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/spaceDecorator.hpp"
64 #include "gc/shared/strongRootsScope.hpp"
65 #include "gc/shared/taskTerminator.hpp"
66 #include "gc/shared/weakProcessor.inline.hpp"
67 #include "gc/shared/workerPolicy.hpp"
68 #include "gc/shared/workerThread.hpp"
69 #include "gc/shared/workerUtils.hpp"
70 #include "logging/log.hpp"
71 #include "memory/iterator.inline.hpp"
72 #include "memory/metaspaceUtils.hpp"
73 #include "memory/resourceArea.hpp"
74 #include "memory/universe.hpp"
75 #include "nmt/memTracker.hpp"
76 #include "oops/access.inline.hpp"
77 #include "oops/flatArrayKlass.inline.hpp"
78 #include "oops/instanceClassLoaderKlass.inline.hpp"
79 #include "oops/instanceKlass.inline.hpp"
80 #include "oops/instanceMirrorKlass.inline.hpp"
81 #include "oops/methodData.hpp"
82 #include "oops/objArrayKlass.inline.hpp"
83 #include "oops/oop.inline.hpp"
84 #include "runtime/atomic.hpp"
85 #include "runtime/handles.inline.hpp"
86 #include "runtime/java.hpp"
87 #include "runtime/safepoint.hpp"
88 #include "runtime/threads.hpp"
89 #include "runtime/vmThread.hpp"
90 #include "services/memoryService.hpp"
91 #include "utilities/align.hpp"
92 #include "utilities/debug.hpp"
93 #include "utilities/events.hpp"
94 #include "utilities/formatBuffer.hpp"
95 #include "utilities/macros.hpp"
96 #include "utilities/stack.inline.hpp"
97 #if INCLUDE_JVMCI
98 #include "jvmci/jvmci.hpp"
99 #endif
100
101 #include <math.h>
102
103 // All sizes are in HeapWords.
104 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
105 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
106 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
107 const size_t ParallelCompactData::RegionSizeBytes =
108 RegionSize << LogHeapWordSize;
109 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
110 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
111 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
112
113 const ParallelCompactData::RegionData::region_sz_t
114 ParallelCompactData::RegionData::dc_shift = 27;
115
116 const ParallelCompactData::RegionData::region_sz_t
117 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
118
119 const ParallelCompactData::RegionData::region_sz_t
120 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
121
122 const ParallelCompactData::RegionData::region_sz_t
123 ParallelCompactData::RegionData::los_mask = ~dc_mask;
124
125 const ParallelCompactData::RegionData::region_sz_t
126 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
127
128 const ParallelCompactData::RegionData::region_sz_t
129 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
130
131 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
132
133 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
134 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
135
136 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
137 HeapWord* destination)
138 {
139 assert(src_region_idx != 0, "invalid src_region_idx");
140 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
141 assert(destination != nullptr, "invalid destination argument");
142
143 _src_region_idx = src_region_idx;
144 _partial_obj_size = partial_obj_size;
145 _destination = destination;
146
147 // These fields may not be updated below, so make sure they're clear.
148 assert(_dest_region_addr == nullptr, "should have been cleared");
149 assert(_first_src_addr == nullptr, "should have been cleared");
150
151 // Determine the number of destination regions for the partial object.
152 HeapWord* const last_word = destination + partial_obj_size - 1;
153 const ParallelCompactData& sd = PSParallelCompact::summary_data();
154 HeapWord* const beg_region_addr = sd.region_align_down(destination);
155 HeapWord* const end_region_addr = sd.region_align_down(last_word);
156
157 if (beg_region_addr == end_region_addr) {
158 // One destination region.
159 _destination_count = 1;
160 if (end_region_addr == destination) {
161 // The destination falls on a region boundary, thus the first word of the
162 // partial object will be the first word copied to the destination region.
163 _dest_region_addr = end_region_addr;
164 _first_src_addr = sd.region_to_addr(src_region_idx);
165 }
166 } else {
167 // Two destination regions. When copied, the partial object will cross a
168 // destination region boundary, so a word somewhere within the partial
169 // object will be the first word copied to the second destination region.
170 _destination_count = 2;
171 _dest_region_addr = end_region_addr;
172 const size_t ofs = pointer_delta(end_region_addr, destination);
173 assert(ofs < _partial_obj_size, "sanity");
174 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
175 }
176 }
177
178 void SplitInfo::clear()
179 {
180 _src_region_idx = 0;
181 _partial_obj_size = 0;
182 _destination = nullptr;
183 _destination_count = 0;
184 _dest_region_addr = nullptr;
185 _first_src_addr = nullptr;
186 assert(!is_valid(), "sanity");
187 }
188
189 #ifdef ASSERT
190 void SplitInfo::verify_clear()
191 {
192 assert(_src_region_idx == 0, "not clear");
193 assert(_partial_obj_size == 0, "not clear");
194 assert(_destination == nullptr, "not clear");
195 assert(_destination_count == 0, "not clear");
196 assert(_dest_region_addr == nullptr, "not clear");
197 assert(_first_src_addr == nullptr, "not clear");
198 }
199 #endif // #ifdef ASSERT
200
201
202 void PSParallelCompact::print_on_error(outputStream* st) {
203 _mark_bitmap.print_on_error(st);
204 }
205
206 ParallelCompactData::ParallelCompactData() :
207 _heap_start(nullptr),
208 DEBUG_ONLY(_heap_end(nullptr) COMMA)
209 _region_vspace(nullptr),
210 _reserved_byte_size(0),
211 _region_data(nullptr),
212 _region_count(0) {}
213
214 bool ParallelCompactData::initialize(MemRegion reserved_heap)
215 {
216 _heap_start = reserved_heap.start();
217 const size_t heap_size = reserved_heap.word_size();
218 DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
219
220 assert(region_align_down(_heap_start) == _heap_start,
221 "region start not aligned");
222
223 return initialize_region_data(heap_size);
224 }
225
226 PSVirtualSpace*
227 ParallelCompactData::create_vspace(size_t count, size_t element_size)
228 {
229 const size_t raw_bytes = count * element_size;
230 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
231 const size_t granularity = os::vm_allocation_granularity();
232 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
233
234 const size_t rs_align = page_sz == os::vm_page_size() ? 0 :
235 MAX2(page_sz, granularity);
236 ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
237 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
238 rs.size(), page_sz);
239
240 MemTracker::record_virtual_memory_tag((address)rs.base(), mtGC);
241
242 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
243 if (vspace != nullptr) {
244 if (vspace->expand_by(_reserved_byte_size)) {
245 return vspace;
246 }
247 delete vspace;
248 // Release memory reserved in the space.
249 rs.release();
250 }
251
252 return nullptr;
253 }
254
255 bool ParallelCompactData::initialize_region_data(size_t heap_size)
256 {
257 assert(is_aligned(heap_size, RegionSize), "precondition");
258
259 const size_t count = heap_size >> Log2RegionSize;
260 _region_vspace = create_vspace(count, sizeof(RegionData));
261 if (_region_vspace != nullptr) {
262 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
263 _region_count = count;
264 return true;
265 }
266 return false;
267 }
268
269 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
270 assert(beg_region <= _region_count, "beg_region out of range");
271 assert(end_region <= _region_count, "end_region out of range");
272
273 const size_t region_cnt = end_region - beg_region;
274 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
275 }
276
277 void
278 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
279 {
280 assert(is_region_aligned(beg), "not RegionSize aligned");
281 assert(is_region_aligned(end), "not RegionSize aligned");
282
283 size_t cur_region = addr_to_region_idx(beg);
284 const size_t end_region = addr_to_region_idx(end);
285 HeapWord* addr = beg;
286 while (cur_region < end_region) {
287 _region_data[cur_region].set_destination(addr);
288 _region_data[cur_region].set_destination_count(0);
289 _region_data[cur_region].set_source_region(cur_region);
290
291 // Update live_obj_size so the region appears completely full.
292 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
293 _region_data[cur_region].set_live_obj_size(live_size);
294
295 ++cur_region;
296 addr += RegionSize;
297 }
298 }
299
300 // Find the point at which a space can be split and, if necessary, record the
301 // split point.
302 //
303 // If the current src region (which overflowed the destination space) doesn't
304 // have a partial object, the split point is at the beginning of the current src
305 // region (an "easy" split, no extra bookkeeping required).
306 //
307 // If the current src region has a partial object, the split point is in the
308 // region where that partial object starts (call it the split_region). If
309 // split_region has a partial object, then the split point is just after that
310 // partial object (a "hard" split where we have to record the split data and
311 // zero the partial_obj_size field). With a "hard" split, we know that the
312 // partial_obj ends within split_region because the partial object that caused
313 // the overflow starts in split_region. If split_region doesn't have a partial
314 // obj, then the split is at the beginning of split_region (another "easy"
315 // split).
316 HeapWord*
317 ParallelCompactData::summarize_split_space(size_t src_region,
318 SplitInfo& split_info,
319 HeapWord* destination,
320 HeapWord* target_end,
321 HeapWord** target_next)
322 {
323 assert(destination <= target_end, "sanity");
324 assert(destination + _region_data[src_region].data_size() > target_end,
325 "region should not fit into target space");
326 assert(is_region_aligned(target_end), "sanity");
327
328 size_t split_region = src_region;
329 HeapWord* split_destination = destination;
330 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
331
332 if (destination + partial_obj_size > target_end) {
333 // The split point is just after the partial object (if any) in the
334 // src_region that contains the start of the object that overflowed the
335 // destination space.
336 //
337 // Find the start of the "overflow" object and set split_region to the
338 // region containing it.
339 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
340 split_region = addr_to_region_idx(overflow_obj);
341
342 // Clear the source_region field of all destination regions whose first word
343 // came from data after the split point (a non-null source_region field
344 // implies a region must be filled).
345 //
346 // An alternative to the simple loop below: clear during post_compact(),
347 // which uses memcpy instead of individual stores, and is easy to
348 // parallelize. (The downside is that it clears the entire RegionData
349 // object as opposed to just one field.)
350 //
351 // post_compact() would have to clear the summary data up to the highest
352 // address that was written during the summary phase, which would be
353 //
354 // max(top, max(new_top, clear_top))
355 //
356 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
357 // to target_end.
358 const RegionData* const sr = region(split_region);
359 const size_t beg_idx =
360 addr_to_region_idx(region_align_up(sr->destination() +
361 sr->partial_obj_size()));
362 const size_t end_idx = addr_to_region_idx(target_end);
363
364 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
365 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
366 _region_data[idx].set_source_region(0);
367 }
368
369 // Set split_destination and partial_obj_size to reflect the split region.
370 split_destination = sr->destination();
371 partial_obj_size = sr->partial_obj_size();
372 }
373
374 // The split is recorded only if a partial object extends onto the region.
375 if (partial_obj_size != 0) {
376 _region_data[split_region].set_partial_obj_size(0);
377 split_info.record(split_region, partial_obj_size, split_destination);
378 }
379
380 // Setup the continuation addresses.
381 *target_next = split_destination + partial_obj_size;
382 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
383
384 if (log_develop_is_enabled(Trace, gc, compaction)) {
385 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
386 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
387 split_type, p2i(source_next), split_region, partial_obj_size);
388 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
389 split_type, p2i(split_destination),
390 addr_to_region_idx(split_destination),
391 p2i(*target_next));
392
393 if (partial_obj_size != 0) {
394 HeapWord* const po_beg = split_info.destination();
395 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
396 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
397 split_type,
398 p2i(po_beg), addr_to_region_idx(po_beg),
399 p2i(po_end), addr_to_region_idx(po_end));
400 }
401 }
402
403 return source_next;
404 }
405
406 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
407 HeapWord** full_region_prefix_end) {
408 size_t cur_region = addr_to_region_idx(space->bottom());
409 const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
410 size_t live_words = 0;
411 if (full_region_prefix_end == nullptr) {
412 for (/* empty */; cur_region < end_region; ++cur_region) {
413 live_words += _region_data[cur_region].data_size();
414 }
415 } else {
416 bool first_set = false;
417 for (/* empty */; cur_region < end_region; ++cur_region) {
418 size_t live_words_in_region = _region_data[cur_region].data_size();
419 if (!first_set && live_words_in_region < RegionSize) {
420 *full_region_prefix_end = region_to_addr(cur_region);
421 first_set = true;
422 }
423 live_words += live_words_in_region;
424 }
425 if (!first_set) {
426 // All regions are full of live objs.
427 assert(is_region_aligned(space->top()), "inv");
428 *full_region_prefix_end = space->top();
429 }
430 assert(*full_region_prefix_end != nullptr, "postcondition");
431 assert(is_region_aligned(*full_region_prefix_end), "inv");
432 assert(*full_region_prefix_end >= space->bottom(), "in-range");
433 assert(*full_region_prefix_end <= space->top(), "in-range");
434 }
435 return live_words;
436 }
437
438 bool ParallelCompactData::summarize(SplitInfo& split_info,
439 HeapWord* source_beg, HeapWord* source_end,
440 HeapWord** source_next,
441 HeapWord* target_beg, HeapWord* target_end,
442 HeapWord** target_next)
443 {
444 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
445 log_develop_trace(gc, compaction)(
446 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
447 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
448 p2i(source_beg), p2i(source_end), p2i(source_next_val),
449 p2i(target_beg), p2i(target_end), p2i(*target_next));
450
451 size_t cur_region = addr_to_region_idx(source_beg);
452 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
453
454 HeapWord *dest_addr = target_beg;
455 while (cur_region < end_region) {
456 // The destination must be set even if the region has no data.
457 _region_data[cur_region].set_destination(dest_addr);
458
459 size_t words = _region_data[cur_region].data_size();
460 if (words > 0) {
461 // If cur_region does not fit entirely into the target space, find a point
462 // at which the source space can be 'split' so that part is copied to the
463 // target space and the rest is copied elsewhere.
464 if (dest_addr + words > target_end) {
465 assert(source_next != nullptr, "source_next is null when splitting");
466 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
467 target_end, target_next);
468 return false;
469 }
470
471 // Compute the destination_count for cur_region, and if necessary, update
472 // source_region for a destination region. The source_region field is
473 // updated if cur_region is the first (left-most) region to be copied to a
474 // destination region.
475 //
476 // The destination_count calculation is a bit subtle. A region that has
477 // data that compacts into itself does not count itself as a destination.
478 // This maintains the invariant that a zero count means the region is
479 // available and can be claimed and then filled.
480 uint destination_count = 0;
481 if (split_info.is_split(cur_region)) {
482 // The current region has been split: the partial object will be copied
483 // to one destination space and the remaining data will be copied to
484 // another destination space. Adjust the initial destination_count and,
485 // if necessary, set the source_region field if the partial object will
486 // cross a destination region boundary.
487 destination_count = split_info.destination_count();
488 if (destination_count == 2) {
489 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
490 _region_data[dest_idx].set_source_region(cur_region);
491 }
492 }
493
494 HeapWord* const last_addr = dest_addr + words - 1;
495 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
496 const size_t dest_region_2 = addr_to_region_idx(last_addr);
497
498 // Initially assume that the destination regions will be the same and
499 // adjust the value below if necessary. Under this assumption, if
500 // cur_region == dest_region_2, then cur_region will be compacted
501 // completely into itself.
502 destination_count += cur_region == dest_region_2 ? 0 : 1;
503 if (dest_region_1 != dest_region_2) {
504 // Destination regions differ; adjust destination_count.
505 destination_count += 1;
506 // Data from cur_region will be copied to the start of dest_region_2.
507 _region_data[dest_region_2].set_source_region(cur_region);
508 } else if (is_region_aligned(dest_addr)) {
509 // Data from cur_region will be copied to the start of the destination
510 // region.
511 _region_data[dest_region_1].set_source_region(cur_region);
512 }
513
514 _region_data[cur_region].set_destination_count(destination_count);
515 dest_addr += words;
516 }
517
518 ++cur_region;
519 }
520
521 *target_next = dest_addr;
522 return true;
523 }
524
525 #ifdef ASSERT
526 void ParallelCompactData::verify_clear()
527 {
528 const size_t* const beg = (const size_t*) _region_vspace->committed_low_addr();
529 const size_t* const end = (const size_t*) _region_vspace->committed_high_addr();
530 for (const size_t* p = beg; p < end; ++p) {
531 assert(*p == 0, "not zero");
532 }
533 }
534 #endif // #ifdef ASSERT
535
536 STWGCTimer PSParallelCompact::_gc_timer;
537 ParallelOldTracer PSParallelCompact::_gc_tracer;
538 elapsedTimer PSParallelCompact::_accumulated_time;
539 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
540 CollectorCounters* PSParallelCompact::_counters = nullptr;
541 ParMarkBitMap PSParallelCompact::_mark_bitmap;
542 ParallelCompactData PSParallelCompact::_summary_data;
543
544 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
545
546 class PCAdjustPointerClosure: public BasicOopIterateClosure {
547 template <typename T>
548 void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
549
550 public:
551 virtual void do_oop(oop* p) { do_oop_work(p); }
552 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
553
554 virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
555 };
556
557 static PCAdjustPointerClosure pc_adjust_pointer_closure;
558
559 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
560
561 void PSParallelCompact::post_initialize() {
562 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
563 _span_based_discoverer.set_span(heap->reserved_region());
564 _ref_processor =
565 new ReferenceProcessor(&_span_based_discoverer,
566 ParallelGCThreads, // mt processing degree
567 ParallelGCThreads, // mt discovery degree
568 false, // concurrent_discovery
569 &_is_alive_closure); // non-header is alive closure
570
571 _counters = new CollectorCounters("Parallel full collection pauses", 1);
572
573 // Initialize static fields in ParCompactionManager.
574 ParCompactionManager::initialize(mark_bitmap());
575 }
576
577 bool PSParallelCompact::initialize_aux_data() {
578 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
579 MemRegion mr = heap->reserved_region();
580 assert(mr.byte_size() != 0, "heap should be reserved");
581
582 initialize_space_info();
583
584 if (!_mark_bitmap.initialize(mr)) {
585 vm_shutdown_during_initialization(
586 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
587 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
588 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
589 return false;
590 }
591
592 if (!_summary_data.initialize(mr)) {
593 vm_shutdown_during_initialization(
594 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
595 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
596 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
597 return false;
598 }
599
600 return true;
601 }
602
603 void PSParallelCompact::initialize_space_info()
604 {
605 memset(&_space_info, 0, sizeof(_space_info));
606
607 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
608 PSYoungGen* young_gen = heap->young_gen();
609
610 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
611 _space_info[eden_space_id].set_space(young_gen->eden_space());
612 _space_info[from_space_id].set_space(young_gen->from_space());
613 _space_info[to_space_id].set_space(young_gen->to_space());
614
615 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
616 }
617
618 void
619 PSParallelCompact::clear_data_covering_space(SpaceId id)
620 {
621 // At this point, top is the value before GC, new_top() is the value that will
622 // be set at the end of GC. The marking bitmap is cleared to top; nothing
623 // should be marked above top. The summary data is cleared to the larger of
624 // top & new_top.
625 MutableSpace* const space = _space_info[id].space();
626 HeapWord* const bot = space->bottom();
627 HeapWord* const top = space->top();
628 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
629
630 _mark_bitmap.clear_range(bot, top);
631
632 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
633 const size_t end_region =
634 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
635 _summary_data.clear_range(beg_region, end_region);
636
637 // Clear the data used to 'split' regions.
638 SplitInfo& split_info = _space_info[id].split_info();
639 if (split_info.is_valid()) {
640 split_info.clear();
641 }
642 DEBUG_ONLY(split_info.verify_clear();)
643 }
644
645 void PSParallelCompact::pre_compact()
646 {
647 // Update the from & to space pointers in space_info, since they are swapped
648 // at each young gen gc. Do the update unconditionally (even though a
649 // promotion failure does not swap spaces) because an unknown number of young
650 // collections will have swapped the spaces an unknown number of times.
651 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
652 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
653 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
654 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
655
656 // Increment the invocation count
657 heap->increment_total_collections(true);
658
659 CodeCache::on_gc_marking_cycle_start();
660
661 heap->print_heap_before_gc();
662 heap->trace_heap_before_gc(&_gc_tracer);
663
664 // Fill in TLABs
665 heap->ensure_parsability(true); // retire TLABs
666
667 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
668 Universe::verify("Before GC");
669 }
670
671 DEBUG_ONLY(mark_bitmap()->verify_clear();)
672 DEBUG_ONLY(summary_data().verify_clear();)
673 }
674
675 void PSParallelCompact::post_compact()
676 {
677 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
678 ParCompactionManager::remove_all_shadow_regions();
679
680 CodeCache::on_gc_marking_cycle_finish();
681 CodeCache::arm_all_nmethods();
682
683 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
684 // Clear the marking bitmap, summary data and split info.
685 clear_data_covering_space(SpaceId(id));
686 {
687 MutableSpace* space = _space_info[id].space();
688 HeapWord* top = space->top();
689 HeapWord* new_top = _space_info[id].new_top();
690 if (ZapUnusedHeapArea && new_top < top) {
691 space->mangle_region(MemRegion(new_top, top));
692 }
693 // Update top(). Must be done after clearing the bitmap and summary data.
694 space->set_top(new_top);
695 }
696 }
697
698 ParCompactionManager::flush_all_string_dedup_requests();
699
700 MutableSpace* const eden_space = _space_info[eden_space_id].space();
701 MutableSpace* const from_space = _space_info[from_space_id].space();
702 MutableSpace* const to_space = _space_info[to_space_id].space();
703
704 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
705 bool eden_empty = eden_space->is_empty();
706
707 // Update heap occupancy information which is used as input to the soft ref
708 // clearing policy at the next gc.
709 Universe::heap()->update_capacity_and_used_at_gc();
710
711 bool young_gen_empty = eden_empty && from_space->is_empty() &&
712 to_space->is_empty();
713
714 PSCardTable* ct = heap->card_table();
715 MemRegion old_mr = heap->old_gen()->committed();
716 if (young_gen_empty) {
717 ct->clear_MemRegion(old_mr);
718 } else {
719 ct->dirty_MemRegion(old_mr);
720 }
721
722 {
723 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
724 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
725 ClassLoaderDataGraph::purge(true /* at_safepoint */);
726 DEBUG_ONLY(MetaspaceUtils::verify();)
727 }
728
729 // Need to clear claim bits for the next mark.
730 ClassLoaderDataGraph::clear_claimed_marks();
731
732 heap->prune_scavengable_nmethods();
733
734 #if COMPILER2_OR_JVMCI
735 DerivedPointerTable::update_pointers();
736 #endif
737
738 // Signal that we have completed a visit to all live objects.
739 Universe::heap()->record_whole_heap_examined_timestamp();
740 }
741
742 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
743 HeapWord* full_region_prefix_end) {
744 const size_t region_size = ParallelCompactData::RegionSize;
745 const ParallelCompactData& sd = summary_data();
746
747 // Iteration starts with the region *after* the full-region-prefix-end.
748 const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
749 // If final region is not full, iteration stops before that region,
750 // because fill_dense_prefix_end assumes that prefix_end <= top.
751 const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
752 assert(start_region <= end_region, "inv");
753
754 size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
755 const RegionData* cur_region = start_region;
756 for (/* empty */; cur_region < end_region; ++cur_region) {
757 assert(region_size >= cur_region->data_size(), "inv");
758 size_t dead_size = region_size - cur_region->data_size();
759 if (max_waste < dead_size) {
760 break;
761 }
762 max_waste -= dead_size;
763 }
764
765 HeapWord* const prefix_end = sd.region_to_addr(cur_region);
766 assert(sd.is_region_aligned(prefix_end), "postcondition");
767 assert(prefix_end >= full_region_prefix_end, "in-range");
768 assert(prefix_end <= old_space->top(), "in-range");
769 return prefix_end;
770 }
771
772 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
773 // Comparing two sizes to decide if filling is required:
774 //
775 // The size of the filler (min-obj-size) is 2 heap words with the default
776 // MinObjAlignment, since both markword and klass take 1 heap word.
777 //
778 // The size of the gap (if any) right before dense-prefix-end is
779 // MinObjAlignment.
780 //
781 // Need to fill in the gap only if it's smaller than min-obj-size, and the
782 // filler obj will extend to next region.
783
784 // Note: If min-fill-size decreases to 1, this whole method becomes redundant.
785 assert(CollectedHeap::min_fill_size() >= 2, "inv");
786 #ifndef _LP64
787 // In 32-bit system, each heap word is 4 bytes, so MinObjAlignment == 2.
788 // The gap is always equal to min-fill-size, so nothing to do.
789 return;
790 #endif
791 if (MinObjAlignment > 1) {
792 return;
793 }
794 assert(CollectedHeap::min_fill_size() == 2, "inv");
795 HeapWord* const dense_prefix_end = dense_prefix(id);
796 assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
797 assert(dense_prefix_end <= space(id)->top(), "precondition");
798 if (dense_prefix_end == space(id)->top()) {
799 // Must not have single-word gap right before prefix-end/top.
800 return;
801 }
802 RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
803
804 if (region_after_dense_prefix->partial_obj_size() != 0 ||
805 _mark_bitmap.is_marked(dense_prefix_end)) {
806 // The region after the dense prefix starts with live bytes.
807 return;
808 }
809
810 HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
811 if (block_start == dense_prefix_end - 1) {
812 assert(!_mark_bitmap.is_marked(block_start), "inv");
813 // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
814 // The filler object will extend into region_after_dense_prefix.
815 const size_t obj_len = 2; // min-fill-size
816 HeapWord* const obj_beg = dense_prefix_end - 1;
817 CollectedHeap::fill_with_object(obj_beg, obj_len);
818 _mark_bitmap.mark_obj(obj_beg);
819 _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
820 region_after_dense_prefix->set_partial_obj_size(1);
821 region_after_dense_prefix->set_partial_obj_addr(obj_beg);
822 assert(start_array(id) != nullptr, "sanity");
823 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
824 }
825 }
826
827 bool PSParallelCompact::check_maximum_compaction(size_t total_live_words,
828 MutableSpace* const old_space,
829 HeapWord* full_region_prefix_end) {
830
831 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
832
833 // Check System.GC
834 bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
835 && GCCause::is_user_requested_gc(heap->gc_cause());
836
837 // Check if all live objs are larger than old-gen.
838 const bool is_old_gen_overflowing = (total_live_words > old_space->capacity_in_words());
839
840 // JVM flags
841 const uint total_invocations = heap->total_full_collections();
842 assert(total_invocations >= _maximum_compaction_gc_num, "sanity");
843 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num;
844 const bool is_interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
845
846 // If all regions in old-gen are full
847 const bool is_region_full =
848 full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
849
850 if (is_max_on_system_gc || is_old_gen_overflowing || is_interval_ended || is_region_full) {
851 _maximum_compaction_gc_num = total_invocations;
852 return true;
853 }
854
855 return false;
856 }
857
858 void PSParallelCompact::summary_phase()
859 {
860 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
861
862 MutableSpace* const old_space = _space_info[old_space_id].space();
863 {
864 size_t total_live_words = 0;
865 HeapWord* full_region_prefix_end = nullptr;
866 {
867 // old-gen
868 size_t live_words = _summary_data.live_words_in_space(old_space,
869 &full_region_prefix_end);
870 total_live_words += live_words;
871 }
872 // young-gen
873 for (uint i = eden_space_id; i < last_space_id; ++i) {
874 const MutableSpace* space = _space_info[i].space();
875 size_t live_words = _summary_data.live_words_in_space(space);
876 total_live_words += live_words;
877 _space_info[i].set_new_top(space->bottom() + live_words);
878 _space_info[i].set_dense_prefix(space->bottom());
879 }
880
881 bool maximum_compaction = check_maximum_compaction(total_live_words,
882 old_space,
883 full_region_prefix_end);
884 HeapWord* dense_prefix_end =
885 maximum_compaction ? full_region_prefix_end
886 : compute_dense_prefix_for_old_space(old_space,
887 full_region_prefix_end);
888 SpaceId id = old_space_id;
889 _space_info[id].set_dense_prefix(dense_prefix_end);
890
891 if (dense_prefix_end != old_space->bottom()) {
892 fill_dense_prefix_end(id);
893 _summary_data.summarize_dense_prefix(old_space->bottom(), dense_prefix_end);
894 }
895 _summary_data.summarize(_space_info[id].split_info(),
896 dense_prefix_end, old_space->top(), nullptr,
897 dense_prefix_end, old_space->end(),
898 _space_info[id].new_top_addr());
899 }
900
901 // Summarize the remaining spaces in the young gen. The initial target space
902 // is the old gen. If a space does not fit entirely into the target, then the
903 // remainder is compacted into the space itself and that space becomes the new
904 // target.
905 SpaceId dst_space_id = old_space_id;
906 HeapWord* dst_space_end = old_space->end();
907 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
908 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
909 const MutableSpace* space = _space_info[id].space();
910 const size_t live = pointer_delta(_space_info[id].new_top(),
911 space->bottom());
912 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
913
914 if (live > 0 && live <= available) {
915 // All the live data will fit.
916 bool done = _summary_data.summarize(_space_info[id].split_info(),
917 space->bottom(), space->top(),
918 nullptr,
919 *new_top_addr, dst_space_end,
920 new_top_addr);
921 assert(done, "space must fit into old gen");
922
923 // Reset the new_top value for the space.
924 _space_info[id].set_new_top(space->bottom());
925 } else if (live > 0) {
926 // Attempt to fit part of the source space into the target space.
927 HeapWord* next_src_addr = nullptr;
928 bool done = _summary_data.summarize(_space_info[id].split_info(),
929 space->bottom(), space->top(),
930 &next_src_addr,
931 *new_top_addr, dst_space_end,
932 new_top_addr);
933 assert(!done, "space should not fit into old gen");
934 assert(next_src_addr != nullptr, "sanity");
935
936 // The source space becomes the new target, so the remainder is compacted
937 // within the space itself.
938 dst_space_id = SpaceId(id);
939 dst_space_end = space->end();
940 new_top_addr = _space_info[id].new_top_addr();
941 done = _summary_data.summarize(_space_info[id].split_info(),
942 next_src_addr, space->top(),
943 nullptr,
944 space->bottom(), dst_space_end,
945 new_top_addr);
946 assert(done, "space must fit when compacted into itself");
947 assert(*new_top_addr <= space->top(), "usage should not grow");
948 }
949 }
950 }
951
952 // This method should contain all heap-specific policy for invoking a full
953 // collection. invoke_no_policy() will only attempt to compact the heap; it
954 // will do nothing further. If we need to bail out for policy reasons, scavenge
955 // before full gc, or any other specialized behavior, it needs to be added here.
956 //
957 // Note that this method should only be called from the vm_thread while at a
958 // safepoint.
959 //
960 // Note that the all_soft_refs_clear flag in the soft ref policy
961 // may be true because this method can be called without intervening
962 // activity. For example when the heap space is tight and full measure
963 // are being taken to free space.
964 bool PSParallelCompact::invoke(bool clear_all_soft_refs) {
965 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
966 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
967 "should be in vm thread");
968
969 SvcGCMarker sgcm(SvcGCMarker::FULL);
970 IsSTWGCActiveMark mark;
971
972 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
973 clear_all_soft_refs = clear_all_soft_refs
974 || heap->soft_ref_policy()->should_clear_all_soft_refs();
975
976 return PSParallelCompact::invoke_no_policy(clear_all_soft_refs);
977 }
978
979 // This method contains no policy. You should probably
980 // be calling invoke() instead.
981 bool PSParallelCompact::invoke_no_policy(bool clear_all_soft_refs) {
982 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
983 assert(ref_processor() != nullptr, "Sanity");
984
985 if (GCLocker::check_active_before_gc()) {
986 return false;
987 }
988
989 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
990
991 GCIdMark gc_id_mark;
992 _gc_timer.register_gc_start();
993 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
994
995 GCCause::Cause gc_cause = heap->gc_cause();
996 PSYoungGen* young_gen = heap->young_gen();
997 PSOldGen* old_gen = heap->old_gen();
998 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
999
1000 // The scope of casr should end after code that can change
1001 // SoftRefPolicy::_should_clear_all_soft_refs.
1002 ClearedAllSoftRefs casr(clear_all_soft_refs,
1003 heap->soft_ref_policy());
1004
1005 // Make sure data structures are sane, make the heap parsable, and do other
1006 // miscellaneous bookkeeping.
1007 pre_compact();
1008
1009 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1010
1011 {
1012 const uint active_workers =
1013 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1014 ParallelScavengeHeap::heap()->workers().active_workers(),
1015 Threads::number_of_non_daemon_threads());
1016 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1017
1018 GCTraceCPUTime tcpu(&_gc_tracer);
1019 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
1020
1021 heap->pre_full_gc_dump(&_gc_timer);
1022
1023 TraceCollectorStats tcs(counters());
1024 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
1025
1026 if (log_is_enabled(Debug, gc, heap, exit)) {
1027 accumulated_time()->start();
1028 }
1029
1030 // Let the size policy know we're starting
1031 size_policy->major_collection_begin();
1032
1033 #if COMPILER2_OR_JVMCI
1034 DerivedPointerTable::clear();
1035 #endif
1036
1037 ref_processor()->start_discovery(clear_all_soft_refs);
1038
1039 ClassUnloadingContext ctx(1 /* num_nmethod_unlink_workers */,
1040 false /* unregister_nmethods_during_purge */,
1041 false /* lock_nmethod_free_separately */);
1042
1043 marking_phase(&_gc_tracer);
1044
1045 summary_phase();
1046
1047 #if COMPILER2_OR_JVMCI
1048 assert(DerivedPointerTable::is_active(), "Sanity");
1049 DerivedPointerTable::set_active(false);
1050 #endif
1051
1052 forward_to_new_addr();
1053
1054 adjust_pointers();
1055
1056 compact();
1057
1058 ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1059
1060 ParCompactionManager::verify_all_region_stack_empty();
1061
1062 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1063 // done before resizing.
1064 post_compact();
1065
1066 // Let the size policy know we're done
1067 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1068
1069 if (UseAdaptiveSizePolicy) {
1070 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1071 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1072 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1073
1074 // Don't check if the size_policy is ready here. Let
1075 // the size_policy check that internally.
1076 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1077 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1078 // Swap the survivor spaces if from_space is empty. The
1079 // resize_young_gen() called below is normally used after
1080 // a successful young GC and swapping of survivor spaces;
1081 // otherwise, it will fail to resize the young gen with
1082 // the current implementation.
1083 if (young_gen->from_space()->is_empty()) {
1084 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1085 young_gen->swap_spaces();
1086 }
1087
1088 // Calculate optimal free space amounts
1089 assert(young_gen->max_gen_size() >
1090 young_gen->from_space()->capacity_in_bytes() +
1091 young_gen->to_space()->capacity_in_bytes(),
1092 "Sizes of space in young gen are out-of-bounds");
1093
1094 size_t young_live = young_gen->used_in_bytes();
1095 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1096 size_t old_live = old_gen->used_in_bytes();
1097 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1098 size_t max_old_gen_size = old_gen->max_gen_size();
1099 size_t max_eden_size = young_gen->max_gen_size() -
1100 young_gen->from_space()->capacity_in_bytes() -
1101 young_gen->to_space()->capacity_in_bytes();
1102
1103 // Used for diagnostics
1104 size_policy->clear_generation_free_space_flags();
1105
1106 size_policy->compute_generations_free_space(young_live,
1107 eden_live,
1108 old_live,
1109 cur_eden,
1110 max_old_gen_size,
1111 max_eden_size,
1112 true /* full gc*/);
1113
1114 size_policy->check_gc_overhead_limit(eden_live,
1115 max_old_gen_size,
1116 max_eden_size,
1117 true /* full gc*/,
1118 gc_cause,
1119 heap->soft_ref_policy());
1120
1121 size_policy->decay_supplemental_growth(true /* full gc*/);
1122
1123 heap->resize_old_gen(
1124 size_policy->calculated_old_free_size_in_bytes());
1125
1126 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1127 size_policy->calculated_survivor_size_in_bytes());
1128 }
1129
1130 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1131 }
1132
1133 if (UsePerfData) {
1134 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1135 counters->update_counters();
1136 counters->update_old_capacity(old_gen->capacity_in_bytes());
1137 counters->update_young_capacity(young_gen->capacity_in_bytes());
1138 }
1139
1140 heap->resize_all_tlabs();
1141
1142 // Resize the metaspace capacity after a collection
1143 MetaspaceGC::compute_new_size();
1144
1145 if (log_is_enabled(Debug, gc, heap, exit)) {
1146 accumulated_time()->stop();
1147 }
1148
1149 heap->print_heap_change(pre_gc_values);
1150
1151 // Track memory usage and detect low memory
1152 MemoryService::track_memory_usage();
1153 heap->update_counters();
1154
1155 heap->post_full_gc_dump(&_gc_timer);
1156 }
1157
1158 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1159 Universe::verify("After GC");
1160 }
1161
1162 heap->print_heap_after_gc();
1163 heap->trace_heap_after_gc(&_gc_tracer);
1164
1165 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1166
1167 _gc_timer.register_gc_end();
1168
1169 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1170 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1171
1172 return true;
1173 }
1174
1175 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1176 private:
1177 uint _worker_id;
1178
1179 public:
1180 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1181 void do_thread(Thread* thread) {
1182 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1183
1184 ResourceMark rm;
1185
1186 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1187
1188 MarkingNMethodClosure mark_and_push_in_blobs(&cm->_mark_and_push_closure,
1189 !NMethodToOopClosure::FixRelocations,
1190 true /* keepalive nmethods */);
1191
1192 thread->oops_do(&cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1193
1194 // Do the real work
1195 cm->follow_marking_stacks();
1196 }
1197 };
1198
1199 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1200 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1201
1202 ParCompactionManager* cm =
1203 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1204
1205 do {
1206 oop obj = nullptr;
1207 ObjArrayTask task;
1208 if (ParCompactionManager::steal_objarray(worker_id, task)) {
1209 cm->follow_array((objArrayOop)task.obj(), task.index());
1210 } else if (ParCompactionManager::steal(worker_id, obj)) {
1211 cm->follow_contents(obj);
1212 }
1213 cm->follow_marking_stacks();
1214 } while (!terminator.offer_termination());
1215 }
1216
1217 class MarkFromRootsTask : public WorkerTask {
1218 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
1219 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1220 TaskTerminator _terminator;
1221 uint _active_workers;
1222
1223 public:
1224 MarkFromRootsTask(uint active_workers) :
1225 WorkerTask("MarkFromRootsTask"),
1226 _strong_roots_scope(active_workers),
1227 _terminator(active_workers, ParCompactionManager::oop_task_queues()),
1228 _active_workers(active_workers) {}
1229
1230 virtual void work(uint worker_id) {
1231 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1232 cm->create_marking_stats_cache();
1233 {
1234 CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1235 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1236
1237 // Do the real work
1238 cm->follow_marking_stacks();
1239 }
1240
1241 {
1242 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
1243 Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1244 }
1245
1246 // Mark from OopStorages
1247 {
1248 _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1249 // Do the real work
1250 cm->follow_marking_stacks();
1251 }
1252
1253 if (_active_workers > 1) {
1254 steal_marking_work(_terminator, worker_id);
1255 }
1256 }
1257 };
1258
1259 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1260 TaskTerminator _terminator;
1261
1262 public:
1263 ParallelCompactRefProcProxyTask(uint max_workers)
1264 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1265 _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
1266
1267 void work(uint worker_id) override {
1268 assert(worker_id < _max_workers, "sanity");
1269 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1270 BarrierEnqueueDiscoveredFieldClosure enqueue;
1271 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1272 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1273 }
1274
1275 void prepare_run_task_hook() override {
1276 _terminator.reset_for_reuse(_queue_count);
1277 }
1278 };
1279
1280 static void flush_marking_stats_cache(const uint num_workers) {
1281 for (uint i = 0; i < num_workers; ++i) {
1282 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1283 cm->flush_and_destroy_marking_stats_cache();
1284 }
1285 }
1286
1287 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1288 // Recursively traverse all live objects and mark them
1289 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1290
1291 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1292
1293 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1294 {
1295 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1296
1297 MarkFromRootsTask task(active_gc_threads);
1298 ParallelScavengeHeap::heap()->workers().run_task(&task);
1299 }
1300
1301 // Process reference objects found during marking
1302 {
1303 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1304
1305 ReferenceProcessorStats stats;
1306 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1307
1308 ref_processor()->set_active_mt_degree(active_gc_threads);
1309 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1310 stats = ref_processor()->process_discovered_references(task, pt);
1311
1312 gc_tracer->report_gc_reference_stats(stats);
1313 pt.print_all_references();
1314 }
1315
1316 {
1317 GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1318
1319 flush_marking_stats_cache(active_gc_threads);
1320 }
1321
1322 // This is the point where the entire marking should have completed.
1323 ParCompactionManager::verify_all_marking_stack_empty();
1324
1325 {
1326 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1327 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1328 is_alive_closure(),
1329 &do_nothing_cl,
1330 1);
1331 }
1332
1333 {
1334 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1335
1336 ClassUnloadingContext* ctx = ClassUnloadingContext::context();
1337
1338 bool unloading_occurred;
1339 {
1340 CodeCache::UnlinkingScope scope(is_alive_closure());
1341
1342 // Follow system dictionary roots and unload classes.
1343 unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1344
1345 // Unload nmethods.
1346 CodeCache::do_unloading(unloading_occurred);
1347 }
1348
1349 {
1350 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1351 // Release unloaded nmethod's memory.
1352 ctx->purge_nmethods();
1353 }
1354 {
1355 GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1356 ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1357 }
1358 {
1359 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1360 ctx->free_nmethods();
1361 }
1362
1363 // Prune dead klasses from subklass/sibling/implementor lists.
1364 Klass::clean_weak_klass_links(unloading_occurred);
1365
1366 // Clean JVMCI metadata handles.
1367 JVMCI_ONLY(JVMCI::do_unloading(unloading_occurred));
1368 }
1369
1370 {
1371 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
1372 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers());
1373 }
1374 #if TASKQUEUE_STATS
1375 ParCompactionManager::oop_task_queues()->print_and_reset_taskqueue_stats("Oop Queue");
1376 ParCompactionManager::_objarray_task_queues->print_and_reset_taskqueue_stats("ObjArrayOop Queue");
1377 #endif
1378 }
1379
1380 template<typename Func>
1381 void PSParallelCompact::adjust_in_space_helper(SpaceId id, volatile uint* claim_counter, Func&& on_stripe) {
1382 MutableSpace* sp = PSParallelCompact::space(id);
1383 HeapWord* const bottom = sp->bottom();
1384 HeapWord* const top = sp->top();
1385 if (bottom == top) {
1386 return;
1387 }
1388
1389 const uint num_regions_per_stripe = 2;
1390 const size_t region_size = ParallelCompactData::RegionSize;
1391 const size_t stripe_size = num_regions_per_stripe * region_size;
1392
1393 while (true) {
1394 uint counter = Atomic::fetch_then_add(claim_counter, num_regions_per_stripe);
1395 HeapWord* cur_stripe = bottom + counter * region_size;
1396 if (cur_stripe >= top) {
1397 break;
1398 }
1399 HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1400 on_stripe(cur_stripe, stripe_end);
1401 }
1402 }
1403
1404 void PSParallelCompact::adjust_in_old_space(volatile uint* claim_counter) {
1405 // Regions in old-space shouldn't be split.
1406 assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1407
1408 auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1409 assert(mark_bitmap()->is_marked(obj_start), "inv");
1410 oop obj = cast_to_oop(obj_start);
1411 return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1412 };
1413
1414 adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1415 assert(_summary_data.is_region_aligned(stripe_start), "inv");
1416 RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1417 HeapWord* obj_start;
1418 if (cur_region->partial_obj_size() != 0) {
1419 obj_start = cur_region->partial_obj_addr();
1420 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1421 } else {
1422 obj_start = stripe_start;
1423 }
1424
1425 while (obj_start < stripe_end) {
1426 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1427 if (obj_start >= stripe_end) {
1428 break;
1429 }
1430 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1431 }
1432 });
1433 }
1434
1435 void PSParallelCompact::adjust_in_young_space(SpaceId id, volatile uint* claim_counter) {
1436 adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1437 HeapWord* obj_start = stripe_start;
1438 while (obj_start < stripe_end) {
1439 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1440 if (obj_start >= stripe_end) {
1441 break;
1442 }
1443 oop obj = cast_to_oop(obj_start);
1444 obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1445 }
1446 });
1447 }
1448
1449 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, volatile uint* claim_counters) {
1450 auto start_time = Ticks::now();
1451 adjust_in_old_space(&claim_counters[0]);
1452 for (uint id = eden_space_id; id < last_space_id; ++id) {
1453 adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1454 }
1455 log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1456 }
1457
1458 class PSAdjustTask final : public WorkerTask {
1459 SubTasksDone _sub_tasks;
1460 WeakProcessor::Task _weak_proc_task;
1461 OopStorageSetStrongParState<false, false> _oop_storage_iter;
1462 uint _nworkers;
1463 volatile uint _claim_counters[PSParallelCompact::last_space_id] = {};
1464
1465 enum PSAdjustSubTask {
1466 PSAdjustSubTask_code_cache,
1467
1468 PSAdjustSubTask_num_elements
1469 };
1470
1471 public:
1472 PSAdjustTask(uint nworkers) :
1473 WorkerTask("PSAdjust task"),
1474 _sub_tasks(PSAdjustSubTask_num_elements),
1475 _weak_proc_task(nworkers),
1476 _nworkers(nworkers) {
1477
1478 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1479 if (nworkers > 1) {
1480 Threads::change_thread_claim_token();
1481 }
1482 }
1483
1484 ~PSAdjustTask() {
1485 Threads::assert_all_threads_claimed();
1486 }
1487
1488 void work(uint worker_id) {
1489 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1490 cm->preserved_marks()->adjust_during_full_gc();
1491 {
1492 // adjust pointers in all spaces
1493 PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1494 }
1495 {
1496 ResourceMark rm;
1497 Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1498 }
1499 _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1500 {
1501 CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1502 ClassLoaderDataGraph::cld_do(&cld_closure);
1503 }
1504 {
1505 AlwaysTrueClosure always_alive;
1506 _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1507 }
1508 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
1509 NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1510 CodeCache::nmethods_do(&adjust_code);
1511 }
1512 _sub_tasks.all_tasks_claimed();
1513 }
1514 };
1515
1516 void PSParallelCompact::adjust_pointers() {
1517 // Adjust the pointers to reflect the new locations
1518 GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1519 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1520 PSAdjustTask task(nworkers);
1521 ParallelScavengeHeap::heap()->workers().run_task(&task);
1522 }
1523
1524 // Split [start, end) evenly for a number of workers and return the
1525 // range for worker_id.
1526 static void split_regions_for_worker(size_t start, size_t end,
1527 uint worker_id, uint num_workers,
1528 size_t* worker_start, size_t* worker_end) {
1529 assert(start < end, "precondition");
1530 assert(num_workers > 0, "precondition");
1531 assert(worker_id < num_workers, "precondition");
1532
1533 size_t num_regions = end - start;
1534 size_t num_regions_per_worker = num_regions / num_workers;
1535 size_t remainder = num_regions % num_workers;
1536 // The first few workers will get one extra.
1537 *worker_start = start + worker_id * num_regions_per_worker
1538 + MIN2(checked_cast<size_t>(worker_id), remainder);
1539 *worker_end = *worker_start + num_regions_per_worker
1540 + (worker_id < remainder ? 1 : 0);
1541 }
1542
1543 void PSParallelCompact::forward_to_new_addr() {
1544 GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1545 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1546
1547 struct ForwardTask final : public WorkerTask {
1548 uint _num_workers;
1549
1550 explicit ForwardTask(uint num_workers) :
1551 WorkerTask("PSForward task"),
1552 _num_workers(num_workers) {}
1553
1554 void work(uint worker_id) override {
1555 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1556 for (uint id = old_space_id; id < last_space_id; ++id) {
1557 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1558 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1559 HeapWord* top = sp->top();
1560
1561 if (dense_prefix_addr == top) {
1562 continue;
1563 }
1564
1565 size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1566 size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1567 size_t start_region;
1568 size_t end_region;
1569 split_regions_for_worker(dense_prefix_region, top_region,
1570 worker_id, _num_workers,
1571 &start_region, &end_region);
1572 for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1573 RegionData* region_ptr = _summary_data.region(cur_region);
1574 size_t live_words = region_ptr->partial_obj_size();
1575
1576 if (live_words == ParallelCompactData::RegionSize) {
1577 // No obj-start
1578 continue;
1579 }
1580
1581 HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1582 HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1583
1584 HeapWord* cur_addr = region_start + live_words;
1585
1586 HeapWord* destination = region_ptr->destination();
1587 while (cur_addr < region_end) {
1588 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, region_end);
1589 if (cur_addr >= region_end) {
1590 break;
1591 }
1592 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1593 HeapWord* new_addr = destination + live_words;
1594 oop obj = cast_to_oop(cur_addr);
1595 if (new_addr != cur_addr) {
1596 cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1597 obj->forward_to(cast_to_oop(new_addr));
1598 }
1599 size_t obj_size = obj->size();
1600 live_words += obj_size;
1601 cur_addr += obj_size;
1602 }
1603 }
1604 }
1605 }
1606 } task(nworkers);
1607
1608 ParallelScavengeHeap::heap()->workers().run_task(&task);
1609 debug_only(verify_forward();)
1610 }
1611
1612 #ifdef ASSERT
1613 void PSParallelCompact::verify_forward() {
1614 HeapWord* old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1615 RegionData* old_region = _summary_data.region(_summary_data.addr_to_region_idx(old_dense_prefix_addr));
1616 HeapWord* bump_ptr = old_region->partial_obj_size() != 0
1617 ? old_dense_prefix_addr + old_region->partial_obj_size()
1618 : old_dense_prefix_addr;
1619 SpaceId bump_ptr_space = old_space_id;
1620
1621 for (uint id = old_space_id; id < last_space_id; ++id) {
1622 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1623 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1624 HeapWord* top = sp->top();
1625 HeapWord* cur_addr = dense_prefix_addr;
1626
1627 while (cur_addr < top) {
1628 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1629 if (cur_addr >= top) {
1630 break;
1631 }
1632 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1633 // Move to the space containing cur_addr
1634 if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1635 bump_ptr = space(space_id(cur_addr))->bottom();
1636 bump_ptr_space = space_id(bump_ptr);
1637 }
1638 oop obj = cast_to_oop(cur_addr);
1639 if (cur_addr != bump_ptr) {
1640 assert(obj->forwardee() == cast_to_oop(bump_ptr), "inv");
1641 }
1642 bump_ptr += obj->size();
1643 cur_addr += obj->size();
1644 }
1645 }
1646 }
1647 #endif
1648
1649 // Helper class to print 8 region numbers per line and then print the total at the end.
1650 class FillableRegionLogger : public StackObj {
1651 private:
1652 Log(gc, compaction) log;
1653 static const int LineLength = 8;
1654 size_t _regions[LineLength];
1655 int _next_index;
1656 bool _enabled;
1657 size_t _total_regions;
1658 public:
1659 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1660 ~FillableRegionLogger() {
1661 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
1662 }
1663
1664 void print_line() {
1665 if (!_enabled || _next_index == 0) {
1666 return;
1667 }
1668 FormatBuffer<> line("Fillable: ");
1669 for (int i = 0; i < _next_index; i++) {
1670 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
1671 }
1672 log.trace("%s", line.buffer());
1673 _next_index = 0;
1674 }
1675
1676 void handle(size_t region) {
1677 if (!_enabled) {
1678 return;
1679 }
1680 _regions[_next_index++] = region;
1681 if (_next_index == LineLength) {
1682 print_line();
1683 }
1684 _total_regions++;
1685 }
1686 };
1687
1688 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1689 {
1690 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1691
1692 // Find the threads that are active
1693 uint worker_id = 0;
1694
1695 // Find all regions that are available (can be filled immediately) and
1696 // distribute them to the thread stacks. The iteration is done in reverse
1697 // order (high to low) so the regions will be removed in ascending order.
1698
1699 const ParallelCompactData& sd = PSParallelCompact::summary_data();
1700
1701 // id + 1 is used to test termination so unsigned can
1702 // be used with an old_space_id == 0.
1703 FillableRegionLogger region_logger;
1704 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
1705 SpaceInfo* const space_info = _space_info + id;
1706 HeapWord* const new_top = space_info->new_top();
1707
1708 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1709 const size_t end_region =
1710 sd.addr_to_region_idx(sd.region_align_up(new_top));
1711
1712 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1713 if (sd.region(cur)->claim_unsafe()) {
1714 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1715 bool result = sd.region(cur)->mark_normal();
1716 assert(result, "Must succeed at this point.");
1717 cm->region_stack()->push(cur);
1718 region_logger.handle(cur);
1719 // Assign regions to tasks in round-robin fashion.
1720 if (++worker_id == parallel_gc_threads) {
1721 worker_id = 0;
1722 }
1723 }
1724 }
1725 region_logger.print_line();
1726 }
1727 }
1728
1729 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1730 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1731
1732 ParCompactionManager* cm =
1733 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1734
1735 // Drain the stacks that have been preloaded with regions
1736 // that are ready to fill.
1737
1738 cm->drain_region_stacks();
1739
1740 guarantee(cm->region_stack()->is_empty(), "Not empty");
1741
1742 size_t region_index = 0;
1743
1744 while (true) {
1745 if (ParCompactionManager::steal(worker_id, region_index)) {
1746 PSParallelCompact::fill_and_update_region(cm, region_index);
1747 cm->drain_region_stacks();
1748 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1749 // Fill and update an unavailable region with the help of a shadow region
1750 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1751 cm->drain_region_stacks();
1752 } else {
1753 if (terminator->offer_termination()) {
1754 break;
1755 }
1756 // Go around again.
1757 }
1758 }
1759 }
1760
1761 class FillDensePrefixAndCompactionTask: public WorkerTask {
1762 uint _num_workers;
1763 TaskTerminator _terminator;
1764
1765 public:
1766 FillDensePrefixAndCompactionTask(uint active_workers) :
1767 WorkerTask("FillDensePrefixAndCompactionTask"),
1768 _num_workers(active_workers),
1769 _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1770 }
1771
1772 virtual void work(uint worker_id) {
1773 {
1774 auto start = Ticks::now();
1775 PSParallelCompact::fill_dead_objs_in_dense_prefix(worker_id, _num_workers);
1776 log_trace(gc, phases)("Fill dense prefix by worker %u: %.3f ms", worker_id, (Ticks::now() - start).seconds() * 1000);
1777 }
1778 compaction_with_stealing_work(&_terminator, worker_id);
1779 }
1780 };
1781
1782 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1783 #ifdef ASSERT
1784 {
1785 assert(start < end, "precondition");
1786 assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1787 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1788 if (start != bottom) {
1789 HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1790 HeapWord* after_obj = obj_start + cast_to_oop(obj_start)->size();
1791 assert(after_obj == start, "precondition");
1792 }
1793 }
1794 #endif
1795
1796 CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1797 HeapWord* addr = start;
1798 do {
1799 size_t size = cast_to_oop(addr)->size();
1800 start_array(old_space_id)->update_for_block(addr, addr + size);
1801 addr += size;
1802 } while (addr < end);
1803 }
1804
1805 void PSParallelCompact::fill_dead_objs_in_dense_prefix(uint worker_id, uint num_workers) {
1806 ParMarkBitMap* bitmap = mark_bitmap();
1807
1808 HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1809 HeapWord* const prefix_end = dense_prefix(old_space_id);
1810
1811 if (bottom == prefix_end) {
1812 return;
1813 }
1814
1815 size_t bottom_region = _summary_data.addr_to_region_idx(bottom);
1816 size_t prefix_end_region = _summary_data.addr_to_region_idx(prefix_end);
1817
1818 size_t start_region;
1819 size_t end_region;
1820 split_regions_for_worker(bottom_region, prefix_end_region,
1821 worker_id, num_workers,
1822 &start_region, &end_region);
1823
1824 if (start_region == end_region) {
1825 return;
1826 }
1827
1828 HeapWord* const start_addr = _summary_data.region_to_addr(start_region);
1829 HeapWord* const end_addr = _summary_data.region_to_addr(end_region);
1830
1831 // Skip live partial obj (if any) from previous region.
1832 HeapWord* cur_addr;
1833 RegionData* start_region_ptr = _summary_data.region(start_region);
1834 if (start_region_ptr->partial_obj_size() != 0) {
1835 HeapWord* partial_obj_start = start_region_ptr->partial_obj_addr();
1836 assert(bitmap->is_marked(partial_obj_start), "inv");
1837 cur_addr = partial_obj_start + cast_to_oop(partial_obj_start)->size();
1838 } else {
1839 cur_addr = start_addr;
1840 }
1841
1842 // end_addr is inclusive to handle regions starting with dead space.
1843 while (cur_addr <= end_addr) {
1844 // Use prefix_end to handle trailing obj in each worker region-chunk.
1845 HeapWord* live_start = bitmap->find_obj_beg(cur_addr, prefix_end);
1846 if (cur_addr != live_start) {
1847 // Only worker 0 handles proceeding dead space.
1848 if (cur_addr != start_addr || worker_id == 0) {
1849 fill_range_in_dense_prefix(cur_addr, live_start);
1850 }
1851 }
1852 if (live_start >= end_addr) {
1853 break;
1854 }
1855 assert(bitmap->is_marked(live_start), "inv");
1856 cur_addr = live_start + cast_to_oop(live_start)->size();
1857 }
1858 }
1859
1860 void PSParallelCompact::compact() {
1861 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1862
1863 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1864
1865 initialize_shadow_regions(active_gc_threads);
1866 prepare_region_draining_tasks(active_gc_threads);
1867
1868 {
1869 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1870
1871 FillDensePrefixAndCompactionTask task(active_gc_threads);
1872 ParallelScavengeHeap::heap()->workers().run_task(&task);
1873
1874 #ifdef ASSERT
1875 verify_filler_in_dense_prefix();
1876
1877 // Verify that all regions have been processed.
1878 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1879 verify_complete(SpaceId(id));
1880 }
1881 #endif
1882 }
1883 }
1884
1885 #ifdef ASSERT
1886 void PSParallelCompact::verify_filler_in_dense_prefix() {
1887 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1888 HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1889 HeapWord* cur_addr = bottom;
1890 while (cur_addr < dense_prefix_end) {
1891 oop obj = cast_to_oop(cur_addr);
1892 oopDesc::verify(obj);
1893 if (!mark_bitmap()->is_marked(cur_addr)) {
1894 Klass* k = cast_to_oop(cur_addr)->klass_without_asserts();
1895 assert(k == Universe::fillerArrayKlass() || k == vmClasses::FillerObject_klass(), "inv");
1896 }
1897 cur_addr += obj->size();
1898 }
1899 }
1900
1901 void PSParallelCompact::verify_complete(SpaceId space_id) {
1902 // All Regions served as compaction targets, from dense_prefix() to
1903 // new_top(), should be marked as filled and all Regions between new_top()
1904 // and top() should be available (i.e., should have been emptied).
1905 ParallelCompactData& sd = summary_data();
1906 SpaceInfo si = _space_info[space_id];
1907 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1908 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1909 const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1910 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1911 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1912
1913 size_t cur_region;
1914 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1915 const RegionData* const c = sd.region(cur_region);
1916 assert(c->completed(), "region %zu not filled: destination_count=%u",
1917 cur_region, c->destination_count());
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 assert(c->available(), "region %zu not empty: destination_count=%u",
1923 cur_region, c->destination_count());
1924 }
1925 }
1926 #endif // #ifdef ASSERT
1927
1928 // Return the SpaceId for the space containing addr. If addr is not in the
1929 // heap, last_space_id is returned. In debug mode it expects the address to be
1930 // in the heap and asserts such.
1931 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1932 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1933
1934 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1935 if (_space_info[id].space()->contains(addr)) {
1936 return SpaceId(id);
1937 }
1938 }
1939
1940 assert(false, "no space contains the addr");
1941 return last_space_id;
1942 }
1943
1944 // Skip over count live words starting from beg, and return the address of the
1945 // next live word. Unless marked, the word corresponding to beg is assumed to
1946 // be dead. Callers must either ensure beg does not correspond to the middle of
1947 // an object, or account for those live words in some other way. Callers must
1948 // also ensure that there are enough live words in the range [beg, end) to skip.
1949 HeapWord*
1950 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1951 {
1952 assert(count > 0, "sanity");
1953
1954 ParMarkBitMap* m = mark_bitmap();
1955 HeapWord* cur_addr = beg;
1956 while (true) {
1957 cur_addr = m->find_obj_beg(cur_addr, end);
1958 assert(cur_addr < end, "inv");
1959 size_t obj_size = cast_to_oop(cur_addr)->size();
1960 // Strictly greater-than
1961 if (obj_size > count) {
1962 return cur_addr + count;
1963 }
1964 count -= obj_size;
1965 cur_addr += obj_size;
1966 }
1967 }
1968
1969 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1970 SpaceId src_space_id,
1971 size_t src_region_idx)
1972 {
1973 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
1974
1975 const SplitInfo& split_info = _space_info[src_space_id].split_info();
1976 if (split_info.dest_region_addr() == dest_addr) {
1977 // The partial object ending at the split point contains the first word to
1978 // be copied to dest_addr.
1979 return split_info.first_src_addr();
1980 }
1981
1982 const ParallelCompactData& sd = summary_data();
1983 ParMarkBitMap* const bitmap = mark_bitmap();
1984 const size_t RegionSize = ParallelCompactData::RegionSize;
1985
1986 assert(sd.is_region_aligned(dest_addr), "not aligned");
1987 const RegionData* const src_region_ptr = sd.region(src_region_idx);
1988 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1989 HeapWord* const src_region_destination = src_region_ptr->destination();
1990
1991 assert(dest_addr >= src_region_destination, "wrong src region");
1992 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1993
1994 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
1995 HeapWord* const src_region_end = src_region_beg + RegionSize;
1996
1997 HeapWord* addr = src_region_beg;
1998 if (dest_addr == src_region_destination) {
1999 // Return the first live word in the source region.
2000 if (partial_obj_size == 0) {
2001 addr = bitmap->find_obj_beg(addr, src_region_end);
2002 assert(addr < src_region_end, "no objects start in src region");
2003 }
2004 return addr;
2005 }
2006
2007 // Must skip some live data.
2008 size_t words_to_skip = dest_addr - src_region_destination;
2009 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2010
2011 if (partial_obj_size >= words_to_skip) {
2012 // All the live words to skip are part of the partial object.
2013 addr += words_to_skip;
2014 if (partial_obj_size == words_to_skip) {
2015 // Find the first live word past the partial object.
2016 addr = bitmap->find_obj_beg(addr, src_region_end);
2017 assert(addr < src_region_end, "wrong src region");
2018 }
2019 return addr;
2020 }
2021
2022 // Skip over the partial object (if any).
2023 if (partial_obj_size != 0) {
2024 words_to_skip -= partial_obj_size;
2025 addr += partial_obj_size;
2026 }
2027
2028 // Skip over live words due to objects that start in the region.
2029 addr = skip_live_words(addr, src_region_end, words_to_skip);
2030 assert(addr < src_region_end, "wrong src region");
2031 return addr;
2032 }
2033
2034 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2035 SpaceId src_space_id,
2036 size_t beg_region,
2037 HeapWord* end_addr)
2038 {
2039 ParallelCompactData& sd = summary_data();
2040
2041 #ifdef ASSERT
2042 MutableSpace* const src_space = _space_info[src_space_id].space();
2043 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2044 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2045 "src_space_id does not match beg_addr");
2046 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2047 "src_space_id does not match end_addr");
2048 #endif // #ifdef ASSERT
2049
2050 RegionData* const beg = sd.region(beg_region);
2051 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2052
2053 // Regions up to new_top() are enqueued if they become available.
2054 HeapWord* const new_top = _space_info[src_space_id].new_top();
2055 RegionData* const enqueue_end =
2056 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2057
2058 for (RegionData* cur = beg; cur < end; ++cur) {
2059 assert(cur->data_size() > 0, "region must have live data");
2060 cur->decrement_destination_count();
2061 if (cur < enqueue_end && cur->available() && cur->claim()) {
2062 if (cur->mark_normal()) {
2063 cm->push_region(sd.region(cur));
2064 } else if (cur->mark_copied()) {
2065 // Try to copy the content of the shadow region back to its corresponding
2066 // heap region if the shadow region is filled. Otherwise, the GC thread
2067 // fills the shadow region will copy the data back (see
2068 // MoveAndUpdateShadowClosure::complete_region).
2069 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2070 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2071 cur->set_completed();
2072 }
2073 }
2074 }
2075 }
2076
2077 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2078 SpaceId& src_space_id,
2079 HeapWord*& src_space_top,
2080 HeapWord* end_addr)
2081 {
2082 typedef ParallelCompactData::RegionData RegionData;
2083
2084 ParallelCompactData& sd = PSParallelCompact::summary_data();
2085 const size_t region_size = ParallelCompactData::RegionSize;
2086
2087 size_t src_region_idx = 0;
2088
2089 // Skip empty regions (if any) up to the top of the space.
2090 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2091 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2092 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2093 const RegionData* const top_region_ptr =
2094 sd.addr_to_region_ptr(top_aligned_up);
2095 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2096 ++src_region_ptr;
2097 }
2098
2099 if (src_region_ptr < top_region_ptr) {
2100 // The next source region is in the current space. Update src_region_idx
2101 // and the source address to match src_region_ptr.
2102 src_region_idx = sd.region(src_region_ptr);
2103 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2104 if (src_region_addr > closure.source()) {
2105 closure.set_source(src_region_addr);
2106 }
2107 return src_region_idx;
2108 }
2109
2110 // Switch to a new source space and find the first non-empty region.
2111 unsigned int space_id = src_space_id + 1;
2112 assert(space_id < last_space_id, "not enough spaces");
2113
2114 HeapWord* const destination = closure.destination();
2115
2116 do {
2117 MutableSpace* space = _space_info[space_id].space();
2118 HeapWord* const bottom = space->bottom();
2119 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2120
2121 // Iterate over the spaces that do not compact into themselves.
2122 if (bottom_cp->destination() != bottom) {
2123 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2124 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2125
2126 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2127 if (src_cp->live_obj_size() > 0) {
2128 // Found it.
2129 assert(src_cp->destination() == destination,
2130 "first live obj in the space must match the destination");
2131 assert(src_cp->partial_obj_size() == 0,
2132 "a space cannot begin with a partial obj");
2133
2134 src_space_id = SpaceId(space_id);
2135 src_space_top = space->top();
2136 const size_t src_region_idx = sd.region(src_cp);
2137 closure.set_source(sd.region_to_addr(src_region_idx));
2138 return src_region_idx;
2139 } else {
2140 assert(src_cp->data_size() == 0, "sanity");
2141 }
2142 }
2143 }
2144 } while (++space_id < last_space_id);
2145
2146 assert(false, "no source region was found");
2147 return 0;
2148 }
2149
2150 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2151 ParallelCompactData& sd = summary_data();
2152 assert(sd.is_region_aligned(region_start_addr), "precondition");
2153
2154 // Use per-region partial_obj_size to locate the end of the obj, that extends to region_start_addr.
2155 SplitInfo& split_info = _space_info[space_id(region_start_addr)].split_info();
2156 size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2157 size_t end_region_idx = sd.region_count();
2158 size_t accumulated_size = 0;
2159 for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2160 if (split_info.is_split(region_idx)) {
2161 accumulated_size += split_info.partial_obj_size();
2162 break;
2163 }
2164 size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2165 accumulated_size += cur_partial_obj_size;
2166 if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2167 break;
2168 }
2169 }
2170 return region_start_addr + accumulated_size;
2171 }
2172
2173 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2174 {
2175 ParMarkBitMap* const bitmap = mark_bitmap();
2176 ParallelCompactData& sd = summary_data();
2177 RegionData* const region_ptr = sd.region(region_idx);
2178
2179 // Get the source region and related info.
2180 size_t src_region_idx = region_ptr->source_region();
2181 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2182 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2183 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2184
2185 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2186
2187 // Adjust src_region_idx to prepare for decrementing destination counts (the
2188 // destination count is not decremented when a region is copied to itself).
2189 if (src_region_idx == region_idx) {
2190 src_region_idx += 1;
2191 }
2192
2193 if (bitmap->is_unmarked(closure.source())) {
2194 // The first source word is in the middle of an object; copy the remainder
2195 // of the object or as much as will fit. The fact that pointer updates were
2196 // deferred will be noted when the object header is processed.
2197 HeapWord* const old_src_addr = closure.source();
2198 {
2199 HeapWord* region_start = sd.region_align_down(closure.source());
2200 HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2201 HeapWord* obj_end;
2202 if (bitmap->is_marked(obj_start)) {
2203 HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2204 HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2205 ? nullptr
2206 : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2207 if (partial_obj_start == obj_start) {
2208 // This obj extends to next region.
2209 obj_end = partial_obj_end(next_region_start);
2210 } else {
2211 // Completely contained in this region; safe to use size().
2212 obj_end = obj_start + cast_to_oop(obj_start)->size();
2213 }
2214 } else {
2215 // This obj extends to current region.
2216 obj_end = partial_obj_end(region_start);
2217 }
2218 size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2219 closure.copy_partial_obj(partial_obj_size);
2220 }
2221
2222 if (closure.is_full()) {
2223 decrement_destination_counts(cm, src_space_id, src_region_idx,
2224 closure.source());
2225 closure.complete_region(dest_addr, region_ptr);
2226 return;
2227 }
2228
2229 HeapWord* const end_addr = sd.region_align_down(closure.source());
2230 if (sd.region_align_down(old_src_addr) != end_addr) {
2231 // The partial object was copied from more than one source region.
2232 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2233
2234 // Move to the next source region, possibly switching spaces as well. All
2235 // args except end_addr may be modified.
2236 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2237 end_addr);
2238 }
2239 }
2240
2241 do {
2242 HeapWord* cur_addr = closure.source();
2243 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2244 src_space_top);
2245 HeapWord* partial_obj_start = (end_addr == src_space_top)
2246 ? nullptr
2247 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2248 // apply closure on objs inside [cur_addr, end_addr)
2249 do {
2250 cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2251 if (cur_addr == end_addr) {
2252 break;
2253 }
2254 size_t obj_size;
2255 if (partial_obj_start == cur_addr) {
2256 obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2257 } else {
2258 // This obj doesn't extend into next region; size() is safe to use.
2259 obj_size = cast_to_oop(cur_addr)->size();
2260 }
2261 closure.do_addr(cur_addr, obj_size);
2262 cur_addr += obj_size;
2263 } while (cur_addr < end_addr && !closure.is_full());
2264
2265 if (closure.is_full()) {
2266 decrement_destination_counts(cm, src_space_id, src_region_idx,
2267 closure.source());
2268 closure.complete_region(dest_addr, region_ptr);
2269 return;
2270 }
2271
2272 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2273
2274 // Move to the next source region, possibly switching spaces as well. All
2275 // args except end_addr may be modified.
2276 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2277 end_addr);
2278 } while (true);
2279 }
2280
2281 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2282 {
2283 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2284 fill_region(cm, cl, region_idx);
2285 }
2286
2287 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2288 {
2289 // Get a shadow region first
2290 ParallelCompactData& sd = summary_data();
2291 RegionData* const region_ptr = sd.region(region_idx);
2292 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2293 // The InvalidShadow return value indicates the corresponding heap region is available,
2294 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2295 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2296 if (shadow_region == ParCompactionManager::InvalidShadow) {
2297 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2298 region_ptr->shadow_to_normal();
2299 return fill_region(cm, cl, region_idx);
2300 } else {
2301 MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2302 return fill_region(cm, cl, region_idx);
2303 }
2304 }
2305
2306 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2307 {
2308 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2309 }
2310
2311 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
2312 {
2313 size_t next = cm->next_shadow_region();
2314 ParallelCompactData& sd = summary_data();
2315 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2316 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2317
2318 while (next < old_new_top) {
2319 if (sd.region(next)->mark_shadow()) {
2320 region_idx = next;
2321 return true;
2322 }
2323 next = cm->move_next_shadow_region_by(active_gc_threads);
2324 }
2325
2326 return false;
2327 }
2328
2329 // The shadow region is an optimization to address region dependencies in full GC. The basic
2330 // idea is making more regions available by temporally storing their live objects in empty
2331 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2332 // GC threads need not wait destination regions to be available before processing sources.
2333 //
2334 // A typical workflow would be:
2335 // After draining its own stack and failing to steal from others, a GC worker would pick an
2336 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2337 // the shadow region by copying live objects from source regions of the unavailable one. Once
2338 // the unavailable region becomes available, the data in the shadow region will be copied back.
2339 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2340 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2341 {
2342 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2343
2344 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2345 SpaceInfo* const space_info = _space_info + id;
2346 MutableSpace* const space = space_info->space();
2347
2348 const size_t beg_region =
2349 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2350 const size_t end_region =
2351 sd.addr_to_region_idx(sd.region_align_down(space->end()));
2352
2353 for (size_t cur = beg_region; cur < end_region; ++cur) {
2354 ParCompactionManager::push_shadow_region(cur);
2355 }
2356 }
2357
2358 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2359 for (uint i = 0; i < parallel_gc_threads; i++) {
2360 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2361 cm->set_next_shadow_region(beg_region + i);
2362 }
2363 }
2364
2365 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2366 {
2367 size_t words = MIN2(partial_obj_size, words_remaining());
2368
2369 // This test is necessary; if omitted, the pointer updates to a partial object
2370 // that crosses the dense prefix boundary could be overwritten.
2371 if (source() != copy_destination()) {
2372 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2373 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2374 }
2375 update_state(words);
2376 }
2377
2378 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2379 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2380 region_ptr->set_completed();
2381 }
2382
2383 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2384 assert(destination() != nullptr, "sanity");
2385 _source = addr;
2386
2387 // The start_array must be updated even if the object is not moving.
2388 if (_start_array != nullptr) {
2389 _start_array->update_for_block(destination(), destination() + words);
2390 }
2391
2392 // Avoid overflow
2393 words = MIN2(words, words_remaining());
2394 assert(words > 0, "inv");
2395
2396 if (copy_destination() != source()) {
2397 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2398 assert(source() != destination(), "inv");
2399 assert(cast_to_oop(source())->is_forwarded(), "inv");
2400 assert(cast_to_oop(source())->forwardee() == cast_to_oop(destination()), "inv");
2401 // Read the klass before the copying, since it might destroy the klass (i.e. overlapping copy)
2402 // and if partial copy, the destination klass may not be copied yet
2403 Klass* klass = cast_to_oop(source())->klass();
2404 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2405 cast_to_oop(copy_destination())->set_mark(Klass::default_prototype_header(klass));
2406 }
2407
2408 update_state(words);
2409 }
2410
2411 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2412 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2413 // Record the shadow region index
2414 region_ptr->set_shadow_region(_shadow);
2415 // Mark the shadow region as filled to indicate the data is ready to be
2416 // copied back
2417 region_ptr->mark_filled();
2418 // Try to copy the content of the shadow region back to its corresponding
2419 // heap region if available; the GC thread that decreases the destination
2420 // count to zero will do the copying otherwise (see
2421 // PSParallelCompact::decrement_destination_counts).
2422 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2423 region_ptr->set_completed();
2424 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2425 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2426 }
2427 }
2428