1 /*
2 * Copyright (c) 2005, 2024, 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 "precompiled.hpp"
26 #include "classfile/classLoaderDataGraph.hpp"
27 #include "classfile/javaClasses.inline.hpp"
28 #include "classfile/stringTable.hpp"
29 #include "classfile/symbolTable.hpp"
30 #include "classfile/systemDictionary.hpp"
31 #include "code/codeCache.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/parallel/objectStartArray.inline.hpp"
34 #include "gc/parallel/parallelArguments.hpp"
35 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
36 #include "gc/parallel/parMarkBitMap.inline.hpp"
37 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
38 #include "gc/parallel/psCompactionManager.inline.hpp"
39 #include "gc/parallel/psOldGen.hpp"
40 #include "gc/parallel/psParallelCompact.inline.hpp"
41 #include "gc/parallel/psPromotionManager.inline.hpp"
42 #include "gc/parallel/psRootType.hpp"
43 #include "gc/parallel/psScavenge.hpp"
44 #include "gc/parallel/psStringDedup.hpp"
45 #include "gc/parallel/psYoungGen.hpp"
46 #include "gc/shared/classUnloadingContext.hpp"
47 #include "gc/shared/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/strongRootsScope.hpp"
64 #include "gc/shared/taskTerminator.hpp"
65 #include "gc/shared/weakProcessor.inline.hpp"
66 #include "gc/shared/workerPolicy.hpp"
67 #include "gc/shared/workerThread.hpp"
68 #include "gc/shared/workerUtils.hpp"
69 #include "logging/log.hpp"
70 #include "memory/iterator.inline.hpp"
71 #include "memory/metaspaceUtils.hpp"
72 #include "memory/resourceArea.hpp"
73 #include "memory/universe.hpp"
74 #include "nmt/memTracker.hpp"
75 #include "oops/access.inline.hpp"
76 #include "oops/instanceClassLoaderKlass.inline.hpp"
77 #include "oops/instanceKlass.inline.hpp"
78 #include "oops/instanceMirrorKlass.inline.hpp"
79 #include "oops/methodData.hpp"
80 #include "oops/objArrayKlass.inline.hpp"
81 #include "oops/oop.inline.hpp"
82 #include "runtime/atomic.hpp"
83 #include "runtime/handles.inline.hpp"
84 #include "runtime/java.hpp"
85 #include "runtime/safepoint.hpp"
86 #include "runtime/threads.hpp"
87 #include "runtime/vmThread.hpp"
88 #include "services/memoryService.hpp"
89 #include "utilities/align.hpp"
90 #include "utilities/debug.hpp"
91 #include "utilities/events.hpp"
92 #include "utilities/formatBuffer.hpp"
93 #include "utilities/macros.hpp"
94 #include "utilities/stack.inline.hpp"
95 #if INCLUDE_JVMCI
96 #include "jvmci/jvmci.hpp"
97 #endif
98
99 #include <math.h>
100
101 // All sizes are in HeapWords.
102 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
103 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
104 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
105 const size_t ParallelCompactData::RegionSizeBytes =
106 RegionSize << LogHeapWordSize;
107 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
108 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
109 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
110
111 const ParallelCompactData::RegionData::region_sz_t
112 ParallelCompactData::RegionData::dc_shift = 27;
113
114 const ParallelCompactData::RegionData::region_sz_t
115 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
116
117 const ParallelCompactData::RegionData::region_sz_t
118 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
119
120 const ParallelCompactData::RegionData::region_sz_t
121 ParallelCompactData::RegionData::los_mask = ~dc_mask;
122
123 const ParallelCompactData::RegionData::region_sz_t
124 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
125
126 const ParallelCompactData::RegionData::region_sz_t
127 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
128
129 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
130
131 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
132 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
133
134 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
135 HeapWord* destination)
136 {
137 assert(src_region_idx != 0, "invalid src_region_idx");
138 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
139 assert(destination != nullptr, "invalid destination argument");
140
141 _src_region_idx = src_region_idx;
142 _partial_obj_size = partial_obj_size;
143 _destination = destination;
144
145 // These fields may not be updated below, so make sure they're clear.
146 assert(_dest_region_addr == nullptr, "should have been cleared");
147 assert(_first_src_addr == nullptr, "should have been cleared");
148
149 // Determine the number of destination regions for the partial object.
150 HeapWord* const last_word = destination + partial_obj_size - 1;
151 const ParallelCompactData& sd = PSParallelCompact::summary_data();
152 HeapWord* const beg_region_addr = sd.region_align_down(destination);
153 HeapWord* const end_region_addr = sd.region_align_down(last_word);
154
155 if (beg_region_addr == end_region_addr) {
156 // One destination region.
157 _destination_count = 1;
158 if (end_region_addr == destination) {
159 // The destination falls on a region boundary, thus the first word of the
160 // partial object will be the first word copied to the destination region.
161 _dest_region_addr = end_region_addr;
162 _first_src_addr = sd.region_to_addr(src_region_idx);
163 }
164 } else {
165 // Two destination regions. When copied, the partial object will cross a
166 // destination region boundary, so a word somewhere within the partial
167 // object will be the first word copied to the second destination region.
168 _destination_count = 2;
169 _dest_region_addr = end_region_addr;
170 const size_t ofs = pointer_delta(end_region_addr, destination);
171 assert(ofs < _partial_obj_size, "sanity");
172 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
173 }
174 }
175
176 void SplitInfo::clear()
177 {
178 _src_region_idx = 0;
179 _partial_obj_size = 0;
180 _destination = nullptr;
181 _destination_count = 0;
182 _dest_region_addr = nullptr;
183 _first_src_addr = nullptr;
184 assert(!is_valid(), "sanity");
185 }
186
187 #ifdef ASSERT
188 void SplitInfo::verify_clear()
189 {
190 assert(_src_region_idx == 0, "not clear");
191 assert(_partial_obj_size == 0, "not clear");
192 assert(_destination == nullptr, "not clear");
193 assert(_destination_count == 0, "not clear");
194 assert(_dest_region_addr == nullptr, "not clear");
195 assert(_first_src_addr == nullptr, "not clear");
196 }
197 #endif // #ifdef ASSERT
198
199
200 void PSParallelCompact::print_on_error(outputStream* st) {
201 _mark_bitmap.print_on_error(st);
202 }
203
204 ParallelCompactData::ParallelCompactData() :
205 _heap_start(nullptr),
206 DEBUG_ONLY(_heap_end(nullptr) COMMA)
207 _region_vspace(nullptr),
208 _reserved_byte_size(0),
209 _region_data(nullptr),
210 _region_count(0) {}
211
212 bool ParallelCompactData::initialize(MemRegion reserved_heap)
213 {
214 _heap_start = reserved_heap.start();
215 const size_t heap_size = reserved_heap.word_size();
216 DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
217
218 assert(region_align_down(_heap_start) == _heap_start,
219 "region start not aligned");
220
221 return initialize_region_data(heap_size);
222 }
223
224 PSVirtualSpace*
225 ParallelCompactData::create_vspace(size_t count, size_t element_size)
226 {
227 const size_t raw_bytes = count * element_size;
228 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
229 const size_t granularity = os::vm_allocation_granularity();
230 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
231
232 const size_t rs_align = page_sz == os::vm_page_size() ? 0 :
233 MAX2(page_sz, granularity);
234 ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
235 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
236 rs.size(), page_sz);
237
238 MemTracker::record_virtual_memory_tag((address)rs.base(), mtGC);
239
240 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
241 if (vspace != nullptr) {
242 if (vspace->expand_by(_reserved_byte_size)) {
243 return vspace;
244 }
245 delete vspace;
246 // Release memory reserved in the space.
247 rs.release();
248 }
249
250 return nullptr;
251 }
252
253 bool ParallelCompactData::initialize_region_data(size_t heap_size)
254 {
255 assert(is_aligned(heap_size, RegionSize), "precondition");
256
257 const size_t count = heap_size >> Log2RegionSize;
258 _region_vspace = create_vspace(count, sizeof(RegionData));
259 if (_region_vspace != nullptr) {
260 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
261 _region_count = count;
262 return true;
263 }
264 return false;
265 }
266
267 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
268 assert(beg_region <= _region_count, "beg_region out of range");
269 assert(end_region <= _region_count, "end_region out of range");
270
271 const size_t region_cnt = end_region - beg_region;
272 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
273 }
274
275 void
276 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
277 {
278 assert(is_region_aligned(beg), "not RegionSize aligned");
279 assert(is_region_aligned(end), "not RegionSize aligned");
280
281 size_t cur_region = addr_to_region_idx(beg);
282 const size_t end_region = addr_to_region_idx(end);
283 HeapWord* addr = beg;
284 while (cur_region < end_region) {
285 _region_data[cur_region].set_destination(addr);
286 _region_data[cur_region].set_destination_count(0);
287 _region_data[cur_region].set_source_region(cur_region);
288
289 // Update live_obj_size so the region appears completely full.
290 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
291 _region_data[cur_region].set_live_obj_size(live_size);
292
293 ++cur_region;
294 addr += RegionSize;
295 }
296 }
297
298 // Find the point at which a space can be split and, if necessary, record the
299 // split point.
300 //
301 // If the current src region (which overflowed the destination space) doesn't
302 // have a partial object, the split point is at the beginning of the current src
303 // region (an "easy" split, no extra bookkeeping required).
304 //
305 // If the current src region has a partial object, the split point is in the
306 // region where that partial object starts (call it the split_region). If
307 // split_region has a partial object, then the split point is just after that
308 // partial object (a "hard" split where we have to record the split data and
309 // zero the partial_obj_size field). With a "hard" split, we know that the
310 // partial_obj ends within split_region because the partial object that caused
311 // the overflow starts in split_region. If split_region doesn't have a partial
312 // obj, then the split is at the beginning of split_region (another "easy"
313 // split).
314 HeapWord*
315 ParallelCompactData::summarize_split_space(size_t src_region,
316 SplitInfo& split_info,
317 HeapWord* destination,
318 HeapWord* target_end,
319 HeapWord** target_next)
320 {
321 assert(destination <= target_end, "sanity");
322 assert(destination + _region_data[src_region].data_size() > target_end,
323 "region should not fit into target space");
324 assert(is_region_aligned(target_end), "sanity");
325
326 size_t split_region = src_region;
327 HeapWord* split_destination = destination;
328 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
329
330 if (destination + partial_obj_size > target_end) {
331 // The split point is just after the partial object (if any) in the
332 // src_region that contains the start of the object that overflowed the
333 // destination space.
334 //
335 // Find the start of the "overflow" object and set split_region to the
336 // region containing it.
337 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
338 split_region = addr_to_region_idx(overflow_obj);
339
340 // Clear the source_region field of all destination regions whose first word
341 // came from data after the split point (a non-null source_region field
342 // implies a region must be filled).
343 //
344 // An alternative to the simple loop below: clear during post_compact(),
345 // which uses memcpy instead of individual stores, and is easy to
346 // parallelize. (The downside is that it clears the entire RegionData
347 // object as opposed to just one field.)
348 //
349 // post_compact() would have to clear the summary data up to the highest
350 // address that was written during the summary phase, which would be
351 //
352 // max(top, max(new_top, clear_top))
353 //
354 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
355 // to target_end.
356 const RegionData* const sr = region(split_region);
357 const size_t beg_idx =
358 addr_to_region_idx(region_align_up(sr->destination() +
359 sr->partial_obj_size()));
360 const size_t end_idx = addr_to_region_idx(target_end);
361
362 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
363 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
364 _region_data[idx].set_source_region(0);
365 }
366
367 // Set split_destination and partial_obj_size to reflect the split region.
368 split_destination = sr->destination();
369 partial_obj_size = sr->partial_obj_size();
370 }
371
372 // The split is recorded only if a partial object extends onto the region.
373 if (partial_obj_size != 0) {
374 _region_data[split_region].set_partial_obj_size(0);
375 split_info.record(split_region, partial_obj_size, split_destination);
376 }
377
378 // Setup the continuation addresses.
379 *target_next = split_destination + partial_obj_size;
380 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
381
382 if (log_develop_is_enabled(Trace, gc, compaction)) {
383 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
384 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
385 split_type, p2i(source_next), split_region, partial_obj_size);
386 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
387 split_type, p2i(split_destination),
388 addr_to_region_idx(split_destination),
389 p2i(*target_next));
390
391 if (partial_obj_size != 0) {
392 HeapWord* const po_beg = split_info.destination();
393 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
394 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
395 split_type,
396 p2i(po_beg), addr_to_region_idx(po_beg),
397 p2i(po_end), addr_to_region_idx(po_end));
398 }
399 }
400
401 return source_next;
402 }
403
404 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
405 HeapWord** full_region_prefix_end) {
406 size_t cur_region = addr_to_region_idx(space->bottom());
407 const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
408 size_t live_words = 0;
409 if (full_region_prefix_end == nullptr) {
410 for (/* empty */; cur_region < end_region; ++cur_region) {
411 live_words += _region_data[cur_region].data_size();
412 }
413 } else {
414 bool first_set = false;
415 for (/* empty */; cur_region < end_region; ++cur_region) {
416 size_t live_words_in_region = _region_data[cur_region].data_size();
417 if (!first_set && live_words_in_region < RegionSize) {
418 *full_region_prefix_end = region_to_addr(cur_region);
419 first_set = true;
420 }
421 live_words += live_words_in_region;
422 }
423 if (!first_set) {
424 // All regions are full of live objs.
425 assert(is_region_aligned(space->top()), "inv");
426 *full_region_prefix_end = space->top();
427 }
428 assert(*full_region_prefix_end != nullptr, "postcondition");
429 assert(is_region_aligned(*full_region_prefix_end), "inv");
430 assert(*full_region_prefix_end >= space->bottom(), "in-range");
431 assert(*full_region_prefix_end <= space->top(), "in-range");
432 }
433 return live_words;
434 }
435
436 bool ParallelCompactData::summarize(SplitInfo& split_info,
437 HeapWord* source_beg, HeapWord* source_end,
438 HeapWord** source_next,
439 HeapWord* target_beg, HeapWord* target_end,
440 HeapWord** target_next)
441 {
442 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
443 log_develop_trace(gc, compaction)(
444 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
445 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
446 p2i(source_beg), p2i(source_end), p2i(source_next_val),
447 p2i(target_beg), p2i(target_end), p2i(*target_next));
448
449 size_t cur_region = addr_to_region_idx(source_beg);
450 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
451
452 HeapWord *dest_addr = target_beg;
453 while (cur_region < end_region) {
454 // The destination must be set even if the region has no data.
455 _region_data[cur_region].set_destination(dest_addr);
456
457 size_t words = _region_data[cur_region].data_size();
458 if (words > 0) {
459 // If cur_region does not fit entirely into the target space, find a point
460 // at which the source space can be 'split' so that part is copied to the
461 // target space and the rest is copied elsewhere.
462 if (dest_addr + words > target_end) {
463 assert(source_next != nullptr, "source_next is null when splitting");
464 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
465 target_end, target_next);
466 return false;
467 }
468
469 // Compute the destination_count for cur_region, and if necessary, update
470 // source_region for a destination region. The source_region field is
471 // updated if cur_region is the first (left-most) region to be copied to a
472 // destination region.
473 //
474 // The destination_count calculation is a bit subtle. A region that has
475 // data that compacts into itself does not count itself as a destination.
476 // This maintains the invariant that a zero count means the region is
477 // available and can be claimed and then filled.
478 uint destination_count = 0;
479 if (split_info.is_split(cur_region)) {
480 // The current region has been split: the partial object will be copied
481 // to one destination space and the remaining data will be copied to
482 // another destination space. Adjust the initial destination_count and,
483 // if necessary, set the source_region field if the partial object will
484 // cross a destination region boundary.
485 destination_count = split_info.destination_count();
486 if (destination_count == 2) {
487 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
488 _region_data[dest_idx].set_source_region(cur_region);
489 }
490 }
491
492 HeapWord* const last_addr = dest_addr + words - 1;
493 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
494 const size_t dest_region_2 = addr_to_region_idx(last_addr);
495
496 // Initially assume that the destination regions will be the same and
497 // adjust the value below if necessary. Under this assumption, if
498 // cur_region == dest_region_2, then cur_region will be compacted
499 // completely into itself.
500 destination_count += cur_region == dest_region_2 ? 0 : 1;
501 if (dest_region_1 != dest_region_2) {
502 // Destination regions differ; adjust destination_count.
503 destination_count += 1;
504 // Data from cur_region will be copied to the start of dest_region_2.
505 _region_data[dest_region_2].set_source_region(cur_region);
506 } else if (is_region_aligned(dest_addr)) {
507 // Data from cur_region will be copied to the start of the destination
508 // region.
509 _region_data[dest_region_1].set_source_region(cur_region);
510 }
511
512 _region_data[cur_region].set_destination_count(destination_count);
513 dest_addr += words;
514 }
515
516 ++cur_region;
517 }
518
519 *target_next = dest_addr;
520 return true;
521 }
522
523 #ifdef ASSERT
524 void ParallelCompactData::verify_clear()
525 {
526 const size_t* const beg = (const size_t*) _region_vspace->committed_low_addr();
527 const size_t* const end = (const size_t*) _region_vspace->committed_high_addr();
528 for (const size_t* p = beg; p < end; ++p) {
529 assert(*p == 0, "not zero");
530 }
531 }
532 #endif // #ifdef ASSERT
533
534 STWGCTimer PSParallelCompact::_gc_timer;
535 ParallelOldTracer PSParallelCompact::_gc_tracer;
536 elapsedTimer PSParallelCompact::_accumulated_time;
537 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
538 CollectorCounters* PSParallelCompact::_counters = nullptr;
539 ParMarkBitMap PSParallelCompact::_mark_bitmap;
540 ParallelCompactData PSParallelCompact::_summary_data;
541
542 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
543
544 class PCAdjustPointerClosure: public BasicOopIterateClosure {
545 template <typename T>
546 void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
547
548 public:
549 virtual void do_oop(oop* p) { do_oop_work(p); }
550 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
551
552 virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
553 };
554
555 static PCAdjustPointerClosure pc_adjust_pointer_closure;
556
557 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
558
559 void PSParallelCompact::post_initialize() {
560 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
561 _span_based_discoverer.set_span(heap->reserved_region());
562 _ref_processor =
563 new ReferenceProcessor(&_span_based_discoverer,
564 ParallelGCThreads, // mt processing degree
565 ParallelGCThreads, // mt discovery degree
566 false, // concurrent_discovery
567 &_is_alive_closure); // non-header is alive closure
568
569 _counters = new CollectorCounters("Parallel full collection pauses", 1);
570
571 // Initialize static fields in ParCompactionManager.
572 ParCompactionManager::initialize(mark_bitmap());
573 }
574
575 bool PSParallelCompact::initialize_aux_data() {
576 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
577 MemRegion mr = heap->reserved_region();
578 assert(mr.byte_size() != 0, "heap should be reserved");
579
580 initialize_space_info();
581
582 if (!_mark_bitmap.initialize(mr)) {
583 vm_shutdown_during_initialization(
584 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
585 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
586 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
587 return false;
588 }
589
590 if (!_summary_data.initialize(mr)) {
591 vm_shutdown_during_initialization(
592 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
593 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
594 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
595 return false;
596 }
597
598 return true;
599 }
600
601 void PSParallelCompact::initialize_space_info()
602 {
603 memset(&_space_info, 0, sizeof(_space_info));
604
605 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
606 PSYoungGen* young_gen = heap->young_gen();
607
608 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
609 _space_info[eden_space_id].set_space(young_gen->eden_space());
610 _space_info[from_space_id].set_space(young_gen->from_space());
611 _space_info[to_space_id].set_space(young_gen->to_space());
612
613 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
614 }
615
616 void
617 PSParallelCompact::clear_data_covering_space(SpaceId id)
618 {
619 // At this point, top is the value before GC, new_top() is the value that will
620 // be set at the end of GC. The marking bitmap is cleared to top; nothing
621 // should be marked above top. The summary data is cleared to the larger of
622 // top & new_top.
623 MutableSpace* const space = _space_info[id].space();
624 HeapWord* const bot = space->bottom();
625 HeapWord* const top = space->top();
626 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
627
628 _mark_bitmap.clear_range(bot, top);
629
630 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
631 const size_t end_region =
632 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
633 _summary_data.clear_range(beg_region, end_region);
634
635 // Clear the data used to 'split' regions.
636 SplitInfo& split_info = _space_info[id].split_info();
637 if (split_info.is_valid()) {
638 split_info.clear();
639 }
640 DEBUG_ONLY(split_info.verify_clear();)
641 }
642
643 void PSParallelCompact::pre_compact()
644 {
645 // Update the from & to space pointers in space_info, since they are swapped
646 // at each young gen gc. Do the update unconditionally (even though a
647 // promotion failure does not swap spaces) because an unknown number of young
648 // collections will have swapped the spaces an unknown number of times.
649 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
650 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
651 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
652 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
653
654 // Increment the invocation count
655 heap->increment_total_collections(true);
656
657 CodeCache::on_gc_marking_cycle_start();
658
659 heap->print_heap_before_gc();
660 heap->trace_heap_before_gc(&_gc_tracer);
661
662 // Fill in TLABs
663 heap->ensure_parsability(true); // retire TLABs
664
665 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
666 Universe::verify("Before GC");
667 }
668
669 DEBUG_ONLY(mark_bitmap()->verify_clear();)
670 DEBUG_ONLY(summary_data().verify_clear();)
671 }
672
673 void PSParallelCompact::post_compact()
674 {
675 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
676 ParCompactionManager::remove_all_shadow_regions();
677
678 CodeCache::on_gc_marking_cycle_finish();
679 CodeCache::arm_all_nmethods();
680
681 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
682 // Clear the marking bitmap, summary data and split info.
683 clear_data_covering_space(SpaceId(id));
684 {
685 MutableSpace* space = _space_info[id].space();
686 HeapWord* top = space->top();
687 HeapWord* new_top = _space_info[id].new_top();
688 if (ZapUnusedHeapArea && new_top < top) {
689 space->mangle_region(MemRegion(new_top, top));
690 }
691 // Update top(). Must be done after clearing the bitmap and summary data.
692 space->set_top(new_top);
693 }
694 }
695
696 ParCompactionManager::flush_all_string_dedup_requests();
697
698 MutableSpace* const eden_space = _space_info[eden_space_id].space();
699 MutableSpace* const from_space = _space_info[from_space_id].space();
700 MutableSpace* const to_space = _space_info[to_space_id].space();
701
702 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
703 bool eden_empty = eden_space->is_empty();
704
705 // Update heap occupancy information which is used as input to the soft ref
706 // clearing policy at the next gc.
707 Universe::heap()->update_capacity_and_used_at_gc();
708
709 bool young_gen_empty = eden_empty && from_space->is_empty() &&
710 to_space->is_empty();
711
712 PSCardTable* ct = heap->card_table();
713 MemRegion old_mr = heap->old_gen()->committed();
714 if (young_gen_empty) {
715 ct->clear_MemRegion(old_mr);
716 } else {
717 ct->dirty_MemRegion(old_mr);
718 }
719
720 {
721 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
722 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
723 ClassLoaderDataGraph::purge(true /* at_safepoint */);
724 DEBUG_ONLY(MetaspaceUtils::verify();)
725 }
726
727 // Need to clear claim bits for the next mark.
728 ClassLoaderDataGraph::clear_claimed_marks();
729
730 heap->prune_scavengable_nmethods();
731
732 #if COMPILER2_OR_JVMCI
733 DerivedPointerTable::update_pointers();
734 #endif
735
736 // Signal that we have completed a visit to all live objects.
737 Universe::heap()->record_whole_heap_examined_timestamp();
738 }
739
740 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
741 HeapWord* full_region_prefix_end) {
742 const size_t region_size = ParallelCompactData::RegionSize;
743 const ParallelCompactData& sd = summary_data();
744
745 // Iteration starts with the region *after* the full-region-prefix-end.
746 const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
747 // If final region is not full, iteration stops before that region,
748 // because fill_dense_prefix_end assumes that prefix_end <= top.
749 const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
750 assert(start_region <= end_region, "inv");
751
752 size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
753 const RegionData* cur_region = start_region;
754 for (/* empty */; cur_region < end_region; ++cur_region) {
755 assert(region_size >= cur_region->data_size(), "inv");
756 size_t dead_size = region_size - cur_region->data_size();
757 if (max_waste < dead_size) {
758 break;
759 }
760 max_waste -= dead_size;
761 }
762
763 HeapWord* const prefix_end = sd.region_to_addr(cur_region);
764 assert(sd.is_region_aligned(prefix_end), "postcondition");
765 assert(prefix_end >= full_region_prefix_end, "in-range");
766 assert(prefix_end <= old_space->top(), "in-range");
767 return prefix_end;
768 }
769
770 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
771 // Comparing two sizes to decide if filling is required:
772 //
773 // The size of the filler (min-obj-size) is 2 heap words with the default
774 // MinObjAlignment, since both markword and klass take 1 heap word.
775 //
776 // The size of the gap (if any) right before dense-prefix-end is
777 // MinObjAlignment.
778 //
779 // Need to fill in the gap only if it's smaller than min-obj-size, and the
780 // filler obj will extend to next region.
781
782 // Note: If min-fill-size decreases to 1, this whole method becomes redundant.
783 assert(CollectedHeap::min_fill_size() >= 2, "inv");
784 #ifndef _LP64
785 // In 32-bit system, each heap word is 4 bytes, so MinObjAlignment == 2.
786 // The gap is always equal to min-fill-size, so nothing to do.
787 return;
788 #endif
789 if (MinObjAlignment > 1) {
790 return;
791 }
792 assert(CollectedHeap::min_fill_size() == 2, "inv");
793 HeapWord* const dense_prefix_end = dense_prefix(id);
794 assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
795 assert(dense_prefix_end <= space(id)->top(), "precondition");
796 if (dense_prefix_end == space(id)->top()) {
797 // Must not have single-word gap right before prefix-end/top.
798 return;
799 }
800 RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
801
802 if (region_after_dense_prefix->partial_obj_size() != 0 ||
803 _mark_bitmap.is_marked(dense_prefix_end)) {
804 // The region after the dense prefix starts with live bytes.
805 return;
806 }
807
808 HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
809 if (block_start == dense_prefix_end - 1) {
810 assert(!_mark_bitmap.is_marked(block_start), "inv");
811 // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
812 // The filler object will extend into region_after_dense_prefix.
813 const size_t obj_len = 2; // min-fill-size
814 HeapWord* const obj_beg = dense_prefix_end - 1;
815 CollectedHeap::fill_with_object(obj_beg, obj_len);
816 _mark_bitmap.mark_obj(obj_beg);
817 _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
818 region_after_dense_prefix->set_partial_obj_size(1);
819 region_after_dense_prefix->set_partial_obj_addr(obj_beg);
820 assert(start_array(id) != nullptr, "sanity");
821 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
822 }
823 }
824
825 bool PSParallelCompact::check_maximum_compaction(size_t total_live_words,
826 MutableSpace* const old_space,
827 HeapWord* full_region_prefix_end) {
828
829 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
830
831 // Check System.GC
832 bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
833 && GCCause::is_user_requested_gc(heap->gc_cause());
834
835 // Check if all live objs are larger than old-gen.
836 const bool is_old_gen_overflowing = (total_live_words > old_space->capacity_in_words());
837
838 // JVM flags
839 const uint total_invocations = heap->total_full_collections();
840 assert(total_invocations >= _maximum_compaction_gc_num, "sanity");
841 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num;
842 const bool is_interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
843
844 // If all regions in old-gen are full
845 const bool is_region_full =
846 full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
847
848 if (is_max_on_system_gc || is_old_gen_overflowing || is_interval_ended || is_region_full) {
849 _maximum_compaction_gc_num = total_invocations;
850 return true;
851 }
852
853 return false;
854 }
855
856 void PSParallelCompact::summary_phase()
857 {
858 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
859
860 MutableSpace* const old_space = _space_info[old_space_id].space();
861 {
862 size_t total_live_words = 0;
863 HeapWord* full_region_prefix_end = nullptr;
864 {
865 // old-gen
866 size_t live_words = _summary_data.live_words_in_space(old_space,
867 &full_region_prefix_end);
868 total_live_words += live_words;
869 }
870 // young-gen
871 for (uint i = eden_space_id; i < last_space_id; ++i) {
872 const MutableSpace* space = _space_info[i].space();
873 size_t live_words = _summary_data.live_words_in_space(space);
874 total_live_words += live_words;
875 _space_info[i].set_new_top(space->bottom() + live_words);
876 _space_info[i].set_dense_prefix(space->bottom());
877 }
878
879 bool maximum_compaction = check_maximum_compaction(total_live_words,
880 old_space,
881 full_region_prefix_end);
882 HeapWord* dense_prefix_end =
883 maximum_compaction ? full_region_prefix_end
884 : compute_dense_prefix_for_old_space(old_space,
885 full_region_prefix_end);
886 SpaceId id = old_space_id;
887 _space_info[id].set_dense_prefix(dense_prefix_end);
888
889 if (dense_prefix_end != old_space->bottom()) {
890 fill_dense_prefix_end(id);
891 _summary_data.summarize_dense_prefix(old_space->bottom(), dense_prefix_end);
892 }
893 _summary_data.summarize(_space_info[id].split_info(),
894 dense_prefix_end, old_space->top(), nullptr,
895 dense_prefix_end, old_space->end(),
896 _space_info[id].new_top_addr());
897 }
898
899 // Summarize the remaining spaces in the young gen. The initial target space
900 // is the old gen. If a space does not fit entirely into the target, then the
901 // remainder is compacted into the space itself and that space becomes the new
902 // target.
903 SpaceId dst_space_id = old_space_id;
904 HeapWord* dst_space_end = old_space->end();
905 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
906 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
907 const MutableSpace* space = _space_info[id].space();
908 const size_t live = pointer_delta(_space_info[id].new_top(),
909 space->bottom());
910 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
911
912 if (live > 0 && live <= available) {
913 // All the live data will fit.
914 bool done = _summary_data.summarize(_space_info[id].split_info(),
915 space->bottom(), space->top(),
916 nullptr,
917 *new_top_addr, dst_space_end,
918 new_top_addr);
919 assert(done, "space must fit into old gen");
920
921 // Reset the new_top value for the space.
922 _space_info[id].set_new_top(space->bottom());
923 } else if (live > 0) {
924 // Attempt to fit part of the source space into the target space.
925 HeapWord* next_src_addr = nullptr;
926 bool done = _summary_data.summarize(_space_info[id].split_info(),
927 space->bottom(), space->top(),
928 &next_src_addr,
929 *new_top_addr, dst_space_end,
930 new_top_addr);
931 assert(!done, "space should not fit into old gen");
932 assert(next_src_addr != nullptr, "sanity");
933
934 // The source space becomes the new target, so the remainder is compacted
935 // within the space itself.
936 dst_space_id = SpaceId(id);
937 dst_space_end = space->end();
938 new_top_addr = _space_info[id].new_top_addr();
939 done = _summary_data.summarize(_space_info[id].split_info(),
940 next_src_addr, space->top(),
941 nullptr,
942 space->bottom(), dst_space_end,
943 new_top_addr);
944 assert(done, "space must fit when compacted into itself");
945 assert(*new_top_addr <= space->top(), "usage should not grow");
946 }
947 }
948 }
949
950 // This method should contain all heap-specific policy for invoking a full
951 // collection. invoke_no_policy() will only attempt to compact the heap; it
952 // will do nothing further. If we need to bail out for policy reasons, scavenge
953 // before full gc, or any other specialized behavior, it needs to be added here.
954 //
955 // Note that this method should only be called from the vm_thread while at a
956 // safepoint.
957 //
958 // Note that the all_soft_refs_clear flag in the soft ref policy
959 // may be true because this method can be called without intervening
960 // activity. For example when the heap space is tight and full measure
961 // are being taken to free space.
962 bool PSParallelCompact::invoke(bool clear_all_soft_refs) {
963 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
964 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
965 "should be in vm thread");
966
967 SvcGCMarker sgcm(SvcGCMarker::FULL);
968 IsSTWGCActiveMark mark;
969
970 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
971 clear_all_soft_refs = clear_all_soft_refs
972 || heap->soft_ref_policy()->should_clear_all_soft_refs();
973
974 return PSParallelCompact::invoke_no_policy(clear_all_soft_refs);
975 }
976
977 // This method contains no policy. You should probably
978 // be calling invoke() instead.
979 bool PSParallelCompact::invoke_no_policy(bool clear_all_soft_refs) {
980 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
981 assert(ref_processor() != nullptr, "Sanity");
982
983 if (GCLocker::check_active_before_gc()) {
984 return false;
985 }
986
987 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
988
989 GCIdMark gc_id_mark;
990 _gc_timer.register_gc_start();
991 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
992
993 GCCause::Cause gc_cause = heap->gc_cause();
994 PSYoungGen* young_gen = heap->young_gen();
995 PSOldGen* old_gen = heap->old_gen();
996 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
997
998 // The scope of casr should end after code that can change
999 // SoftRefPolicy::_should_clear_all_soft_refs.
1000 ClearedAllSoftRefs casr(clear_all_soft_refs,
1001 heap->soft_ref_policy());
1002
1003 // Make sure data structures are sane, make the heap parsable, and do other
1004 // miscellaneous bookkeeping.
1005 pre_compact();
1006
1007 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1008
1009 {
1010 const uint active_workers =
1011 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1012 ParallelScavengeHeap::heap()->workers().active_workers(),
1013 Threads::number_of_non_daemon_threads());
1014 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1015
1016 GCTraceCPUTime tcpu(&_gc_tracer);
1017 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
1018
1019 heap->pre_full_gc_dump(&_gc_timer);
1020
1021 TraceCollectorStats tcs(counters());
1022 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
1023
1024 if (log_is_enabled(Debug, gc, heap, exit)) {
1025 accumulated_time()->start();
1026 }
1027
1028 // Let the size policy know we're starting
1029 size_policy->major_collection_begin();
1030
1031 #if COMPILER2_OR_JVMCI
1032 DerivedPointerTable::clear();
1033 #endif
1034
1035 ref_processor()->start_discovery(clear_all_soft_refs);
1036
1037 ClassUnloadingContext ctx(1 /* num_nmethod_unlink_workers */,
1038 false /* unregister_nmethods_during_purge */,
1039 false /* lock_nmethod_free_separately */);
1040
1041 marking_phase(&_gc_tracer);
1042
1043 summary_phase();
1044
1045 #if COMPILER2_OR_JVMCI
1046 assert(DerivedPointerTable::is_active(), "Sanity");
1047 DerivedPointerTable::set_active(false);
1048 #endif
1049
1050 forward_to_new_addr();
1051
1052 adjust_pointers();
1053
1054 compact();
1055
1056 ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1057
1058 ParCompactionManager::verify_all_region_stack_empty();
1059
1060 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1061 // done before resizing.
1062 post_compact();
1063
1064 // Let the size policy know we're done
1065 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1066
1067 if (UseAdaptiveSizePolicy) {
1068 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1069 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1070 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1071
1072 // Don't check if the size_policy is ready here. Let
1073 // the size_policy check that internally.
1074 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1075 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1076 // Swap the survivor spaces if from_space is empty. The
1077 // resize_young_gen() called below is normally used after
1078 // a successful young GC and swapping of survivor spaces;
1079 // otherwise, it will fail to resize the young gen with
1080 // the current implementation.
1081 if (young_gen->from_space()->is_empty()) {
1082 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1083 young_gen->swap_spaces();
1084 }
1085
1086 // Calculate optimal free space amounts
1087 assert(young_gen->max_gen_size() >
1088 young_gen->from_space()->capacity_in_bytes() +
1089 young_gen->to_space()->capacity_in_bytes(),
1090 "Sizes of space in young gen are out-of-bounds");
1091
1092 size_t young_live = young_gen->used_in_bytes();
1093 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1094 size_t old_live = old_gen->used_in_bytes();
1095 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1096 size_t max_old_gen_size = old_gen->max_gen_size();
1097 size_t max_eden_size = young_gen->max_gen_size() -
1098 young_gen->from_space()->capacity_in_bytes() -
1099 young_gen->to_space()->capacity_in_bytes();
1100
1101 // Used for diagnostics
1102 size_policy->clear_generation_free_space_flags();
1103
1104 size_policy->compute_generations_free_space(young_live,
1105 eden_live,
1106 old_live,
1107 cur_eden,
1108 max_old_gen_size,
1109 max_eden_size,
1110 true /* full gc*/);
1111
1112 size_policy->check_gc_overhead_limit(eden_live,
1113 max_old_gen_size,
1114 max_eden_size,
1115 true /* full gc*/,
1116 gc_cause,
1117 heap->soft_ref_policy());
1118
1119 size_policy->decay_supplemental_growth(true /* full gc*/);
1120
1121 heap->resize_old_gen(
1122 size_policy->calculated_old_free_size_in_bytes());
1123
1124 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1125 size_policy->calculated_survivor_size_in_bytes());
1126 }
1127
1128 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1129 }
1130
1131 if (UsePerfData) {
1132 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1133 counters->update_counters();
1134 counters->update_old_capacity(old_gen->capacity_in_bytes());
1135 counters->update_young_capacity(young_gen->capacity_in_bytes());
1136 }
1137
1138 heap->resize_all_tlabs();
1139
1140 // Resize the metaspace capacity after a collection
1141 MetaspaceGC::compute_new_size();
1142
1143 if (log_is_enabled(Debug, gc, heap, exit)) {
1144 accumulated_time()->stop();
1145 }
1146
1147 heap->print_heap_change(pre_gc_values);
1148
1149 // Track memory usage and detect low memory
1150 MemoryService::track_memory_usage();
1151 heap->update_counters();
1152
1153 heap->post_full_gc_dump(&_gc_timer);
1154 }
1155
1156 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1157 Universe::verify("After GC");
1158 }
1159
1160 heap->print_heap_after_gc();
1161 heap->trace_heap_after_gc(&_gc_tracer);
1162
1163 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1164
1165 _gc_timer.register_gc_end();
1166
1167 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1168 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1169
1170 return true;
1171 }
1172
1173 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1174 private:
1175 uint _worker_id;
1176
1177 public:
1178 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1179 void do_thread(Thread* thread) {
1180 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1181
1182 ResourceMark rm;
1183
1184 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1185
1186 MarkingNMethodClosure mark_and_push_in_blobs(&cm->_mark_and_push_closure,
1187 !NMethodToOopClosure::FixRelocations,
1188 true /* keepalive nmethods */);
1189
1190 thread->oops_do(&cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1191
1192 // Do the real work
1193 cm->follow_marking_stacks();
1194 }
1195 };
1196
1197 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1198 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1199
1200 ParCompactionManager* cm =
1201 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1202
1203 do {
1204 oop obj = nullptr;
1205 ObjArrayTask task;
1206 if (ParCompactionManager::steal_objarray(worker_id, task)) {
1207 cm->follow_array((objArrayOop)task.obj(), task.index());
1208 } else if (ParCompactionManager::steal(worker_id, obj)) {
1209 cm->follow_contents(obj);
1210 }
1211 cm->follow_marking_stacks();
1212 } while (!terminator.offer_termination());
1213 }
1214
1215 class MarkFromRootsTask : public WorkerTask {
1216 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
1217 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1218 TaskTerminator _terminator;
1219 uint _active_workers;
1220
1221 public:
1222 MarkFromRootsTask(uint active_workers) :
1223 WorkerTask("MarkFromRootsTask"),
1224 _strong_roots_scope(active_workers),
1225 _terminator(active_workers, ParCompactionManager::oop_task_queues()),
1226 _active_workers(active_workers) {}
1227
1228 virtual void work(uint worker_id) {
1229 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1230 cm->create_marking_stats_cache();
1231 {
1232 CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1233 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1234
1235 // Do the real work
1236 cm->follow_marking_stacks();
1237 }
1238
1239 {
1240 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
1241 Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1242 }
1243
1244 // Mark from OopStorages
1245 {
1246 _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1247 // Do the real work
1248 cm->follow_marking_stacks();
1249 }
1250
1251 if (_active_workers > 1) {
1252 steal_marking_work(_terminator, worker_id);
1253 }
1254 }
1255 };
1256
1257 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1258 TaskTerminator _terminator;
1259
1260 public:
1261 ParallelCompactRefProcProxyTask(uint max_workers)
1262 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1263 _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
1264
1265 void work(uint worker_id) override {
1266 assert(worker_id < _max_workers, "sanity");
1267 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1268 BarrierEnqueueDiscoveredFieldClosure enqueue;
1269 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1270 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1271 }
1272
1273 void prepare_run_task_hook() override {
1274 _terminator.reset_for_reuse(_queue_count);
1275 }
1276 };
1277
1278 static void flush_marking_stats_cache(const uint num_workers) {
1279 for (uint i = 0; i < num_workers; ++i) {
1280 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1281 cm->flush_and_destroy_marking_stats_cache();
1282 }
1283 }
1284
1285 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1286 // Recursively traverse all live objects and mark them
1287 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1288
1289 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1290
1291 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1292 {
1293 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1294
1295 MarkFromRootsTask task(active_gc_threads);
1296 ParallelScavengeHeap::heap()->workers().run_task(&task);
1297 }
1298
1299 // Process reference objects found during marking
1300 {
1301 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1302
1303 ReferenceProcessorStats stats;
1304 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1305
1306 ref_processor()->set_active_mt_degree(active_gc_threads);
1307 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1308 stats = ref_processor()->process_discovered_references(task, pt);
1309
1310 gc_tracer->report_gc_reference_stats(stats);
1311 pt.print_all_references();
1312 }
1313
1314 {
1315 GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1316
1317 flush_marking_stats_cache(active_gc_threads);
1318 }
1319
1320 // This is the point where the entire marking should have completed.
1321 ParCompactionManager::verify_all_marking_stack_empty();
1322
1323 {
1324 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1325 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1326 is_alive_closure(),
1327 &do_nothing_cl,
1328 1);
1329 }
1330
1331 {
1332 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1333
1334 ClassUnloadingContext* ctx = ClassUnloadingContext::context();
1335
1336 bool unloading_occurred;
1337 {
1338 CodeCache::UnlinkingScope scope(is_alive_closure());
1339
1340 // Follow system dictionary roots and unload classes.
1341 unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1342
1343 // Unload nmethods.
1344 CodeCache::do_unloading(unloading_occurred);
1345 }
1346
1347 {
1348 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1349 // Release unloaded nmethod's memory.
1350 ctx->purge_nmethods();
1351 }
1352 {
1353 GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1354 ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1355 }
1356 {
1357 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1358 ctx->free_nmethods();
1359 }
1360
1361 // Prune dead klasses from subklass/sibling/implementor lists.
1362 Klass::clean_weak_klass_links(unloading_occurred);
1363
1364 // Clean JVMCI metadata handles.
1365 JVMCI_ONLY(JVMCI::do_unloading(unloading_occurred));
1366 }
1367
1368 {
1369 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
1370 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers());
1371 }
1372 #if TASKQUEUE_STATS
1373 ParCompactionManager::oop_task_queues()->print_and_reset_taskqueue_stats("Oop Queue");
1374 ParCompactionManager::_objarray_task_queues->print_and_reset_taskqueue_stats("ObjArrayOop Queue");
1375 #endif
1376 }
1377
1378 template<typename Func>
1379 void PSParallelCompact::adjust_in_space_helper(SpaceId id, volatile uint* claim_counter, Func&& on_stripe) {
1380 MutableSpace* sp = PSParallelCompact::space(id);
1381 HeapWord* const bottom = sp->bottom();
1382 HeapWord* const top = sp->top();
1383 if (bottom == top) {
1384 return;
1385 }
1386
1387 const uint num_regions_per_stripe = 2;
1388 const size_t region_size = ParallelCompactData::RegionSize;
1389 const size_t stripe_size = num_regions_per_stripe * region_size;
1390
1391 while (true) {
1392 uint counter = Atomic::fetch_then_add(claim_counter, num_regions_per_stripe);
1393 HeapWord* cur_stripe = bottom + counter * region_size;
1394 if (cur_stripe >= top) {
1395 break;
1396 }
1397 HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1398 on_stripe(cur_stripe, stripe_end);
1399 }
1400 }
1401
1402 void PSParallelCompact::adjust_in_old_space(volatile uint* claim_counter) {
1403 // Regions in old-space shouldn't be split.
1404 assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1405
1406 auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1407 assert(mark_bitmap()->is_marked(obj_start), "inv");
1408 oop obj = cast_to_oop(obj_start);
1409 return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1410 };
1411
1412 adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1413 assert(_summary_data.is_region_aligned(stripe_start), "inv");
1414 RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1415 HeapWord* obj_start;
1416 if (cur_region->partial_obj_size() != 0) {
1417 obj_start = cur_region->partial_obj_addr();
1418 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1419 } else {
1420 obj_start = stripe_start;
1421 }
1422
1423 while (obj_start < stripe_end) {
1424 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1425 if (obj_start >= stripe_end) {
1426 break;
1427 }
1428 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1429 }
1430 });
1431 }
1432
1433 void PSParallelCompact::adjust_in_young_space(SpaceId id, volatile uint* claim_counter) {
1434 adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1435 HeapWord* obj_start = stripe_start;
1436 while (obj_start < stripe_end) {
1437 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1438 if (obj_start >= stripe_end) {
1439 break;
1440 }
1441 oop obj = cast_to_oop(obj_start);
1442 obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1443 }
1444 });
1445 }
1446
1447 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, volatile uint* claim_counters) {
1448 auto start_time = Ticks::now();
1449 adjust_in_old_space(&claim_counters[0]);
1450 for (uint id = eden_space_id; id < last_space_id; ++id) {
1451 adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1452 }
1453 log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1454 }
1455
1456 class PSAdjustTask final : public WorkerTask {
1457 SubTasksDone _sub_tasks;
1458 WeakProcessor::Task _weak_proc_task;
1459 OopStorageSetStrongParState<false, false> _oop_storage_iter;
1460 uint _nworkers;
1461 volatile uint _claim_counters[PSParallelCompact::last_space_id] = {};
1462
1463 enum PSAdjustSubTask {
1464 PSAdjustSubTask_code_cache,
1465
1466 PSAdjustSubTask_num_elements
1467 };
1468
1469 public:
1470 PSAdjustTask(uint nworkers) :
1471 WorkerTask("PSAdjust task"),
1472 _sub_tasks(PSAdjustSubTask_num_elements),
1473 _weak_proc_task(nworkers),
1474 _nworkers(nworkers) {
1475
1476 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1477 if (nworkers > 1) {
1478 Threads::change_thread_claim_token();
1479 }
1480 }
1481
1482 ~PSAdjustTask() {
1483 Threads::assert_all_threads_claimed();
1484 }
1485
1486 void work(uint worker_id) {
1487 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1488 cm->preserved_marks()->adjust_during_full_gc();
1489 {
1490 // adjust pointers in all spaces
1491 PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1492 }
1493 {
1494 ResourceMark rm;
1495 Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1496 }
1497 _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1498 {
1499 CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1500 ClassLoaderDataGraph::cld_do(&cld_closure);
1501 }
1502 {
1503 AlwaysTrueClosure always_alive;
1504 _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1505 }
1506 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
1507 NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1508 CodeCache::nmethods_do(&adjust_code);
1509 }
1510 _sub_tasks.all_tasks_claimed();
1511 }
1512 };
1513
1514 void PSParallelCompact::adjust_pointers() {
1515 // Adjust the pointers to reflect the new locations
1516 GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1517 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1518 PSAdjustTask task(nworkers);
1519 ParallelScavengeHeap::heap()->workers().run_task(&task);
1520 }
1521
1522 // Split [start, end) evenly for a number of workers and return the
1523 // range for worker_id.
1524 static void split_regions_for_worker(size_t start, size_t end,
1525 uint worker_id, uint num_workers,
1526 size_t* worker_start, size_t* worker_end) {
1527 assert(start < end, "precondition");
1528 assert(num_workers > 0, "precondition");
1529 assert(worker_id < num_workers, "precondition");
1530
1531 size_t num_regions = end - start;
1532 size_t num_regions_per_worker = num_regions / num_workers;
1533 size_t remainder = num_regions % num_workers;
1534 // The first few workers will get one extra.
1535 *worker_start = start + worker_id * num_regions_per_worker
1536 + MIN2(checked_cast<size_t>(worker_id), remainder);
1537 *worker_end = *worker_start + num_regions_per_worker
1538 + (worker_id < remainder ? 1 : 0);
1539 }
1540
1541 void PSParallelCompact::forward_to_new_addr() {
1542 GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1543 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1544
1545 struct ForwardTask final : public WorkerTask {
1546 uint _num_workers;
1547
1548 explicit ForwardTask(uint num_workers) :
1549 WorkerTask("PSForward task"),
1550 _num_workers(num_workers) {}
1551
1552 void work(uint worker_id) override {
1553 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1554 for (uint id = old_space_id; id < last_space_id; ++id) {
1555 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1556 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1557 HeapWord* top = sp->top();
1558
1559 if (dense_prefix_addr == top) {
1560 continue;
1561 }
1562
1563 size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1564 size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1565 size_t start_region;
1566 size_t end_region;
1567 split_regions_for_worker(dense_prefix_region, top_region,
1568 worker_id, _num_workers,
1569 &start_region, &end_region);
1570 for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1571 RegionData* region_ptr = _summary_data.region(cur_region);
1572 size_t live_words = region_ptr->partial_obj_size();
1573
1574 if (live_words == ParallelCompactData::RegionSize) {
1575 // No obj-start
1576 continue;
1577 }
1578
1579 HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1580 HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1581
1582 HeapWord* cur_addr = region_start + live_words;
1583
1584 HeapWord* destination = region_ptr->destination();
1585 while (cur_addr < region_end) {
1586 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, region_end);
1587 if (cur_addr >= region_end) {
1588 break;
1589 }
1590 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1591 HeapWord* new_addr = destination + live_words;
1592 oop obj = cast_to_oop(cur_addr);
1593 if (new_addr != cur_addr) {
1594 cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1595 obj->forward_to(cast_to_oop(new_addr));
1596 }
1597 size_t obj_size = obj->size();
1598 live_words += obj_size;
1599 cur_addr += obj_size;
1600 }
1601 }
1602 }
1603 }
1604 } task(nworkers);
1605
1606 ParallelScavengeHeap::heap()->workers().run_task(&task);
1607 debug_only(verify_forward();)
1608 }
1609
1610 #ifdef ASSERT
1611 void PSParallelCompact::verify_forward() {
1612 HeapWord* old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1613 RegionData* old_region = _summary_data.region(_summary_data.addr_to_region_idx(old_dense_prefix_addr));
1614 HeapWord* bump_ptr = old_region->partial_obj_size() != 0
1615 ? old_dense_prefix_addr + old_region->partial_obj_size()
1616 : old_dense_prefix_addr;
1617 SpaceId bump_ptr_space = old_space_id;
1618
1619 for (uint id = old_space_id; id < last_space_id; ++id) {
1620 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1621 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1622 HeapWord* top = sp->top();
1623 HeapWord* cur_addr = dense_prefix_addr;
1624
1625 while (cur_addr < top) {
1626 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1627 if (cur_addr >= top) {
1628 break;
1629 }
1630 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1631 // Move to the space containing cur_addr
1632 if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1633 bump_ptr = space(space_id(cur_addr))->bottom();
1634 bump_ptr_space = space_id(bump_ptr);
1635 }
1636 oop obj = cast_to_oop(cur_addr);
1637 if (cur_addr != bump_ptr) {
1638 assert(obj->forwardee() == cast_to_oop(bump_ptr), "inv");
1639 }
1640 bump_ptr += obj->size();
1641 cur_addr += obj->size();
1642 }
1643 }
1644 }
1645 #endif
1646
1647 // Helper class to print 8 region numbers per line and then print the total at the end.
1648 class FillableRegionLogger : public StackObj {
1649 private:
1650 Log(gc, compaction) log;
1651 static const int LineLength = 8;
1652 size_t _regions[LineLength];
1653 int _next_index;
1654 bool _enabled;
1655 size_t _total_regions;
1656 public:
1657 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1658 ~FillableRegionLogger() {
1659 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
1660 }
1661
1662 void print_line() {
1663 if (!_enabled || _next_index == 0) {
1664 return;
1665 }
1666 FormatBuffer<> line("Fillable: ");
1667 for (int i = 0; i < _next_index; i++) {
1668 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
1669 }
1670 log.trace("%s", line.buffer());
1671 _next_index = 0;
1672 }
1673
1674 void handle(size_t region) {
1675 if (!_enabled) {
1676 return;
1677 }
1678 _regions[_next_index++] = region;
1679 if (_next_index == LineLength) {
1680 print_line();
1681 }
1682 _total_regions++;
1683 }
1684 };
1685
1686 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1687 {
1688 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1689
1690 // Find the threads that are active
1691 uint worker_id = 0;
1692
1693 // Find all regions that are available (can be filled immediately) and
1694 // distribute them to the thread stacks. The iteration is done in reverse
1695 // order (high to low) so the regions will be removed in ascending order.
1696
1697 const ParallelCompactData& sd = PSParallelCompact::summary_data();
1698
1699 // id + 1 is used to test termination so unsigned can
1700 // be used with an old_space_id == 0.
1701 FillableRegionLogger region_logger;
1702 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
1703 SpaceInfo* const space_info = _space_info + id;
1704 HeapWord* const new_top = space_info->new_top();
1705
1706 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1707 const size_t end_region =
1708 sd.addr_to_region_idx(sd.region_align_up(new_top));
1709
1710 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1711 if (sd.region(cur)->claim_unsafe()) {
1712 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1713 bool result = sd.region(cur)->mark_normal();
1714 assert(result, "Must succeed at this point.");
1715 cm->region_stack()->push(cur);
1716 region_logger.handle(cur);
1717 // Assign regions to tasks in round-robin fashion.
1718 if (++worker_id == parallel_gc_threads) {
1719 worker_id = 0;
1720 }
1721 }
1722 }
1723 region_logger.print_line();
1724 }
1725 }
1726
1727 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1728 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1729
1730 ParCompactionManager* cm =
1731 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1732
1733 // Drain the stacks that have been preloaded with regions
1734 // that are ready to fill.
1735
1736 cm->drain_region_stacks();
1737
1738 guarantee(cm->region_stack()->is_empty(), "Not empty");
1739
1740 size_t region_index = 0;
1741
1742 while (true) {
1743 if (ParCompactionManager::steal(worker_id, region_index)) {
1744 PSParallelCompact::fill_and_update_region(cm, region_index);
1745 cm->drain_region_stacks();
1746 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1747 // Fill and update an unavailable region with the help of a shadow region
1748 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1749 cm->drain_region_stacks();
1750 } else {
1751 if (terminator->offer_termination()) {
1752 break;
1753 }
1754 // Go around again.
1755 }
1756 }
1757 }
1758
1759 class FillDensePrefixAndCompactionTask: public WorkerTask {
1760 uint _num_workers;
1761 TaskTerminator _terminator;
1762
1763 public:
1764 FillDensePrefixAndCompactionTask(uint active_workers) :
1765 WorkerTask("FillDensePrefixAndCompactionTask"),
1766 _num_workers(active_workers),
1767 _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1768 }
1769
1770 virtual void work(uint worker_id) {
1771 {
1772 auto start = Ticks::now();
1773 PSParallelCompact::fill_dead_objs_in_dense_prefix(worker_id, _num_workers);
1774 log_trace(gc, phases)("Fill dense prefix by worker %u: %.3f ms", worker_id, (Ticks::now() - start).seconds() * 1000);
1775 }
1776 compaction_with_stealing_work(&_terminator, worker_id);
1777 }
1778 };
1779
1780 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1781 #ifdef ASSERT
1782 {
1783 assert(start < end, "precondition");
1784 assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1785 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1786 if (start != bottom) {
1787 HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1788 HeapWord* after_obj = obj_start + cast_to_oop(obj_start)->size();
1789 assert(after_obj == start, "precondition");
1790 }
1791 }
1792 #endif
1793
1794 CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1795 HeapWord* addr = start;
1796 do {
1797 size_t size = cast_to_oop(addr)->size();
1798 start_array(old_space_id)->update_for_block(addr, addr + size);
1799 addr += size;
1800 } while (addr < end);
1801 }
1802
1803 void PSParallelCompact::fill_dead_objs_in_dense_prefix(uint worker_id, uint num_workers) {
1804 ParMarkBitMap* bitmap = mark_bitmap();
1805
1806 HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1807 HeapWord* const prefix_end = dense_prefix(old_space_id);
1808
1809 if (bottom == prefix_end) {
1810 return;
1811 }
1812
1813 size_t bottom_region = _summary_data.addr_to_region_idx(bottom);
1814 size_t prefix_end_region = _summary_data.addr_to_region_idx(prefix_end);
1815
1816 size_t start_region;
1817 size_t end_region;
1818 split_regions_for_worker(bottom_region, prefix_end_region,
1819 worker_id, num_workers,
1820 &start_region, &end_region);
1821
1822 if (start_region == end_region) {
1823 return;
1824 }
1825
1826 HeapWord* const start_addr = _summary_data.region_to_addr(start_region);
1827 HeapWord* const end_addr = _summary_data.region_to_addr(end_region);
1828
1829 // Skip live partial obj (if any) from previous region.
1830 HeapWord* cur_addr;
1831 RegionData* start_region_ptr = _summary_data.region(start_region);
1832 if (start_region_ptr->partial_obj_size() != 0) {
1833 HeapWord* partial_obj_start = start_region_ptr->partial_obj_addr();
1834 assert(bitmap->is_marked(partial_obj_start), "inv");
1835 cur_addr = partial_obj_start + cast_to_oop(partial_obj_start)->size();
1836 } else {
1837 cur_addr = start_addr;
1838 }
1839
1840 // end_addr is inclusive to handle regions starting with dead space.
1841 while (cur_addr <= end_addr) {
1842 // Use prefix_end to handle trailing obj in each worker region-chunk.
1843 HeapWord* live_start = bitmap->find_obj_beg(cur_addr, prefix_end);
1844 if (cur_addr != live_start) {
1845 // Only worker 0 handles proceeding dead space.
1846 if (cur_addr != start_addr || worker_id == 0) {
1847 fill_range_in_dense_prefix(cur_addr, live_start);
1848 }
1849 }
1850 if (live_start >= end_addr) {
1851 break;
1852 }
1853 assert(bitmap->is_marked(live_start), "inv");
1854 cur_addr = live_start + cast_to_oop(live_start)->size();
1855 }
1856 }
1857
1858 void PSParallelCompact::compact() {
1859 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1860
1861 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1862
1863 initialize_shadow_regions(active_gc_threads);
1864 prepare_region_draining_tasks(active_gc_threads);
1865
1866 {
1867 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1868
1869 FillDensePrefixAndCompactionTask task(active_gc_threads);
1870 ParallelScavengeHeap::heap()->workers().run_task(&task);
1871
1872 #ifdef ASSERT
1873 verify_filler_in_dense_prefix();
1874
1875 // Verify that all regions have been processed.
1876 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1877 verify_complete(SpaceId(id));
1878 }
1879 #endif
1880 }
1881 }
1882
1883 #ifdef ASSERT
1884 void PSParallelCompact::verify_filler_in_dense_prefix() {
1885 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1886 HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1887 HeapWord* cur_addr = bottom;
1888 while (cur_addr < dense_prefix_end) {
1889 oop obj = cast_to_oop(cur_addr);
1890 oopDesc::verify(obj);
1891 if (!mark_bitmap()->is_marked(cur_addr)) {
1892 Klass* k = cast_to_oop(cur_addr)->klass_without_asserts();
1893 assert(k == Universe::fillerArrayKlass() || k == vmClasses::FillerObject_klass(), "inv");
1894 }
1895 cur_addr += obj->size();
1896 }
1897 }
1898
1899 void PSParallelCompact::verify_complete(SpaceId space_id) {
1900 // All Regions served as compaction targets, from dense_prefix() to
1901 // new_top(), should be marked as filled and all Regions between new_top()
1902 // and top() should be available (i.e., should have been emptied).
1903 ParallelCompactData& sd = summary_data();
1904 SpaceInfo si = _space_info[space_id];
1905 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1906 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1907 const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1908 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1909 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1910
1911 size_t cur_region;
1912 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1913 const RegionData* const c = sd.region(cur_region);
1914 if (!c->completed()) {
1915 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
1916 cur_region, c->destination_count());
1917 }
1918 }
1919
1920 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1921 const RegionData* const c = sd.region(cur_region);
1922 if (!c->available()) {
1923 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
1924 cur_region, c->destination_count());
1925 }
1926 }
1927 }
1928 #endif // #ifdef ASSERT
1929
1930 // Return the SpaceId for the space containing addr. If addr is not in the
1931 // heap, last_space_id is returned. In debug mode it expects the address to be
1932 // in the heap and asserts such.
1933 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1934 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1935
1936 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1937 if (_space_info[id].space()->contains(addr)) {
1938 return SpaceId(id);
1939 }
1940 }
1941
1942 assert(false, "no space contains the addr");
1943 return last_space_id;
1944 }
1945
1946 // Skip over count live words starting from beg, and return the address of the
1947 // next live word. Unless marked, the word corresponding to beg is assumed to
1948 // be dead. Callers must either ensure beg does not correspond to the middle of
1949 // an object, or account for those live words in some other way. Callers must
1950 // also ensure that there are enough live words in the range [beg, end) to skip.
1951 HeapWord*
1952 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1953 {
1954 assert(count > 0, "sanity");
1955
1956 ParMarkBitMap* m = mark_bitmap();
1957 HeapWord* cur_addr = beg;
1958 while (true) {
1959 cur_addr = m->find_obj_beg(cur_addr, end);
1960 assert(cur_addr < end, "inv");
1961 size_t obj_size = cast_to_oop(cur_addr)->size();
1962 // Strictly greater-than
1963 if (obj_size > count) {
1964 return cur_addr + count;
1965 }
1966 count -= obj_size;
1967 cur_addr += obj_size;
1968 }
1969 }
1970
1971 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1972 SpaceId src_space_id,
1973 size_t src_region_idx)
1974 {
1975 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
1976
1977 const SplitInfo& split_info = _space_info[src_space_id].split_info();
1978 if (split_info.dest_region_addr() == dest_addr) {
1979 // The partial object ending at the split point contains the first word to
1980 // be copied to dest_addr.
1981 return split_info.first_src_addr();
1982 }
1983
1984 const ParallelCompactData& sd = summary_data();
1985 ParMarkBitMap* const bitmap = mark_bitmap();
1986 const size_t RegionSize = ParallelCompactData::RegionSize;
1987
1988 assert(sd.is_region_aligned(dest_addr), "not aligned");
1989 const RegionData* const src_region_ptr = sd.region(src_region_idx);
1990 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1991 HeapWord* const src_region_destination = src_region_ptr->destination();
1992
1993 assert(dest_addr >= src_region_destination, "wrong src region");
1994 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1995
1996 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
1997 HeapWord* const src_region_end = src_region_beg + RegionSize;
1998
1999 HeapWord* addr = src_region_beg;
2000 if (dest_addr == src_region_destination) {
2001 // Return the first live word in the source region.
2002 if (partial_obj_size == 0) {
2003 addr = bitmap->find_obj_beg(addr, src_region_end);
2004 assert(addr < src_region_end, "no objects start in src region");
2005 }
2006 return addr;
2007 }
2008
2009 // Must skip some live data.
2010 size_t words_to_skip = dest_addr - src_region_destination;
2011 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2012
2013 if (partial_obj_size >= words_to_skip) {
2014 // All the live words to skip are part of the partial object.
2015 addr += words_to_skip;
2016 if (partial_obj_size == words_to_skip) {
2017 // Find the first live word past the partial object.
2018 addr = bitmap->find_obj_beg(addr, src_region_end);
2019 assert(addr < src_region_end, "wrong src region");
2020 }
2021 return addr;
2022 }
2023
2024 // Skip over the partial object (if any).
2025 if (partial_obj_size != 0) {
2026 words_to_skip -= partial_obj_size;
2027 addr += partial_obj_size;
2028 }
2029
2030 // Skip over live words due to objects that start in the region.
2031 addr = skip_live_words(addr, src_region_end, words_to_skip);
2032 assert(addr < src_region_end, "wrong src region");
2033 return addr;
2034 }
2035
2036 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2037 SpaceId src_space_id,
2038 size_t beg_region,
2039 HeapWord* end_addr)
2040 {
2041 ParallelCompactData& sd = summary_data();
2042
2043 #ifdef ASSERT
2044 MutableSpace* const src_space = _space_info[src_space_id].space();
2045 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2046 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2047 "src_space_id does not match beg_addr");
2048 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2049 "src_space_id does not match end_addr");
2050 #endif // #ifdef ASSERT
2051
2052 RegionData* const beg = sd.region(beg_region);
2053 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2054
2055 // Regions up to new_top() are enqueued if they become available.
2056 HeapWord* const new_top = _space_info[src_space_id].new_top();
2057 RegionData* const enqueue_end =
2058 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2059
2060 for (RegionData* cur = beg; cur < end; ++cur) {
2061 assert(cur->data_size() > 0, "region must have live data");
2062 cur->decrement_destination_count();
2063 if (cur < enqueue_end && cur->available() && cur->claim()) {
2064 if (cur->mark_normal()) {
2065 cm->push_region(sd.region(cur));
2066 } else if (cur->mark_copied()) {
2067 // Try to copy the content of the shadow region back to its corresponding
2068 // heap region if the shadow region is filled. Otherwise, the GC thread
2069 // fills the shadow region will copy the data back (see
2070 // MoveAndUpdateShadowClosure::complete_region).
2071 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2072 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2073 cur->set_completed();
2074 }
2075 }
2076 }
2077 }
2078
2079 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2080 SpaceId& src_space_id,
2081 HeapWord*& src_space_top,
2082 HeapWord* end_addr)
2083 {
2084 typedef ParallelCompactData::RegionData RegionData;
2085
2086 ParallelCompactData& sd = PSParallelCompact::summary_data();
2087 const size_t region_size = ParallelCompactData::RegionSize;
2088
2089 size_t src_region_idx = 0;
2090
2091 // Skip empty regions (if any) up to the top of the space.
2092 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2093 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2094 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2095 const RegionData* const top_region_ptr =
2096 sd.addr_to_region_ptr(top_aligned_up);
2097 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2098 ++src_region_ptr;
2099 }
2100
2101 if (src_region_ptr < top_region_ptr) {
2102 // The next source region is in the current space. Update src_region_idx
2103 // and the source address to match src_region_ptr.
2104 src_region_idx = sd.region(src_region_ptr);
2105 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2106 if (src_region_addr > closure.source()) {
2107 closure.set_source(src_region_addr);
2108 }
2109 return src_region_idx;
2110 }
2111
2112 // Switch to a new source space and find the first non-empty region.
2113 unsigned int space_id = src_space_id + 1;
2114 assert(space_id < last_space_id, "not enough spaces");
2115
2116 HeapWord* const destination = closure.destination();
2117
2118 do {
2119 MutableSpace* space = _space_info[space_id].space();
2120 HeapWord* const bottom = space->bottom();
2121 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2122
2123 // Iterate over the spaces that do not compact into themselves.
2124 if (bottom_cp->destination() != bottom) {
2125 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2126 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2127
2128 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2129 if (src_cp->live_obj_size() > 0) {
2130 // Found it.
2131 assert(src_cp->destination() == destination,
2132 "first live obj in the space must match the destination");
2133 assert(src_cp->partial_obj_size() == 0,
2134 "a space cannot begin with a partial obj");
2135
2136 src_space_id = SpaceId(space_id);
2137 src_space_top = space->top();
2138 const size_t src_region_idx = sd.region(src_cp);
2139 closure.set_source(sd.region_to_addr(src_region_idx));
2140 return src_region_idx;
2141 } else {
2142 assert(src_cp->data_size() == 0, "sanity");
2143 }
2144 }
2145 }
2146 } while (++space_id < last_space_id);
2147
2148 assert(false, "no source region was found");
2149 return 0;
2150 }
2151
2152 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2153 ParallelCompactData& sd = summary_data();
2154 assert(sd.is_region_aligned(region_start_addr), "precondition");
2155
2156 // Use per-region partial_obj_size to locate the end of the obj, that extends to region_start_addr.
2157 SplitInfo& split_info = _space_info[space_id(region_start_addr)].split_info();
2158 size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2159 size_t end_region_idx = sd.region_count();
2160 size_t accumulated_size = 0;
2161 for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2162 if (split_info.is_split(region_idx)) {
2163 accumulated_size += split_info.partial_obj_size();
2164 break;
2165 }
2166 size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2167 accumulated_size += cur_partial_obj_size;
2168 if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2169 break;
2170 }
2171 }
2172 return region_start_addr + accumulated_size;
2173 }
2174
2175 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2176 {
2177 ParMarkBitMap* const bitmap = mark_bitmap();
2178 ParallelCompactData& sd = summary_data();
2179 RegionData* const region_ptr = sd.region(region_idx);
2180
2181 // Get the source region and related info.
2182 size_t src_region_idx = region_ptr->source_region();
2183 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2184 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2185 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2186
2187 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2188
2189 // Adjust src_region_idx to prepare for decrementing destination counts (the
2190 // destination count is not decremented when a region is copied to itself).
2191 if (src_region_idx == region_idx) {
2192 src_region_idx += 1;
2193 }
2194
2195 if (bitmap->is_unmarked(closure.source())) {
2196 // The first source word is in the middle of an object; copy the remainder
2197 // of the object or as much as will fit. The fact that pointer updates were
2198 // deferred will be noted when the object header is processed.
2199 HeapWord* const old_src_addr = closure.source();
2200 {
2201 HeapWord* region_start = sd.region_align_down(closure.source());
2202 HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2203 HeapWord* obj_end;
2204 if (bitmap->is_marked(obj_start)) {
2205 HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2206 HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2207 ? nullptr
2208 : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2209 if (partial_obj_start == obj_start) {
2210 // This obj extends to next region.
2211 obj_end = partial_obj_end(next_region_start);
2212 } else {
2213 // Completely contained in this region; safe to use size().
2214 obj_end = obj_start + cast_to_oop(obj_start)->size();
2215 }
2216 } else {
2217 // This obj extends to current region.
2218 obj_end = partial_obj_end(region_start);
2219 }
2220 size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2221 closure.copy_partial_obj(partial_obj_size);
2222 }
2223
2224 if (closure.is_full()) {
2225 decrement_destination_counts(cm, src_space_id, src_region_idx,
2226 closure.source());
2227 closure.complete_region(dest_addr, region_ptr);
2228 return;
2229 }
2230
2231 HeapWord* const end_addr = sd.region_align_down(closure.source());
2232 if (sd.region_align_down(old_src_addr) != end_addr) {
2233 // The partial object was copied from more than one source region.
2234 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2235
2236 // Move to the next source region, possibly switching spaces as well. All
2237 // args except end_addr may be modified.
2238 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2239 end_addr);
2240 }
2241 }
2242
2243 do {
2244 HeapWord* cur_addr = closure.source();
2245 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2246 src_space_top);
2247 HeapWord* partial_obj_start = (end_addr == src_space_top)
2248 ? nullptr
2249 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2250 // apply closure on objs inside [cur_addr, end_addr)
2251 do {
2252 cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2253 if (cur_addr == end_addr) {
2254 break;
2255 }
2256 size_t obj_size;
2257 if (partial_obj_start == cur_addr) {
2258 obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2259 } else {
2260 // This obj doesn't extend into next region; size() is safe to use.
2261 obj_size = cast_to_oop(cur_addr)->size();
2262 }
2263 closure.do_addr(cur_addr, obj_size);
2264 cur_addr += obj_size;
2265 } while (cur_addr < end_addr && !closure.is_full());
2266
2267 if (closure.is_full()) {
2268 decrement_destination_counts(cm, src_space_id, src_region_idx,
2269 closure.source());
2270 closure.complete_region(dest_addr, region_ptr);
2271 return;
2272 }
2273
2274 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2275
2276 // Move to the next source region, possibly switching spaces as well. All
2277 // args except end_addr may be modified.
2278 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2279 end_addr);
2280 } while (true);
2281 }
2282
2283 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2284 {
2285 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2286 fill_region(cm, cl, region_idx);
2287 }
2288
2289 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2290 {
2291 // Get a shadow region first
2292 ParallelCompactData& sd = summary_data();
2293 RegionData* const region_ptr = sd.region(region_idx);
2294 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2295 // The InvalidShadow return value indicates the corresponding heap region is available,
2296 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2297 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2298 if (shadow_region == ParCompactionManager::InvalidShadow) {
2299 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2300 region_ptr->shadow_to_normal();
2301 return fill_region(cm, cl, region_idx);
2302 } else {
2303 MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2304 return fill_region(cm, cl, region_idx);
2305 }
2306 }
2307
2308 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2309 {
2310 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2311 }
2312
2313 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
2314 {
2315 size_t next = cm->next_shadow_region();
2316 ParallelCompactData& sd = summary_data();
2317 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2318 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2319
2320 while (next < old_new_top) {
2321 if (sd.region(next)->mark_shadow()) {
2322 region_idx = next;
2323 return true;
2324 }
2325 next = cm->move_next_shadow_region_by(active_gc_threads);
2326 }
2327
2328 return false;
2329 }
2330
2331 // The shadow region is an optimization to address region dependencies in full GC. The basic
2332 // idea is making more regions available by temporally storing their live objects in empty
2333 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2334 // GC threads need not wait destination regions to be available before processing sources.
2335 //
2336 // A typical workflow would be:
2337 // After draining its own stack and failing to steal from others, a GC worker would pick an
2338 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2339 // the shadow region by copying live objects from source regions of the unavailable one. Once
2340 // the unavailable region becomes available, the data in the shadow region will be copied back.
2341 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2342 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2343 {
2344 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2345
2346 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2347 SpaceInfo* const space_info = _space_info + id;
2348 MutableSpace* const space = space_info->space();
2349
2350 const size_t beg_region =
2351 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2352 const size_t end_region =
2353 sd.addr_to_region_idx(sd.region_align_down(space->end()));
2354
2355 for (size_t cur = beg_region; cur < end_region; ++cur) {
2356 ParCompactionManager::push_shadow_region(cur);
2357 }
2358 }
2359
2360 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2361 for (uint i = 0; i < parallel_gc_threads; i++) {
2362 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2363 cm->set_next_shadow_region(beg_region + i);
2364 }
2365 }
2366
2367 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2368 {
2369 size_t words = MIN2(partial_obj_size, words_remaining());
2370
2371 // This test is necessary; if omitted, the pointer updates to a partial object
2372 // that crosses the dense prefix boundary could be overwritten.
2373 if (source() != copy_destination()) {
2374 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2375 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2376 }
2377 update_state(words);
2378 }
2379
2380 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2381 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2382 region_ptr->set_completed();
2383 }
2384
2385 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2386 assert(destination() != nullptr, "sanity");
2387 _source = addr;
2388
2389 // The start_array must be updated even if the object is not moving.
2390 if (_start_array != nullptr) {
2391 _start_array->update_for_block(destination(), destination() + words);
2392 }
2393
2394 // Avoid overflow
2395 words = MIN2(words, words_remaining());
2396 assert(words > 0, "inv");
2397
2398 if (copy_destination() != source()) {
2399 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2400 assert(source() != destination(), "inv");
2401 assert(cast_to_oop(source())->is_forwarded(), "inv");
2402 assert(cast_to_oop(source())->forwardee() == cast_to_oop(destination()), "inv");
2403 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2404 cast_to_oop(copy_destination())->init_mark();
2405 }
2406
2407 update_state(words);
2408 }
2409
2410 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2411 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2412 // Record the shadow region index
2413 region_ptr->set_shadow_region(_shadow);
2414 // Mark the shadow region as filled to indicate the data is ready to be
2415 // copied back
2416 region_ptr->mark_filled();
2417 // Try to copy the content of the shadow region back to its corresponding
2418 // heap region if available; the GC thread that decreases the destination
2419 // count to zero will do the copying otherwise (see
2420 // PSParallelCompact::decrement_destination_counts).
2421 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2422 region_ptr->set_completed();
2423 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2424 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2425 }
2426 }
2427