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/isGCActiveMark.hpp"
55 #include "gc/shared/oopStorage.inline.hpp"
56 #include "gc/shared/oopStorageSet.inline.hpp"
57 #include "gc/shared/oopStorageSetParState.inline.hpp"
58 #include "gc/shared/preservedMarks.inline.hpp"
59 #include "gc/shared/referencePolicy.hpp"
60 #include "gc/shared/referenceProcessor.hpp"
61 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
62 #include "gc/shared/strongRootsScope.hpp"
63 #include "gc/shared/taskTerminator.hpp"
64 #include "gc/shared/weakProcessor.inline.hpp"
65 #include "gc/shared/workerPolicy.hpp"
66 #include "gc/shared/workerThread.hpp"
67 #include "gc/shared/workerUtils.hpp"
68 #include "logging/log.hpp"
69 #include "memory/iterator.inline.hpp"
70 #include "memory/metaspaceUtils.hpp"
71 #include "memory/resourceArea.hpp"
72 #include "memory/universe.hpp"
73 #include "nmt/memTracker.hpp"
74 #include "oops/access.inline.hpp"
75 #include "oops/instanceClassLoaderKlass.inline.hpp"
76 #include "oops/instanceKlass.inline.hpp"
77 #include "oops/instanceMirrorKlass.inline.hpp"
78 #include "oops/methodData.hpp"
79 #include "oops/objArrayKlass.inline.hpp"
80 #include "oops/oop.inline.hpp"
81 #include "runtime/atomic.hpp"
82 #include "runtime/handles.inline.hpp"
83 #include "runtime/java.hpp"
84 #include "runtime/safepoint.hpp"
85 #include "runtime/threads.hpp"
86 #include "runtime/vmThread.hpp"
87 #include "services/memoryService.hpp"
88 #include "utilities/align.hpp"
89 #include "utilities/debug.hpp"
90 #include "utilities/events.hpp"
91 #include "utilities/formatBuffer.hpp"
92 #include "utilities/macros.hpp"
93 #include "utilities/stack.inline.hpp"
94 #if INCLUDE_JVMCI
95 #include "jvmci/jvmci.hpp"
96 #endif
97
98 #include <math.h>
99
100 // All sizes are in HeapWords.
101 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
102 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
103 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
104 const size_t ParallelCompactData::RegionSizeBytes =
105 RegionSize << LogHeapWordSize;
106 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
107 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
108 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
109
110 const ParallelCompactData::RegionData::region_sz_t
111 ParallelCompactData::RegionData::dc_shift = 27;
112
113 const ParallelCompactData::RegionData::region_sz_t
114 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
115
116 const ParallelCompactData::RegionData::region_sz_t
117 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
118
119 const ParallelCompactData::RegionData::region_sz_t
120 ParallelCompactData::RegionData::los_mask = ~dc_mask;
121
122 const ParallelCompactData::RegionData::region_sz_t
123 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
124
125 const ParallelCompactData::RegionData::region_sz_t
126 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
127
128 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
129
130 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
131 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
132
133 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
134 HeapWord* destination)
135 {
136 assert(src_region_idx != 0, "invalid src_region_idx");
137 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
138 assert(destination != nullptr, "invalid destination argument");
139
140 _src_region_idx = src_region_idx;
141 _partial_obj_size = partial_obj_size;
142 _destination = destination;
143
144 // These fields may not be updated below, so make sure they're clear.
145 assert(_dest_region_addr == nullptr, "should have been cleared");
146 assert(_first_src_addr == nullptr, "should have been cleared");
147
148 // Determine the number of destination regions for the partial object.
149 HeapWord* const last_word = destination + partial_obj_size - 1;
150 const ParallelCompactData& sd = PSParallelCompact::summary_data();
151 HeapWord* const beg_region_addr = sd.region_align_down(destination);
152 HeapWord* const end_region_addr = sd.region_align_down(last_word);
153
154 if (beg_region_addr == end_region_addr) {
155 // One destination region.
156 _destination_count = 1;
157 if (end_region_addr == destination) {
158 // The destination falls on a region boundary, thus the first word of the
159 // partial object will be the first word copied to the destination region.
160 _dest_region_addr = end_region_addr;
161 _first_src_addr = sd.region_to_addr(src_region_idx);
162 }
163 } else {
164 // Two destination regions. When copied, the partial object will cross a
165 // destination region boundary, so a word somewhere within the partial
166 // object will be the first word copied to the second destination region.
167 _destination_count = 2;
168 _dest_region_addr = end_region_addr;
169 const size_t ofs = pointer_delta(end_region_addr, destination);
170 assert(ofs < _partial_obj_size, "sanity");
171 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
172 }
173 }
174
175 void SplitInfo::clear()
176 {
177 _src_region_idx = 0;
178 _partial_obj_size = 0;
179 _destination = nullptr;
180 _destination_count = 0;
181 _dest_region_addr = nullptr;
182 _first_src_addr = nullptr;
183 assert(!is_valid(), "sanity");
184 }
185
186 #ifdef ASSERT
187 void SplitInfo::verify_clear()
188 {
189 assert(_src_region_idx == 0, "not clear");
190 assert(_partial_obj_size == 0, "not clear");
191 assert(_destination == nullptr, "not clear");
192 assert(_destination_count == 0, "not clear");
193 assert(_dest_region_addr == nullptr, "not clear");
194 assert(_first_src_addr == nullptr, "not clear");
195 }
196 #endif // #ifdef ASSERT
197
198
199 void PSParallelCompact::print_on_error(outputStream* st) {
200 _mark_bitmap.print_on_error(st);
201 }
202
203 ParallelCompactData::ParallelCompactData() :
204 _heap_start(nullptr),
205 DEBUG_ONLY(_heap_end(nullptr) COMMA)
206 _region_vspace(nullptr),
207 _reserved_byte_size(0),
208 _region_data(nullptr),
209 _region_count(0) {}
210
211 bool ParallelCompactData::initialize(MemRegion reserved_heap)
212 {
213 _heap_start = reserved_heap.start();
214 const size_t heap_size = reserved_heap.word_size();
215 DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
216
217 assert(region_align_down(_heap_start) == _heap_start,
218 "region start not aligned");
219
220 return initialize_region_data(heap_size);
221 }
222
223 PSVirtualSpace*
224 ParallelCompactData::create_vspace(size_t count, size_t element_size)
225 {
226 const size_t raw_bytes = count * element_size;
227 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
228 const size_t granularity = os::vm_allocation_granularity();
229 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
230
231 const size_t rs_align = page_sz == os::vm_page_size() ? 0 :
232 MAX2(page_sz, granularity);
233 ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
234 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
235 rs.size(), page_sz);
236
237 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
238
239 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
240 if (vspace != 0) {
241 if (vspace->expand_by(_reserved_byte_size)) {
242 return vspace;
243 }
244 delete vspace;
245 // Release memory reserved in the space.
246 rs.release();
247 }
248
249 return 0;
250 }
251
252 bool ParallelCompactData::initialize_region_data(size_t heap_size)
253 {
254 assert(is_aligned(heap_size, RegionSize), "precondition");
255
256 const size_t count = heap_size >> Log2RegionSize;
257 _region_vspace = create_vspace(count, sizeof(RegionData));
258 if (_region_vspace != 0) {
259 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
260 _region_count = count;
261 return true;
262 }
263 return false;
264 }
265
266 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
267 assert(beg_region <= _region_count, "beg_region out of range");
268 assert(end_region <= _region_count, "end_region out of range");
269
270 const size_t region_cnt = end_region - beg_region;
271 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
272 }
273
274 void
275 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
276 {
277 assert(is_region_aligned(beg), "not RegionSize aligned");
278 assert(is_region_aligned(end), "not RegionSize aligned");
279
280 size_t cur_region = addr_to_region_idx(beg);
281 const size_t end_region = addr_to_region_idx(end);
282 HeapWord* addr = beg;
283 while (cur_region < end_region) {
284 _region_data[cur_region].set_destination(addr);
285 _region_data[cur_region].set_destination_count(0);
286 _region_data[cur_region].set_source_region(cur_region);
287
288 // Update live_obj_size so the region appears completely full.
289 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
290 _region_data[cur_region].set_live_obj_size(live_size);
291
292 ++cur_region;
293 addr += RegionSize;
294 }
295 }
296
297 // Find the point at which a space can be split and, if necessary, record the
298 // split point.
299 //
300 // If the current src region (which overflowed the destination space) doesn't
301 // have a partial object, the split point is at the beginning of the current src
302 // region (an "easy" split, no extra bookkeeping required).
303 //
304 // If the current src region has a partial object, the split point is in the
305 // region where that partial object starts (call it the split_region). If
306 // split_region has a partial object, then the split point is just after that
307 // partial object (a "hard" split where we have to record the split data and
308 // zero the partial_obj_size field). With a "hard" split, we know that the
309 // partial_obj ends within split_region because the partial object that caused
310 // the overflow starts in split_region. If split_region doesn't have a partial
311 // obj, then the split is at the beginning of split_region (another "easy"
312 // split).
313 HeapWord*
314 ParallelCompactData::summarize_split_space(size_t src_region,
315 SplitInfo& split_info,
316 HeapWord* destination,
317 HeapWord* target_end,
318 HeapWord** target_next)
319 {
320 assert(destination <= target_end, "sanity");
321 assert(destination + _region_data[src_region].data_size() > target_end,
322 "region should not fit into target space");
323 assert(is_region_aligned(target_end), "sanity");
324
325 size_t split_region = src_region;
326 HeapWord* split_destination = destination;
327 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
328
329 if (destination + partial_obj_size > target_end) {
330 // The split point is just after the partial object (if any) in the
331 // src_region that contains the start of the object that overflowed the
332 // destination space.
333 //
334 // Find the start of the "overflow" object and set split_region to the
335 // region containing it.
336 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
337 split_region = addr_to_region_idx(overflow_obj);
338
339 // Clear the source_region field of all destination regions whose first word
340 // came from data after the split point (a non-null source_region field
341 // implies a region must be filled).
342 //
343 // An alternative to the simple loop below: clear during post_compact(),
344 // which uses memcpy instead of individual stores, and is easy to
345 // parallelize. (The downside is that it clears the entire RegionData
346 // object as opposed to just one field.)
347 //
348 // post_compact() would have to clear the summary data up to the highest
349 // address that was written during the summary phase, which would be
350 //
351 // max(top, max(new_top, clear_top))
352 //
353 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
354 // to target_end.
355 const RegionData* const sr = region(split_region);
356 const size_t beg_idx =
357 addr_to_region_idx(region_align_up(sr->destination() +
358 sr->partial_obj_size()));
359 const size_t end_idx = addr_to_region_idx(target_end);
360
361 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
362 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
363 _region_data[idx].set_source_region(0);
364 }
365
366 // Set split_destination and partial_obj_size to reflect the split region.
367 split_destination = sr->destination();
368 partial_obj_size = sr->partial_obj_size();
369 }
370
371 // The split is recorded only if a partial object extends onto the region.
372 if (partial_obj_size != 0) {
373 _region_data[split_region].set_partial_obj_size(0);
374 split_info.record(split_region, partial_obj_size, split_destination);
375 }
376
377 // Setup the continuation addresses.
378 *target_next = split_destination + partial_obj_size;
379 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
380
381 if (log_develop_is_enabled(Trace, gc, compaction)) {
382 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
383 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
384 split_type, p2i(source_next), split_region, partial_obj_size);
385 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
386 split_type, p2i(split_destination),
387 addr_to_region_idx(split_destination),
388 p2i(*target_next));
389
390 if (partial_obj_size != 0) {
391 HeapWord* const po_beg = split_info.destination();
392 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
393 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
394 split_type,
395 p2i(po_beg), addr_to_region_idx(po_beg),
396 p2i(po_end), addr_to_region_idx(po_end));
397 }
398 }
399
400 return source_next;
401 }
402
403 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
404 HeapWord** full_region_prefix_end) {
405 size_t cur_region = addr_to_region_idx(space->bottom());
406 const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
407 size_t live_words = 0;
408 if (full_region_prefix_end == nullptr) {
409 for (/* empty */; cur_region < end_region; ++cur_region) {
410 live_words += _region_data[cur_region].data_size();
411 }
412 } else {
413 bool first_set = false;
414 for (/* empty */; cur_region < end_region; ++cur_region) {
415 size_t live_words_in_region = _region_data[cur_region].data_size();
416 if (!first_set && live_words_in_region < RegionSize) {
417 *full_region_prefix_end = region_to_addr(cur_region);
418 first_set = true;
419 }
420 live_words += live_words_in_region;
421 }
422 if (!first_set) {
423 // All regions are full of live objs.
424 assert(is_region_aligned(space->top()), "inv");
425 *full_region_prefix_end = space->top();
426 }
427 assert(*full_region_prefix_end != nullptr, "postcondition");
428 assert(is_region_aligned(*full_region_prefix_end), "inv");
429 assert(*full_region_prefix_end >= space->bottom(), "in-range");
430 assert(*full_region_prefix_end <= space->top(), "in-range");
431 }
432 return live_words;
433 }
434
435 bool ParallelCompactData::summarize(SplitInfo& split_info,
436 HeapWord* source_beg, HeapWord* source_end,
437 HeapWord** source_next,
438 HeapWord* target_beg, HeapWord* target_end,
439 HeapWord** target_next)
440 {
441 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
442 log_develop_trace(gc, compaction)(
443 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
444 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
445 p2i(source_beg), p2i(source_end), p2i(source_next_val),
446 p2i(target_beg), p2i(target_end), p2i(*target_next));
447
448 size_t cur_region = addr_to_region_idx(source_beg);
449 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
450
451 HeapWord *dest_addr = target_beg;
452 while (cur_region < end_region) {
453 // The destination must be set even if the region has no data.
454 _region_data[cur_region].set_destination(dest_addr);
455
456 size_t words = _region_data[cur_region].data_size();
457 if (words > 0) {
458 // If cur_region does not fit entirely into the target space, find a point
459 // at which the source space can be 'split' so that part is copied to the
460 // target space and the rest is copied elsewhere.
461 if (dest_addr + words > target_end) {
462 assert(source_next != nullptr, "source_next is null when splitting");
463 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
464 target_end, target_next);
465 return false;
466 }
467
468 // Compute the destination_count for cur_region, and if necessary, update
469 // source_region for a destination region. The source_region field is
470 // updated if cur_region is the first (left-most) region to be copied to a
471 // destination region.
472 //
473 // The destination_count calculation is a bit subtle. A region that has
474 // data that compacts into itself does not count itself as a destination.
475 // This maintains the invariant that a zero count means the region is
476 // available and can be claimed and then filled.
477 uint destination_count = 0;
478 if (split_info.is_split(cur_region)) {
479 // The current region has been split: the partial object will be copied
480 // to one destination space and the remaining data will be copied to
481 // another destination space. Adjust the initial destination_count and,
482 // if necessary, set the source_region field if the partial object will
483 // cross a destination region boundary.
484 destination_count = split_info.destination_count();
485 if (destination_count == 2) {
486 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
487 _region_data[dest_idx].set_source_region(cur_region);
488 }
489 }
490
491 HeapWord* const last_addr = dest_addr + words - 1;
492 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
493 const size_t dest_region_2 = addr_to_region_idx(last_addr);
494
495 // Initially assume that the destination regions will be the same and
496 // adjust the value below if necessary. Under this assumption, if
497 // cur_region == dest_region_2, then cur_region will be compacted
498 // completely into itself.
499 destination_count += cur_region == dest_region_2 ? 0 : 1;
500 if (dest_region_1 != dest_region_2) {
501 // Destination regions differ; adjust destination_count.
502 destination_count += 1;
503 // Data from cur_region will be copied to the start of dest_region_2.
504 _region_data[dest_region_2].set_source_region(cur_region);
505 } else if (is_region_aligned(dest_addr)) {
506 // Data from cur_region will be copied to the start of the destination
507 // region.
508 _region_data[dest_region_1].set_source_region(cur_region);
509 }
510
511 _region_data[cur_region].set_destination_count(destination_count);
512 dest_addr += words;
513 }
514
515 ++cur_region;
516 }
517
518 *target_next = dest_addr;
519 return true;
520 }
521
522 #ifdef ASSERT
523 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
524 {
525 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
526 const size_t* const end = (const size_t*)vspace->committed_high_addr();
527 for (const size_t* p = beg; p < end; ++p) {
528 assert(*p == 0, "not zero");
529 }
530 }
531
532 void ParallelCompactData::verify_clear()
533 {
534 verify_clear(_region_vspace);
535 }
536 #endif // #ifdef ASSERT
537
538 STWGCTimer PSParallelCompact::_gc_timer;
539 ParallelOldTracer PSParallelCompact::_gc_tracer;
540 elapsedTimer PSParallelCompact::_accumulated_time;
541 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
542 CollectorCounters* PSParallelCompact::_counters = nullptr;
543 ParMarkBitMap PSParallelCompact::_mark_bitmap;
544 ParallelCompactData PSParallelCompact::_summary_data;
545
546 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
547
548 class PCAdjustPointerClosure: public BasicOopIterateClosure {
549 template <typename T>
550 void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
551
552 public:
553 virtual void do_oop(oop* p) { do_oop_work(p); }
554 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
555
556 virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
557 };
558
559 static PCAdjustPointerClosure pc_adjust_pointer_closure;
560
561 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
562
563 void PSParallelCompact::post_initialize() {
564 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
565 _span_based_discoverer.set_span(heap->reserved_region());
566 _ref_processor =
567 new ReferenceProcessor(&_span_based_discoverer,
568 ParallelGCThreads, // mt processing degree
569 ParallelGCThreads, // mt discovery degree
570 false, // concurrent_discovery
571 &_is_alive_closure); // non-header is alive closure
572
573 _counters = new CollectorCounters("Parallel full collection pauses", 1);
574
575 // Initialize static fields in ParCompactionManager.
576 ParCompactionManager::initialize(mark_bitmap());
577 }
578
579 bool PSParallelCompact::initialize_aux_data() {
580 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
581 MemRegion mr = heap->reserved_region();
582 assert(mr.byte_size() != 0, "heap should be reserved");
583
584 initialize_space_info();
585
586 if (!_mark_bitmap.initialize(mr)) {
587 vm_shutdown_during_initialization(
588 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
589 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
590 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
591 return false;
592 }
593
594 if (!_summary_data.initialize(mr)) {
595 vm_shutdown_during_initialization(
596 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
597 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
598 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
599 return false;
600 }
601
602 return true;
603 }
604
605 void PSParallelCompact::initialize_space_info()
606 {
607 memset(&_space_info, 0, sizeof(_space_info));
608
609 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
610 PSYoungGen* young_gen = heap->young_gen();
611
612 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
613 _space_info[eden_space_id].set_space(young_gen->eden_space());
614 _space_info[from_space_id].set_space(young_gen->from_space());
615 _space_info[to_space_id].set_space(young_gen->to_space());
616
617 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
618 }
619
620 void
621 PSParallelCompact::clear_data_covering_space(SpaceId id)
622 {
623 // At this point, top is the value before GC, new_top() is the value that will
624 // be set at the end of GC. The marking bitmap is cleared to top; nothing
625 // should be marked above top. The summary data is cleared to the larger of
626 // top & new_top.
627 MutableSpace* const space = _space_info[id].space();
628 HeapWord* const bot = space->bottom();
629 HeapWord* const top = space->top();
630 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
631
632 _mark_bitmap.clear_range(bot, top);
633
634 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
635 const size_t end_region =
636 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
637 _summary_data.clear_range(beg_region, end_region);
638
639 // Clear the data used to 'split' regions.
640 SplitInfo& split_info = _space_info[id].split_info();
641 if (split_info.is_valid()) {
642 split_info.clear();
643 }
644 DEBUG_ONLY(split_info.verify_clear();)
645 }
646
647 void PSParallelCompact::pre_compact()
648 {
649 // Update the from & to space pointers in space_info, since they are swapped
650 // at each young gen gc. Do the update unconditionally (even though a
651 // promotion failure does not swap spaces) because an unknown number of young
652 // collections will have swapped the spaces an unknown number of times.
653 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
654 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
655 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
656 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
657
658 // Increment the invocation count
659 heap->increment_total_collections(true);
660
661 CodeCache::on_gc_marking_cycle_start();
662
663 heap->print_heap_before_gc();
664 heap->trace_heap_before_gc(&_gc_tracer);
665
666 // Fill in TLABs
667 heap->ensure_parsability(true); // retire TLABs
668
669 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
670 Universe::verify("Before GC");
671 }
672
673 DEBUG_ONLY(mark_bitmap()->verify_clear();)
674 DEBUG_ONLY(summary_data().verify_clear();)
675 }
676
677 void PSParallelCompact::post_compact()
678 {
679 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
680 ParCompactionManager::remove_all_shadow_regions();
681
682 CodeCache::on_gc_marking_cycle_finish();
683 CodeCache::arm_all_nmethods();
684
685 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
686 // Clear the marking bitmap, summary data and split info.
687 clear_data_covering_space(SpaceId(id));
688 {
689 MutableSpace* space = _space_info[id].space();
690 HeapWord* top = space->top();
691 HeapWord* new_top = _space_info[id].new_top();
692 if (ZapUnusedHeapArea && new_top < top) {
693 space->mangle_region(MemRegion(new_top, top));
694 }
695 // Update top(). Must be done after clearing the bitmap and summary data.
696 space->set_top(new_top);
697 }
698 }
699
700 ParCompactionManager::flush_all_string_dedup_requests();
701
702 MutableSpace* const eden_space = _space_info[eden_space_id].space();
703 MutableSpace* const from_space = _space_info[from_space_id].space();
704 MutableSpace* const to_space = _space_info[to_space_id].space();
705
706 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
707 bool eden_empty = eden_space->is_empty();
708
709 // Update heap occupancy information which is used as input to the soft ref
710 // clearing policy at the next gc.
711 Universe::heap()->update_capacity_and_used_at_gc();
712
713 bool young_gen_empty = eden_empty && from_space->is_empty() &&
714 to_space->is_empty();
715
716 PSCardTable* ct = heap->card_table();
717 MemRegion old_mr = heap->old_gen()->committed();
718 if (young_gen_empty) {
719 ct->clear_MemRegion(old_mr);
720 } else {
721 ct->dirty_MemRegion(old_mr);
722 }
723
724 {
725 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
726 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
727 ClassLoaderDataGraph::purge(true /* at_safepoint */);
728 DEBUG_ONLY(MetaspaceUtils::verify();)
729 }
730
731 // Need to clear claim bits for the next mark.
732 ClassLoaderDataGraph::clear_claimed_marks();
733
734 heap->prune_scavengable_nmethods();
735
736 #if COMPILER2_OR_JVMCI
737 DerivedPointerTable::update_pointers();
738 #endif
739
740 // Signal that we have completed a visit to all live objects.
741 Universe::heap()->record_whole_heap_examined_timestamp();
742 }
743
744 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
745 HeapWord* full_region_prefix_end) {
746 const size_t region_size = ParallelCompactData::RegionSize;
747 const ParallelCompactData& sd = summary_data();
748
749 // Iteration starts with the region *after* the full-region-prefix-end.
750 const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
751 // If final region is not full, iteration stops before that region,
752 // because fill_dense_prefix_end assumes that prefix_end <= top.
753 const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
754 assert(start_region <= end_region, "inv");
755
756 size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
757 const RegionData* cur_region = start_region;
758 for (/* empty */; cur_region < end_region; ++cur_region) {
759 assert(region_size >= cur_region->data_size(), "inv");
760 size_t dead_size = region_size - cur_region->data_size();
761 if (max_waste < dead_size) {
762 break;
763 }
764 max_waste -= dead_size;
765 }
766
767 HeapWord* const prefix_end = sd.region_to_addr(cur_region);
768 assert(sd.is_region_aligned(prefix_end), "postcondition");
769 assert(prefix_end >= full_region_prefix_end, "in-range");
770 assert(prefix_end <= old_space->top(), "in-range");
771 return prefix_end;
772 }
773
774 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
775 // Comparing two sizes to decide if filling is required:
776 //
777 // The size of the filler (min-obj-size) is 2 heap words with the default
778 // MinObjAlignment, since both markword and klass take 1 heap word.
779 //
780 // The size of the gap (if any) right before dense-prefix-end is
781 // MinObjAlignment.
782 //
783 // Need to fill in the gap only if it's smaller than min-obj-size, and the
784 // filler obj will extend to next region.
785
786 // Note: If min-fill-size decreases to 1, this whole method becomes redundant.
787 assert(CollectedHeap::min_fill_size() >= 2, "inv");
788 #ifndef _LP64
789 // In 32-bit system, each heap word is 4 bytes, so MinObjAlignment == 2.
790 // The gap is always equal to min-fill-size, so nothing to do.
791 return;
792 #endif
793 if (MinObjAlignment > 1) {
794 return;
795 }
796 assert(CollectedHeap::min_fill_size() == 2, "inv");
797 HeapWord* const dense_prefix_end = dense_prefix(id);
798 assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
799 assert(dense_prefix_end <= space(id)->top(), "precondition");
800 if (dense_prefix_end == space(id)->top()) {
801 // Must not have single-word gap right before prefix-end/top.
802 return;
803 }
804 RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
805
806 if (region_after_dense_prefix->partial_obj_size() != 0 ||
807 _mark_bitmap.is_marked(dense_prefix_end)) {
808 // The region after the dense prefix starts with live bytes.
809 return;
810 }
811
812 HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
813 if (block_start == dense_prefix_end - 1) {
814 assert(!_mark_bitmap.is_marked(block_start), "inv");
815 // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
816 // The filler object will extend into region_after_dense_prefix.
817 const size_t obj_len = 2; // min-fill-size
818 HeapWord* const obj_beg = dense_prefix_end - 1;
819 CollectedHeap::fill_with_object(obj_beg, obj_len);
820 _mark_bitmap.mark_obj(obj_beg);
821 _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
822 region_after_dense_prefix->set_partial_obj_size(1);
823 region_after_dense_prefix->set_partial_obj_addr(obj_beg);
824 assert(start_array(id) != nullptr, "sanity");
825 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
826 }
827 }
828
829 bool PSParallelCompact::reassess_maximum_compaction(bool maximum_compaction,
830 size_t total_live_words,
831 MutableSpace* const old_space,
832 HeapWord* full_region_prefix_end) {
833 // Check if all live objs are larger than old-gen.
834 const bool is_old_gen_overflowing = (total_live_words > old_space->capacity_in_words());
835
836 // JVM flags
837 const uint total_invocations = ParallelScavengeHeap::heap()->total_full_collections();
838 assert(total_invocations >= _maximum_compaction_gc_num, "sanity");
839 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num;
840 const bool is_interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
841
842 // If all regions in old-gen are full
843 const bool is_region_full =
844 full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
845
846 if (maximum_compaction || is_old_gen_overflowing || is_interval_ended || is_region_full) {
847 _maximum_compaction_gc_num = total_invocations;
848 return true;
849 }
850
851 return false;
852 }
853
854 void PSParallelCompact::summary_phase(bool maximum_compaction)
855 {
856 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
857
858 MutableSpace* const old_space = _space_info[old_space_id].space();
859 {
860 size_t total_live_words = 0;
861 HeapWord* full_region_prefix_end = nullptr;
862 {
863 // old-gen
864 size_t live_words = _summary_data.live_words_in_space(old_space,
865 &full_region_prefix_end);
866 total_live_words += live_words;
867 }
868 // young-gen
869 for (uint i = eden_space_id; i < last_space_id; ++i) {
870 const MutableSpace* space = _space_info[i].space();
871 size_t live_words = _summary_data.live_words_in_space(space);
872 total_live_words += live_words;
873 _space_info[i].set_new_top(space->bottom() + live_words);
874 _space_info[i].set_dense_prefix(space->bottom());
875 }
876
877 maximum_compaction = reassess_maximum_compaction(maximum_compaction,
878 total_live_words,
879 old_space,
880 full_region_prefix_end);
881 HeapWord* dense_prefix_end =
882 maximum_compaction ? full_region_prefix_end
883 : compute_dense_prefix_for_old_space(old_space,
884 full_region_prefix_end);
885 SpaceId id = old_space_id;
886 _space_info[id].set_dense_prefix(dense_prefix_end);
887
888 if (dense_prefix_end != old_space->bottom()) {
889 fill_dense_prefix_end(id);
890 _summary_data.summarize_dense_prefix(old_space->bottom(), dense_prefix_end);
891 }
892 _summary_data.summarize(_space_info[id].split_info(),
893 dense_prefix_end, old_space->top(), nullptr,
894 dense_prefix_end, old_space->end(),
895 _space_info[id].new_top_addr());
896 }
897
898 // Summarize the remaining spaces in the young gen. The initial target space
899 // is the old gen. If a space does not fit entirely into the target, then the
900 // remainder is compacted into the space itself and that space becomes the new
901 // target.
902 SpaceId dst_space_id = old_space_id;
903 HeapWord* dst_space_end = old_space->end();
904 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
905 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
906 const MutableSpace* space = _space_info[id].space();
907 const size_t live = pointer_delta(_space_info[id].new_top(),
908 space->bottom());
909 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
910
911 if (live > 0 && live <= available) {
912 // All the live data will fit.
913 bool done = _summary_data.summarize(_space_info[id].split_info(),
914 space->bottom(), space->top(),
915 nullptr,
916 *new_top_addr, dst_space_end,
917 new_top_addr);
918 assert(done, "space must fit into old gen");
919
920 // Reset the new_top value for the space.
921 _space_info[id].set_new_top(space->bottom());
922 } else if (live > 0) {
923 // Attempt to fit part of the source space into the target space.
924 HeapWord* next_src_addr = nullptr;
925 bool done = _summary_data.summarize(_space_info[id].split_info(),
926 space->bottom(), space->top(),
927 &next_src_addr,
928 *new_top_addr, dst_space_end,
929 new_top_addr);
930 assert(!done, "space should not fit into old gen");
931 assert(next_src_addr != nullptr, "sanity");
932
933 // The source space becomes the new target, so the remainder is compacted
934 // within the space itself.
935 dst_space_id = SpaceId(id);
936 dst_space_end = space->end();
937 new_top_addr = _space_info[id].new_top_addr();
938 done = _summary_data.summarize(_space_info[id].split_info(),
939 next_src_addr, space->top(),
940 nullptr,
941 space->bottom(), dst_space_end,
942 new_top_addr);
943 assert(done, "space must fit when compacted into itself");
944 assert(*new_top_addr <= space->top(), "usage should not grow");
945 }
946 }
947 }
948
949 // This method should contain all heap-specific policy for invoking a full
950 // collection. invoke_no_policy() will only attempt to compact the heap; it
951 // will do nothing further. If we need to bail out for policy reasons, scavenge
952 // before full gc, or any other specialized behavior, it needs to be added here.
953 //
954 // Note that this method should only be called from the vm_thread while at a
955 // safepoint.
956 //
957 // Note that the all_soft_refs_clear flag in the soft ref policy
958 // may be true because this method can be called without intervening
959 // activity. For example when the heap space is tight and full measure
960 // are being taken to free space.
961 bool PSParallelCompact::invoke(bool maximum_heap_compaction) {
962 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
963 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
964 "should be in vm thread");
965
966 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
967 assert(!heap->is_stw_gc_active(), "not reentrant");
968
969 IsSTWGCActiveMark mark;
970
971 const bool clear_all_soft_refs =
972 heap->soft_ref_policy()->should_clear_all_soft_refs();
973
974 return PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
975 maximum_heap_compaction);
976 }
977
978 // This method contains no policy. You should probably
979 // be calling invoke() instead.
980 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
981 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
982 assert(ref_processor() != nullptr, "Sanity");
983
984 if (GCLocker::check_active_before_gc()) {
985 return false;
986 }
987
988 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
989
990 GCIdMark gc_id_mark;
991 _gc_timer.register_gc_start();
992 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
993
994 GCCause::Cause gc_cause = heap->gc_cause();
995 PSYoungGen* young_gen = heap->young_gen();
996 PSOldGen* old_gen = heap->old_gen();
997 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
998
999 // The scope of casr should end after code that can change
1000 // SoftRefPolicy::_should_clear_all_soft_refs.
1001 ClearedAllSoftRefs casr(maximum_heap_compaction,
1002 heap->soft_ref_policy());
1003
1004 // Make sure data structures are sane, make the heap parsable, and do other
1005 // miscellaneous bookkeeping.
1006 pre_compact();
1007
1008 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1009
1010 {
1011 const uint active_workers =
1012 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1013 ParallelScavengeHeap::heap()->workers().active_workers(),
1014 Threads::number_of_non_daemon_threads());
1015 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1016
1017 GCTraceCPUTime tcpu(&_gc_tracer);
1018 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
1019
1020 heap->pre_full_gc_dump(&_gc_timer);
1021
1022 TraceCollectorStats tcs(counters());
1023 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
1024
1025 if (log_is_enabled(Debug, gc, heap, exit)) {
1026 accumulated_time()->start();
1027 }
1028
1029 // Let the size policy know we're starting
1030 size_policy->major_collection_begin();
1031
1032 #if COMPILER2_OR_JVMCI
1033 DerivedPointerTable::clear();
1034 #endif
1035
1036 ref_processor()->start_discovery(maximum_heap_compaction);
1037
1038 ClassUnloadingContext ctx(1 /* num_nmethod_unlink_workers */,
1039 false /* unregister_nmethods_during_purge */,
1040 false /* lock_nmethod_free_separately */);
1041
1042 marking_phase(&_gc_tracer);
1043
1044 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1045 && GCCause::is_user_requested_gc(gc_cause);
1046 summary_phase(maximum_heap_compaction || max_on_system_gc);
1047
1048 #if COMPILER2_OR_JVMCI
1049 assert(DerivedPointerTable::is_active(), "Sanity");
1050 DerivedPointerTable::set_active(false);
1051 #endif
1052
1053 forward_to_new_addr();
1054
1055 adjust_pointers();
1056
1057 compact();
1058
1059 ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1060
1061 ParCompactionManager::verify_all_region_stack_empty();
1062
1063 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1064 // done before resizing.
1065 post_compact();
1066
1067 // Let the size policy know we're done
1068 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1069
1070 if (UseAdaptiveSizePolicy) {
1071 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1072 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1073 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1074
1075 // Don't check if the size_policy is ready here. Let
1076 // the size_policy check that internally.
1077 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1078 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1079 // Swap the survivor spaces if from_space is empty. The
1080 // resize_young_gen() called below is normally used after
1081 // a successful young GC and swapping of survivor spaces;
1082 // otherwise, it will fail to resize the young gen with
1083 // the current implementation.
1084 if (young_gen->from_space()->is_empty()) {
1085 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1086 young_gen->swap_spaces();
1087 }
1088
1089 // Calculate optimal free space amounts
1090 assert(young_gen->max_gen_size() >
1091 young_gen->from_space()->capacity_in_bytes() +
1092 young_gen->to_space()->capacity_in_bytes(),
1093 "Sizes of space in young gen are out-of-bounds");
1094
1095 size_t young_live = young_gen->used_in_bytes();
1096 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1097 size_t old_live = old_gen->used_in_bytes();
1098 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1099 size_t max_old_gen_size = old_gen->max_gen_size();
1100 size_t max_eden_size = young_gen->max_gen_size() -
1101 young_gen->from_space()->capacity_in_bytes() -
1102 young_gen->to_space()->capacity_in_bytes();
1103
1104 // Used for diagnostics
1105 size_policy->clear_generation_free_space_flags();
1106
1107 size_policy->compute_generations_free_space(young_live,
1108 eden_live,
1109 old_live,
1110 cur_eden,
1111 max_old_gen_size,
1112 max_eden_size,
1113 true /* full gc*/);
1114
1115 size_policy->check_gc_overhead_limit(eden_live,
1116 max_old_gen_size,
1117 max_eden_size,
1118 true /* full gc*/,
1119 gc_cause,
1120 heap->soft_ref_policy());
1121
1122 size_policy->decay_supplemental_growth(true /* full gc*/);
1123
1124 heap->resize_old_gen(
1125 size_policy->calculated_old_free_size_in_bytes());
1126
1127 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1128 size_policy->calculated_survivor_size_in_bytes());
1129 }
1130
1131 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1132 }
1133
1134 if (UsePerfData) {
1135 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1136 counters->update_counters();
1137 counters->update_old_capacity(old_gen->capacity_in_bytes());
1138 counters->update_young_capacity(young_gen->capacity_in_bytes());
1139 }
1140
1141 heap->resize_all_tlabs();
1142
1143 // Resize the metaspace capacity after a collection
1144 MetaspaceGC::compute_new_size();
1145
1146 if (log_is_enabled(Debug, gc, heap, exit)) {
1147 accumulated_time()->stop();
1148 }
1149
1150 heap->print_heap_change(pre_gc_values);
1151
1152 // Track memory usage and detect low memory
1153 MemoryService::track_memory_usage();
1154 heap->update_counters();
1155
1156 heap->post_full_gc_dump(&_gc_timer);
1157 }
1158
1159 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1160 Universe::verify("After GC");
1161 }
1162
1163 heap->print_heap_after_gc();
1164 heap->trace_heap_after_gc(&_gc_tracer);
1165
1166 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1167
1168 _gc_timer.register_gc_end();
1169
1170 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1171 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1172
1173 return true;
1174 }
1175
1176 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1177 private:
1178 uint _worker_id;
1179
1180 public:
1181 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1182 void do_thread(Thread* thread) {
1183 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1184
1185 ResourceMark rm;
1186
1187 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1188
1189 PCMarkAndPushClosure mark_and_push_closure(cm);
1190 MarkingNMethodClosure mark_and_push_in_blobs(&mark_and_push_closure, !NMethodToOopClosure::FixRelocations, true /* keepalive nmethods */);
1191
1192 thread->oops_do(&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 PCMarkAndPushClosure mark_and_push_closure(cm);
1234
1235 {
1236 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1237 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1238
1239 // Do the real work
1240 cm->follow_marking_stacks();
1241 }
1242
1243 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
1244 Threads::possibly_parallel_threads_do(true /* is_par */, &closure);
1245
1246 // Mark from OopStorages
1247 {
1248 _oop_storage_set_par_state.oops_do(&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 PCMarkAndPushClosure keep_alive(cm);
1271 BarrierEnqueueDiscoveredFieldClosure enqueue;
1272 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1273 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
1274 }
1275
1276 void prepare_run_task_hook() override {
1277 _terminator.reset_for_reuse(_queue_count);
1278 }
1279 };
1280
1281 static void flush_marking_stats_cache(const uint num_workers) {
1282 for (uint i = 0; i < num_workers; ++i) {
1283 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1284 cm->flush_and_destroy_marking_stats_cache();
1285 }
1286 }
1287
1288 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1289 // Recursively traverse all live objects and mark them
1290 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1291
1292 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1293
1294 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1295 {
1296 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1297
1298 MarkFromRootsTask task(active_gc_threads);
1299 ParallelScavengeHeap::heap()->workers().run_task(&task);
1300 }
1301
1302 // Process reference objects found during marking
1303 {
1304 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1305
1306 ReferenceProcessorStats stats;
1307 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1308
1309 ref_processor()->set_active_mt_degree(active_gc_threads);
1310 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1311 stats = ref_processor()->process_discovered_references(task, pt);
1312
1313 gc_tracer->report_gc_reference_stats(stats);
1314 pt.print_all_references();
1315 }
1316
1317 {
1318 GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1319
1320 flush_marking_stats_cache(active_gc_threads);
1321 }
1322
1323 // This is the point where the entire marking should have completed.
1324 ParCompactionManager::verify_all_marking_stack_empty();
1325
1326 {
1327 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1328 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1329 is_alive_closure(),
1330 &do_nothing_cl,
1331 1);
1332 }
1333
1334 {
1335 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1336
1337 ClassUnloadingContext* ctx = ClassUnloadingContext::context();
1338
1339 bool unloading_occurred;
1340 {
1341 CodeCache::UnlinkingScope scope(is_alive_closure());
1342
1343 // Follow system dictionary roots and unload classes.
1344 unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1345
1346 // Unload nmethods.
1347 CodeCache::do_unloading(unloading_occurred);
1348 }
1349
1350 {
1351 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1352 // Release unloaded nmethod's memory.
1353 ctx->purge_nmethods();
1354 }
1355 {
1356 GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1357 ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1358 }
1359 {
1360 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1361 ctx->free_nmethods();
1362 }
1363
1364 // Prune dead klasses from subklass/sibling/implementor lists.
1365 Klass::clean_weak_klass_links(unloading_occurred);
1366
1367 // Clean JVMCI metadata handles.
1368 JVMCI_ONLY(JVMCI::do_unloading(unloading_occurred));
1369 }
1370
1371 {
1372 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
1373 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers());
1374 }
1375 #if TASKQUEUE_STATS
1376 ParCompactionManager::oop_task_queues()->print_and_reset_taskqueue_stats("Oop Queue");
1377 ParCompactionManager::_objarray_task_queues->print_and_reset_taskqueue_stats("ObjArrayOop Queue");
1378 #endif
1379 }
1380
1381 template<typename Func>
1382 void PSParallelCompact::adjust_in_space_helper(SpaceId id, volatile uint* claim_counter, Func&& on_stripe) {
1383 MutableSpace* sp = PSParallelCompact::space(id);
1384 HeapWord* const bottom = sp->bottom();
1385 HeapWord* const top = sp->top();
1386 if (bottom == top) {
1387 return;
1388 }
1389
1390 const uint num_regions_per_stripe = 2;
1391 const size_t region_size = ParallelCompactData::RegionSize;
1392 const size_t stripe_size = num_regions_per_stripe * region_size;
1393
1394 while (true) {
1395 uint counter = Atomic::fetch_then_add(claim_counter, num_regions_per_stripe);
1396 HeapWord* cur_stripe = bottom + counter * region_size;
1397 if (cur_stripe >= top) {
1398 break;
1399 }
1400 HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1401 on_stripe(cur_stripe, stripe_end);
1402 }
1403 }
1404
1405 void PSParallelCompact::adjust_in_old_space(volatile uint* claim_counter) {
1406 // Regions in old-space shouldn't be split.
1407 assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1408
1409 auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1410 assert(mark_bitmap()->is_marked(obj_start), "inv");
1411 oop obj = cast_to_oop(obj_start);
1412 return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1413 };
1414
1415 adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1416 assert(_summary_data.is_region_aligned(stripe_start), "inv");
1417 RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1418 HeapWord* obj_start;
1419 if (cur_region->partial_obj_size() != 0) {
1420 obj_start = cur_region->partial_obj_addr();
1421 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1422 } else {
1423 obj_start = stripe_start;
1424 }
1425
1426 while (obj_start < stripe_end) {
1427 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1428 if (obj_start >= stripe_end) {
1429 break;
1430 }
1431 obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1432 }
1433 });
1434 }
1435
1436 void PSParallelCompact::adjust_in_young_space(SpaceId id, volatile uint* claim_counter) {
1437 adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1438 HeapWord* obj_start = stripe_start;
1439 while (obj_start < stripe_end) {
1440 obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1441 if (obj_start >= stripe_end) {
1442 break;
1443 }
1444 oop obj = cast_to_oop(obj_start);
1445 obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1446 }
1447 });
1448 }
1449
1450 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, volatile uint* claim_counters) {
1451 auto start_time = Ticks::now();
1452 adjust_in_old_space(&claim_counters[0]);
1453 for (uint id = eden_space_id; id < last_space_id; ++id) {
1454 adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1455 }
1456 log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1457 }
1458
1459 class PSAdjustTask final : public WorkerTask {
1460 SubTasksDone _sub_tasks;
1461 WeakProcessor::Task _weak_proc_task;
1462 OopStorageSetStrongParState<false, false> _oop_storage_iter;
1463 uint _nworkers;
1464 volatile uint _claim_counters[PSParallelCompact::last_space_id] = {};
1465
1466 enum PSAdjustSubTask {
1467 PSAdjustSubTask_code_cache,
1468
1469 PSAdjustSubTask_num_elements
1470 };
1471
1472 public:
1473 PSAdjustTask(uint nworkers) :
1474 WorkerTask("PSAdjust task"),
1475 _sub_tasks(PSAdjustSubTask_num_elements),
1476 _weak_proc_task(nworkers),
1477 _nworkers(nworkers) {
1478
1479 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1480 if (nworkers > 1) {
1481 Threads::change_thread_claim_token();
1482 }
1483 }
1484
1485 ~PSAdjustTask() {
1486 Threads::assert_all_threads_claimed();
1487 }
1488
1489 void work(uint worker_id) {
1490 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1491 cm->preserved_marks()->adjust_during_full_gc();
1492 {
1493 // adjust pointers in all spaces
1494 PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1495 }
1496 {
1497 ResourceMark rm;
1498 Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1499 }
1500 _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1501 {
1502 CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1503 ClassLoaderDataGraph::cld_do(&cld_closure);
1504 }
1505 {
1506 AlwaysTrueClosure always_alive;
1507 _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1508 }
1509 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
1510 NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1511 CodeCache::nmethods_do(&adjust_code);
1512 }
1513 _sub_tasks.all_tasks_claimed();
1514 }
1515 };
1516
1517 void PSParallelCompact::adjust_pointers() {
1518 // Adjust the pointers to reflect the new locations
1519 GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1520 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1521 PSAdjustTask task(nworkers);
1522 ParallelScavengeHeap::heap()->workers().run_task(&task);
1523 }
1524
1525 // Split [start, end) evenly for a number of workers and return the
1526 // range for worker_id.
1527 static void split_regions_for_worker(size_t start, size_t end,
1528 uint worker_id, uint num_workers,
1529 size_t* worker_start, size_t* worker_end) {
1530 assert(start < end, "precondition");
1531 assert(num_workers > 0, "precondition");
1532 assert(worker_id < num_workers, "precondition");
1533
1534 size_t num_regions = end - start;
1535 size_t num_regions_per_worker = num_regions / num_workers;
1536 size_t remainder = num_regions % num_workers;
1537 // The first few workers will get one extra.
1538 *worker_start = start + worker_id * num_regions_per_worker
1539 + MIN2(checked_cast<size_t>(worker_id), remainder);
1540 *worker_end = *worker_start + num_regions_per_worker
1541 + (worker_id < remainder ? 1 : 0);
1542 }
1543
1544 void PSParallelCompact::forward_to_new_addr() {
1545 GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1546 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1547
1548 struct ForwardTask final : public WorkerTask {
1549 uint _num_workers;
1550
1551 explicit ForwardTask(uint num_workers) :
1552 WorkerTask("PSForward task"),
1553 _num_workers(num_workers) {}
1554
1555 void work(uint worker_id) override {
1556 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1557 for (uint id = old_space_id; id < last_space_id; ++id) {
1558 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1559 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1560 HeapWord* top = sp->top();
1561
1562 if (dense_prefix_addr == top) {
1563 continue;
1564 }
1565
1566 size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1567 size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1568 size_t start_region;
1569 size_t end_region;
1570 split_regions_for_worker(dense_prefix_region, top_region,
1571 worker_id, _num_workers,
1572 &start_region, &end_region);
1573 for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1574 RegionData* region_ptr = _summary_data.region(cur_region);
1575 size_t live_words = region_ptr->partial_obj_size();
1576
1577 if (live_words == ParallelCompactData::RegionSize) {
1578 // No obj-start
1579 continue;
1580 }
1581
1582 HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1583 HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1584
1585 HeapWord* cur_addr = region_start + live_words;
1586
1587 HeapWord* destination = region_ptr->destination();
1588 while (cur_addr < region_end) {
1589 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, region_end);
1590 if (cur_addr >= region_end) {
1591 break;
1592 }
1593 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1594 HeapWord* new_addr = destination + live_words;
1595 oop obj = cast_to_oop(cur_addr);
1596 if (new_addr != cur_addr) {
1597 cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1598 obj->forward_to(cast_to_oop(new_addr));
1599 }
1600 size_t obj_size = obj->size();
1601 live_words += obj_size;
1602 cur_addr += obj_size;
1603 }
1604 }
1605 }
1606 }
1607 } task(nworkers);
1608
1609 ParallelScavengeHeap::heap()->workers().run_task(&task);
1610 debug_only(verify_forward();)
1611 }
1612
1613 #ifdef ASSERT
1614 void PSParallelCompact::verify_forward() {
1615 HeapWord* old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1616 RegionData* old_region = _summary_data.region(_summary_data.addr_to_region_idx(old_dense_prefix_addr));
1617 HeapWord* bump_ptr = old_region->partial_obj_size() != 0
1618 ? old_dense_prefix_addr + old_region->partial_obj_size()
1619 : old_dense_prefix_addr;
1620 SpaceId bump_ptr_space = old_space_id;
1621
1622 for (uint id = old_space_id; id < last_space_id; ++id) {
1623 MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1624 HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1625 HeapWord* top = sp->top();
1626 HeapWord* cur_addr = dense_prefix_addr;
1627
1628 while (cur_addr < top) {
1629 cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1630 if (cur_addr >= top) {
1631 break;
1632 }
1633 assert(mark_bitmap()->is_marked(cur_addr), "inv");
1634 // Move to the space containing cur_addr
1635 if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1636 bump_ptr = space(space_id(cur_addr))->bottom();
1637 bump_ptr_space = space_id(bump_ptr);
1638 }
1639 oop obj = cast_to_oop(cur_addr);
1640 if (cur_addr != bump_ptr) {
1641 assert(obj->forwardee() == cast_to_oop(bump_ptr), "inv");
1642 }
1643 bump_ptr += obj->size();
1644 cur_addr += obj->size();
1645 }
1646 }
1647 }
1648 #endif
1649
1650 // Helper class to print 8 region numbers per line and then print the total at the end.
1651 class FillableRegionLogger : public StackObj {
1652 private:
1653 Log(gc, compaction) log;
1654 static const int LineLength = 8;
1655 size_t _regions[LineLength];
1656 int _next_index;
1657 bool _enabled;
1658 size_t _total_regions;
1659 public:
1660 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1661 ~FillableRegionLogger() {
1662 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
1663 }
1664
1665 void print_line() {
1666 if (!_enabled || _next_index == 0) {
1667 return;
1668 }
1669 FormatBuffer<> line("Fillable: ");
1670 for (int i = 0; i < _next_index; i++) {
1671 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
1672 }
1673 log.trace("%s", line.buffer());
1674 _next_index = 0;
1675 }
1676
1677 void handle(size_t region) {
1678 if (!_enabled) {
1679 return;
1680 }
1681 _regions[_next_index++] = region;
1682 if (_next_index == LineLength) {
1683 print_line();
1684 }
1685 _total_regions++;
1686 }
1687 };
1688
1689 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1690 {
1691 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1692
1693 // Find the threads that are active
1694 uint worker_id = 0;
1695
1696 // Find all regions that are available (can be filled immediately) and
1697 // distribute them to the thread stacks. The iteration is done in reverse
1698 // order (high to low) so the regions will be removed in ascending order.
1699
1700 const ParallelCompactData& sd = PSParallelCompact::summary_data();
1701
1702 // id + 1 is used to test termination so unsigned can
1703 // be used with an old_space_id == 0.
1704 FillableRegionLogger region_logger;
1705 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
1706 SpaceInfo* const space_info = _space_info + id;
1707 HeapWord* const new_top = space_info->new_top();
1708
1709 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1710 const size_t end_region =
1711 sd.addr_to_region_idx(sd.region_align_up(new_top));
1712
1713 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1714 if (sd.region(cur)->claim_unsafe()) {
1715 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1716 bool result = sd.region(cur)->mark_normal();
1717 assert(result, "Must succeed at this point.");
1718 cm->region_stack()->push(cur);
1719 region_logger.handle(cur);
1720 // Assign regions to tasks in round-robin fashion.
1721 if (++worker_id == parallel_gc_threads) {
1722 worker_id = 0;
1723 }
1724 }
1725 }
1726 region_logger.print_line();
1727 }
1728 }
1729
1730 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1731 assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1732
1733 ParCompactionManager* cm =
1734 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1735
1736 // Drain the stacks that have been preloaded with regions
1737 // that are ready to fill.
1738
1739 cm->drain_region_stacks();
1740
1741 guarantee(cm->region_stack()->is_empty(), "Not empty");
1742
1743 size_t region_index = 0;
1744
1745 while (true) {
1746 if (ParCompactionManager::steal(worker_id, region_index)) {
1747 PSParallelCompact::fill_and_update_region(cm, region_index);
1748 cm->drain_region_stacks();
1749 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1750 // Fill and update an unavailable region with the help of a shadow region
1751 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1752 cm->drain_region_stacks();
1753 } else {
1754 if (terminator->offer_termination()) {
1755 break;
1756 }
1757 // Go around again.
1758 }
1759 }
1760 }
1761
1762 class FillDensePrefixAndCompactionTask: public WorkerTask {
1763 uint _num_workers;
1764 TaskTerminator _terminator;
1765
1766 public:
1767 FillDensePrefixAndCompactionTask(uint active_workers) :
1768 WorkerTask("FillDensePrefixAndCompactionTask"),
1769 _num_workers(active_workers),
1770 _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1771 }
1772
1773 virtual void work(uint worker_id) {
1774 {
1775 auto start = Ticks::now();
1776 PSParallelCompact::fill_dead_objs_in_dense_prefix(worker_id, _num_workers);
1777 log_trace(gc, phases)("Fill dense prefix by worker %u: %.3f ms", worker_id, (Ticks::now() - start).seconds() * 1000);
1778 }
1779 compaction_with_stealing_work(&_terminator, worker_id);
1780 }
1781 };
1782
1783 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1784 #ifdef ASSERT
1785 {
1786 assert(start < end, "precondition");
1787 assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1788 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1789 if (start != bottom) {
1790 HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1791 HeapWord* after_obj = obj_start + cast_to_oop(obj_start)->size();
1792 assert(after_obj == start, "precondition");
1793 }
1794 }
1795 #endif
1796
1797 CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1798 HeapWord* addr = start;
1799 do {
1800 size_t size = cast_to_oop(addr)->size();
1801 start_array(old_space_id)->update_for_block(addr, addr + size);
1802 addr += size;
1803 } while (addr < end);
1804 }
1805
1806 void PSParallelCompact::fill_dead_objs_in_dense_prefix(uint worker_id, uint num_workers) {
1807 ParMarkBitMap* bitmap = mark_bitmap();
1808
1809 HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1810 HeapWord* const prefix_end = dense_prefix(old_space_id);
1811
1812 if (bottom == prefix_end) {
1813 return;
1814 }
1815
1816 size_t bottom_region = _summary_data.addr_to_region_idx(bottom);
1817 size_t prefix_end_region = _summary_data.addr_to_region_idx(prefix_end);
1818
1819 size_t start_region;
1820 size_t end_region;
1821 split_regions_for_worker(bottom_region, prefix_end_region,
1822 worker_id, num_workers,
1823 &start_region, &end_region);
1824
1825 if (start_region == end_region) {
1826 return;
1827 }
1828
1829 HeapWord* const start_addr = _summary_data.region_to_addr(start_region);
1830 HeapWord* const end_addr = _summary_data.region_to_addr(end_region);
1831
1832 // Skip live partial obj (if any) from previous region.
1833 HeapWord* cur_addr;
1834 RegionData* start_region_ptr = _summary_data.region(start_region);
1835 if (start_region_ptr->partial_obj_size() != 0) {
1836 HeapWord* partial_obj_start = start_region_ptr->partial_obj_addr();
1837 assert(bitmap->is_marked(partial_obj_start), "inv");
1838 cur_addr = partial_obj_start + cast_to_oop(partial_obj_start)->size();
1839 } else {
1840 cur_addr = start_addr;
1841 }
1842
1843 // end_addr is inclusive to handle regions starting with dead space.
1844 while (cur_addr <= end_addr) {
1845 // Use prefix_end to handle trailing obj in each worker region-chunk.
1846 HeapWord* live_start = bitmap->find_obj_beg(cur_addr, prefix_end);
1847 if (cur_addr != live_start) {
1848 // Only worker 0 handles proceeding dead space.
1849 if (cur_addr != start_addr || worker_id == 0) {
1850 fill_range_in_dense_prefix(cur_addr, live_start);
1851 }
1852 }
1853 if (live_start >= end_addr) {
1854 break;
1855 }
1856 assert(bitmap->is_marked(live_start), "inv");
1857 cur_addr = live_start + cast_to_oop(live_start)->size();
1858 }
1859 }
1860
1861 void PSParallelCompact::compact() {
1862 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1863
1864 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1865
1866 initialize_shadow_regions(active_gc_threads);
1867 prepare_region_draining_tasks(active_gc_threads);
1868
1869 {
1870 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1871
1872 FillDensePrefixAndCompactionTask task(active_gc_threads);
1873 ParallelScavengeHeap::heap()->workers().run_task(&task);
1874
1875 #ifdef ASSERT
1876 verify_filler_in_dense_prefix();
1877
1878 // Verify that all regions have been processed.
1879 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1880 verify_complete(SpaceId(id));
1881 }
1882 #endif
1883 }
1884 }
1885
1886 #ifdef ASSERT
1887 void PSParallelCompact::verify_filler_in_dense_prefix() {
1888 HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1889 HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1890 HeapWord* cur_addr = bottom;
1891 while (cur_addr < dense_prefix_end) {
1892 oop obj = cast_to_oop(cur_addr);
1893 oopDesc::verify(obj);
1894 if (!mark_bitmap()->is_marked(cur_addr)) {
1895 Klass* k = cast_to_oop(cur_addr)->klass_without_asserts();
1896 assert(k == Universe::fillerArrayKlass() || k == vmClasses::FillerObject_klass(), "inv");
1897 }
1898 cur_addr += obj->size();
1899 }
1900 }
1901
1902 void PSParallelCompact::verify_complete(SpaceId space_id) {
1903 // All Regions served as compaction targets, from dense_prefix() to
1904 // new_top(), should be marked as filled and all Regions between new_top()
1905 // and top() should be available (i.e., should have been emptied).
1906 ParallelCompactData& sd = summary_data();
1907 SpaceInfo si = _space_info[space_id];
1908 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1909 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1910 const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1911 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1912 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1913
1914 size_t cur_region;
1915 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1916 const RegionData* const c = sd.region(cur_region);
1917 if (!c->completed()) {
1918 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
1919 cur_region, c->destination_count());
1920 }
1921 }
1922
1923 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1924 const RegionData* const c = sd.region(cur_region);
1925 if (!c->available()) {
1926 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
1927 cur_region, c->destination_count());
1928 }
1929 }
1930 }
1931 #endif // #ifdef ASSERT
1932
1933 // Return the SpaceId for the space containing addr. If addr is not in the
1934 // heap, last_space_id is returned. In debug mode it expects the address to be
1935 // in the heap and asserts such.
1936 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1937 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1938
1939 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1940 if (_space_info[id].space()->contains(addr)) {
1941 return SpaceId(id);
1942 }
1943 }
1944
1945 assert(false, "no space contains the addr");
1946 return last_space_id;
1947 }
1948
1949 // Skip over count live words starting from beg, and return the address of the
1950 // next live word. Unless marked, the word corresponding to beg is assumed to
1951 // be dead. Callers must either ensure beg does not correspond to the middle of
1952 // an object, or account for those live words in some other way. Callers must
1953 // also ensure that there are enough live words in the range [beg, end) to skip.
1954 HeapWord*
1955 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1956 {
1957 assert(count > 0, "sanity");
1958
1959 ParMarkBitMap* m = mark_bitmap();
1960 HeapWord* cur_addr = beg;
1961 while (true) {
1962 cur_addr = m->find_obj_beg(cur_addr, end);
1963 assert(cur_addr < end, "inv");
1964 size_t obj_size = cast_to_oop(cur_addr)->size();
1965 // Strictly greater-than
1966 if (obj_size > count) {
1967 return cur_addr + count;
1968 }
1969 count -= obj_size;
1970 cur_addr += obj_size;
1971 }
1972 }
1973
1974 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1975 SpaceId src_space_id,
1976 size_t src_region_idx)
1977 {
1978 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
1979
1980 const SplitInfo& split_info = _space_info[src_space_id].split_info();
1981 if (split_info.dest_region_addr() == dest_addr) {
1982 // The partial object ending at the split point contains the first word to
1983 // be copied to dest_addr.
1984 return split_info.first_src_addr();
1985 }
1986
1987 const ParallelCompactData& sd = summary_data();
1988 ParMarkBitMap* const bitmap = mark_bitmap();
1989 const size_t RegionSize = ParallelCompactData::RegionSize;
1990
1991 assert(sd.is_region_aligned(dest_addr), "not aligned");
1992 const RegionData* const src_region_ptr = sd.region(src_region_idx);
1993 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1994 HeapWord* const src_region_destination = src_region_ptr->destination();
1995
1996 assert(dest_addr >= src_region_destination, "wrong src region");
1997 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1998
1999 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2000 HeapWord* const src_region_end = src_region_beg + RegionSize;
2001
2002 HeapWord* addr = src_region_beg;
2003 if (dest_addr == src_region_destination) {
2004 // Return the first live word in the source region.
2005 if (partial_obj_size == 0) {
2006 addr = bitmap->find_obj_beg(addr, src_region_end);
2007 assert(addr < src_region_end, "no objects start in src region");
2008 }
2009 return addr;
2010 }
2011
2012 // Must skip some live data.
2013 size_t words_to_skip = dest_addr - src_region_destination;
2014 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2015
2016 if (partial_obj_size >= words_to_skip) {
2017 // All the live words to skip are part of the partial object.
2018 addr += words_to_skip;
2019 if (partial_obj_size == words_to_skip) {
2020 // Find the first live word past the partial object.
2021 addr = bitmap->find_obj_beg(addr, src_region_end);
2022 assert(addr < src_region_end, "wrong src region");
2023 }
2024 return addr;
2025 }
2026
2027 // Skip over the partial object (if any).
2028 if (partial_obj_size != 0) {
2029 words_to_skip -= partial_obj_size;
2030 addr += partial_obj_size;
2031 }
2032
2033 // Skip over live words due to objects that start in the region.
2034 addr = skip_live_words(addr, src_region_end, words_to_skip);
2035 assert(addr < src_region_end, "wrong src region");
2036 return addr;
2037 }
2038
2039 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2040 SpaceId src_space_id,
2041 size_t beg_region,
2042 HeapWord* end_addr)
2043 {
2044 ParallelCompactData& sd = summary_data();
2045
2046 #ifdef ASSERT
2047 MutableSpace* const src_space = _space_info[src_space_id].space();
2048 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2049 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2050 "src_space_id does not match beg_addr");
2051 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2052 "src_space_id does not match end_addr");
2053 #endif // #ifdef ASSERT
2054
2055 RegionData* const beg = sd.region(beg_region);
2056 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2057
2058 // Regions up to new_top() are enqueued if they become available.
2059 HeapWord* const new_top = _space_info[src_space_id].new_top();
2060 RegionData* const enqueue_end =
2061 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2062
2063 for (RegionData* cur = beg; cur < end; ++cur) {
2064 assert(cur->data_size() > 0, "region must have live data");
2065 cur->decrement_destination_count();
2066 if (cur < enqueue_end && cur->available() && cur->claim()) {
2067 if (cur->mark_normal()) {
2068 cm->push_region(sd.region(cur));
2069 } else if (cur->mark_copied()) {
2070 // Try to copy the content of the shadow region back to its corresponding
2071 // heap region if the shadow region is filled. Otherwise, the GC thread
2072 // fills the shadow region will copy the data back (see
2073 // MoveAndUpdateShadowClosure::complete_region).
2074 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2075 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2076 cur->set_completed();
2077 }
2078 }
2079 }
2080 }
2081
2082 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2083 SpaceId& src_space_id,
2084 HeapWord*& src_space_top,
2085 HeapWord* end_addr)
2086 {
2087 typedef ParallelCompactData::RegionData RegionData;
2088
2089 ParallelCompactData& sd = PSParallelCompact::summary_data();
2090 const size_t region_size = ParallelCompactData::RegionSize;
2091
2092 size_t src_region_idx = 0;
2093
2094 // Skip empty regions (if any) up to the top of the space.
2095 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2096 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2097 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2098 const RegionData* const top_region_ptr =
2099 sd.addr_to_region_ptr(top_aligned_up);
2100 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2101 ++src_region_ptr;
2102 }
2103
2104 if (src_region_ptr < top_region_ptr) {
2105 // The next source region is in the current space. Update src_region_idx
2106 // and the source address to match src_region_ptr.
2107 src_region_idx = sd.region(src_region_ptr);
2108 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2109 if (src_region_addr > closure.source()) {
2110 closure.set_source(src_region_addr);
2111 }
2112 return src_region_idx;
2113 }
2114
2115 // Switch to a new source space and find the first non-empty region.
2116 unsigned int space_id = src_space_id + 1;
2117 assert(space_id < last_space_id, "not enough spaces");
2118
2119 HeapWord* const destination = closure.destination();
2120
2121 do {
2122 MutableSpace* space = _space_info[space_id].space();
2123 HeapWord* const bottom = space->bottom();
2124 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2125
2126 // Iterate over the spaces that do not compact into themselves.
2127 if (bottom_cp->destination() != bottom) {
2128 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2129 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2130
2131 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2132 if (src_cp->live_obj_size() > 0) {
2133 // Found it.
2134 assert(src_cp->destination() == destination,
2135 "first live obj in the space must match the destination");
2136 assert(src_cp->partial_obj_size() == 0,
2137 "a space cannot begin with a partial obj");
2138
2139 src_space_id = SpaceId(space_id);
2140 src_space_top = space->top();
2141 const size_t src_region_idx = sd.region(src_cp);
2142 closure.set_source(sd.region_to_addr(src_region_idx));
2143 return src_region_idx;
2144 } else {
2145 assert(src_cp->data_size() == 0, "sanity");
2146 }
2147 }
2148 }
2149 } while (++space_id < last_space_id);
2150
2151 assert(false, "no source region was found");
2152 return 0;
2153 }
2154
2155 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2156 ParallelCompactData& sd = summary_data();
2157 assert(sd.is_region_aligned(region_start_addr), "precondition");
2158
2159 // Use per-region partial_obj_size to locate the end of the obj, that extends to region_start_addr.
2160 SplitInfo& split_info = _space_info[space_id(region_start_addr)].split_info();
2161 size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2162 size_t end_region_idx = sd.region_count();
2163 size_t accumulated_size = 0;
2164 for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2165 if (split_info.is_split(region_idx)) {
2166 accumulated_size += split_info.partial_obj_size();
2167 break;
2168 }
2169 size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2170 accumulated_size += cur_partial_obj_size;
2171 if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2172 break;
2173 }
2174 }
2175 return region_start_addr + accumulated_size;
2176 }
2177
2178 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2179 {
2180 ParMarkBitMap* const bitmap = mark_bitmap();
2181 ParallelCompactData& sd = summary_data();
2182 RegionData* const region_ptr = sd.region(region_idx);
2183
2184 // Get the source region and related info.
2185 size_t src_region_idx = region_ptr->source_region();
2186 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2187 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2188 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2189
2190 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2191
2192 // Adjust src_region_idx to prepare for decrementing destination counts (the
2193 // destination count is not decremented when a region is copied to itself).
2194 if (src_region_idx == region_idx) {
2195 src_region_idx += 1;
2196 }
2197
2198 if (bitmap->is_unmarked(closure.source())) {
2199 // The first source word is in the middle of an object; copy the remainder
2200 // of the object or as much as will fit. The fact that pointer updates were
2201 // deferred will be noted when the object header is processed.
2202 HeapWord* const old_src_addr = closure.source();
2203 {
2204 HeapWord* region_start = sd.region_align_down(closure.source());
2205 HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2206 HeapWord* obj_end;
2207 if (bitmap->is_marked(obj_start)) {
2208 HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2209 HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2210 ? nullptr
2211 : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2212 if (partial_obj_start == obj_start) {
2213 // This obj extends to next region.
2214 obj_end = partial_obj_end(next_region_start);
2215 } else {
2216 // Completely contained in this region; safe to use size().
2217 obj_end = obj_start + cast_to_oop(obj_start)->size();
2218 }
2219 } else {
2220 // This obj extends to current region.
2221 obj_end = partial_obj_end(region_start);
2222 }
2223 size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2224 closure.copy_partial_obj(partial_obj_size);
2225 }
2226
2227 if (closure.is_full()) {
2228 decrement_destination_counts(cm, src_space_id, src_region_idx,
2229 closure.source());
2230 closure.complete_region(dest_addr, region_ptr);
2231 return;
2232 }
2233
2234 HeapWord* const end_addr = sd.region_align_down(closure.source());
2235 if (sd.region_align_down(old_src_addr) != end_addr) {
2236 // The partial object was copied from more than one source region.
2237 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2238
2239 // Move to the next source region, possibly switching spaces as well. All
2240 // args except end_addr may be modified.
2241 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2242 end_addr);
2243 }
2244 }
2245
2246 do {
2247 HeapWord* cur_addr = closure.source();
2248 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2249 src_space_top);
2250 HeapWord* partial_obj_start = (end_addr == src_space_top)
2251 ? nullptr
2252 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2253 // apply closure on objs inside [cur_addr, end_addr)
2254 do {
2255 cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2256 if (cur_addr == end_addr) {
2257 break;
2258 }
2259 size_t obj_size;
2260 if (partial_obj_start == cur_addr) {
2261 obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2262 } else {
2263 // This obj doesn't extend into next region; size() is safe to use.
2264 obj_size = cast_to_oop(cur_addr)->size();
2265 }
2266 closure.do_addr(cur_addr, obj_size);
2267 cur_addr += obj_size;
2268 } while (cur_addr < end_addr && !closure.is_full());
2269
2270 if (closure.is_full()) {
2271 decrement_destination_counts(cm, src_space_id, src_region_idx,
2272 closure.source());
2273 closure.complete_region(dest_addr, region_ptr);
2274 return;
2275 }
2276
2277 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2278
2279 // Move to the next source region, possibly switching spaces as well. All
2280 // args except end_addr may be modified.
2281 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2282 end_addr);
2283 } while (true);
2284 }
2285
2286 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2287 {
2288 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2289 fill_region(cm, cl, region_idx);
2290 }
2291
2292 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2293 {
2294 // Get a shadow region first
2295 ParallelCompactData& sd = summary_data();
2296 RegionData* const region_ptr = sd.region(region_idx);
2297 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2298 // The InvalidShadow return value indicates the corresponding heap region is available,
2299 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2300 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2301 if (shadow_region == ParCompactionManager::InvalidShadow) {
2302 MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2303 region_ptr->shadow_to_normal();
2304 return fill_region(cm, cl, region_idx);
2305 } else {
2306 MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2307 return fill_region(cm, cl, region_idx);
2308 }
2309 }
2310
2311 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2312 {
2313 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2314 }
2315
2316 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
2317 {
2318 size_t next = cm->next_shadow_region();
2319 ParallelCompactData& sd = summary_data();
2320 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2321 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2322
2323 while (next < old_new_top) {
2324 if (sd.region(next)->mark_shadow()) {
2325 region_idx = next;
2326 return true;
2327 }
2328 next = cm->move_next_shadow_region_by(active_gc_threads);
2329 }
2330
2331 return false;
2332 }
2333
2334 // The shadow region is an optimization to address region dependencies in full GC. The basic
2335 // idea is making more regions available by temporally storing their live objects in empty
2336 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2337 // GC threads need not wait destination regions to be available before processing sources.
2338 //
2339 // A typical workflow would be:
2340 // After draining its own stack and failing to steal from others, a GC worker would pick an
2341 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2342 // the shadow region by copying live objects from source regions of the unavailable one. Once
2343 // the unavailable region becomes available, the data in the shadow region will be copied back.
2344 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2345 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2346 {
2347 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2348
2349 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2350 SpaceInfo* const space_info = _space_info + id;
2351 MutableSpace* const space = space_info->space();
2352
2353 const size_t beg_region =
2354 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2355 const size_t end_region =
2356 sd.addr_to_region_idx(sd.region_align_down(space->end()));
2357
2358 for (size_t cur = beg_region; cur < end_region; ++cur) {
2359 ParCompactionManager::push_shadow_region(cur);
2360 }
2361 }
2362
2363 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2364 for (uint i = 0; i < parallel_gc_threads; i++) {
2365 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2366 cm->set_next_shadow_region(beg_region + i);
2367 }
2368 }
2369
2370 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2371 {
2372 size_t words = MIN2(partial_obj_size, words_remaining());
2373
2374 // This test is necessary; if omitted, the pointer updates to a partial object
2375 // that crosses the dense prefix boundary could be overwritten.
2376 if (source() != copy_destination()) {
2377 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2378 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2379 }
2380 update_state(words);
2381 }
2382
2383 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2384 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2385 region_ptr->set_completed();
2386 }
2387
2388 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2389 assert(destination() != nullptr, "sanity");
2390 _source = addr;
2391
2392 // The start_array must be updated even if the object is not moving.
2393 if (_start_array != nullptr) {
2394 _start_array->update_for_block(destination(), destination() + words);
2395 }
2396
2397 // Avoid overflow
2398 words = MIN2(words, words_remaining());
2399 assert(words > 0, "inv");
2400
2401 if (copy_destination() != source()) {
2402 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2403 assert(source() != destination(), "inv");
2404 assert(cast_to_oop(source())->is_forwarded(), "inv");
2405 assert(cast_to_oop(source())->forwardee() == cast_to_oop(destination()), "inv");
2406 Copy::aligned_conjoint_words(source(), copy_destination(), words);
2407 cast_to_oop(copy_destination())->init_mark();
2408 }
2409
2410 update_state(words);
2411 }
2412
2413 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2414 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2415 // Record the shadow region index
2416 region_ptr->set_shadow_region(_shadow);
2417 // Mark the shadow region as filled to indicate the data is ready to be
2418 // copied back
2419 region_ptr->mark_filled();
2420 // Try to copy the content of the shadow region back to its corresponding
2421 // heap region if available; the GC thread that decreases the destination
2422 // count to zero will do the copying otherwise (see
2423 // PSParallelCompact::decrement_destination_counts).
2424 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2425 region_ptr->set_completed();
2426 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2427 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2428 }
2429 }
2430