num;
855 const bool is_interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
856
857 // If all regions in old-gen are full
858 const bool is_region_full =
859 full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
860
861 if (is_max_on_system_gc || is_old_gen_overflowing || is_interval_ended || is_region_full) {
862 _maximum_compaction_gc_num = total_invocations;
863 return true;
864 }
865
866 return false;
867 }
868
869 void PSParallelCompact::summary_phase()
870 {
871 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
872
873 MutableSpace* const old_space = _space_info[old_space_id].space();
874 {
875 size_t total_live_words = 0;
876 HeapWord* full_region_prefix_end = nullptr;
877 {
878 // old-gen
879 size_t live_words = _summary_data.live_words_in_space(old_space,
880 &full_region_prefix_end);
881 total_live_words += live_words;
882 }
883 // young-gen
884 for (uint i = eden_space_id; i < last_space_id; ++i) {
885 const MutableSpace* space = _space_info[i].space();
886 size_t live_words = _summary_data.live_words_in_space(space);
887 total_live_words += live_words;
888 _space_info[i].set_new_top(space->bottom() + live_words);
889 _space_info[i].set_dense_prefix(space->bottom());
890 }
891
892 bool maximum_compaction = check_maximum_compaction(total_live_words,
893 old_space,
894 full_region_prefix_end);
895 HeapWord* dense_prefix_end = maximum_compaction
896 ? full_region_prefix_end
897 : compute_dense_prefix_for_old_space(old_space,
898 full_region_prefix_end);
899 SpaceId id = old_space_id;
900 _space_info[id].set_dense_prefix(dense_prefix_end);
901
902 if (dense_prefix_end != old_space->bottom()) {
903 fill_dense_prefix_end(id);
904 _summary_data.summarize_dense_prefix(old_space->bottom(), dense_prefix_end);
905 }
906
907 // Compacting objs in [dense_prefix_end, old_space->top())
908 _summary_data.summarize(_space_info[id].split_info(),
909 dense_prefix_end, old_space->top(), nullptr,
910 dense_prefix_end, old_space->end(),
911 _space_info[id].new_top_addr());
912 }
913
914 // Summarize the remaining spaces in the young gen. The initial target space
915 // is the old gen. If a space does not fit entirely into the target, then the
916 // remainder is compacted into the space itself and that space becomes the new
917 // target.
918 SpaceId dst_space_id = old_space_id;
919 HeapWord* dst_space_end = old_space->end();
920 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
921 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
922 const MutableSpace* space = _space_info[id].space();
923 const size_t live = pointer_delta(_space_info[id].new_top(),
924 space->bottom());
925 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
926
927 if (live > 0 && live <= available) {
928 // All the live data will fit.
929 bool done = _summary_data.summarize(_space_info[id].split_info(),
930 space->bottom(), space->top(),
931 nullptr,
932 *new_top_addr, dst_space_end,
933 new_top_addr);
934 assert(done, "space must fit into old gen");
935
936 // Reset the new_top value for the space.
937 _space_info[id].set_new_top(space->bottom());
938 } else if (live > 0) {
939 // Attempt to fit part of the source space into the target space.
940 HeapWord* next_src_addr = nullptr;
941 bool done = _summary_data.summarize(_space_info[id].split_info(),
942 space->bottom(), space->top(),
943 &next_src_addr,
944 *new_top_addr, dst_space_end,
945 new_top_addr);
946 assert(!done, "space should not fit into old gen");
947 assert(next_src_addr != nullptr, "sanity");
948
949 // The source space becomes the new target, so the remainder is compacted
950 // within the space itself.
951 dst_space_id = SpaceId(id);
952 dst_space_end = space->end();
953 new_top_addr = _space_info[id].new_top_addr();
954 done = _summary_data.summarize(_space_info[id].split_info(),
955 next_src_addr, space->top(),
956 nullptr,
957 space->bottom(), dst_space_end,
958 new_top_addr);
959 assert(done, "space must fit when compacted into itself");
960 assert(*new_top_addr <= space->top(), "usage should not grow");
961 }
962 }
963 }
964
965 // This method should contain all heap-specific policy for invoking a full
966 // collection. invoke_no_policy() will only attempt to compact the heap; it
967 // will do nothing further. If we need to bail out for policy reasons, scavenge
968 // before full gc, or any other specialized behavior, it needs to be added here.
969 //
970 // Note that this method should only be called from the vm_thread while at a
971 // safepoint.
972 //
973 // Note that the all_soft_refs_clear flag in the soft ref policy
974 // may be true because this method can be called without intervening
975 // activity. For example when the heap space is tight and full measure
976 // are being taken to free space.
977 bool PSParallelCompact::invoke(bool clear_all_soft_refs) {
978 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
979 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
980 "should be in vm thread");
981
982 SvcGCMarker sgcm(SvcGCMarker::FULL);
983 IsSTWGCActiveMark mark;
984
985 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
986 clear_all_soft_refs = clear_all_soft_refs
987 || heap->soft_ref_policy()->should_clear_all_soft_refs();
988
989 return PSParallelCompact::invoke_no_policy(clear_all_soft_refs);
990 }
991
992 // This method contains no policy. You should probably
993 // be calling invoke() instead.
994 bool PSParallelCompact::invoke_no_policy(bool clear_all_soft_refs) {
995 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
996 assert(ref_processor() != nullptr, "Sanity");
997
998 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
999
1000 GCIdMark gc_id_mark;
1001 _gc_timer.register_gc_start();
1002 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1003
1004 GCCause::Cause gc_cause = heap->gc_cause();
1005 PSYoungGen* young_gen = heap->young_gen();
1006 PSOldGen* old_gen = heap->old_gen();
1007 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1008
1009 // The scope of casr should end after code that can change
1010 // SoftRefPolicy::_should_clear_all_soft_refs.
1011 ClearedAllSoftRefs casr(clear_all_soft_refs,
1012 heap->soft_ref_policy());
1013
1014 // Make sure data structures are sane, make the heap parsable, and do other
1015 // miscellaneous bookkeeping.
1016 pre_compact();
1017
1018 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1019
1020 {
1021 const uint active_workers =
1022 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1023 ParallelScavengeHeap::heap()->workers().active_workers(),
1024 Threads::number_of_non_daemon_threads());
1025 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1026
1027 GCTraceCPUTime tcpu(&_gc_tracer);
1028 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
1029
1030 heap->pre_full_gc_dump(&_gc_timer);
1031
1032 TraceCollectorStats tcs(counters());
1033 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
1034
1035 if (log_is_enabled(Debug, gc, heap, exit)) {
1036 accumulated_time()->start();
1037 }
1038
1039 // Let the size policy know we're starting
1040 size_policy->major_collection_begin();
1041
1042 #if COMPILER2_OR_JVMCI
1043 DerivedPointerTable::clear();
1044 #endif
1045
1046 ref_processor()->start_discovery(clear_all_soft_refs);
1047
1048 ClassUnloadingContext ctx(1 /* num_nmethod_unlink_workers */,
1049 false /* unregister_nmethods_during_purge */,
1050 false /* lock_nmethod_free_separately */);
1051
1052 marking_phase(&_gc_tracer);
1053
1054 summary_phase();
1055
1056 #if COMPILER2_OR_JVMCI
1057 assert(DerivedPointerTable::is_active(), "Sanity");
1058 DerivedPointerTable::set_active(false);
1059 #endif
1060
1061 forward_to_new_addr();
1062
1063 adjust_pointers();
1064
1065 compact();
1066
1067 ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1068
1069 ParCompactionManager::verify_all_region_stack_empty();
1070
1071 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1072 // done before resizing.
1073 post_compact();
1074
1075 // Let the size policy know we're done
1076 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1077
1078 if (UseAdaptiveSizePolicy) {
1079 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1080 log_trace(gc, ergo)("old_gen_capacity: %zu young_gen_capacity: %zu",
1081 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1082
1083 // Don't check if the size_policy is ready here. Let
1084 // the size_policy check that internally.
1085 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1086 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1087 // Swap the survivor spaces if from_space is empty. The
1088 // resize_young_gen() called below is normally used after
1089 // a successful young GC and swapping of survivor spaces;
1090 // otherwise, it will fail to resize the young gen with
1091 // the current implementation.
1092 if (young_gen->from_space()->is_empty()) {
1093 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1094 young_gen->swap_spaces();
1095 }
1096
1097 // Calculate optimal free space amounts
1098 assert(young_gen->max_gen_size() >
1099 young_gen->from_space()->capacity_in_bytes() +
1100 young_gen->to_space()->capacity_in_bytes(),
1101 "Sizes of space in young gen are out-of-bounds");
1102
1103 size_t young_live = young_gen->used_in_bytes();
1104 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1105 size_t old_live = old_gen->used_in_bytes();
1106 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1107 size_t max_old_gen_size = old_gen->max_gen_size();
1108 size_t max_eden_size = young_gen->max_gen_size() -
1109 young_gen->from_space()->capacity_in_bytes() -
1110 young_gen->to_space()->capacity_in_bytes();
1111
1112 // Used for diagnostics
1113 size_policy->clear_generation_free_space_flags();
1114
1115 size_policy->compute_generations_free_space(young_live,
1116 eden_live,
1117 old_live,
1118 cur_eden,
1119 max_old_gen_size,
1120 max_eden_size,
1121 true /* full gc*/);
1122
1123 size_policy->check_gc_overhead_limit(eden_live,
1124 max_old_gen_size,
1125 max_eden_size,
1126 true /* full gc*/,
1127 gc_cause,
1128 heap->soft_ref_policy());
1129
1130 size_policy->decay_supplemental_growth(true /* full gc*/);
1131
1132 heap->resize_old_gen(
1133 size_policy->calculated_old_free_size_in_bytes());
1134
1135 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1136 size_policy->calculated_survivor_size_in_bytes());
1137 }
1138
1139 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1140 }
1141
1142 if (UsePerfData) {
1143 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1144 counters->update_counters();
1145 counters->update_old_capacity(old_gen->capacity_in_bytes());
1146 counters->update_young_capacity(young_gen->capacity_in_bytes());
1147 }
1148
1149 heap->resize_all_tlabs();
1150
1151 // Resize the metaspace capacity after a collection
1152 MetaspaceGC::compute_new_size();
1153
1154 if (log_is_enabled(Debug, gc, heap, exit)) {
1155 accumulated_time()->stop();
1156 }
1157
1158 heap->print_heap_change(pre_gc_values);
1159
1160 // Track memory usage and detect low memory
1161 MemoryService::track_memory_usage();
1162 heap->update_counters();
1163
1164 heap->post_full_gc_dump(&_gc_timer);
1165 }
1166
1167 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1168 Universe::verify("After GC");
1169 }
1170
1171 heap->print_heap_after_gc();
1172 heap->trace_heap_after_gc(&_gc_tracer);
1173
1174 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1175
1176 _gc_timer.register_gc_end();
1177
1178 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1179 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1180
1181 return true;
1182 }
1183
1184 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1185 private:
1186 uint _worker_id;
1187
1188 public:
1189 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1190 void do_thread(Thread* thread) {
1191 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1192
1193 ResourceMark rm;
1194
1195 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1196
1197 MarkingNMethodClosure mark_and_push_in_blobs(&cm->_mark_and_push_closure,
1198 !NMethodToOopClosure::FixRelocations,
1199 true /* keepalive nmethods */);
1200
1201 thread->oops_do(&cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1202
1203 // Do the real work
1204 cm->follow_marking_stacks();
1205 }
1206 };
1207
1208 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1209 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1210
1211 ParCompactionManager* cm =
1212 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1213
1214 do {
1215 ScannerTask task;
1216 if (ParCompactionManager::steal(worker_id, task)) {
1217 cm->follow_contents(task, true);
1218 }
1219 cm->follow_marking_stacks();
1220 } while (!terminator.offer_termination());
1221 }
1222
1223 class MarkFromRootsTask : public WorkerTask {
1224 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
1225 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1226 TaskTerminator _terminator;
1227 uint _active_workers;
1228
1229 public:
1230 MarkFromRootsTask(uint active_workers) :
1231 WorkerTask("MarkFromRootsTask"),
1232 _strong_roots_scope(active_workers),
1233 _terminator(active_workers, ParCompactionManager::marking_stacks()),
1234 _active_workers(active_workers) {}
1235
1236 virtual void work(uint worker_id) {
1237 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1238 cm->create_marking_stats_cache();
1239 {
1240 CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1241 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1242
1243 // Do the real work
1244 cm->follow_marking_stacks();
1245 }
1246
1247 {
1248 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
1249 Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1250 }
1251
1252 // Mark from OopStorages
1253 {
1254 _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1255 // Do the real work
1256 cm->follow_marking_stacks();
1257 }
1258
1259 if (_active_workers > 1) {
1260 steal_marking_work(_terminator, worker_id);
1261 }
1262 }
1263 };
1264
1265 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1266 TaskTerminator _terminator;
1267
1268 public:
1269 ParallelCompactRefProcProxyTask(uint max_workers)
1270 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1271 _terminator(_max_workers, ParCompactionManager::marking_stacks()) {}
1272
1273 void work(uint worker_id) override {
1274 assert(worker_id < _max_workers, "sanity");
1275 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1276 BarrierEnqueueDiscoveredFieldClosure enqueue;
1277 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1278 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1279 }
1280
1281 void prepare_run_task_hook() override {
1282 _terminator.reset_for_reuse(_queue_count);
1283 }
1284 };
1285
1286 static void flush_marking_stats_cache(const uint num_workers) {
1287 for (uint i = 0; i < num_workers; ++i) {
1288 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1289 cm->flush_and_destroy_marking_stats_cache();
1290 }
1291 }
1292
1293 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1294 // Recursively traverse all live objects and mark them
1295 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1296
1297 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1298
1299 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1300 {
1301 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1302
1303 MarkFromRootsTask task(active_gc_threads);
1304 ParallelScavengeHeap::heap()->workers().run_task(&task);
1305 }
1306
1307 // Process reference objects found during marking
1308 {
1309 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1310
1311 ReferenceProcessorStats stats;
1312 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1313
1314 ref_processor()->set_active_mt_degree(active_gc_threads);
1315 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1316 stats = ref_processor()->process_discovered_references(task, pt);
1317
1318 gc_tracer->report_gc_reference_stats(stats);
1319 pt.print_all_references();
1320 }
1321
1322 {
1323 GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1324
1325 flush_marking_stats_cache(active_gc_threads);
1326 }
1327
1328 // This is the point where the entire marking should have completed.
1329 ParCompactionManager::verify_all_marking_stack_empty();
1330
1331 {
1332 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1333 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1334 is_alive_closure(),
1335 &do_nothing_cl,
1336 1);
1337 }
1338
1339 {
1340 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1341
1342 ClassUnloadingContext* ctx = ClassUnloadingContext::context();
1343
1344 bool unloading_occurred;
1345 {
1346 CodeCache::UnlinkingScope scope(is_alive_closure());
1347
1348 // Follow system dictionary roots and unload classes.
1349 unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1350
1351 // Unload nmethods.
1352 CodeCache::do_unloading(unloading_occurred);
1353 }
1354
1355 {
1356 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1357 // Release unloaded nmethod's memory.
1358 ctx->purge_nmethods();
1359 }
1360 {
1361 GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1362 ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1363 }
1364 {
1365 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1366 ctx->free_nmethods();
1367 }
1368
1369 // Prune dead klasses from subklass/sibling/implementor lists.
1370 Klass::clean_weak_klass_links(unloading_occurred);
1371
1372 // Clean JVMCI metadata handles.
1373 JVMCI_ONLY(JVMCI::do_unloading(unloading_occurred));
1374 }
1375
1376 {
1377 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
1378 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers());
1379 }
1380 #if TASKQUEUE_STATS
1381 ParCompactionManager::print_and_reset_taskqueue_stats();
1382 #endif
1383 }
1384
1385 template<typename Func>
1386 void PSParallelCompact::adjust_in_space_helper(SpaceId id, volatile uint* claim_counter, Func&& on_stripe) {
1387 MutableSpace* sp = PSParallelCompact::space(id);
1388 HeapWord* const bottom = sp->bottom();
1389 HeapWord* const top = sp->top();
1390 if (bottom == top) {
1391 return;
1392 }
1393
1394 const uint num_regions_per_stripe = 2;
1395 const size_t region_size = ParallelCompactData::RegionSize;
1396 const size_t stripe_size = num_regions_per_stripe * region_size;
1397
1398 while (true) {
1399 uint counter = Atomic::fetch_then_add(claim_counter, num_regions_per_stripe);
1400 HeapWord* cur_stripe = bottom + counter * region_size;
1401 if (cur_stripe >= top) {
1402 break;
1403 }
1404 HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1405 on_stripe(cur_stripe, stripe_end);
1406 }
1407 }
1408
1409 void PSParallelCompact::adjust_in_old_space(volatile uint* claim_counter) {
1410 // Regions in old-space shouldn't be split.
1411 assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1412
1413 auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1414 assert(mark_bitmap()->is_marked(obj_start), "inv");
1415 oop obj = cast_to_oop(obj_start);
1416 return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1417 };
1418
1419 adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1420 assert(_summary_data.is_region_aligned(stripe_start), "inv");
1421 RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1422 HeapWord* obj_start;
1423 if (cur_region->partial_obj_size() != 0) {
1424 obj_start = cur_region->partial_obj_addr();
1425 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1426 } else {
1427 obj_start = stripe_start;
1428 }
1429
1430 while (obj_start < stripe_end) {
1431 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1432 if (obj_start >= stripe_end) {
1433 break;
1434 }
1435 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1436 }
1437 });
1438 }
1439
1440 void PSParallelCompact::adjust_in_young_space(SpaceId id, volatile uint* claim_counter) {
1441 adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1442 HeapWord* obj_start = stripe_start;
1443 while (obj_start < stripe_end) {
1444 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1445 if (obj_start >= stripe_end) {
1446 break;
1447 }
1448 oop obj = cast_to_oop(obj_start);
1449 obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1450 }
1451 });
1452 }
1453
1454 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, volatile uint* claim_counters) {
1455 auto start_time = Ticks::now();
1456 adjust_in_old_space(&claim_counters[0]);
1457 for (uint id = eden_space_id; id < last_space_id; ++id) {
1458 adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1459 }
1460 log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1461 }
1462
1463 class PSAdjustTask final : public WorkerTask {
1464 SubTasksDone _sub_tasks;
1465 WeakProcessor::Task _weak_proc_task;
1466 OopStorageSetStrongParState<false, false> _oop_storage_iter;
1467 uint _nworkers;
1468 volatile uint _claim_counters[PSParallelCompact::last_space_id] = {};
1469
1470 enum PSAdjustSubTask {
1471 PSAdjustSubTask_code_cache,
1472
1473 PSAdjustSubTask_num_elements
1474 };
1475
1476 public:
1477 PSAdjustTask(uint nworkers) :
1478 WorkerTask("PSAdjust task"),
1479 _sub_tasks(PSAdjustSubTask_num_elements),
1480 _weak_proc_task(nworkers),
1481 _nworkers(nworkers) {
1482
1483 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1484 if (nworkers > 1) {
1485 Threads::change_thread_claim_token();
1486 }
1487 }
1488
1489 ~PSAdjustTask() {
1490 Threads::assert_all_threads_claimed();
1491 }
1492
1493 void work(uint worker_id) {
1494 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1495 cm->preserved_marks()->adjust_during_full_gc();
1496 {
1497 // adjust pointers in all spaces
1498 PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1499 }
1500 {
1501 ResourceMark rm;
1502 Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1503 }
1504 _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1505 {
1506 CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1507 ClassLoaderDataGraph::cld_do(&cld_closure);
1508 }
1509 {
1510 AlwaysTrueClosure always_alive;
1511 _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1512 }
1513 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
1514 NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1515 CodeCache::nmethods_do(&adjust_code);
1516 }
1517 _sub_tasks.all_tasks_claimed();
1518 }
1519 };
1520
1521 void PSParallelCompact::adjust_pointers() {
1522 // Adjust the pointers to reflect the new locations
1523 GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1524 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1525 PSAdjustTask task(nworkers);
1526 ParallelScavengeHeap::heap()->workers().run_task(&task);
1527 }
1528
1529 // Split [start, end) evenly for a number of workers and return the
1530 // range for worker_id.
1531 static void split_regions_for_worker(size_t start, size_t end,
1532 uint worker_id, uint num_workers,
1533 size_t* worker_start, size_t* worker_end) {
1534 assert(start < end, "precondition");
1535 assert(num_workers > 0, "precondition");
1536 assert(worker_id < num_workers, "precondition");
1537
1538 size_t num_regions = end - start;
1539 size_t num_regions_per_worker = num_regions / num_workers;
1540 size_t remainder = num_regions % num_workers;
1541 // The first few workers will get one extra.
1542 *worker_start = start + worker_id * num_regions_per_worker
1543 + MIN2(checked_cast<size_t>(worker_id), remainder);
1544 *worker_end = *worker_start + num_regions_per_worker
1545 + (worker_id < remainder ? 1 : 0);
1546 }
1547
1548 void PSParallelCompact::forward_to_new_addr() {
1549 GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1550 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1551
1552 struct ForwardTask final : public WorkerTask {
1553 uint _num_workers;
1554
1555 explicit ForwardTask(uint num_workers) :
1556 WorkerTask("PSForward task"),
1557 _num_workers(num_workers) {}
1558
1559 static void forward_objs_in_range(ParCompactionManager* cm,
1560 HeapWord* start,
1561 HeapWord* end,
1562 HeapWord* destination) {
1563 HeapWord* cur_addr = start;
1564 HeapWord* new_addr = destination;
1565
1566 while (cur_addr < end) {
1567 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, end);
1568 if (cur_addr >= end) {
1569 return;
1570 }
1571 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1572 oop obj = cast_to_oop(cur_addr);
1573 if (new_addr != cur_addr) {
1574 cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1575 FullGCForwarding::forward_to(obj, cast_to_oop(new_addr));
1576 }
1577 size_t obj_size = obj->size();
1578 new_addr += obj_size;
1579 cur_addr += obj_size;
1580 }
1581 }
1582
1583 void work(uint worker_id) override {
1584 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1585 for (uint id = old_space_id; id < last_space_id; ++id) {
1586 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1587 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1588 HeapWord* top = sp->top();
1589
1590 if (dense_prefix_addr == top) {
1591 continue;
1592 }
1593
1594 const SplitInfo& split_info = _space_info[SpaceId(id)].split_info();
1595
1596 size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1597 size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1598 size_t start_region;
1599 size_t end_region;
1600 split_regions_for_worker(dense_prefix_region, top_region,
1601 worker_id, _num_workers,
1602 &start_region, &end_region);
1603 for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1604 RegionData* region_ptr = _summary_data.region(cur_region);
1605 size_t partial_obj_size = region_ptr->partial_obj_size();
1606
1607 if (partial_obj_size == ParallelCompactData::RegionSize) {
1608 // No obj-start
1609 continue;
1610 }
1611
1612 HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1613 HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1614
1615 if (split_info.is_split(cur_region)) {
1616 // Part 1: will be relocated to space-1
1617 HeapWord* preceding_destination = split_info.preceding_destination();
1618 HeapWord* split_point = split_info.split_point();
1619 forward_objs_in_range(cm, region_start + partial_obj_size, split_point, preceding_destination + partial_obj_size);
1620
1621 // Part 2: will be relocated to space-2
1622 HeapWord* destination = region_ptr->destination();
1623 forward_objs_in_range(cm, split_point, region_end, destination);
1624 } else {
1625 HeapWord* destination = region_ptr->destination();
1626 forward_objs_in_range(cm, region_start + partial_obj_size, region_end, destination + partial_obj_size);
1627 }
1628 }
1629 }
1630 }
1631 } task(nworkers);
1632
1633 ParallelScavengeHeap::heap()->workers().run_task(&task);
1634 DEBUG_ONLY(verify_forward();)
1635 }
1636
1637 #ifdef ASSERT
1638 void PSParallelCompact::verify_forward() {
1639 HeapWord* old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1640 RegionData* old_region = _summary_data.region(_summary_data.addr_to_region_idx(old_dense_prefix_addr));
1641 HeapWord* bump_ptr = old_region->partial_obj_size() != 0
1642 ? old_dense_prefix_addr + old_region->partial_obj_size()
1643 : old_dense_prefix_addr;
1644 SpaceId bump_ptr_space = old_space_id;
1645
1646 for (uint id = old_space_id; id < last_space_id; ++id) {
1647 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1648 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1649 HeapWord* top = sp->top();
1650 HeapWord* cur_addr = dense_prefix_addr;
1651
1652 while (cur_addr < top) {
1653 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1654 if (cur_addr >= top) {
1655 break;
1656 }
1657 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1658 assert(bump_ptr <= _space_info[bump_ptr_space].new_top(), "inv");
1659 // Move to the space containing cur_addr
1660 if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1661 bump_ptr = space(space_id(cur_addr))->bottom();
1662 bump_ptr_space = space_id(bump_ptr);
1663 }
1664 oop obj = cast_to_oop(cur_addr);
1665 if (cur_addr == bump_ptr) {
1666 assert(!FullGCForwarding::is_forwarded(obj), "inv");
1667 } else {
1668 assert(FullGCForwarding::forwardee(obj) == cast_to_oop(bump_ptr), "inv");
1669 }
1670 bump_ptr += obj->size();
1671 cur_addr += obj->size();
1672 }
1673 }
1674 }
1675 #endif
1676
1677 // Helper class to print 8 region numbers per line and then print the total at the end.
1678 class FillableRegionLogger : public StackObj {
1679 private:
1680 Log(gc, compaction) log;
1681 static const int LineLength = 8;
1682 size_t _regions[LineLength];
1683 int _next_index;
1684 bool _enabled;
1685 size_t _total_regions;
1686 public:
1687 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1688 ~FillableRegionLogger() {
1689 log.trace("%zu initially fillable regions", _total_regions);
1690 }
1691
1692 void print_line() {
1693 if (!_enabled || _next_index == 0) {
1694 return;
1695 }
1696 FormatBuffer<> line("Fillable: ");
1697 for (int i = 0; i < _next_index; i++) {
1698 line.append(" %7zu", _regions[i]);
1699 }
1700 log.trace("%s", line.buffer());
1701 _next_index = 0;
1702 }
1703
1704 void handle(size_t region) {
1705 if (!_enabled) {
1706 return;
1707 }
1708 _regions[_next_index++] = region;
1709 if (_next_index == LineLength) {
1710 print_line();
1711 }
1712 _total_regions++;
1713 }
1714 };
1715
1716 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1717 {
1718 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1719
1720 // Find the threads that are active
1721 uint worker_id = 0;
1722
1723 // Find all regions that are available (can be filled immediately) and
1724 // distribute them to the thread stacks. The iteration is done in reverse
1725 // order (high to low) so the regions will be removed in ascending order.
1726
1727 const ParallelCompactData& sd = PSParallelCompact::summary_data();
1728
1729 // id + 1 is used to test termination so unsigned can
1730 // be used with an old_space_id == 0.
1731 FillableRegionLogger region_logger;
1732 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
1733 SpaceInfo* const space_info = _space_info + id;
1734 HeapWord* const new_top = space_info->new_top();
1735
1736 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1737 const size_t end_region =
1738 sd.addr_to_region_idx(sd.region_align_up(new_top));
1739
1740 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1741 if (sd.region(cur)->claim_unsafe()) {
1742 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1743 bool result = sd.region(cur)->mark_normal();
1744 assert(result, "Must succeed at this point.");
1745 cm->region_stack()->push(cur);
1746 region_logger.handle(cur);
1747 // Assign regions to tasks in round-robin fashion.
1748 if (++worker_id == parallel_gc_threads) {
1749 worker_id = 0;
1750 }
1751 }
1752 }
1753 region_logger.print_line();
1754 }
1755 }
1756
1757 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1758 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1759
1760 ParCompactionManager* cm =
1761 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1762
1763 // Drain the stacks that have been preloaded with regions
1764 // that are ready to fill.
1765
1766 cm->drain_region_stacks();
1767
1768 guarantee(cm->region_stack()->is_empty(), "Not empty");
1769
1770 size_t region_index = 0;
1771
1772 while (true) {
1773 if (ParCompactionManager::steal(worker_id, region_index)) {
1774 PSParallelCompact::fill_and_update_region(cm, region_index);
1775 cm->drain_region_stacks();
1776 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1777 // Fill and update an unavailable region with the help of a shadow region
1778 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1779 cm->drain_region_stacks();
1780 } else {
1781 if (terminator->offer_termination()) {
1782 break;
1783 }
1784 // Go around again.
1785 }
1786 }
1787 }
1788
1789 class FillDensePrefixAndCompactionTask: public WorkerTask {
1790 uint _num_workers;
1791 TaskTerminator _terminator;
1792
1793 public:
1794 FillDensePrefixAndCompactionTask(uint active_workers) :
1795 WorkerTask("FillDensePrefixAndCompactionTask"),
1796 _num_workers(active_workers),
1797 _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1798 }
1799
1800 virtual void work(uint worker_id) {
1801 {
1802 auto start = Ticks::now();
1803 PSParallelCompact::fill_dead_objs_in_dense_prefix(worker_id, _num_workers);
1804 log_trace(gc, phases)("Fill dense prefix by worker %u: %.3f ms", worker_id, (Ticks::now() - start).seconds() * 1000);
1805 }
1806 compaction_with_stealing_work(&_terminator, worker_id);
1807 }
1808 };
1809
1810 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1811 #ifdef ASSERT
1812 {
1813 assert(start < end, "precondition");
1814 assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1815 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1816 if (start != bottom) {
1817 HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1818 HeapWord* after_obj = obj_start + cast_to_oop(obj_start)->size();
1819 assert(after_obj == start, "precondition");
1820 }
1821 }
1822 #endif
1823
1824 CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1825 HeapWord* addr = start;
1826 do {
1827 size_t size = cast_to_oop(addr)->size();
1828 start_array(old_space_id)->update_for_block(addr, addr + size);
1829 addr += size;
1830 } while (addr < end);
1831 }
1832
1833 void PSParallelCompact::fill_dead_objs_in_dense_prefix(uint worker_id, uint num_workers) {
1834 ParMarkBitMap* bitmap = mark_bitmap();
1835
1836 HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1837 HeapWord* const prefix_end = dense_prefix(old_space_id);
1838
1839 if (bottom == prefix_end) {
1840 return;
1841 }
1842
1843 size_t bottom_region = _summary_data.addr_to_region_idx(bottom);
1844 size_t prefix_end_region = _summary_data.addr_to_region_idx(prefix_end);
1845
1846 size_t start_region;
1847 size_t end_region;
1848 split_regions_for_worker(bottom_region, prefix_end_region,
1849 worker_id, num_workers,
1850 &start_region, &end_region);
1851
1852 if (start_region == end_region) {
1853 return;
1854 }
1855
1856 HeapWord* const start_addr = _summary_data.region_to_addr(start_region);
1857 HeapWord* const end_addr = _summary_data.region_to_addr(end_region);
1858
1859 // Skip live partial obj (if any) from previous region.
1860 HeapWord* cur_addr;
1861 RegionData* start_region_ptr = _summary_data.region(start_region);
1862 if (start_region_ptr->partial_obj_size() != 0) {
1863 HeapWord* partial_obj_start = start_region_ptr->partial_obj_addr();
1864 assert(bitmap->is_marked(partial_obj_start), "inv");
1865 cur_addr = partial_obj_start + cast_to_oop(partial_obj_start)->size();
1866 } else {
1867 cur_addr = start_addr;
1868 }
1869
1870 // end_addr is inclusive to handle regions starting with dead space.
1871 while (cur_addr <= end_addr) {
1872 // Use prefix_end to handle trailing obj in each worker region-chunk.
1873 HeapWord* live_start = bitmap->find_obj_beg(cur_addr, prefix_end);
1874 if (cur_addr != live_start) {
1875 // Only worker 0 handles proceeding dead space.
1876 if (cur_addr != start_addr || worker_id == 0) {
1877 fill_range_in_dense_prefix(cur_addr, live_start);
1878 }
1879 }
1880 if (live_start >= end_addr) {
1881 break;
1882 }
1883 assert(bitmap->is_marked(live_start), "inv");
1884 cur_addr = live_start + cast_to_oop(live_start)->size();
1885 }
1886 }
1887
1888 void PSParallelCompact::compact() {
1889 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1890
1891 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1892
1893 initialize_shadow_regions(active_gc_threads);
1894 prepare_region_draining_tasks(active_gc_threads);
1895
1896 {
1897 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1898
1899 FillDensePrefixAndCompactionTask task(active_gc_threads);
1900 ParallelScavengeHeap::heap()->workers().run_task(&task);
1901
1902 #ifdef ASSERT
1903 verify_filler_in_dense_prefix();
1904
1905 // Verify that all regions have been processed.
1906 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1907 verify_complete(SpaceId(id));
1908 }
1909 #endif
1910 }
1911 }
1912
1913 #ifdef ASSERT
1914 void PSParallelCompact::verify_filler_in_dense_prefix() {
1915 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1916 HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1917 HeapWord* cur_addr = bottom;
1918 while (cur_addr < dense_prefix_end) {
1919 oop obj = cast_to_oop(cur_addr);
1920 oopDesc::verify(obj);
1921 if (!mark_bitmap()->is_marked(cur_addr)) {
1922 Klass* k = cast_to_oop(cur_addr)->klass();
1923 assert(k == Universe::fillerArrayKlass() || k == vmClasses::FillerObject_klass(), "inv");
1924 }
1925 cur_addr += obj->size();
1926 }
1927 }
1928
1929 void PSParallelCompact::verify_complete(SpaceId space_id) {
1930 // All Regions served as compaction targets, from dense_prefix() to
1931 // new_top(), should be marked as filled and all Regions between new_top()
1932 // and top() should be available (i.e., should have been emptied).
1933 ParallelCompactData& sd = summary_data();
1934 SpaceInfo si = _space_info[space_id];
1935 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1936 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1937 const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1938 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1939 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1940
1941 size_t cur_region;
1942 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1943 const RegionData* const c = sd.region(cur_region);
1944 assert(c->completed(), "region %zu not filled: destination_count=%u",
1945 cur_region, c->destination_count());
1946 }
1947
1948 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1949 const RegionData* const c = sd.region(cur_region);
1950 assert(c->available(), "region %zu not empty: destination_count=%u",
1951 cur_region, c->destination_count());
1952 }
1953 }
1954 #endif // #ifdef ASSERT
1955
1956 // Return the SpaceId for the space containing addr. If addr is not in the
1957 // heap, last_space_id is returned. In debug mode it expects the address to be
1958 // in the heap and asserts such.
1959 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1960 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1961
1962 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1963 if (_space_info[id].space()->contains(addr)) {
1964 return SpaceId(id);
1965 }
1966 }
1967
1968 assert(false, "no space contains the addr");
1969 return last_space_id;
1970 }
1971
1972 // Skip over count live words starting from beg, and return the address of the
1973 // next live word. Callers must also ensure that there are enough live words in
1974 // the range [beg, end) to skip.
1975 HeapWord* PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1976 {
1977 ParMarkBitMap* m = mark_bitmap();
1978 HeapWord* cur_addr = beg;
1979 while (true) {
1980 cur_addr = m->find_obj_beg(cur_addr, end);
1981 assert(cur_addr < end, "inv");
1982 size_t obj_size = cast_to_oop(cur_addr)->size();
1983 // Strictly greater-than
1984 if (obj_size > count) {
1985 return cur_addr + count;
1986 }
1987 count -= obj_size;
1988 cur_addr += obj_size;
1989 }
1990 }
1991
1992 // On starting to fill a destination region (dest-region), we need to know the
1993 // location of the word that will be at the start of the dest-region after
1994 // compaction. A dest-region can have one or more source regions, but only the
1995 // first source-region contains this location. This location is retrieved by
1996 // calling `first_src_addr` on a dest-region.
1997 // Conversely, a source-region has a dest-region which holds the destination of
1998 // the first live word on this source-region, based on which the destination
1999 // for the rest of live words can be derived.
2000 //
2001 // Note:
2002 // There is some complication due to space-boundary-fragmentation (an obj can't
2003 // cross space-boundary) -- a source-region may be split and behave like two
2004 // distinct regions with their own dest-region, as depicted below.
2005 //
2006 // source-region: region-n
2007 //
2008 // **********************
2009 // | A|A~~~~B|B |
2010 // **********************
2011 // n-1 n n+1
2012 //
2013 // AA, BB denote two live objs. ~~~~ denotes unknown number of live objs.
2014 //
2015 // Assuming the dest-region for region-n is the final region before
2016 // old-space-end and its first-live-word is the middle of AA, the heap content
2017 // will look like the following after compaction:
2018 //
2019 // ************** *************
2020 // A|A~~~~ | |BB |
2021 // ************** *************
2022 // ^ ^
2023 // | old-space-end | eden-space-start
2024 //
2025 // Therefore, in this example, region-n will have two dest-regions:
2026 // 1. the final region in old-space
2027 // 2. the first region in eden-space.
2028 // To handle this special case, we introduce the concept of split-region, whose
2029 // contents are relocated to two spaces. `SplitInfo` captures all necessary
2030 // info about the split, the first part, spliting-point, and the second part.
2031 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2032 SpaceId src_space_id,
2033 size_t src_region_idx)
2034 {
2035 const size_t RegionSize = ParallelCompactData::RegionSize;
2036 const ParallelCompactData& sd = summary_data();
2037 assert(sd.is_region_aligned(dest_addr), "precondition");
2038
2039 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2040 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2041
2042 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2043 HeapWord* const src_region_destination = src_region_ptr->destination();
2044
2045 HeapWord* const region_start = sd.region_to_addr(src_region_idx);
2046 HeapWord* const region_end = sd.region_to_addr(src_region_idx) + RegionSize;
2047
2048 // Identify the actual destination for the first live words on this region,
2049 // taking split-region into account.
2050 HeapWord* region_start_destination;
2051 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2052 if (split_info.is_split(src_region_idx)) {
2053 // The second part of this split region; use the recorded split point.
2054 if (dest_addr == src_region_destination) {
2055 return split_info.split_point();
2056 }
2057 region_start_destination = split_info.preceding_destination();
2058 } else {
2059 region_start_destination = src_region_destination;
2060 }
2061
2062 // Calculate the offset to be skipped
2063 size_t words_to_skip = pointer_delta(dest_addr, region_start_destination);
2064
2065 HeapWord* result;
2066 if (partial_obj_size > words_to_skip) {
2067 result = region_start + words_to_skip;
2068 } else {
2069 words_to_skip -= partial_obj_size;
2070 result = skip_live_words(region_start + partial_obj_size, region_end, words_to_skip);
2071 }
2072
2073 if (split_info.is_split(src_region_idx)) {
2074 assert(result < split_info.split_point(), "postcondition");
2075 } else {
2076 assert(result < region_end, "postcondition");
2077 }
2078
2079 return result;
2080 }
2081
2082 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2083 SpaceId src_space_id,
2084 size_t beg_region,
2085 HeapWord* end_addr)
2086 {
2087 ParallelCompactData& sd = summary_data();
2088
2089 #ifdef ASSERT
2090 MutableSpace* const src_space = _space_info[src_space_id].space();
2091 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2092 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2093 "src_space_id does not match beg_addr");
2094 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2095 "src_space_id does not match end_addr");
2096 #endif // #ifdef ASSERT
2097
2098 RegionData* const beg = sd.region(beg_region);
2099 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2100
2101 // Regions up to new_top() are enqueued if they become available.
2102 HeapWord* const new_top = _space_info[src_space_id].new_top();
2103 RegionData* const enqueue_end =
2104 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2105
2106 for (RegionData* cur = beg; cur < end; ++cur) {
2107 assert(cur->data_size() > 0, "region must have live data");
2108 cur->decrement_destination_count();
2109 if (cur < enqueue_end && cur->available() && cur->claim()) {
2110 if (cur->mark_normal()) {
2111 cm->push_region(sd.region(cur));
2112 } else if (cur->mark_copied()) {
2113 // Try to copy the content of the shadow region back to its corresponding
2114 // heap region if the shadow region is filled. Otherwise, the GC thread
2115 // fills the shadow region will copy the data back (see
2116 // MoveAndUpdateShadowClosure::complete_region).
2117 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2118 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2119 cur->set_completed();
2120 }
2121 }
2122 }
2123 }
2124
2125 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2126 SpaceId& src_space_id,
2127 HeapWord*& src_space_top,
2128 HeapWord* end_addr)
2129 {
2130 ParallelCompactData& sd = PSParallelCompact::summary_data();
2131
2132 size_t src_region_idx = 0;
2133
2134 // Skip empty regions (if any) up to the top of the space.
2135 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2136 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2137 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2138 const RegionData* const top_region_ptr = sd.addr_to_region_ptr(top_aligned_up);
2139
2140 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2141 ++src_region_ptr;
2142 }
2143
2144 if (src_region_ptr < top_region_ptr) {
2145 // Found the first non-empty region in the same space.
2146 src_region_idx = sd.region(src_region_ptr);
2147 closure.set_source(sd.region_to_addr(src_region_idx));
2148 return src_region_idx;
2149 }
2150
2151 // Switch to a new source space and find the first non-empty region.
2152 uint space_id = src_space_id + 1;
2153 assert(space_id < last_space_id, "not enough spaces");
2154
2155 for (/* empty */; space_id < last_space_id; ++space_id) {
2156 HeapWord* bottom = _space_info[space_id].space()->bottom();
2157 HeapWord* top = _space_info[space_id].space()->top();
2158 // Skip empty space
2159 if (bottom == top) {
2160 continue;
2161 }
2162
2163 // Identify the first region that contains live words in this space
2164 size_t cur_region = sd.addr_to_region_idx(bottom);
2165 size_t end_region = sd.addr_to_region_idx(sd.region_align_up(top));
2166
2167 for (/* empty */ ; cur_region < end_region; ++cur_region) {
2168 RegionData* cur = sd.region(cur_region);
2169 if (cur->live_obj_size() > 0) {
2170 HeapWord* region_start_addr = sd.region_to_addr(cur_region);
2171
2172 src_space_id = SpaceId(space_id);
2173 src_space_top = top;
2174 closure.set_source(region_start_addr);
2175 return cur_region;
2176 }
2177 }
2178 }
2179
2180 ShouldNotReachHere();
2181 }
2182
2183 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2184 ParallelCompactData& sd = summary_data();
2185 assert(sd.is_region_aligned(region_start_addr), "precondition");
2186
2187 // Use per-region partial_obj_size to locate the end of the obj, that extends
2188 // to region_start_addr.
2189 size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2190 size_t end_region_idx = sd.region_count();
2191 size_t accumulated_size = 0;
2192 for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2193 size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2194 accumulated_size += cur_partial_obj_size;
2195 if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2196 break;
2197 }
2198 }
2199 return region_start_addr + accumulated_size;
2200 }
2201
2202 // Use region_idx as the destination region, and evacuate all live objs on its
2203 // source regions to this destination region.
2204 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2205 {
2206 ParMarkBitMap* const bitmap = mark_bitmap();
2207 ParallelCompactData& sd = summary_data();
2208 RegionData* const region_ptr = sd.region(region_idx);
2209
2210 // Get the source region and related info.
2211 size_t src_region_idx = region_ptr->source_region();
2212 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2213 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2214 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2215
2216 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2217
2218 // Adjust src_region_idx to prepare for decrementing destination counts (the
2219 // destination count is not decremented when a region is copied to itself).
2220 if (src_region_idx == region_idx) {
2221 src_region_idx += 1;
2222 }
2223
2224 // source-region:
2225 //
2226 // **********
2227 // | ~~~ |
2228 // **********
2229 // ^
2230 // |-- closure.source() / first_src_addr
2231 //
2232 //
2233 // ~~~ : live words
2234 //
2235 // destination-region:
2236 //
2237 // **********
2238 // | |
2239 // **********
2240 // ^
2241 // |-- region-start
2242 if (bitmap->is_unmarked(closure.source())) {
2243 // An object overflows the previous destination region, so this
2244 // destination region should copy the remainder of the object or as much as
2245 // will fit.
2246 HeapWord* const old_src_addr = closure.source();
2247 {
2248 HeapWord* region_start = sd.region_align_down(closure.source());
2249 HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2250 HeapWord* obj_end;
2251 if (obj_start != closure.source()) {
2252 assert(bitmap->is_marked(obj_start), "inv");
2253 // Found the actual obj-start, try to find the obj-end using either
2254 // size() if this obj is completely contained in the current region.
2255 HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2256 HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2257 ? nullptr
2258 : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2259 // This obj extends to next region iff partial_obj_addr of the *next*
2260 // region is the same as obj-start.
2261 if (partial_obj_start == obj_start) {
2262 // This obj extends to next region.
2263 obj_end = partial_obj_end(next_region_start);
2264 } else {
2265 // Completely contained in this region; safe to use size().
2266 obj_end = obj_start + cast_to_oop(obj_start)->size();
2267 }
2268 } else {
2269 // This obj extends to current region.
2270 obj_end = partial_obj_end(region_start);
2271 }
2272 size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2273 closure.copy_partial_obj(partial_obj_size);
2274 }
2275
2276 if (closure.is_full()) {
2277 decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2278 closure.complete_region(dest_addr, region_ptr);
2279 return;
2280 }
2281
2282 // Finished copying without using up the current destination-region
2283 HeapWord* const end_addr = sd.region_align_down(closure.source());
2284 if (sd.region_align_down(old_src_addr) != end_addr) {
2285 assert(sd.region_align_up(old_src_addr) == end_addr, "only one region");
2286 // The partial object was copied from more than one source region.
2287 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2288
2289 // Move to the next source region, possibly switching spaces as well. All
2290 // args except end_addr may be modified.
2291 src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2292 }
2293 }
2294
2295 // Handle the rest obj-by-obj, where we know obj-start.
2296 do {
2297 HeapWord* cur_addr = closure.source();
2298 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2299 src_space_top);
2300 // To handle the case where the final obj in source region extends to next region.
2301 HeapWord* final_obj_start = (end_addr == src_space_top)
2302 ? nullptr
2303 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2304 // Apply closure on objs inside [cur_addr, end_addr)
2305 do {
2306 cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2307 if (cur_addr == end_addr) {
2308 break;
2309 }
2310 size_t obj_size;
2311 if (final_obj_start == cur_addr) {
2312 obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2313 } else {
2314 // This obj doesn't extend into next region; size() is safe to use.
2315 obj_size = cast_to_oop(cur_addr)->size();
2316 }
2317 closure.do_addr(cur_addr, obj_size);
2318 cur_addr += obj_size;
2319 } while (cur_addr < end_addr && !closure.is_full());
2320
2321 if (closure.is_full()) {
2322 decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2323 closure.complete_region(dest_addr, region_ptr);
2324 return;
2325 }
2326
2327 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2328
2329 // Move to the next source region, possibly switching spaces as well. All
2330 // args except end_addr may be modified.
2331 src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2332 } while (true);
2333 }
2334
2335 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2336 {
2337 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2338 fill_region(cm, cl, region_idx);
2339 }
2340
2341 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2342 {
2343 // Get a shadow region first
2344 ParallelCompactData& sd = summary_data();
2345 RegionData* const region_ptr = sd.region(region_idx);
2346 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2347 // The InvalidShadow return value indicates the corresponding heap region is available,
2348 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2349 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2350 if (shadow_region == ParCompactionManager::InvalidShadow) {
2351 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2352 region_ptr->shadow_to_normal();
2353 return fill_region(cm, cl, region_idx);
2354 } else {
2355 MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2356 return fill_region(cm, cl, region_idx);
2357 }
2358 }
2359
2360 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2361 {
2362 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2363 }
2364
2365 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
2366 {
2367 size_t next = cm->next_shadow_region();
2368 ParallelCompactData& sd = summary_data();
2369 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2370 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2371
2372 while (next < old_new_top) {
2373 if (sd.region(next)->mark_shadow()) {
2374 region_idx = next;
2375 return true;
2376 }
2377 next = cm->move_next_shadow_region_by(active_gc_threads);
2378 }
2379
2380 return false;
2381 }
2382
2383 // The shadow region is an optimization to address region dependencies in full GC. The basic
2384 // idea is making more regions available by temporally storing their live objects in empty
2385 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2386 // GC threads need not wait destination regions to be available before processing sources.
2387 //
2388 // A typical workflow would be:
2389 // After draining its own stack and failing to steal from others, a GC worker would pick an
2390 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2391 // the shadow region by copying live objects from source regions of the unavailable one. Once
2392 // the unavailable region becomes available, the data in the shadow region will be copied back.
2393 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2394 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2395 {
2396 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2397
2398 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2399 SpaceInfo* const space_info = _space_info + id;
2400 MutableSpace* const space = space_info->space();
2401
2402 const size_t beg_region =
2403 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2404 const size_t end_region =
2405 sd.addr_to_region_idx(sd.region_align_down(space->end()));
2406
2407 for (size_t cur = beg_region; cur < end_region; ++cur) {
2408 ParCompactionManager::push_shadow_region(cur);
2409 }
2410 }
2411
2412 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2413 for (uint i = 0; i < parallel_gc_threads; i++) {
2414 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2415 cm->set_next_shadow_region(beg_region + i);
2416 }
2417 }
2418
2419 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2420 {
2421 size_t words = MIN2(partial_obj_size, words_remaining());
2422
2423 // This test is necessary; if omitted, the pointer updates to a partial object
2424 // that crosses the dense prefix boundary could be overwritten.
2425 if (source() != copy_destination()) {
2426 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2427 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2428 }
2429 update_state(words);
2430 }
2431
2432 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2433 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2434 region_ptr->set_completed();
2435 }
2436
2437 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2438 assert(destination() != nullptr, "sanity");
2439 _source = addr;
2440
2441 // The start_array must be updated even if the object is not moving.
2442 if (_start_array != nullptr) {
2443 _start_array->update_for_block(destination(), destination() + words);
2444 }
2445
2446 // Avoid overflow
2447 words = MIN2(words, words_remaining());
2448 assert(words > 0, "inv");
2449
2450 if (copy_destination() != source()) {
2451 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2452 assert(source() != destination(), "inv");
2453 assert(FullGCForwarding::is_forwarded(cast_to_oop(source())), "inv");
2454 assert(FullGCForwarding::forwardee(cast_to_oop(source())) == cast_to_oop(destination()), "inv");
2455 // Read the klass before the copying, since it might destroy the klass (i.e. overlapping copy)
2456 // and if partial copy, the destination klass may not be copied yet
2457 Klass* klass = cast_to_oop(source())->klass();
2458 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2459 cast_to_oop(copy_destination())->set_mark(Klass::default_prototype_header(klass));
2460 }
2461
2462 update_state(words);
2463 }
2464
2465 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2466 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2467 // Record the shadow region index
2468 region_ptr->set_shadow_region(_shadow);
2469 // Mark the shadow region as filled to indicate the data is ready to be
2470 // copied back
2471 region_ptr->mark_filled();
2472 // Try to copy the content of the shadow region back to its corresponding
2473 // heap region if available; the GC thread that decreases the destination
2474 // count to zero will do the copying otherwise (see
2475 // PSParallelCompact::decrement_destination_counts).
2476 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2477 region_ptr->set_completed();
2478 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2479 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2480 }
2481 }