1 /*
2 * Copyright (c) 2005, 2023, 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/parallelArguments.hpp"
34 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
35 #include "gc/parallel/parMarkBitMap.inline.hpp"
36 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
37 #include "gc/parallel/psCompactionManager.inline.hpp"
38 #include "gc/parallel/psOldGen.hpp"
39 #include "gc/parallel/psParallelCompact.inline.hpp"
40 #include "gc/parallel/psPromotionManager.inline.hpp"
41 #include "gc/parallel/psRootType.hpp"
42 #include "gc/parallel/psScavenge.hpp"
43 #include "gc/parallel/psStringDedup.hpp"
44 #include "gc/parallel/psYoungGen.hpp"
45 #include "gc/shared/classUnloadingContext.hpp"
46 #include "gc/shared/gcCause.hpp"
47 #include "gc/shared/gcHeapSummary.hpp"
48 #include "gc/shared/gcId.hpp"
49 #include "gc/shared/gcLocker.hpp"
50 #include "gc/shared/gcTimer.hpp"
51 #include "gc/shared/gcTrace.hpp"
52 #include "gc/shared/gcTraceTime.inline.hpp"
53 #include "gc/shared/isGCActiveMark.hpp"
54 #include "gc/shared/oopStorage.inline.hpp"
55 #include "gc/shared/oopStorageSet.inline.hpp"
56 #include "gc/shared/oopStorageSetParState.inline.hpp"
57 #include "gc/shared/referencePolicy.hpp"
58 #include "gc/shared/referenceProcessor.hpp"
59 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
60 #include "gc/shared/spaceDecorator.inline.hpp"
61 #include "gc/shared/taskTerminator.hpp"
62 #include "gc/shared/weakProcessor.inline.hpp"
63 #include "gc/shared/workerPolicy.hpp"
64 #include "gc/shared/workerThread.hpp"
65 #include "gc/shared/workerUtils.hpp"
66 #include "logging/log.hpp"
67 #include "memory/iterator.inline.hpp"
68 #include "memory/metaspaceUtils.hpp"
69 #include "memory/resourceArea.hpp"
70 #include "memory/universe.hpp"
71 #include "nmt/memTracker.hpp"
72 #include "oops/access.inline.hpp"
73 #include "oops/flatArrayKlass.inline.hpp"
74 #include "oops/instanceClassLoaderKlass.inline.hpp"
75 #include "oops/instanceKlass.inline.hpp"
76 #include "oops/instanceMirrorKlass.inline.hpp"
77 #include "oops/methodData.hpp"
78 #include "oops/objArrayKlass.inline.hpp"
79 #include "oops/oop.inline.hpp"
80 #include "runtime/atomic.hpp"
81 #include "runtime/handles.inline.hpp"
82 #include "runtime/java.hpp"
83 #include "runtime/safepoint.hpp"
84 #include "runtime/threads.hpp"
85 #include "runtime/vmThread.hpp"
86 #include "services/memoryService.hpp"
87 #include "utilities/align.hpp"
88 #include "utilities/debug.hpp"
89 #include "utilities/events.hpp"
90 #include "utilities/formatBuffer.hpp"
91 #include "utilities/macros.hpp"
92 #include "utilities/stack.inline.hpp"
93 #if INCLUDE_JVMCI
94 #include "jvmci/jvmci.hpp"
95 #endif
96
97 #include <math.h>
98
99 // All sizes are in HeapWords.
100 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
101 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
102 const size_t ParallelCompactData::RegionSizeBytes =
103 RegionSize << LogHeapWordSize;
104 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
105 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
106 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
107
108 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
109 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
110 const size_t ParallelCompactData::BlockSizeBytes =
111 BlockSize << LogHeapWordSize;
112 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
113 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
114 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
115
116 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
117 const size_t ParallelCompactData::Log2BlocksPerRegion =
118 Log2RegionSize - Log2BlockSize;
119
120 const ParallelCompactData::RegionData::region_sz_t
121 ParallelCompactData::RegionData::dc_shift = 27;
122
123 const ParallelCompactData::RegionData::region_sz_t
124 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
125
126 const ParallelCompactData::RegionData::region_sz_t
127 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
128
129 const ParallelCompactData::RegionData::region_sz_t
130 ParallelCompactData::RegionData::los_mask = ~dc_mask;
131
132 const ParallelCompactData::RegionData::region_sz_t
133 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
134
135 const ParallelCompactData::RegionData::region_sz_t
136 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
137
138 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
139
140 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
141 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
142
143 double PSParallelCompact::_dwl_mean;
144 double PSParallelCompact::_dwl_std_dev;
145 double PSParallelCompact::_dwl_first_term;
146 double PSParallelCompact::_dwl_adjustment;
147 #ifdef ASSERT
148 bool PSParallelCompact::_dwl_initialized = false;
149 #endif // #ifdef ASSERT
150
151 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
152 HeapWord* destination)
153 {
154 assert(src_region_idx != 0, "invalid src_region_idx");
155 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
156 assert(destination != nullptr, "invalid destination argument");
157
158 _src_region_idx = src_region_idx;
159 _partial_obj_size = partial_obj_size;
160 _destination = destination;
161
162 // These fields may not be updated below, so make sure they're clear.
163 assert(_dest_region_addr == nullptr, "should have been cleared");
164 assert(_first_src_addr == nullptr, "should have been cleared");
165
166 // Determine the number of destination regions for the partial object.
167 HeapWord* const last_word = destination + partial_obj_size - 1;
168 const ParallelCompactData& sd = PSParallelCompact::summary_data();
169 HeapWord* const beg_region_addr = sd.region_align_down(destination);
170 HeapWord* const end_region_addr = sd.region_align_down(last_word);
171
172 if (beg_region_addr == end_region_addr) {
173 // One destination region.
174 _destination_count = 1;
175 if (end_region_addr == destination) {
176 // The destination falls on a region boundary, thus the first word of the
177 // partial object will be the first word copied to the destination region.
178 _dest_region_addr = end_region_addr;
179 _first_src_addr = sd.region_to_addr(src_region_idx);
180 }
181 } else {
182 // Two destination regions. When copied, the partial object will cross a
183 // destination region boundary, so a word somewhere within the partial
184 // object will be the first word copied to the second destination region.
185 _destination_count = 2;
186 _dest_region_addr = end_region_addr;
187 const size_t ofs = pointer_delta(end_region_addr, destination);
188 assert(ofs < _partial_obj_size, "sanity");
189 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
190 }
191 }
192
193 void SplitInfo::clear()
194 {
195 _src_region_idx = 0;
196 _partial_obj_size = 0;
197 _destination = nullptr;
198 _destination_count = 0;
199 _dest_region_addr = nullptr;
200 _first_src_addr = nullptr;
201 assert(!is_valid(), "sanity");
202 }
203
204 #ifdef ASSERT
205 void SplitInfo::verify_clear()
206 {
207 assert(_src_region_idx == 0, "not clear");
208 assert(_partial_obj_size == 0, "not clear");
209 assert(_destination == nullptr, "not clear");
210 assert(_destination_count == 0, "not clear");
211 assert(_dest_region_addr == nullptr, "not clear");
212 assert(_first_src_addr == nullptr, "not clear");
213 }
214 #endif // #ifdef ASSERT
215
216
217 void PSParallelCompact::print_on_error(outputStream* st) {
218 _mark_bitmap.print_on_error(st);
219 }
220
221 #ifndef PRODUCT
222 const char* PSParallelCompact::space_names[] = {
223 "old ", "eden", "from", "to "
224 };
225
226 void PSParallelCompact::print_region_ranges() {
227 if (!log_develop_is_enabled(Trace, gc, compaction)) {
228 return;
229 }
230 Log(gc, compaction) log;
231 ResourceMark rm;
232 LogStream ls(log.trace());
233 Universe::print_on(&ls);
234 log.trace("space bottom top end new_top");
235 log.trace("------ ---------- ---------- ---------- ----------");
236
237 for (unsigned int id = 0; id < last_space_id; ++id) {
238 const MutableSpace* space = _space_info[id].space();
239 log.trace("%u %s "
240 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
241 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
242 id, space_names[id],
243 summary_data().addr_to_region_idx(space->bottom()),
244 summary_data().addr_to_region_idx(space->top()),
245 summary_data().addr_to_region_idx(space->end()),
246 summary_data().addr_to_region_idx(_space_info[id].new_top()));
247 }
248 }
249
250 void
251 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
252 {
253 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
254 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
255
256 ParallelCompactData& sd = PSParallelCompact::summary_data();
257 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
258 log_develop_trace(gc, compaction)(
259 REGION_IDX_FORMAT " " PTR_FORMAT " "
260 REGION_IDX_FORMAT " " PTR_FORMAT " "
261 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
262 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
263 i, p2i(c->data_location()), dci, p2i(c->destination()),
264 c->partial_obj_size(), c->live_obj_size(),
265 c->data_size(), c->source_region(), c->destination_count());
266
267 #undef REGION_IDX_FORMAT
268 #undef REGION_DATA_FORMAT
269 }
270
271 void
272 print_generic_summary_data(ParallelCompactData& summary_data,
273 HeapWord* const beg_addr,
274 HeapWord* const end_addr)
275 {
276 size_t total_words = 0;
277 size_t i = summary_data.addr_to_region_idx(beg_addr);
278 const size_t last = summary_data.addr_to_region_idx(end_addr);
279 HeapWord* pdest = 0;
280
281 while (i < last) {
282 ParallelCompactData::RegionData* c = summary_data.region(i);
283 if (c->data_size() != 0 || c->destination() != pdest) {
284 print_generic_summary_region(i, c);
285 total_words += c->data_size();
286 pdest = c->destination();
287 }
288 ++i;
289 }
290
291 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
292 }
293
294 void
295 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
296 HeapWord* const beg_addr,
297 HeapWord* const end_addr) {
298 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
299 }
300
301 void
302 print_generic_summary_data(ParallelCompactData& summary_data,
303 SpaceInfo* space_info)
304 {
305 if (!log_develop_is_enabled(Trace, gc, compaction)) {
306 return;
307 }
308
309 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
310 const MutableSpace* space = space_info[id].space();
311 print_generic_summary_data(summary_data, space->bottom(),
312 MAX2(space->top(), space_info[id].new_top()));
313 }
314 }
315
316 void
317 print_initial_summary_data(ParallelCompactData& summary_data,
318 const MutableSpace* space) {
319 if (space->top() == space->bottom()) {
320 return;
321 }
322
323 const size_t region_size = ParallelCompactData::RegionSize;
324 typedef ParallelCompactData::RegionData RegionData;
325 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
326 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
327 const RegionData* c = summary_data.region(end_region - 1);
328 HeapWord* end_addr = c->destination() + c->data_size();
329 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
330
331 // Print (and count) the full regions at the beginning of the space.
332 size_t full_region_count = 0;
333 size_t i = summary_data.addr_to_region_idx(space->bottom());
334 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
335 ParallelCompactData::RegionData* c = summary_data.region(i);
336 log_develop_trace(gc, compaction)(
337 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
338 i, p2i(c->destination()),
339 c->partial_obj_size(), c->live_obj_size(),
340 c->data_size(), c->source_region(), c->destination_count());
341 ++full_region_count;
342 ++i;
343 }
344
345 size_t live_to_right = live_in_space - full_region_count * region_size;
346
347 double max_reclaimed_ratio = 0.0;
348 size_t max_reclaimed_ratio_region = 0;
349 size_t max_dead_to_right = 0;
350 size_t max_live_to_right = 0;
351
352 // Print the 'reclaimed ratio' for regions while there is something live in
353 // the region or to the right of it. The remaining regions are empty (and
354 // uninteresting), and computing the ratio will result in division by 0.
355 while (i < end_region && live_to_right > 0) {
356 c = summary_data.region(i);
357 HeapWord* const region_addr = summary_data.region_to_addr(i);
358 const size_t used_to_right = pointer_delta(space->top(), region_addr);
359 const size_t dead_to_right = used_to_right - live_to_right;
360 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
361
362 if (reclaimed_ratio > max_reclaimed_ratio) {
363 max_reclaimed_ratio = reclaimed_ratio;
364 max_reclaimed_ratio_region = i;
365 max_dead_to_right = dead_to_right;
366 max_live_to_right = live_to_right;
367 }
368
369 ParallelCompactData::RegionData* c = summary_data.region(i);
370 log_develop_trace(gc, compaction)(
371 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
372 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
373 i, p2i(c->destination()),
374 c->partial_obj_size(), c->live_obj_size(),
375 c->data_size(), c->source_region(), c->destination_count(),
376 reclaimed_ratio, dead_to_right, live_to_right);
377
378
379 live_to_right -= c->data_size();
380 ++i;
381 }
382
383 // Any remaining regions are empty. Print one more if there is one.
384 if (i < end_region) {
385 ParallelCompactData::RegionData* c = summary_data.region(i);
386 log_develop_trace(gc, compaction)(
387 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
388 i, p2i(c->destination()),
389 c->partial_obj_size(), c->live_obj_size(),
390 c->data_size(), c->source_region(), c->destination_count());
391 }
392
393 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
394 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
395 }
396
397 void
398 print_initial_summary_data(ParallelCompactData& summary_data,
399 SpaceInfo* space_info) {
400 if (!log_develop_is_enabled(Trace, gc, compaction)) {
401 return;
402 }
403
404 unsigned int id = PSParallelCompact::old_space_id;
405 const MutableSpace* space;
406 do {
407 space = space_info[id].space();
408 print_initial_summary_data(summary_data, space);
409 } while (++id < PSParallelCompact::eden_space_id);
410
411 do {
412 space = space_info[id].space();
413 print_generic_summary_data(summary_data, space->bottom(), space->top());
414 } while (++id < PSParallelCompact::last_space_id);
415 }
416 #endif // #ifndef PRODUCT
417
418 ParallelCompactData::ParallelCompactData() :
419 _region_start(nullptr),
420 DEBUG_ONLY(_region_end(nullptr) COMMA)
421 _region_vspace(nullptr),
422 _reserved_byte_size(0),
423 _region_data(nullptr),
424 _region_count(0),
425 _block_vspace(nullptr),
426 _block_data(nullptr),
427 _block_count(0) {}
428
429 bool ParallelCompactData::initialize(MemRegion covered_region)
430 {
431 _region_start = covered_region.start();
432 const size_t region_size = covered_region.word_size();
433 DEBUG_ONLY(_region_end = _region_start + region_size;)
434
435 assert(region_align_down(_region_start) == _region_start,
436 "region start not aligned");
437
438 bool result = initialize_region_data(region_size) && initialize_block_data();
439 return result;
440 }
441
442 PSVirtualSpace*
443 ParallelCompactData::create_vspace(size_t count, size_t element_size)
444 {
445 const size_t raw_bytes = count * element_size;
446 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
447 const size_t granularity = os::vm_allocation_granularity();
448 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
449
450 const size_t rs_align = page_sz == os::vm_page_size() ? 0 :
451 MAX2(page_sz, granularity);
452 ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
453 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
454 rs.size(), page_sz);
455
456 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
457
458 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
459 if (vspace != 0) {
460 if (vspace->expand_by(_reserved_byte_size)) {
461 return vspace;
462 }
463 delete vspace;
464 // Release memory reserved in the space.
465 rs.release();
466 }
467
468 return 0;
469 }
470
471 bool ParallelCompactData::initialize_region_data(size_t region_size)
472 {
473 assert((region_size & RegionSizeOffsetMask) == 0,
474 "region size not a multiple of RegionSize");
475
476 const size_t count = region_size >> Log2RegionSize;
477 _region_vspace = create_vspace(count, sizeof(RegionData));
478 if (_region_vspace != 0) {
479 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
480 _region_count = count;
481 return true;
482 }
483 return false;
484 }
485
486 bool ParallelCompactData::initialize_block_data()
487 {
488 assert(_region_count != 0, "region data must be initialized first");
489 const size_t count = _region_count << Log2BlocksPerRegion;
490 _block_vspace = create_vspace(count, sizeof(BlockData));
491 if (_block_vspace != 0) {
492 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
493 _block_count = count;
494 return true;
495 }
496 return false;
497 }
498
499 void ParallelCompactData::clear()
500 {
501 memset(_region_data, 0, _region_vspace->committed_size());
502 memset(_block_data, 0, _block_vspace->committed_size());
503 }
504
505 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
506 assert(beg_region <= _region_count, "beg_region out of range");
507 assert(end_region <= _region_count, "end_region out of range");
508 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
509
510 const size_t region_cnt = end_region - beg_region;
511 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
512
513 const size_t beg_block = beg_region * BlocksPerRegion;
514 const size_t block_cnt = region_cnt * BlocksPerRegion;
515 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
516 }
517
518 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
519 {
520 const RegionData* cur_cp = region(region_idx);
521 const RegionData* const end_cp = region(region_count() - 1);
522
523 HeapWord* result = region_to_addr(region_idx);
524 if (cur_cp < end_cp) {
525 do {
526 result += cur_cp->partial_obj_size();
527 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
528 }
529 return result;
530 }
531
532 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
533 {
534 const size_t obj_ofs = pointer_delta(addr, _region_start);
535 const size_t beg_region = obj_ofs >> Log2RegionSize;
536 // end_region is inclusive
537 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
538
539 if (beg_region == end_region) {
540 // All in one region.
541 _region_data[beg_region].add_live_obj(len);
542 return;
543 }
544
545 // First region.
546 const size_t beg_ofs = region_offset(addr);
547 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
548
549 // Middle regions--completely spanned by this object.
550 for (size_t region = beg_region + 1; region < end_region; ++region) {
551 _region_data[region].set_partial_obj_size(RegionSize);
552 _region_data[region].set_partial_obj_addr(addr);
553 }
554
555 // Last region.
556 const size_t end_ofs = region_offset(addr + len - 1);
557 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
558 _region_data[end_region].set_partial_obj_addr(addr);
559 }
560
561 void
562 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
563 {
564 assert(is_region_aligned(beg), "not RegionSize aligned");
565 assert(is_region_aligned(end), "not RegionSize aligned");
566
567 size_t cur_region = addr_to_region_idx(beg);
568 const size_t end_region = addr_to_region_idx(end);
569 HeapWord* addr = beg;
570 while (cur_region < end_region) {
571 _region_data[cur_region].set_destination(addr);
572 _region_data[cur_region].set_destination_count(0);
573 _region_data[cur_region].set_source_region(cur_region);
574 _region_data[cur_region].set_data_location(addr);
575
576 // Update live_obj_size so the region appears completely full.
577 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
578 _region_data[cur_region].set_live_obj_size(live_size);
579
580 ++cur_region;
581 addr += RegionSize;
582 }
583 }
584
585 // Find the point at which a space can be split and, if necessary, record the
586 // split point.
587 //
588 // If the current src region (which overflowed the destination space) doesn't
589 // have a partial object, the split point is at the beginning of the current src
590 // region (an "easy" split, no extra bookkeeping required).
591 //
592 // If the current src region has a partial object, the split point is in the
593 // region where that partial object starts (call it the split_region). If
594 // split_region has a partial object, then the split point is just after that
595 // partial object (a "hard" split where we have to record the split data and
596 // zero the partial_obj_size field). With a "hard" split, we know that the
597 // partial_obj ends within split_region because the partial object that caused
598 // the overflow starts in split_region. If split_region doesn't have a partial
599 // obj, then the split is at the beginning of split_region (another "easy"
600 // split).
601 HeapWord*
602 ParallelCompactData::summarize_split_space(size_t src_region,
603 SplitInfo& split_info,
604 HeapWord* destination,
605 HeapWord* target_end,
606 HeapWord** target_next)
607 {
608 assert(destination <= target_end, "sanity");
609 assert(destination + _region_data[src_region].data_size() > target_end,
610 "region should not fit into target space");
611 assert(is_region_aligned(target_end), "sanity");
612
613 size_t split_region = src_region;
614 HeapWord* split_destination = destination;
615 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
616
617 if (destination + partial_obj_size > target_end) {
618 // The split point is just after the partial object (if any) in the
619 // src_region that contains the start of the object that overflowed the
620 // destination space.
621 //
622 // Find the start of the "overflow" object and set split_region to the
623 // region containing it.
624 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
625 split_region = addr_to_region_idx(overflow_obj);
626
627 // Clear the source_region field of all destination regions whose first word
628 // came from data after the split point (a non-null source_region field
629 // implies a region must be filled).
630 //
631 // An alternative to the simple loop below: clear during post_compact(),
632 // which uses memcpy instead of individual stores, and is easy to
633 // parallelize. (The downside is that it clears the entire RegionData
634 // object as opposed to just one field.)
635 //
636 // post_compact() would have to clear the summary data up to the highest
637 // address that was written during the summary phase, which would be
638 //
639 // max(top, max(new_top, clear_top))
640 //
641 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
642 // to target_end.
643 const RegionData* const sr = region(split_region);
644 const size_t beg_idx =
645 addr_to_region_idx(region_align_up(sr->destination() +
646 sr->partial_obj_size()));
647 const size_t end_idx = addr_to_region_idx(target_end);
648
649 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
650 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
651 _region_data[idx].set_source_region(0);
652 }
653
654 // Set split_destination and partial_obj_size to reflect the split region.
655 split_destination = sr->destination();
656 partial_obj_size = sr->partial_obj_size();
657 }
658
659 // The split is recorded only if a partial object extends onto the region.
660 if (partial_obj_size != 0) {
661 _region_data[split_region].set_partial_obj_size(0);
662 split_info.record(split_region, partial_obj_size, split_destination);
663 }
664
665 // Setup the continuation addresses.
666 *target_next = split_destination + partial_obj_size;
667 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
668
669 if (log_develop_is_enabled(Trace, gc, compaction)) {
670 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
671 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
672 split_type, p2i(source_next), split_region, partial_obj_size);
673 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
674 split_type, p2i(split_destination),
675 addr_to_region_idx(split_destination),
676 p2i(*target_next));
677
678 if (partial_obj_size != 0) {
679 HeapWord* const po_beg = split_info.destination();
680 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
681 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
682 split_type,
683 p2i(po_beg), addr_to_region_idx(po_beg),
684 p2i(po_end), addr_to_region_idx(po_end));
685 }
686 }
687
688 return source_next;
689 }
690
691 bool ParallelCompactData::summarize(SplitInfo& split_info,
692 HeapWord* source_beg, HeapWord* source_end,
693 HeapWord** source_next,
694 HeapWord* target_beg, HeapWord* target_end,
695 HeapWord** target_next)
696 {
697 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
698 log_develop_trace(gc, compaction)(
699 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
700 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
701 p2i(source_beg), p2i(source_end), p2i(source_next_val),
702 p2i(target_beg), p2i(target_end), p2i(*target_next));
703
704 size_t cur_region = addr_to_region_idx(source_beg);
705 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
706
707 HeapWord *dest_addr = target_beg;
708 while (cur_region < end_region) {
709 // The destination must be set even if the region has no data.
710 _region_data[cur_region].set_destination(dest_addr);
711
712 size_t words = _region_data[cur_region].data_size();
713 if (words > 0) {
714 // If cur_region does not fit entirely into the target space, find a point
715 // at which the source space can be 'split' so that part is copied to the
716 // target space and the rest is copied elsewhere.
717 if (dest_addr + words > target_end) {
718 assert(source_next != nullptr, "source_next is null when splitting");
719 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
720 target_end, target_next);
721 return false;
722 }
723
724 // Compute the destination_count for cur_region, and if necessary, update
725 // source_region for a destination region. The source_region field is
726 // updated if cur_region is the first (left-most) region to be copied to a
727 // destination region.
728 //
729 // The destination_count calculation is a bit subtle. A region that has
730 // data that compacts into itself does not count itself as a destination.
731 // This maintains the invariant that a zero count means the region is
732 // available and can be claimed and then filled.
733 uint destination_count = 0;
734 if (split_info.is_split(cur_region)) {
735 // The current region has been split: the partial object will be copied
736 // to one destination space and the remaining data will be copied to
737 // another destination space. Adjust the initial destination_count and,
738 // if necessary, set the source_region field if the partial object will
739 // cross a destination region boundary.
740 destination_count = split_info.destination_count();
741 if (destination_count == 2) {
742 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
743 _region_data[dest_idx].set_source_region(cur_region);
744 }
745 }
746
747 HeapWord* const last_addr = dest_addr + words - 1;
748 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
749 const size_t dest_region_2 = addr_to_region_idx(last_addr);
750
751 // Initially assume that the destination regions will be the same and
752 // adjust the value below if necessary. Under this assumption, if
753 // cur_region == dest_region_2, then cur_region will be compacted
754 // completely into itself.
755 destination_count += cur_region == dest_region_2 ? 0 : 1;
756 if (dest_region_1 != dest_region_2) {
757 // Destination regions differ; adjust destination_count.
758 destination_count += 1;
759 // Data from cur_region will be copied to the start of dest_region_2.
760 _region_data[dest_region_2].set_source_region(cur_region);
761 } else if (is_region_aligned(dest_addr)) {
762 // Data from cur_region will be copied to the start of the destination
763 // region.
764 _region_data[dest_region_1].set_source_region(cur_region);
765 }
766
767 _region_data[cur_region].set_destination_count(destination_count);
768 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
769 dest_addr += words;
770 }
771
772 ++cur_region;
773 }
774
775 *target_next = dest_addr;
776 return true;
777 }
778
779 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const {
780 assert(addr != nullptr, "Should detect null oop earlier");
781 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
782 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
783
784 // Region covering the object.
785 RegionData* const region_ptr = addr_to_region_ptr(addr);
786 HeapWord* result = region_ptr->destination();
787
788 // If the entire Region is live, the new location is region->destination + the
789 // offset of the object within in the Region.
790
791 // Run some performance tests to determine if this special case pays off. It
792 // is worth it for pointers into the dense prefix. If the optimization to
793 // avoid pointer updates in regions that only point to the dense prefix is
794 // ever implemented, this should be revisited.
795 if (region_ptr->data_size() == RegionSize) {
796 result += region_offset(addr);
797 return result;
798 }
799
800 // Otherwise, the new location is region->destination + block offset + the
801 // number of live words in the Block that are (a) to the left of addr and (b)
802 // due to objects that start in the Block.
803
804 // Fill in the block table if necessary. This is unsynchronized, so multiple
805 // threads may fill the block table for a region (harmless, since it is
806 // idempotent).
807 if (!region_ptr->blocks_filled()) {
808 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
809 region_ptr->set_blocks_filled();
810 }
811
812 HeapWord* const search_start = block_align_down(addr);
813 const size_t block_offset = addr_to_block_ptr(addr)->offset();
814
815 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
816 const size_t live = bitmap->live_words_in_range(cm, search_start, cast_to_oop(addr));
817 result += block_offset + live;
818 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
819 return result;
820 }
821
822 #ifdef ASSERT
823 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
824 {
825 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
826 const size_t* const end = (const size_t*)vspace->committed_high_addr();
827 for (const size_t* p = beg; p < end; ++p) {
828 assert(*p == 0, "not zero");
829 }
830 }
831
832 void ParallelCompactData::verify_clear()
833 {
834 verify_clear(_region_vspace);
835 verify_clear(_block_vspace);
836 }
837 #endif // #ifdef ASSERT
838
839 STWGCTimer PSParallelCompact::_gc_timer;
840 ParallelOldTracer PSParallelCompact::_gc_tracer;
841 elapsedTimer PSParallelCompact::_accumulated_time;
842 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
843 CollectorCounters* PSParallelCompact::_counters = nullptr;
844 ParMarkBitMap PSParallelCompact::_mark_bitmap;
845 ParallelCompactData PSParallelCompact::_summary_data;
846
847 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
848
849 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
850
851 void PSParallelCompact::post_initialize() {
852 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
853 _span_based_discoverer.set_span(heap->reserved_region());
854 _ref_processor =
855 new ReferenceProcessor(&_span_based_discoverer,
856 ParallelGCThreads, // mt processing degree
857 ParallelGCThreads, // mt discovery degree
858 false, // concurrent_discovery
859 &_is_alive_closure); // non-header is alive closure
860
861 _counters = new CollectorCounters("Parallel full collection pauses", 1);
862
863 // Initialize static fields in ParCompactionManager.
864 ParCompactionManager::initialize(mark_bitmap());
865 }
866
867 bool PSParallelCompact::initialize() {
868 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
869 MemRegion mr = heap->reserved_region();
870
871 // Was the old gen get allocated successfully?
872 if (!heap->old_gen()->is_allocated()) {
873 return false;
874 }
875
876 initialize_space_info();
877 initialize_dead_wood_limiter();
878
879 if (!_mark_bitmap.initialize(mr)) {
880 vm_shutdown_during_initialization(
881 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
882 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
883 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
884 return false;
885 }
886
887 if (!_summary_data.initialize(mr)) {
888 vm_shutdown_during_initialization(
889 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
890 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
891 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
892 return false;
893 }
894
895 return true;
896 }
897
898 void PSParallelCompact::initialize_space_info()
899 {
900 memset(&_space_info, 0, sizeof(_space_info));
901
902 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
903 PSYoungGen* young_gen = heap->young_gen();
904
905 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
906 _space_info[eden_space_id].set_space(young_gen->eden_space());
907 _space_info[from_space_id].set_space(young_gen->from_space());
908 _space_info[to_space_id].set_space(young_gen->to_space());
909
910 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
911 }
912
913 void PSParallelCompact::initialize_dead_wood_limiter()
914 {
915 const size_t max = 100;
916 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
917 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
918 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
919 DEBUG_ONLY(_dwl_initialized = true;)
920 _dwl_adjustment = normal_distribution(1.0);
921 }
922
923 void
924 PSParallelCompact::clear_data_covering_space(SpaceId id)
925 {
926 // At this point, top is the value before GC, new_top() is the value that will
927 // be set at the end of GC. The marking bitmap is cleared to top; nothing
928 // should be marked above top. The summary data is cleared to the larger of
929 // top & new_top.
930 MutableSpace* const space = _space_info[id].space();
931 HeapWord* const bot = space->bottom();
932 HeapWord* const top = space->top();
933 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
934
935 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
936 const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
937 _mark_bitmap.clear_range(beg_bit, end_bit);
938
939 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
940 const size_t end_region =
941 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
942 _summary_data.clear_range(beg_region, end_region);
943
944 // Clear the data used to 'split' regions.
945 SplitInfo& split_info = _space_info[id].split_info();
946 if (split_info.is_valid()) {
947 split_info.clear();
948 }
949 DEBUG_ONLY(split_info.verify_clear();)
950 }
951
952 void PSParallelCompact::pre_compact()
953 {
954 // Update the from & to space pointers in space_info, since they are swapped
955 // at each young gen gc. Do the update unconditionally (even though a
956 // promotion failure does not swap spaces) because an unknown number of young
957 // collections will have swapped the spaces an unknown number of times.
958 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
959 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
960 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
961 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
962
963 // Increment the invocation count
964 heap->increment_total_collections(true);
965
966 CodeCache::on_gc_marking_cycle_start();
967
968 heap->print_heap_before_gc();
969 heap->trace_heap_before_gc(&_gc_tracer);
970
971 // Fill in TLABs
972 heap->ensure_parsability(true); // retire TLABs
973
974 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
975 Universe::verify("Before GC");
976 }
977
978 // Verify object start arrays
979 if (VerifyObjectStartArray &&
980 VerifyBeforeGC) {
981 heap->old_gen()->verify_object_start_array();
982 }
983
984 DEBUG_ONLY(mark_bitmap()->verify_clear();)
985 DEBUG_ONLY(summary_data().verify_clear();)
986
987 ParCompactionManager::reset_all_bitmap_query_caches();
988 }
989
990 void PSParallelCompact::post_compact()
991 {
992 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
993 ParCompactionManager::remove_all_shadow_regions();
994
995 CodeCache::on_gc_marking_cycle_finish();
996 CodeCache::arm_all_nmethods();
997
998 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
999 // Clear the marking bitmap, summary data and split info.
1000 clear_data_covering_space(SpaceId(id));
1001 // Update top(). Must be done after clearing the bitmap and summary data.
1002 _space_info[id].publish_new_top();
1003 }
1004
1005 ParCompactionManager::flush_all_string_dedup_requests();
1006
1007 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1008 MutableSpace* const from_space = _space_info[from_space_id].space();
1009 MutableSpace* const to_space = _space_info[to_space_id].space();
1010
1011 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1012 bool eden_empty = eden_space->is_empty();
1013
1014 // Update heap occupancy information which is used as input to the soft ref
1015 // clearing policy at the next gc.
1016 Universe::heap()->update_capacity_and_used_at_gc();
1017
1018 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1019 to_space->is_empty();
1020
1021 PSCardTable* ct = heap->card_table();
1022 MemRegion old_mr = heap->old_gen()->committed();
1023 if (young_gen_empty) {
1024 ct->clear_MemRegion(old_mr);
1025 } else {
1026 ct->dirty_MemRegion(old_mr);
1027 }
1028
1029 {
1030 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1031 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
1032 ClassLoaderDataGraph::purge(true /* at_safepoint */);
1033 DEBUG_ONLY(MetaspaceUtils::verify();)
1034 }
1035
1036 // Need to clear claim bits for the next mark.
1037 ClassLoaderDataGraph::clear_claimed_marks();
1038
1039 heap->prune_scavengable_nmethods();
1040
1041 #if COMPILER2_OR_JVMCI
1042 DerivedPointerTable::update_pointers();
1043 #endif
1044
1045 if (ZapUnusedHeapArea) {
1046 heap->gen_mangle_unused_area();
1047 }
1048
1049 // Signal that we have completed a visit to all live objects.
1050 Universe::heap()->record_whole_heap_examined_timestamp();
1051 }
1052
1053 HeapWord*
1054 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1055 bool maximum_compaction)
1056 {
1057 const size_t region_size = ParallelCompactData::RegionSize;
1058 const ParallelCompactData& sd = summary_data();
1059
1060 const MutableSpace* const space = _space_info[id].space();
1061 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1062 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1063 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1064
1065 // Skip full regions at the beginning of the space--they are necessarily part
1066 // of the dense prefix.
1067 size_t full_count = 0;
1068 const RegionData* cp;
1069 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1070 ++full_count;
1071 }
1072
1073 const uint total_invocations = ParallelScavengeHeap::heap()->total_full_collections();
1074 assert(total_invocations >= _maximum_compaction_gc_num, "sanity");
1075 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num;
1076 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1077 if (maximum_compaction || cp == end_cp || interval_ended) {
1078 _maximum_compaction_gc_num = total_invocations;
1079 return sd.region_to_addr(cp);
1080 }
1081
1082 HeapWord* const new_top = _space_info[id].new_top();
1083 const size_t space_live = pointer_delta(new_top, space->bottom());
1084 const size_t space_used = space->used_in_words();
1085 const size_t space_capacity = space->capacity_in_words();
1086
1087 const double cur_density = double(space_live) / space_capacity;
1088 const double deadwood_density =
1089 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1090 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1091
1092 log_develop_debug(gc, compaction)(
1093 "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1094 cur_density, deadwood_density, deadwood_goal);
1095 log_develop_debug(gc, compaction)(
1096 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1097 "space_cap=" SIZE_FORMAT,
1098 space_live, space_used,
1099 space_capacity);
1100
1101 // XXX - Use binary search?
1102 HeapWord* dense_prefix = sd.region_to_addr(cp);
1103 const RegionData* full_cp = cp;
1104 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1105 while (cp < end_cp) {
1106 HeapWord* region_destination = cp->destination();
1107 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1108
1109 log_develop_trace(gc, compaction)(
1110 "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1111 "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1112 sd.region(cp), p2i(region_destination),
1113 p2i(dense_prefix), cur_deadwood);
1114
1115 if (cur_deadwood >= deadwood_goal) {
1116 // Found the region that has the correct amount of deadwood to the left.
1117 // This typically occurs after crossing a fairly sparse set of regions, so
1118 // iterate backwards over those sparse regions, looking for the region
1119 // that has the lowest density of live objects 'to the right.'
1120 size_t space_to_left = sd.region(cp) * region_size;
1121 size_t live_to_left = space_to_left - cur_deadwood;
1122 size_t space_to_right = space_capacity - space_to_left;
1123 size_t live_to_right = space_live - live_to_left;
1124 double density_to_right = double(live_to_right) / space_to_right;
1125 while (cp > full_cp) {
1126 --cp;
1127 const size_t prev_region_live_to_right = live_to_right -
1128 cp->data_size();
1129 const size_t prev_region_space_to_right = space_to_right + region_size;
1130 double prev_region_density_to_right =
1131 double(prev_region_live_to_right) / prev_region_space_to_right;
1132 if (density_to_right <= prev_region_density_to_right) {
1133 return dense_prefix;
1134 }
1135
1136 log_develop_trace(gc, compaction)(
1137 "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1138 "pc_d2r=%10.8f",
1139 sd.region(cp), density_to_right,
1140 prev_region_density_to_right);
1141
1142 dense_prefix -= region_size;
1143 live_to_right = prev_region_live_to_right;
1144 space_to_right = prev_region_space_to_right;
1145 density_to_right = prev_region_density_to_right;
1146 }
1147 return dense_prefix;
1148 }
1149
1150 dense_prefix += region_size;
1151 ++cp;
1152 }
1153
1154 return dense_prefix;
1155 }
1156
1157 #ifndef PRODUCT
1158 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1159 const SpaceId id,
1160 const bool maximum_compaction,
1161 HeapWord* const addr)
1162 {
1163 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1164 RegionData* const cp = summary_data().region(region_idx);
1165 const MutableSpace* const space = _space_info[id].space();
1166 HeapWord* const new_top = _space_info[id].new_top();
1167
1168 const size_t space_live = pointer_delta(new_top, space->bottom());
1169 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1170 const size_t space_cap = space->capacity_in_words();
1171 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1172 const size_t live_to_right = new_top - cp->destination();
1173 const size_t dead_to_right = space->top() - addr - live_to_right;
1174
1175 log_develop_debug(gc, compaction)(
1176 "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1177 "spl=" SIZE_FORMAT " "
1178 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1179 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1180 "ratio=%10.8f",
1181 algorithm, p2i(addr), region_idx,
1182 space_live,
1183 dead_to_left, dead_to_left_pct,
1184 dead_to_right, live_to_right,
1185 double(dead_to_right) / live_to_right);
1186 }
1187 #endif // #ifndef PRODUCT
1188
1189 // Return a fraction indicating how much of the generation can be treated as
1190 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1191 // based on the density of live objects in the generation to determine a limit,
1192 // which is then adjusted so the return value is min_percent when the density is
1193 // 1.
1194 //
1195 // The following table shows some return values for a different values of the
1196 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1197 // min_percent is 1.
1198 //
1199 // fraction allowed as dead wood
1200 // -----------------------------------------------------------------
1201 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1202 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1203 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1204 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1205 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1206 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1207 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1208 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1209 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1210 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1211 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1212 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1213 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1214 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1215 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1216 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1217 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1218 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1219 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1220 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1221 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1222 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1223 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1224
1225 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1226 {
1227 assert(_dwl_initialized, "uninitialized");
1228
1229 // The raw limit is the value of the normal distribution at x = density.
1230 const double raw_limit = normal_distribution(density);
1231
1232 // Adjust the raw limit so it becomes the minimum when the density is 1.
1233 //
1234 // First subtract the adjustment value (which is simply the precomputed value
1235 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1236 // Then add the minimum value, so the minimum is returned when the density is
1237 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1238 const double min = double(min_percent) / 100.0;
1239 const double limit = raw_limit - _dwl_adjustment + min;
1240 return MAX2(limit, 0.0);
1241 }
1242
1243 ParallelCompactData::RegionData*
1244 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1245 const RegionData* end)
1246 {
1247 const size_t region_size = ParallelCompactData::RegionSize;
1248 ParallelCompactData& sd = summary_data();
1249 size_t left = sd.region(beg);
1250 size_t right = end > beg ? sd.region(end) - 1 : left;
1251
1252 // Binary search.
1253 while (left < right) {
1254 // Equivalent to (left + right) / 2, but does not overflow.
1255 const size_t middle = left + (right - left) / 2;
1256 RegionData* const middle_ptr = sd.region(middle);
1257 HeapWord* const dest = middle_ptr->destination();
1258 HeapWord* const addr = sd.region_to_addr(middle);
1259 assert(dest != nullptr, "sanity");
1260 assert(dest <= addr, "must move left");
1261
1262 if (middle > left && dest < addr) {
1263 right = middle - 1;
1264 } else if (middle < right && middle_ptr->data_size() == region_size) {
1265 left = middle + 1;
1266 } else {
1267 return middle_ptr;
1268 }
1269 }
1270 return sd.region(left);
1271 }
1272
1273 ParallelCompactData::RegionData*
1274 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1275 const RegionData* end,
1276 size_t dead_words)
1277 {
1278 ParallelCompactData& sd = summary_data();
1279 size_t left = sd.region(beg);
1280 size_t right = end > beg ? sd.region(end) - 1 : left;
1281
1282 // Binary search.
1283 while (left < right) {
1284 // Equivalent to (left + right) / 2, but does not overflow.
1285 const size_t middle = left + (right - left) / 2;
1286 RegionData* const middle_ptr = sd.region(middle);
1287 HeapWord* const dest = middle_ptr->destination();
1288 HeapWord* const addr = sd.region_to_addr(middle);
1289 assert(dest != nullptr, "sanity");
1290 assert(dest <= addr, "must move left");
1291
1292 const size_t dead_to_left = pointer_delta(addr, dest);
1293 if (middle > left && dead_to_left > dead_words) {
1294 right = middle - 1;
1295 } else if (middle < right && dead_to_left < dead_words) {
1296 left = middle + 1;
1297 } else {
1298 return middle_ptr;
1299 }
1300 }
1301 return sd.region(left);
1302 }
1303
1304 // The result is valid during the summary phase, after the initial summarization
1305 // of each space into itself, and before final summarization.
1306 inline double
1307 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1308 HeapWord* const bottom,
1309 HeapWord* const top,
1310 HeapWord* const new_top)
1311 {
1312 ParallelCompactData& sd = summary_data();
1313
1314 assert(cp != nullptr, "sanity");
1315 assert(bottom != nullptr, "sanity");
1316 assert(top != nullptr, "sanity");
1317 assert(new_top != nullptr, "sanity");
1318 assert(top >= new_top, "summary data problem?");
1319 assert(new_top > bottom, "space is empty; should not be here");
1320 assert(new_top >= cp->destination(), "sanity");
1321 assert(top >= sd.region_to_addr(cp), "sanity");
1322
1323 HeapWord* const destination = cp->destination();
1324 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1325 const size_t compacted_region_live = pointer_delta(new_top, destination);
1326 const size_t compacted_region_used = pointer_delta(top,
1327 sd.region_to_addr(cp));
1328 const size_t reclaimable = compacted_region_used - compacted_region_live;
1329
1330 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1331 return double(reclaimable) / divisor;
1332 }
1333
1334 // Return the address of the end of the dense prefix, a.k.a. the start of the
1335 // compacted region. The address is always on a region boundary.
1336 //
1337 // Completely full regions at the left are skipped, since no compaction can
1338 // occur in those regions. Then the maximum amount of dead wood to allow is
1339 // computed, based on the density (amount live / capacity) of the generation;
1340 // the region with approximately that amount of dead space to the left is
1341 // identified as the limit region. Regions between the last completely full
1342 // region and the limit region are scanned and the one that has the best
1343 // (maximum) reclaimed_ratio() is selected.
1344 HeapWord*
1345 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1346 bool maximum_compaction)
1347 {
1348 const size_t region_size = ParallelCompactData::RegionSize;
1349 const ParallelCompactData& sd = summary_data();
1350
1351 const MutableSpace* const space = _space_info[id].space();
1352 HeapWord* const top = space->top();
1353 HeapWord* const top_aligned_up = sd.region_align_up(top);
1354 HeapWord* const new_top = _space_info[id].new_top();
1355 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1356 HeapWord* const bottom = space->bottom();
1357 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1358 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1359 const RegionData* const new_top_cp =
1360 sd.addr_to_region_ptr(new_top_aligned_up);
1361
1362 // Skip full regions at the beginning of the space--they are necessarily part
1363 // of the dense prefix.
1364 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1365 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1366 space->is_empty(), "no dead space allowed to the left");
1367 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1368 "region must have dead space");
1369
1370 // The gc number is saved whenever a maximum compaction is done, and used to
1371 // determine when the maximum compaction interval has expired. This avoids
1372 // successive max compactions for different reasons.
1373 const uint total_invocations = ParallelScavengeHeap::heap()->total_full_collections();
1374 assert(total_invocations >= _maximum_compaction_gc_num, "sanity");
1375 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num;
1376 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1377 total_invocations == HeapFirstMaximumCompactionCount;
1378 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1379 _maximum_compaction_gc_num = total_invocations;
1380 return sd.region_to_addr(full_cp);
1381 }
1382
1383 const size_t space_live = pointer_delta(new_top, bottom);
1384 const size_t space_used = space->used_in_words();
1385 const size_t space_capacity = space->capacity_in_words();
1386
1387 const double density = double(space_live) / double(space_capacity);
1388 const size_t min_percent_free = MarkSweepDeadRatio;
1389 const double limiter = dead_wood_limiter(density, min_percent_free);
1390 const size_t dead_wood_max = space_used - space_live;
1391 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1392 dead_wood_max);
1393
1394 log_develop_debug(gc, compaction)(
1395 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1396 "space_cap=" SIZE_FORMAT,
1397 space_live, space_used,
1398 space_capacity);
1399 log_develop_debug(gc, compaction)(
1400 "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1401 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1402 density, min_percent_free, limiter,
1403 dead_wood_max, dead_wood_limit);
1404
1405 // Locate the region with the desired amount of dead space to the left.
1406 const RegionData* const limit_cp =
1407 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1408
1409 // Scan from the first region with dead space to the limit region and find the
1410 // one with the best (largest) reclaimed ratio.
1411 double best_ratio = 0.0;
1412 const RegionData* best_cp = full_cp;
1413 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1414 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1415 if (tmp_ratio > best_ratio) {
1416 best_cp = cp;
1417 best_ratio = tmp_ratio;
1418 }
1419 }
1420
1421 return sd.region_to_addr(best_cp);
1422 }
1423
1424 void PSParallelCompact::summarize_spaces_quick()
1425 {
1426 for (unsigned int i = 0; i < last_space_id; ++i) {
1427 const MutableSpace* space = _space_info[i].space();
1428 HeapWord** nta = _space_info[i].new_top_addr();
1429 bool result = _summary_data.summarize(_space_info[i].split_info(),
1430 space->bottom(), space->top(), nullptr,
1431 space->bottom(), space->end(), nta);
1432 assert(result, "space must fit into itself");
1433 _space_info[i].set_dense_prefix(space->bottom());
1434 }
1435 }
1436
1437 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1438 {
1439 HeapWord* const dense_prefix_end = dense_prefix(id);
1440 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1441 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1442 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1443 // Only enough dead space is filled so that any remaining dead space to the
1444 // left is larger than the minimum filler object. (The remainder is filled
1445 // during the copy/update phase.)
1446 //
1447 // The size of the dead space to the right of the boundary is not a
1448 // concern, since compaction will be able to use whatever space is
1449 // available.
1450 //
1451 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1452 // surrounds the space to be filled with an object.
1453 //
1454 // In the 32-bit VM, each bit represents two 32-bit words:
1455 // +---+
1456 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1457 // end_bits: ... x x x | 0 | || 0 x x ...
1458 // +---+
1459 //
1460 // In the 64-bit VM, each bit represents one 64-bit word:
1461 // +------------+
1462 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1463 // end_bits: ... x x 1 | 0 || 0 | x x ...
1464 // +------------+
1465 // +-------+
1466 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1467 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1468 // +-------+
1469 // +-----------+
1470 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1471 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1472 // +-----------+
1473 // +-------+
1474 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1475 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1476 // +-------+
1477
1478 // Initially assume case a, c or e will apply.
1479 size_t obj_len = CollectedHeap::min_fill_size();
1480 HeapWord* obj_beg = dense_prefix_end - obj_len;
1481
1482 #ifdef _LP64
1483 if (MinObjAlignment > 1) { // object alignment > heap word size
1484 // Cases a, c or e.
1485 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1486 // Case b above.
1487 obj_beg = dense_prefix_end - 1;
1488 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1489 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1490 // Case d above.
1491 obj_beg = dense_prefix_end - 3;
1492 obj_len = 3;
1493 }
1494 #endif // #ifdef _LP64
1495
1496 CollectedHeap::fill_with_object(obj_beg, obj_len);
1497 _mark_bitmap.mark_obj(obj_beg, obj_len);
1498 _summary_data.add_obj(obj_beg, obj_len);
1499 assert(start_array(id) != nullptr, "sanity");
1500 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
1501 }
1502 }
1503
1504 void
1505 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1506 {
1507 assert(id < last_space_id, "id out of range");
1508 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1509 "should have been reset in summarize_spaces_quick()");
1510
1511 const MutableSpace* space = _space_info[id].space();
1512 if (_space_info[id].new_top() != space->bottom()) {
1513 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1514 _space_info[id].set_dense_prefix(dense_prefix_end);
1515
1516 #ifndef PRODUCT
1517 if (log_is_enabled(Debug, gc, compaction)) {
1518 print_dense_prefix_stats("ratio", id, maximum_compaction,
1519 dense_prefix_end);
1520 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1521 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1522 }
1523 #endif // #ifndef PRODUCT
1524
1525 // Recompute the summary data, taking into account the dense prefix. If
1526 // every last byte will be reclaimed, then the existing summary data which
1527 // compacts everything can be left in place.
1528 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1529 // If dead space crosses the dense prefix boundary, it is (at least
1530 // partially) filled with a dummy object, marked live and added to the
1531 // summary data. This simplifies the copy/update phase and must be done
1532 // before the final locations of objects are determined, to prevent
1533 // leaving a fragment of dead space that is too small to fill.
1534 fill_dense_prefix_end(id);
1535
1536 // Compute the destination of each Region, and thus each object.
1537 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1538 _summary_data.summarize(_space_info[id].split_info(),
1539 dense_prefix_end, space->top(), nullptr,
1540 dense_prefix_end, space->end(),
1541 _space_info[id].new_top_addr());
1542 }
1543 }
1544
1545 if (log_develop_is_enabled(Trace, gc, compaction)) {
1546 const size_t region_size = ParallelCompactData::RegionSize;
1547 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1548 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1549 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1550 HeapWord* const new_top = _space_info[id].new_top();
1551 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1552 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1553 log_develop_trace(gc, compaction)(
1554 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1555 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1556 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1557 id, space->capacity_in_words(), p2i(dense_prefix_end),
1558 dp_region, dp_words / region_size,
1559 cr_words / region_size, p2i(new_top));
1560 }
1561 }
1562
1563 #ifndef PRODUCT
1564 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1565 HeapWord* dst_beg, HeapWord* dst_end,
1566 SpaceId src_space_id,
1567 HeapWord* src_beg, HeapWord* src_end)
1568 {
1569 log_develop_trace(gc, compaction)(
1570 "Summarizing %d [%s] into %d [%s]: "
1571 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1572 SIZE_FORMAT "-" SIZE_FORMAT " "
1573 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1574 SIZE_FORMAT "-" SIZE_FORMAT,
1575 src_space_id, space_names[src_space_id],
1576 dst_space_id, space_names[dst_space_id],
1577 p2i(src_beg), p2i(src_end),
1578 _summary_data.addr_to_region_idx(src_beg),
1579 _summary_data.addr_to_region_idx(src_end),
1580 p2i(dst_beg), p2i(dst_end),
1581 _summary_data.addr_to_region_idx(dst_beg),
1582 _summary_data.addr_to_region_idx(dst_end));
1583 }
1584 #endif // #ifndef PRODUCT
1585
1586 void PSParallelCompact::summary_phase(bool maximum_compaction)
1587 {
1588 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1589
1590 // Quick summarization of each space into itself, to see how much is live.
1591 summarize_spaces_quick();
1592
1593 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1594 NOT_PRODUCT(print_region_ranges());
1595 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1596
1597 // The amount of live data that will end up in old space (assuming it fits).
1598 size_t old_space_total_live = 0;
1599 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1600 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1601 _space_info[id].space()->bottom());
1602 }
1603
1604 MutableSpace* const old_space = _space_info[old_space_id].space();
1605 const size_t old_capacity = old_space->capacity_in_words();
1606 if (old_space_total_live > old_capacity) {
1607 // XXX - should also try to expand
1608 maximum_compaction = true;
1609 }
1610
1611 // Old generations.
1612 summarize_space(old_space_id, maximum_compaction);
1613
1614 // Summarize the remaining spaces in the young gen. The initial target space
1615 // is the old gen. If a space does not fit entirely into the target, then the
1616 // remainder is compacted into the space itself and that space becomes the new
1617 // target.
1618 SpaceId dst_space_id = old_space_id;
1619 HeapWord* dst_space_end = old_space->end();
1620 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1621 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1622 const MutableSpace* space = _space_info[id].space();
1623 const size_t live = pointer_delta(_space_info[id].new_top(),
1624 space->bottom());
1625 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1626
1627 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1628 SpaceId(id), space->bottom(), space->top());)
1629 if (live > 0 && live <= available) {
1630 // All the live data will fit.
1631 bool done = _summary_data.summarize(_space_info[id].split_info(),
1632 space->bottom(), space->top(),
1633 nullptr,
1634 *new_top_addr, dst_space_end,
1635 new_top_addr);
1636 assert(done, "space must fit into old gen");
1637
1638 // Reset the new_top value for the space.
1639 _space_info[id].set_new_top(space->bottom());
1640 } else if (live > 0) {
1641 // Attempt to fit part of the source space into the target space.
1642 HeapWord* next_src_addr = nullptr;
1643 bool done = _summary_data.summarize(_space_info[id].split_info(),
1644 space->bottom(), space->top(),
1645 &next_src_addr,
1646 *new_top_addr, dst_space_end,
1647 new_top_addr);
1648 assert(!done, "space should not fit into old gen");
1649 assert(next_src_addr != nullptr, "sanity");
1650
1651 // The source space becomes the new target, so the remainder is compacted
1652 // within the space itself.
1653 dst_space_id = SpaceId(id);
1654 dst_space_end = space->end();
1655 new_top_addr = _space_info[id].new_top_addr();
1656 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1657 space->bottom(), dst_space_end,
1658 SpaceId(id), next_src_addr, space->top());)
1659 done = _summary_data.summarize(_space_info[id].split_info(),
1660 next_src_addr, space->top(),
1661 nullptr,
1662 space->bottom(), dst_space_end,
1663 new_top_addr);
1664 assert(done, "space must fit when compacted into itself");
1665 assert(*new_top_addr <= space->top(), "usage should not grow");
1666 }
1667 }
1668
1669 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1670 NOT_PRODUCT(print_region_ranges());
1671 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1672 }
1673
1674 // This method should contain all heap-specific policy for invoking a full
1675 // collection. invoke_no_policy() will only attempt to compact the heap; it
1676 // will do nothing further. If we need to bail out for policy reasons, scavenge
1677 // before full gc, or any other specialized behavior, it needs to be added here.
1678 //
1679 // Note that this method should only be called from the vm_thread while at a
1680 // safepoint.
1681 //
1682 // Note that the all_soft_refs_clear flag in the soft ref policy
1683 // may be true because this method can be called without intervening
1684 // activity. For example when the heap space is tight and full measure
1685 // are being taken to free space.
1686 bool PSParallelCompact::invoke(bool maximum_heap_compaction) {
1687 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1688 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1689 "should be in vm thread");
1690
1691 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1692 assert(!heap->is_gc_active(), "not reentrant");
1693
1694 IsGCActiveMark mark;
1695
1696 if (ScavengeBeforeFullGC) {
1697 PSScavenge::invoke_no_policy();
1698 }
1699
1700 const bool clear_all_soft_refs =
1701 heap->soft_ref_policy()->should_clear_all_soft_refs();
1702
1703 return PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1704 maximum_heap_compaction);
1705 }
1706
1707 // This method contains no policy. You should probably
1708 // be calling invoke() instead.
1709 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1710 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1711 assert(ref_processor() != nullptr, "Sanity");
1712
1713 if (GCLocker::check_active_before_gc()) {
1714 return false;
1715 }
1716
1717 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1718
1719 GCIdMark gc_id_mark;
1720 _gc_timer.register_gc_start();
1721 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1722
1723 GCCause::Cause gc_cause = heap->gc_cause();
1724 PSYoungGen* young_gen = heap->young_gen();
1725 PSOldGen* old_gen = heap->old_gen();
1726 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1727
1728 // The scope of casr should end after code that can change
1729 // SoftRefPolicy::_should_clear_all_soft_refs.
1730 ClearedAllSoftRefs casr(maximum_heap_compaction,
1731 heap->soft_ref_policy());
1732
1733 if (ZapUnusedHeapArea) {
1734 // Save information needed to minimize mangling
1735 heap->record_gen_tops_before_GC();
1736 }
1737
1738 // Make sure data structures are sane, make the heap parsable, and do other
1739 // miscellaneous bookkeeping.
1740 pre_compact();
1741
1742 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1743
1744 {
1745 const uint active_workers =
1746 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1747 ParallelScavengeHeap::heap()->workers().active_workers(),
1748 Threads::number_of_non_daemon_threads());
1749 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1750
1751 GCTraceCPUTime tcpu(&_gc_tracer);
1752 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
1753
1754 heap->pre_full_gc_dump(&_gc_timer);
1755
1756 TraceCollectorStats tcs(counters());
1757 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
1758
1759 if (log_is_enabled(Debug, gc, heap, exit)) {
1760 accumulated_time()->start();
1761 }
1762
1763 // Let the size policy know we're starting
1764 size_policy->major_collection_begin();
1765
1766 #if COMPILER2_OR_JVMCI
1767 DerivedPointerTable::clear();
1768 #endif
1769
1770 ref_processor()->start_discovery(maximum_heap_compaction);
1771
1772 ClassUnloadingContext ctx(1 /* num_nmethod_unlink_workers */,
1773 false /* lock_codeblob_free_separately */);
1774
1775 marking_phase(&_gc_tracer);
1776
1777 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1778 && GCCause::is_user_requested_gc(gc_cause);
1779 summary_phase(maximum_heap_compaction || max_on_system_gc);
1780
1781 #if COMPILER2_OR_JVMCI
1782 assert(DerivedPointerTable::is_active(), "Sanity");
1783 DerivedPointerTable::set_active(false);
1784 #endif
1785
1786 // adjust_roots() updates Universe::_intArrayKlassObj which is
1787 // needed by the compaction for filling holes in the dense prefix.
1788 adjust_roots();
1789
1790 compact();
1791
1792 ParCompactionManager::verify_all_region_stack_empty();
1793
1794 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1795 // done before resizing.
1796 post_compact();
1797
1798 // Let the size policy know we're done
1799 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1800
1801 if (UseAdaptiveSizePolicy) {
1802 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1803 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1804 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1805
1806 // Don't check if the size_policy is ready here. Let
1807 // the size_policy check that internally.
1808 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1809 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1810 // Swap the survivor spaces if from_space is empty. The
1811 // resize_young_gen() called below is normally used after
1812 // a successful young GC and swapping of survivor spaces;
1813 // otherwise, it will fail to resize the young gen with
1814 // the current implementation.
1815 if (young_gen->from_space()->is_empty()) {
1816 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1817 young_gen->swap_spaces();
1818 }
1819
1820 // Calculate optimal free space amounts
1821 assert(young_gen->max_gen_size() >
1822 young_gen->from_space()->capacity_in_bytes() +
1823 young_gen->to_space()->capacity_in_bytes(),
1824 "Sizes of space in young gen are out-of-bounds");
1825
1826 size_t young_live = young_gen->used_in_bytes();
1827 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1828 size_t old_live = old_gen->used_in_bytes();
1829 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1830 size_t max_old_gen_size = old_gen->max_gen_size();
1831 size_t max_eden_size = young_gen->max_gen_size() -
1832 young_gen->from_space()->capacity_in_bytes() -
1833 young_gen->to_space()->capacity_in_bytes();
1834
1835 // Used for diagnostics
1836 size_policy->clear_generation_free_space_flags();
1837
1838 size_policy->compute_generations_free_space(young_live,
1839 eden_live,
1840 old_live,
1841 cur_eden,
1842 max_old_gen_size,
1843 max_eden_size,
1844 true /* full gc*/);
1845
1846 size_policy->check_gc_overhead_limit(eden_live,
1847 max_old_gen_size,
1848 max_eden_size,
1849 true /* full gc*/,
1850 gc_cause,
1851 heap->soft_ref_policy());
1852
1853 size_policy->decay_supplemental_growth(true /* full gc*/);
1854
1855 heap->resize_old_gen(
1856 size_policy->calculated_old_free_size_in_bytes());
1857
1858 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1859 size_policy->calculated_survivor_size_in_bytes());
1860 }
1861
1862 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1863 }
1864
1865 if (UsePerfData) {
1866 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1867 counters->update_counters();
1868 counters->update_old_capacity(old_gen->capacity_in_bytes());
1869 counters->update_young_capacity(young_gen->capacity_in_bytes());
1870 }
1871
1872 heap->resize_all_tlabs();
1873
1874 // Resize the metaspace capacity after a collection
1875 MetaspaceGC::compute_new_size();
1876
1877 if (log_is_enabled(Debug, gc, heap, exit)) {
1878 accumulated_time()->stop();
1879 }
1880
1881 heap->print_heap_change(pre_gc_values);
1882
1883 // Track memory usage and detect low memory
1884 MemoryService::track_memory_usage();
1885 heap->update_counters();
1886
1887 heap->post_full_gc_dump(&_gc_timer);
1888 }
1889
1890 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1891 Universe::verify("After GC");
1892 }
1893
1894 // Re-verify object start arrays
1895 if (VerifyObjectStartArray &&
1896 VerifyAfterGC) {
1897 old_gen->verify_object_start_array();
1898 }
1899
1900 if (ZapUnusedHeapArea) {
1901 old_gen->object_space()->check_mangled_unused_area_complete();
1902 }
1903
1904 heap->print_heap_after_gc();
1905 heap->trace_heap_after_gc(&_gc_tracer);
1906
1907 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1908
1909 _gc_timer.register_gc_end();
1910
1911 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1912 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1913
1914 return true;
1915 }
1916
1917 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1918 private:
1919 uint _worker_id;
1920
1921 public:
1922 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1923 void do_thread(Thread* thread) {
1924 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1925
1926 ResourceMark rm;
1927
1928 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1929
1930 PCMarkAndPushClosure mark_and_push_closure(cm);
1931 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations, true /* keepalive nmethods */);
1932
1933 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1934
1935 // Do the real work
1936 cm->follow_marking_stacks();
1937 }
1938 };
1939
1940 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1941 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1942
1943 ParCompactionManager* cm =
1944 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1945
1946 do {
1947 oop obj = nullptr;
1948 ObjArrayTask task;
1949 if (ParCompactionManager::steal_objarray(worker_id, task)) {
1950 cm->follow_array((objArrayOop)task.obj(), task.index());
1951 } else if (ParCompactionManager::steal(worker_id, obj)) {
1952 cm->follow_contents(obj);
1953 }
1954 cm->follow_marking_stacks();
1955 } while (!terminator.offer_termination());
1956 }
1957
1958 class MarkFromRootsTask : public WorkerTask {
1959 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
1960 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1961 TaskTerminator _terminator;
1962 uint _active_workers;
1963
1964 public:
1965 MarkFromRootsTask(uint active_workers) :
1966 WorkerTask("MarkFromRootsTask"),
1967 _strong_roots_scope(active_workers),
1968 _terminator(active_workers, ParCompactionManager::oop_task_queues()),
1969 _active_workers(active_workers) {}
1970
1971 virtual void work(uint worker_id) {
1972 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1973 PCMarkAndPushClosure mark_and_push_closure(cm);
1974
1975 {
1976 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1977 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1978
1979 // Do the real work
1980 cm->follow_marking_stacks();
1981 }
1982
1983 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
1984 Threads::possibly_parallel_threads_do(true /* is_par */, &closure);
1985
1986 // Mark from OopStorages
1987 {
1988 _oop_storage_set_par_state.oops_do(&mark_and_push_closure);
1989 // Do the real work
1990 cm->follow_marking_stacks();
1991 }
1992
1993 if (_active_workers > 1) {
1994 steal_marking_work(_terminator, worker_id);
1995 }
1996 }
1997 };
1998
1999 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2000 TaskTerminator _terminator;
2001
2002 public:
2003 ParallelCompactRefProcProxyTask(uint max_workers)
2004 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2005 _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2006
2007 void work(uint worker_id) override {
2008 assert(worker_id < _max_workers, "sanity");
2009 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2010 PCMarkAndPushClosure keep_alive(cm);
2011 BarrierEnqueueDiscoveredFieldClosure enqueue;
2012 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2013 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2014 }
2015
2016 void prepare_run_task_hook() override {
2017 _terminator.reset_for_reuse(_queue_count);
2018 }
2019 };
2020
2021 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
2022 // Recursively traverse all live objects and mark them
2023 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2024
2025 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2026
2027 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
2028 {
2029 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2030
2031 MarkFromRootsTask task(active_gc_threads);
2032 ParallelScavengeHeap::heap()->workers().run_task(&task);
2033 }
2034
2035 // Process reference objects found during marking
2036 {
2037 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2038
2039 ReferenceProcessorStats stats;
2040 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2041
2042 ref_processor()->set_active_mt_degree(active_gc_threads);
2043 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2044 stats = ref_processor()->process_discovered_references(task, pt);
2045
2046 gc_tracer->report_gc_reference_stats(stats);
2047 pt.print_all_references();
2048 }
2049
2050 // This is the point where the entire marking should have completed.
2051 ParCompactionManager::verify_all_marking_stack_empty();
2052
2053 {
2054 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2055 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2056 is_alive_closure(),
2057 &do_nothing_cl,
2058 1);
2059 }
2060
2061 {
2062 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2063
2064 ClassUnloadingContext* ctx = ClassUnloadingContext::context();
2065
2066 bool unloading_occurred;
2067 {
2068 CodeCache::UnlinkingScope scope(is_alive_closure());
2069
2070 // Follow system dictionary roots and unload classes.
2071 unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
2072
2073 // Unload nmethods.
2074 CodeCache::do_unloading(unloading_occurred);
2075 }
2076
2077 {
2078 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
2079 // Release unloaded nmethod's memory.
2080 ctx->purge_nmethods();
2081 }
2082 {
2083 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
2084 ctx->free_code_blobs();
2085 }
2086
2087 // Prune dead klasses from subklass/sibling/implementor lists.
2088 Klass::clean_weak_klass_links(unloading_occurred);
2089
2090 // Clean JVMCI metadata handles.
2091 JVMCI_ONLY(JVMCI::do_unloading(unloading_occurred));
2092 }
2093
2094 {
2095 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
2096 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers());
2097 }
2098 #if TASKQUEUE_STATS
2099 ParCompactionManager::oop_task_queues()->print_and_reset_taskqueue_stats("Oop Queue");
2100 ParCompactionManager::_objarray_task_queues->print_and_reset_taskqueue_stats("ObjArrayOop Queue");
2101 #endif
2102 }
2103
2104 class PSAdjustTask final : public WorkerTask {
2105 SubTasksDone _sub_tasks;
2106 WeakProcessor::Task _weak_proc_task;
2107 OopStorageSetStrongParState<false, false> _oop_storage_iter;
2108 uint _nworkers;
2109
2110 enum PSAdjustSubTask {
2111 PSAdjustSubTask_code_cache,
2112
2113 PSAdjustSubTask_num_elements
2114 };
2115
2116 public:
2117 PSAdjustTask(uint nworkers) :
2118 WorkerTask("PSAdjust task"),
2119 _sub_tasks(PSAdjustSubTask_num_elements),
2120 _weak_proc_task(nworkers),
2121 _nworkers(nworkers) {
2122
2123 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
2124 if (nworkers > 1) {
2125 Threads::change_thread_claim_token();
2126 }
2127 }
2128
2129 ~PSAdjustTask() {
2130 Threads::assert_all_threads_claimed();
2131 }
2132
2133 void work(uint worker_id) {
2134 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2135 PCAdjustPointerClosure adjust(cm);
2136 {
2137 ResourceMark rm;
2138 Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2139 }
2140 _oop_storage_iter.oops_do(&adjust);
2141 {
2142 CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_stw_fullgc_adjust);
2143 ClassLoaderDataGraph::cld_do(&cld_closure);
2144 }
2145 {
2146 AlwaysTrueClosure always_alive;
2147 _weak_proc_task.work(worker_id, &always_alive, &adjust);
2148 }
2149 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2150 CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2151 CodeCache::blobs_do(&adjust_code);
2152 }
2153 _sub_tasks.all_tasks_claimed();
2154 }
2155 };
2156
2157 void PSParallelCompact::adjust_roots() {
2158 // Adjust the pointers to reflect the new locations
2159 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2160 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2161 PSAdjustTask task(nworkers);
2162 ParallelScavengeHeap::heap()->workers().run_task(&task);
2163 }
2164
2165 // Helper class to print 8 region numbers per line and then print the total at the end.
2166 class FillableRegionLogger : public StackObj {
2167 private:
2168 Log(gc, compaction) log;
2169 static const int LineLength = 8;
2170 size_t _regions[LineLength];
2171 int _next_index;
2172 bool _enabled;
2173 size_t _total_regions;
2174 public:
2175 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2176 ~FillableRegionLogger() {
2177 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2178 }
2179
2180 void print_line() {
2181 if (!_enabled || _next_index == 0) {
2182 return;
2183 }
2184 FormatBuffer<> line("Fillable: ");
2185 for (int i = 0; i < _next_index; i++) {
2186 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2187 }
2188 log.trace("%s", line.buffer());
2189 _next_index = 0;
2190 }
2191
2192 void handle(size_t region) {
2193 if (!_enabled) {
2194 return;
2195 }
2196 _regions[_next_index++] = region;
2197 if (_next_index == LineLength) {
2198 print_line();
2199 }
2200 _total_regions++;
2201 }
2202 };
2203
2204 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2205 {
2206 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2207
2208 // Find the threads that are active
2209 uint worker_id = 0;
2210
2211 // Find all regions that are available (can be filled immediately) and
2212 // distribute them to the thread stacks. The iteration is done in reverse
2213 // order (high to low) so the regions will be removed in ascending order.
2214
2215 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2216
2217 // id + 1 is used to test termination so unsigned can
2218 // be used with an old_space_id == 0.
2219 FillableRegionLogger region_logger;
2220 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2221 SpaceInfo* const space_info = _space_info + id;
2222 HeapWord* const new_top = space_info->new_top();
2223
2224 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2225 const size_t end_region =
2226 sd.addr_to_region_idx(sd.region_align_up(new_top));
2227
2228 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2229 if (sd.region(cur)->claim_unsafe()) {
2230 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2231 bool result = sd.region(cur)->mark_normal();
2232 assert(result, "Must succeed at this point.");
2233 cm->region_stack()->push(cur);
2234 region_logger.handle(cur);
2235 // Assign regions to tasks in round-robin fashion.
2236 if (++worker_id == parallel_gc_threads) {
2237 worker_id = 0;
2238 }
2239 }
2240 }
2241 region_logger.print_line();
2242 }
2243 }
2244
2245 class TaskQueue : StackObj {
2246 volatile uint _counter;
2247 uint _size;
2248 uint _insert_index;
2249 PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2250 public:
2251 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(nullptr) {
2252 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2253 }
2254 ~TaskQueue() {
2255 assert(_counter >= _insert_index, "not all queue elements were claimed");
2256 FREE_C_HEAP_ARRAY(T, _backing_array);
2257 }
2258
2259 void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2260 assert(_insert_index < _size, "too small backing array");
2261 _backing_array[_insert_index++] = value;
2262 }
2263
2264 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2265 uint claimed = Atomic::fetch_then_add(&_counter, 1u);
2266 if (claimed < _insert_index) {
2267 reference = _backing_array[claimed];
2268 return true;
2269 } else {
2270 return false;
2271 }
2272 }
2273 };
2274
2275 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2276
2277 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2278 uint parallel_gc_threads) {
2279 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2280
2281 ParallelCompactData& sd = PSParallelCompact::summary_data();
2282
2283 // Iterate over all the spaces adding tasks for updating
2284 // regions in the dense prefix. Assume that 1 gc thread
2285 // will work on opening the gaps and the remaining gc threads
2286 // will work on the dense prefix.
2287 unsigned int space_id;
2288 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2289 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2290 const MutableSpace* const space = _space_info[space_id].space();
2291
2292 if (dense_prefix_end == space->bottom()) {
2293 // There is no dense prefix for this space.
2294 continue;
2295 }
2296
2297 // The dense prefix is before this region.
2298 size_t region_index_end_dense_prefix =
2299 sd.addr_to_region_idx(dense_prefix_end);
2300 RegionData* const dense_prefix_cp =
2301 sd.region(region_index_end_dense_prefix);
2302 assert(dense_prefix_end == space->end() ||
2303 dense_prefix_cp->available() ||
2304 dense_prefix_cp->claimed(),
2305 "The region after the dense prefix should always be ready to fill");
2306
2307 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2308
2309 // Is there dense prefix work?
2310 size_t total_dense_prefix_regions =
2311 region_index_end_dense_prefix - region_index_start;
2312 // How many regions of the dense prefix should be given to
2313 // each thread?
2314 if (total_dense_prefix_regions > 0) {
2315 uint tasks_for_dense_prefix = 1;
2316 if (total_dense_prefix_regions <=
2317 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2318 // Don't over partition. This assumes that
2319 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2320 // so there are not many regions to process.
2321 tasks_for_dense_prefix = parallel_gc_threads;
2322 } else {
2323 // Over partition
2324 tasks_for_dense_prefix = parallel_gc_threads *
2325 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2326 }
2327 size_t regions_per_thread = total_dense_prefix_regions /
2328 tasks_for_dense_prefix;
2329 // Give each thread at least 1 region.
2330 if (regions_per_thread == 0) {
2331 regions_per_thread = 1;
2332 }
2333
2334 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2335 if (region_index_start >= region_index_end_dense_prefix) {
2336 break;
2337 }
2338 // region_index_end is not processed
2339 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2340 region_index_end_dense_prefix);
2341 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2342 region_index_start,
2343 region_index_end));
2344 region_index_start = region_index_end;
2345 }
2346 }
2347 // This gets any part of the dense prefix that did not
2348 // fit evenly.
2349 if (region_index_start < region_index_end_dense_prefix) {
2350 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2351 region_index_start,
2352 region_index_end_dense_prefix));
2353 }
2354 }
2355 }
2356
2357 #ifdef ASSERT
2358 // Write a histogram of the number of times the block table was filled for a
2359 // region.
2360 void PSParallelCompact::write_block_fill_histogram()
2361 {
2362 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2363 return;
2364 }
2365
2366 Log(gc, compaction) log;
2367 ResourceMark rm;
2368 LogStream ls(log.trace());
2369 outputStream* out = &ls;
2370
2371 typedef ParallelCompactData::RegionData rd_t;
2372 ParallelCompactData& sd = summary_data();
2373
2374 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2375 MutableSpace* const spc = _space_info[id].space();
2376 if (spc->bottom() != spc->top()) {
2377 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2378 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2379 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2380
2381 size_t histo[5] = { 0, 0, 0, 0, 0 };
2382 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2383 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2384
2385 for (const rd_t* cur = beg; cur < end; ++cur) {
2386 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2387 }
2388 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2389 for (size_t i = 0; i < histo_len; ++i) {
2390 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2391 histo[i], 100.0 * histo[i] / region_cnt);
2392 }
2393 out->cr();
2394 }
2395 }
2396 }
2397 #endif // #ifdef ASSERT
2398
2399 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2400 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2401
2402 ParCompactionManager* cm =
2403 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2404
2405 // Drain the stacks that have been preloaded with regions
2406 // that are ready to fill.
2407
2408 cm->drain_region_stacks();
2409
2410 guarantee(cm->region_stack()->is_empty(), "Not empty");
2411
2412 size_t region_index = 0;
2413
2414 while (true) {
2415 if (ParCompactionManager::steal(worker_id, region_index)) {
2416 PSParallelCompact::fill_and_update_region(cm, region_index);
2417 cm->drain_region_stacks();
2418 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2419 // Fill and update an unavailable region with the help of a shadow region
2420 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2421 cm->drain_region_stacks();
2422 } else {
2423 if (terminator->offer_termination()) {
2424 break;
2425 }
2426 // Go around again.
2427 }
2428 }
2429 }
2430
2431 class UpdateDensePrefixAndCompactionTask: public WorkerTask {
2432 TaskQueue& _tq;
2433 TaskTerminator _terminator;
2434
2435 public:
2436 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2437 WorkerTask("UpdateDensePrefixAndCompactionTask"),
2438 _tq(tq),
2439 _terminator(active_workers, ParCompactionManager::region_task_queues()) {
2440 }
2441 virtual void work(uint worker_id) {
2442 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2443
2444 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2445 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2446 task._space_id,
2447 task._region_index_start,
2448 task._region_index_end);
2449 }
2450
2451 // Once a thread has drained it's stack, it should try to steal regions from
2452 // other threads.
2453 compaction_with_stealing_work(&_terminator, worker_id);
2454
2455 // At this point all regions have been compacted, so it's now safe
2456 // to update the deferred objects that cross region boundaries.
2457 cm->drain_deferred_objects();
2458 }
2459 };
2460
2461 void PSParallelCompact::compact() {
2462 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2463
2464 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2465 PSOldGen* old_gen = heap->old_gen();
2466 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2467
2468 // for [0..last_space_id)
2469 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2470 // push
2471 // push
2472 //
2473 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2474 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2475 initialize_shadow_regions(active_gc_threads);
2476 prepare_region_draining_tasks(active_gc_threads);
2477 enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2478
2479 {
2480 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2481
2482 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2483 ParallelScavengeHeap::heap()->workers().run_task(&task);
2484
2485 #ifdef ASSERT
2486 // Verify that all regions have been processed.
2487 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2488 verify_complete(SpaceId(id));
2489 }
2490 #endif
2491 }
2492
2493 DEBUG_ONLY(write_block_fill_histogram());
2494 }
2495
2496 #ifdef ASSERT
2497 void PSParallelCompact::verify_complete(SpaceId space_id) {
2498 // All Regions between space bottom() to new_top() should be marked as filled
2499 // and all Regions between new_top() and top() should be available (i.e.,
2500 // should have been emptied).
2501 ParallelCompactData& sd = summary_data();
2502 SpaceInfo si = _space_info[space_id];
2503 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2504 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2505 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2506 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2507 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2508
2509 bool issued_a_warning = false;
2510
2511 size_t cur_region;
2512 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2513 const RegionData* const c = sd.region(cur_region);
2514 if (!c->completed()) {
2515 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2516 cur_region, c->destination_count());
2517 issued_a_warning = true;
2518 }
2519 }
2520
2521 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2522 const RegionData* const c = sd.region(cur_region);
2523 if (!c->available()) {
2524 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2525 cur_region, c->destination_count());
2526 issued_a_warning = true;
2527 }
2528 }
2529
2530 if (issued_a_warning) {
2531 print_region_ranges();
2532 }
2533 }
2534 #endif // #ifdef ASSERT
2535
2536 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2537 _start_array->update_for_block(addr, addr + cast_to_oop(addr)->size());
2538 compaction_manager()->update_contents(cast_to_oop(addr));
2539 }
2540
2541 // Update interior oops in the ranges of regions [beg_region, end_region).
2542 void
2543 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2544 SpaceId space_id,
2545 size_t beg_region,
2546 size_t end_region) {
2547 ParallelCompactData& sd = summary_data();
2548 ParMarkBitMap* const mbm = mark_bitmap();
2549
2550 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2551 HeapWord* const end_addr = sd.region_to_addr(end_region);
2552 assert(beg_region <= end_region, "bad region range");
2553 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2554
2555 #ifdef ASSERT
2556 // Claim the regions to avoid triggering an assert when they are marked as
2557 // filled.
2558 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2559 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2560 }
2561 #endif // #ifdef ASSERT
2562
2563 if (beg_addr != space(space_id)->bottom()) {
2564 // Find the first live object or block of dead space that *starts* in this
2565 // range of regions. If a partial object crosses onto the region, skip it;
2566 // it will be marked for 'deferred update' when the object head is
2567 // processed. If dead space crosses onto the region, it is also skipped; it
2568 // will be filled when the prior region is processed. If neither of those
2569 // apply, the first word in the region is the start of a live object or dead
2570 // space.
2571 assert(beg_addr > space(space_id)->bottom(), "sanity");
2572 const RegionData* const cp = sd.region(beg_region);
2573 if (cp->partial_obj_size() != 0) {
2574 beg_addr = sd.partial_obj_end(beg_region);
2575 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2576 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2577 }
2578 }
2579
2580 if (beg_addr < end_addr) {
2581 // A live object or block of dead space starts in this range of Regions.
2582 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2583
2584 // Create closures and iterate.
2585 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2586 FillClosure fill_closure(cm, space_id);
2587 ParMarkBitMap::IterationStatus status;
2588 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2589 dense_prefix_end);
2590 if (status == ParMarkBitMap::incomplete) {
2591 update_closure.do_addr(update_closure.source());
2592 }
2593 }
2594
2595 // Mark the regions as filled.
2596 RegionData* const beg_cp = sd.region(beg_region);
2597 RegionData* const end_cp = sd.region(end_region);
2598 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2599 cp->set_completed();
2600 }
2601 }
2602
2603 // Return the SpaceId for the space containing addr. If addr is not in the
2604 // heap, last_space_id is returned. In debug mode it expects the address to be
2605 // in the heap and asserts such.
2606 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2607 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2608
2609 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2610 if (_space_info[id].space()->contains(addr)) {
2611 return SpaceId(id);
2612 }
2613 }
2614
2615 assert(false, "no space contains the addr");
2616 return last_space_id;
2617 }
2618
2619 void PSParallelCompact::update_deferred_object(ParCompactionManager* cm, HeapWord *addr) {
2620 #ifdef ASSERT
2621 ParallelCompactData& sd = summary_data();
2622 size_t region_idx = sd.addr_to_region_idx(addr);
2623 assert(sd.region(region_idx)->completed(), "first region must be completed before deferred updates");
2624 assert(sd.region(region_idx + 1)->completed(), "second region must be completed before deferred updates");
2625 #endif
2626
2627 const SpaceInfo* const space_info = _space_info + space_id(addr);
2628 ObjectStartArray* const start_array = space_info->start_array();
2629 if (start_array != nullptr) {
2630 start_array->update_for_block(addr, addr + cast_to_oop(addr)->size());
2631 }
2632
2633 cm->update_contents(cast_to_oop(addr));
2634 assert(oopDesc::is_oop(cast_to_oop(addr)), "Expected an oop at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2635 }
2636
2637 // Skip over count live words starting from beg, and return the address of the
2638 // next live word. Unless marked, the word corresponding to beg is assumed to
2639 // be dead. Callers must either ensure beg does not correspond to the middle of
2640 // an object, or account for those live words in some other way. Callers must
2641 // also ensure that there are enough live words in the range [beg, end) to skip.
2642 HeapWord*
2643 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2644 {
2645 assert(count > 0, "sanity");
2646
2647 ParMarkBitMap* m = mark_bitmap();
2648 idx_t bits_to_skip = m->words_to_bits(count);
2649 idx_t cur_beg = m->addr_to_bit(beg);
2650 const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2651
2652 do {
2653 cur_beg = m->find_obj_beg(cur_beg, search_end);
2654 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2655 const size_t obj_bits = cur_end - cur_beg + 1;
2656 if (obj_bits > bits_to_skip) {
2657 return m->bit_to_addr(cur_beg + bits_to_skip);
2658 }
2659 bits_to_skip -= obj_bits;
2660 cur_beg = cur_end + 1;
2661 } while (bits_to_skip > 0);
2662
2663 // Skipping the desired number of words landed just past the end of an object.
2664 // Find the start of the next object.
2665 cur_beg = m->find_obj_beg(cur_beg, search_end);
2666 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2667 return m->bit_to_addr(cur_beg);
2668 }
2669
2670 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2671 SpaceId src_space_id,
2672 size_t src_region_idx)
2673 {
2674 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2675
2676 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2677 if (split_info.dest_region_addr() == dest_addr) {
2678 // The partial object ending at the split point contains the first word to
2679 // be copied to dest_addr.
2680 return split_info.first_src_addr();
2681 }
2682
2683 const ParallelCompactData& sd = summary_data();
2684 ParMarkBitMap* const bitmap = mark_bitmap();
2685 const size_t RegionSize = ParallelCompactData::RegionSize;
2686
2687 assert(sd.is_region_aligned(dest_addr), "not aligned");
2688 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2689 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2690 HeapWord* const src_region_destination = src_region_ptr->destination();
2691
2692 assert(dest_addr >= src_region_destination, "wrong src region");
2693 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2694
2695 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2696 HeapWord* const src_region_end = src_region_beg + RegionSize;
2697
2698 HeapWord* addr = src_region_beg;
2699 if (dest_addr == src_region_destination) {
2700 // Return the first live word in the source region.
2701 if (partial_obj_size == 0) {
2702 addr = bitmap->find_obj_beg(addr, src_region_end);
2703 assert(addr < src_region_end, "no objects start in src region");
2704 }
2705 return addr;
2706 }
2707
2708 // Must skip some live data.
2709 size_t words_to_skip = dest_addr - src_region_destination;
2710 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2711
2712 if (partial_obj_size >= words_to_skip) {
2713 // All the live words to skip are part of the partial object.
2714 addr += words_to_skip;
2715 if (partial_obj_size == words_to_skip) {
2716 // Find the first live word past the partial object.
2717 addr = bitmap->find_obj_beg(addr, src_region_end);
2718 assert(addr < src_region_end, "wrong src region");
2719 }
2720 return addr;
2721 }
2722
2723 // Skip over the partial object (if any).
2724 if (partial_obj_size != 0) {
2725 words_to_skip -= partial_obj_size;
2726 addr += partial_obj_size;
2727 }
2728
2729 // Skip over live words due to objects that start in the region.
2730 addr = skip_live_words(addr, src_region_end, words_to_skip);
2731 assert(addr < src_region_end, "wrong src region");
2732 return addr;
2733 }
2734
2735 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2736 SpaceId src_space_id,
2737 size_t beg_region,
2738 HeapWord* end_addr)
2739 {
2740 ParallelCompactData& sd = summary_data();
2741
2742 #ifdef ASSERT
2743 MutableSpace* const src_space = _space_info[src_space_id].space();
2744 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2745 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2746 "src_space_id does not match beg_addr");
2747 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2748 "src_space_id does not match end_addr");
2749 #endif // #ifdef ASSERT
2750
2751 RegionData* const beg = sd.region(beg_region);
2752 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2753
2754 // Regions up to new_top() are enqueued if they become available.
2755 HeapWord* const new_top = _space_info[src_space_id].new_top();
2756 RegionData* const enqueue_end =
2757 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2758
2759 for (RegionData* cur = beg; cur < end; ++cur) {
2760 assert(cur->data_size() > 0, "region must have live data");
2761 cur->decrement_destination_count();
2762 if (cur < enqueue_end && cur->available() && cur->claim()) {
2763 if (cur->mark_normal()) {
2764 cm->push_region(sd.region(cur));
2765 } else if (cur->mark_copied()) {
2766 // Try to copy the content of the shadow region back to its corresponding
2767 // heap region if the shadow region is filled. Otherwise, the GC thread
2768 // fills the shadow region will copy the data back (see
2769 // MoveAndUpdateShadowClosure::complete_region).
2770 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2771 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2772 cur->set_completed();
2773 }
2774 }
2775 }
2776 }
2777
2778 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2779 SpaceId& src_space_id,
2780 HeapWord*& src_space_top,
2781 HeapWord* end_addr)
2782 {
2783 typedef ParallelCompactData::RegionData RegionData;
2784
2785 ParallelCompactData& sd = PSParallelCompact::summary_data();
2786 const size_t region_size = ParallelCompactData::RegionSize;
2787
2788 size_t src_region_idx = 0;
2789
2790 // Skip empty regions (if any) up to the top of the space.
2791 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2792 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2793 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2794 const RegionData* const top_region_ptr =
2795 sd.addr_to_region_ptr(top_aligned_up);
2796 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2797 ++src_region_ptr;
2798 }
2799
2800 if (src_region_ptr < top_region_ptr) {
2801 // The next source region is in the current space. Update src_region_idx
2802 // and the source address to match src_region_ptr.
2803 src_region_idx = sd.region(src_region_ptr);
2804 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2805 if (src_region_addr > closure.source()) {
2806 closure.set_source(src_region_addr);
2807 }
2808 return src_region_idx;
2809 }
2810
2811 // Switch to a new source space and find the first non-empty region.
2812 unsigned int space_id = src_space_id + 1;
2813 assert(space_id < last_space_id, "not enough spaces");
2814
2815 HeapWord* const destination = closure.destination();
2816
2817 do {
2818 MutableSpace* space = _space_info[space_id].space();
2819 HeapWord* const bottom = space->bottom();
2820 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2821
2822 // Iterate over the spaces that do not compact into themselves.
2823 if (bottom_cp->destination() != bottom) {
2824 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2825 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2826
2827 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2828 if (src_cp->live_obj_size() > 0) {
2829 // Found it.
2830 assert(src_cp->destination() == destination,
2831 "first live obj in the space must match the destination");
2832 assert(src_cp->partial_obj_size() == 0,
2833 "a space cannot begin with a partial obj");
2834
2835 src_space_id = SpaceId(space_id);
2836 src_space_top = space->top();
2837 const size_t src_region_idx = sd.region(src_cp);
2838 closure.set_source(sd.region_to_addr(src_region_idx));
2839 return src_region_idx;
2840 } else {
2841 assert(src_cp->data_size() == 0, "sanity");
2842 }
2843 }
2844 }
2845 } while (++space_id < last_space_id);
2846
2847 assert(false, "no source region was found");
2848 return 0;
2849 }
2850
2851 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2852 {
2853 typedef ParMarkBitMap::IterationStatus IterationStatus;
2854 ParMarkBitMap* const bitmap = mark_bitmap();
2855 ParallelCompactData& sd = summary_data();
2856 RegionData* const region_ptr = sd.region(region_idx);
2857
2858 // Get the source region and related info.
2859 size_t src_region_idx = region_ptr->source_region();
2860 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2861 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2862 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2863
2864 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2865
2866 // Adjust src_region_idx to prepare for decrementing destination counts (the
2867 // destination count is not decremented when a region is copied to itself).
2868 if (src_region_idx == region_idx) {
2869 src_region_idx += 1;
2870 }
2871
2872 if (bitmap->is_unmarked(closure.source())) {
2873 // The first source word is in the middle of an object; copy the remainder
2874 // of the object or as much as will fit. The fact that pointer updates were
2875 // deferred will be noted when the object header is processed.
2876 HeapWord* const old_src_addr = closure.source();
2877 closure.copy_partial_obj();
2878 if (closure.is_full()) {
2879 decrement_destination_counts(cm, src_space_id, src_region_idx,
2880 closure.source());
2881 closure.complete_region(cm, dest_addr, region_ptr);
2882 return;
2883 }
2884
2885 HeapWord* const end_addr = sd.region_align_down(closure.source());
2886 if (sd.region_align_down(old_src_addr) != end_addr) {
2887 // The partial object was copied from more than one source region.
2888 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2889
2890 // Move to the next source region, possibly switching spaces as well. All
2891 // args except end_addr may be modified.
2892 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2893 end_addr);
2894 }
2895 }
2896
2897 do {
2898 HeapWord* const cur_addr = closure.source();
2899 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2900 src_space_top);
2901 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2902
2903 if (status == ParMarkBitMap::incomplete) {
2904 // The last obj that starts in the source region does not end in the
2905 // region.
2906 assert(closure.source() < end_addr, "sanity");
2907 HeapWord* const obj_beg = closure.source();
2908 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2909 src_space_top);
2910 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2911 if (obj_end < range_end) {
2912 // The end was found; the entire object will fit.
2913 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2914 assert(status != ParMarkBitMap::would_overflow, "sanity");
2915 } else {
2916 // The end was not found; the object will not fit.
2917 assert(range_end < src_space_top, "obj cannot cross space boundary");
2918 status = ParMarkBitMap::would_overflow;
2919 }
2920 }
2921
2922 if (status == ParMarkBitMap::would_overflow) {
2923 // The last object did not fit. Note that interior oop updates were
2924 // deferred, then copy enough of the object to fill the region.
2925 cm->push_deferred_object(closure.destination());
2926 status = closure.copy_until_full(); // copies from closure.source()
2927
2928 decrement_destination_counts(cm, src_space_id, src_region_idx,
2929 closure.source());
2930 closure.complete_region(cm, dest_addr, region_ptr);
2931 return;
2932 }
2933
2934 if (status == ParMarkBitMap::full) {
2935 decrement_destination_counts(cm, src_space_id, src_region_idx,
2936 closure.source());
2937 closure.complete_region(cm, dest_addr, region_ptr);
2938 return;
2939 }
2940
2941 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2942
2943 // Move to the next source region, possibly switching spaces as well. All
2944 // args except end_addr may be modified.
2945 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2946 end_addr);
2947 } while (true);
2948 }
2949
2950 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2951 {
2952 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
2953 fill_region(cm, cl, region_idx);
2954 }
2955
2956 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2957 {
2958 // Get a shadow region first
2959 ParallelCompactData& sd = summary_data();
2960 RegionData* const region_ptr = sd.region(region_idx);
2961 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2962 // The InvalidShadow return value indicates the corresponding heap region is available,
2963 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2964 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2965 if (shadow_region == ParCompactionManager::InvalidShadow) {
2966 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
2967 region_ptr->shadow_to_normal();
2968 return fill_region(cm, cl, region_idx);
2969 } else {
2970 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
2971 return fill_region(cm, cl, region_idx);
2972 }
2973 }
2974
2975 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2976 {
2977 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2978 }
2979
2980 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
2981 {
2982 size_t next = cm->next_shadow_region();
2983 ParallelCompactData& sd = summary_data();
2984 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2985 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2986
2987 while (next < old_new_top) {
2988 if (sd.region(next)->mark_shadow()) {
2989 region_idx = next;
2990 return true;
2991 }
2992 next = cm->move_next_shadow_region_by(active_gc_threads);
2993 }
2994
2995 return false;
2996 }
2997
2998 // The shadow region is an optimization to address region dependencies in full GC. The basic
2999 // idea is making more regions available by temporally storing their live objects in empty
3000 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3001 // GC threads need not wait destination regions to be available before processing sources.
3002 //
3003 // A typical workflow would be:
3004 // After draining its own stack and failing to steal from others, a GC worker would pick an
3005 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3006 // the shadow region by copying live objects from source regions of the unavailable one. Once
3007 // the unavailable region becomes available, the data in the shadow region will be copied back.
3008 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3009 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3010 {
3011 const ParallelCompactData& sd = PSParallelCompact::summary_data();
3012
3013 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3014 SpaceInfo* const space_info = _space_info + id;
3015 MutableSpace* const space = space_info->space();
3016
3017 const size_t beg_region =
3018 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3019 const size_t end_region =
3020 sd.addr_to_region_idx(sd.region_align_down(space->end()));
3021
3022 for (size_t cur = beg_region; cur < end_region; ++cur) {
3023 ParCompactionManager::push_shadow_region(cur);
3024 }
3025 }
3026
3027 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3028 for (uint i = 0; i < parallel_gc_threads; i++) {
3029 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3030 cm->set_next_shadow_region(beg_region + i);
3031 }
3032 }
3033
3034 void PSParallelCompact::fill_blocks(size_t region_idx)
3035 {
3036 // Fill in the block table elements for the specified region. Each block
3037 // table element holds the number of live words in the region that are to the
3038 // left of the first object that starts in the block. Thus only blocks in
3039 // which an object starts need to be filled.
3040 //
3041 // The algorithm scans the section of the bitmap that corresponds to the
3042 // region, keeping a running total of the live words. When an object start is
3043 // found, if it's the first to start in the block that contains it, the
3044 // current total is written to the block table element.
3045 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3046 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3047 const size_t RegionSize = ParallelCompactData::RegionSize;
3048
3049 ParallelCompactData& sd = summary_data();
3050 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3051 if (partial_obj_size >= RegionSize) {
3052 return; // No objects start in this region.
3053 }
3054
3055 // Ensure the first loop iteration decides that the block has changed.
3056 size_t cur_block = sd.block_count();
3057
3058 const ParMarkBitMap* const bitmap = mark_bitmap();
3059
3060 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3061 assert((size_t)1 << Log2BitsPerBlock ==
3062 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3063
3064 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3065 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3066 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3067 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3068 while (beg_bit < range_end) {
3069 const size_t new_block = beg_bit >> Log2BitsPerBlock;
3070 if (new_block != cur_block) {
3071 cur_block = new_block;
3072 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3073 }
3074
3075 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3076 if (end_bit < range_end - 1) {
3077 live_bits += end_bit - beg_bit + 1;
3078 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3079 } else {
3080 return;
3081 }
3082 }
3083 }
3084
3085 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3086 {
3087 if (source() != copy_destination()) {
3088 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3089 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3090 }
3091 update_state(words_remaining());
3092 assert(is_full(), "sanity");
3093 return ParMarkBitMap::full;
3094 }
3095
3096 void MoveAndUpdateClosure::copy_partial_obj()
3097 {
3098 size_t words = words_remaining();
3099
3100 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3101 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3102 if (end_addr < range_end) {
3103 words = bitmap()->obj_size(source(), end_addr);
3104 }
3105
3106 // This test is necessary; if omitted, the pointer updates to a partial object
3107 // that crosses the dense prefix boundary could be overwritten.
3108 if (source() != copy_destination()) {
3109 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3110 Copy::aligned_conjoint_words(source(), copy_destination(), words);
3111 }
3112 update_state(words);
3113 }
3114
3115 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3116 PSParallelCompact::RegionData *region_ptr) {
3117 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3118 region_ptr->set_completed();
3119 }
3120
3121 ParMarkBitMapClosure::IterationStatus
3122 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3123 assert(destination() != nullptr, "sanity");
3124 assert(bitmap()->obj_size(addr) == words, "bad size");
3125
3126 _source = addr;
3127 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3128 destination(), "wrong destination");
3129
3130 if (words > words_remaining()) {
3131 return ParMarkBitMap::would_overflow;
3132 }
3133
3134 // The start_array must be updated even if the object is not moving.
3135 if (_start_array != nullptr) {
3136 _start_array->update_for_block(destination(), destination() + words);
3137 }
3138
3139 if (copy_destination() != source()) {
3140 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3141 Copy::aligned_conjoint_words(source(), copy_destination(), words);
3142 }
3143
3144 oop moved_oop = cast_to_oop(copy_destination());
3145 compaction_manager()->update_contents(moved_oop);
3146 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or null at " PTR_FORMAT, p2i(moved_oop));
3147
3148 update_state(words);
3149 assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3150 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3151 }
3152
3153 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3154 PSParallelCompact::RegionData *region_ptr) {
3155 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3156 // Record the shadow region index
3157 region_ptr->set_shadow_region(_shadow);
3158 // Mark the shadow region as filled to indicate the data is ready to be
3159 // copied back
3160 region_ptr->mark_filled();
3161 // Try to copy the content of the shadow region back to its corresponding
3162 // heap region if available; the GC thread that decreases the destination
3163 // count to zero will do the copying otherwise (see
3164 // PSParallelCompact::decrement_destination_counts).
3165 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3166 region_ptr->set_completed();
3167 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3168 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3169 }
3170 }
3171
3172 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3173 ParCompactionManager* cm,
3174 PSParallelCompact::SpaceId space_id) :
3175 ParMarkBitMapClosure(mbm, cm),
3176 _start_array(PSParallelCompact::start_array(space_id))
3177 {
3178 }
3179
3180 // Updates the references in the object to their new values.
3181 ParMarkBitMapClosure::IterationStatus
3182 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3183 do_addr(addr);
3184 return ParMarkBitMap::incomplete;
3185 }
3186
3187 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3188 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3189 _start_array(PSParallelCompact::start_array(space_id))
3190 {
3191 assert(space_id == PSParallelCompact::old_space_id,
3192 "cannot use FillClosure in the young gen");
3193 }
3194
3195 ParMarkBitMapClosure::IterationStatus
3196 FillClosure::do_addr(HeapWord* addr, size_t size) {
3197 CollectedHeap::fill_with_objects(addr, size);
3198 HeapWord* const end = addr + size;
3199 do {
3200 size_t size = cast_to_oop(addr)->size();
3201 _start_array->update_for_block(addr, addr + size);
3202 addr += size;
3203 } while (addr < end);
3204 return ParMarkBitMap::incomplete;
3205 }