1 /*
2 * Copyright (c) 2005, 2019, 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 "aot/aotLoader.hpp"
27 #include "classfile/classLoaderDataGraph.hpp"
28 #include "classfile/javaClasses.inline.hpp"
29 #include "classfile/stringTable.hpp"
30 #include "classfile/symbolTable.hpp"
31 #include "classfile/systemDictionary.hpp"
32 #include "code/codeCache.hpp"
33 #include "gc/parallel/gcTaskManager.hpp"
34 #include "gc/parallel/parallelArguments.hpp"
35 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
36 #include "gc/parallel/parMarkBitMap.inline.hpp"
37 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
38 #include "gc/parallel/psCompactionManager.inline.hpp"
39 #include "gc/parallel/psOldGen.hpp"
40 #include "gc/parallel/psParallelCompact.inline.hpp"
41 #include "gc/parallel/psPromotionManager.inline.hpp"
42 #include "gc/parallel/psRootType.inline.hpp"
43 #include "gc/parallel/psScavenge.hpp"
44 #include "gc/parallel/psYoungGen.hpp"
45 #include "gc/shared/gcCause.hpp"
46 #include "gc/shared/gcHeapSummary.hpp"
47 #include "gc/shared/gcId.hpp"
48 #include "gc/shared/gcLocker.hpp"
49 #include "gc/shared/gcTimer.hpp"
50 #include "gc/shared/gcTrace.hpp"
51 #include "gc/shared/gcTraceTime.inline.hpp"
52 #include "gc/shared/isGCActiveMark.hpp"
53 #include "gc/shared/referencePolicy.hpp"
54 #include "gc/shared/referenceProcessor.hpp"
55 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
56 #include "gc/shared/spaceDecorator.hpp"
57 #include "gc/shared/weakProcessor.hpp"
58 #include "gc/shared/workerPolicy.hpp"
59 #include "gc/shared/workgroup.hpp"
60 #if INCLUDE_JVMCI
61 #include "jvmci/jvmci.hpp"
62 #endif
63 #include "logging/log.hpp"
64 #include "memory/iterator.inline.hpp"
65 #include "memory/resourceArea.hpp"
66 #include "memory/universe.hpp"
67 #include "oops/access.inline.hpp"
68 #include "oops/instanceClassLoaderKlass.inline.hpp"
69 #include "oops/instanceKlass.inline.hpp"
70 #include "oops/instanceMirrorKlass.inline.hpp"
71 #include "oops/methodData.hpp"
72 #include "oops/objArrayKlass.inline.hpp"
73 #include "oops/oop.inline.hpp"
74 #include "runtime/atomic.hpp"
75 #include "runtime/handles.inline.hpp"
76 #include "runtime/safepoint.hpp"
77 #include "runtime/vmThread.hpp"
78 #include "services/management.hpp"
79 #include "services/memTracker.hpp"
80 #include "services/memoryService.hpp"
81 #include "utilities/align.hpp"
82 #include "utilities/debug.hpp"
83 #include "utilities/events.hpp"
84 #include "utilities/formatBuffer.hpp"
85 #include "utilities/macros.hpp"
86 #include "utilities/stack.inline.hpp"
87
88 #include <math.h>
89
90 // All sizes are in HeapWords.
91 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
92 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
93 const size_t ParallelCompactData::RegionSizeBytes =
94 RegionSize << LogHeapWordSize;
95 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
96 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
97 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
98
99 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
100 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
101 const size_t ParallelCompactData::BlockSizeBytes =
102 BlockSize << LogHeapWordSize;
103 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
104 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
105 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
106
107 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
108 const size_t ParallelCompactData::Log2BlocksPerRegion =
109 Log2RegionSize - Log2BlockSize;
110
111 const ParallelCompactData::RegionData::region_sz_t
112 ParallelCompactData::RegionData::dc_shift = 27;
113
114 const ParallelCompactData::RegionData::region_sz_t
115 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
116
117 const ParallelCompactData::RegionData::region_sz_t
118 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
119
120 const ParallelCompactData::RegionData::region_sz_t
121 ParallelCompactData::RegionData::los_mask = ~dc_mask;
122
123 const ParallelCompactData::RegionData::region_sz_t
124 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
125
126 const ParallelCompactData::RegionData::region_sz_t
127 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
128
129 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
130
131 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
132 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
133
134 double PSParallelCompact::_dwl_mean;
135 double PSParallelCompact::_dwl_std_dev;
136 double PSParallelCompact::_dwl_first_term;
137 double PSParallelCompact::_dwl_adjustment;
138 #ifdef ASSERT
139 bool PSParallelCompact::_dwl_initialized = false;
140 #endif // #ifdef ASSERT
141
142 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
143 HeapWord* destination)
144 {
145 assert(src_region_idx != 0, "invalid src_region_idx");
146 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
147 assert(destination != NULL, "invalid destination argument");
148
149 _src_region_idx = src_region_idx;
150 _partial_obj_size = partial_obj_size;
151 _destination = destination;
152
153 // These fields may not be updated below, so make sure they're clear.
154 assert(_dest_region_addr == NULL, "should have been cleared");
155 assert(_first_src_addr == NULL, "should have been cleared");
156
157 // Determine the number of destination regions for the partial object.
158 HeapWord* const last_word = destination + partial_obj_size - 1;
159 const ParallelCompactData& sd = PSParallelCompact::summary_data();
160 HeapWord* const beg_region_addr = sd.region_align_down(destination);
161 HeapWord* const end_region_addr = sd.region_align_down(last_word);
162
163 if (beg_region_addr == end_region_addr) {
164 // One destination region.
165 _destination_count = 1;
166 if (end_region_addr == destination) {
167 // The destination falls on a region boundary, thus the first word of the
168 // partial object will be the first word copied to the destination region.
169 _dest_region_addr = end_region_addr;
170 _first_src_addr = sd.region_to_addr(src_region_idx);
171 }
172 } else {
173 // Two destination regions. When copied, the partial object will cross a
174 // destination region boundary, so a word somewhere within the partial
175 // object will be the first word copied to the second destination region.
176 _destination_count = 2;
177 _dest_region_addr = end_region_addr;
178 const size_t ofs = pointer_delta(end_region_addr, destination);
179 assert(ofs < _partial_obj_size, "sanity");
180 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
181 }
182 }
183
184 void SplitInfo::clear()
185 {
186 _src_region_idx = 0;
187 _partial_obj_size = 0;
188 _destination = NULL;
189 _destination_count = 0;
190 _dest_region_addr = NULL;
191 _first_src_addr = NULL;
192 assert(!is_valid(), "sanity");
193 }
194
195 #ifdef ASSERT
196 void SplitInfo::verify_clear()
197 {
198 assert(_src_region_idx == 0, "not clear");
199 assert(_partial_obj_size == 0, "not clear");
200 assert(_destination == NULL, "not clear");
201 assert(_destination_count == 0, "not clear");
202 assert(_dest_region_addr == NULL, "not clear");
203 assert(_first_src_addr == NULL, "not clear");
204 }
205 #endif // #ifdef ASSERT
206
207
208 void PSParallelCompact::print_on_error(outputStream* st) {
209 _mark_bitmap.print_on_error(st);
210 }
211
212 #ifndef PRODUCT
213 const char* PSParallelCompact::space_names[] = {
214 "old ", "eden", "from", "to "
215 };
216
217 void PSParallelCompact::print_region_ranges() {
218 if (!log_develop_is_enabled(Trace, gc, compaction)) {
219 return;
220 }
221 Log(gc, compaction) log;
222 ResourceMark rm;
223 LogStream ls(log.trace());
224 Universe::print_on(&ls);
225 log.trace("space bottom top end new_top");
226 log.trace("------ ---------- ---------- ---------- ----------");
227
228 for (unsigned int id = 0; id < last_space_id; ++id) {
229 const MutableSpace* space = _space_info[id].space();
230 log.trace("%u %s "
231 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
232 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
233 id, space_names[id],
234 summary_data().addr_to_region_idx(space->bottom()),
235 summary_data().addr_to_region_idx(space->top()),
236 summary_data().addr_to_region_idx(space->end()),
237 summary_data().addr_to_region_idx(_space_info[id].new_top()));
238 }
239 }
240
241 void
242 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
243 {
244 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
245 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
246
247 ParallelCompactData& sd = PSParallelCompact::summary_data();
248 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
249 log_develop_trace(gc, compaction)(
250 REGION_IDX_FORMAT " " PTR_FORMAT " "
251 REGION_IDX_FORMAT " " PTR_FORMAT " "
252 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
253 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
254 i, p2i(c->data_location()), dci, p2i(c->destination()),
255 c->partial_obj_size(), c->live_obj_size(),
256 c->data_size(), c->source_region(), c->destination_count());
257
258 #undef REGION_IDX_FORMAT
259 #undef REGION_DATA_FORMAT
260 }
261
262 void
263 print_generic_summary_data(ParallelCompactData& summary_data,
264 HeapWord* const beg_addr,
265 HeapWord* const end_addr)
266 {
267 size_t total_words = 0;
268 size_t i = summary_data.addr_to_region_idx(beg_addr);
269 const size_t last = summary_data.addr_to_region_idx(end_addr);
270 HeapWord* pdest = 0;
271
272 while (i < last) {
273 ParallelCompactData::RegionData* c = summary_data.region(i);
274 if (c->data_size() != 0 || c->destination() != pdest) {
275 print_generic_summary_region(i, c);
276 total_words += c->data_size();
277 pdest = c->destination();
278 }
279 ++i;
280 }
281
282 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
283 }
284
285 void
286 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
287 HeapWord* const beg_addr,
288 HeapWord* const end_addr) {
289 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
290 }
291
292 void
293 print_generic_summary_data(ParallelCompactData& summary_data,
294 SpaceInfo* space_info)
295 {
296 if (!log_develop_is_enabled(Trace, gc, compaction)) {
297 return;
298 }
299
300 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
301 const MutableSpace* space = space_info[id].space();
302 print_generic_summary_data(summary_data, space->bottom(),
303 MAX2(space->top(), space_info[id].new_top()));
304 }
305 }
306
307 void
308 print_initial_summary_data(ParallelCompactData& summary_data,
309 const MutableSpace* space) {
310 if (space->top() == space->bottom()) {
311 return;
312 }
313
314 const size_t region_size = ParallelCompactData::RegionSize;
315 typedef ParallelCompactData::RegionData RegionData;
316 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
317 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
318 const RegionData* c = summary_data.region(end_region - 1);
319 HeapWord* end_addr = c->destination() + c->data_size();
320 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
321
322 // Print (and count) the full regions at the beginning of the space.
323 size_t full_region_count = 0;
324 size_t i = summary_data.addr_to_region_idx(space->bottom());
325 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
326 ParallelCompactData::RegionData* c = summary_data.region(i);
327 log_develop_trace(gc, compaction)(
328 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
329 i, p2i(c->destination()),
330 c->partial_obj_size(), c->live_obj_size(),
331 c->data_size(), c->source_region(), c->destination_count());
332 ++full_region_count;
333 ++i;
334 }
335
336 size_t live_to_right = live_in_space - full_region_count * region_size;
337
338 double max_reclaimed_ratio = 0.0;
339 size_t max_reclaimed_ratio_region = 0;
340 size_t max_dead_to_right = 0;
341 size_t max_live_to_right = 0;
342
343 // Print the 'reclaimed ratio' for regions while there is something live in
344 // the region or to the right of it. The remaining regions are empty (and
345 // uninteresting), and computing the ratio will result in division by 0.
346 while (i < end_region && live_to_right > 0) {
347 c = summary_data.region(i);
348 HeapWord* const region_addr = summary_data.region_to_addr(i);
349 const size_t used_to_right = pointer_delta(space->top(), region_addr);
350 const size_t dead_to_right = used_to_right - live_to_right;
351 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
352
353 if (reclaimed_ratio > max_reclaimed_ratio) {
354 max_reclaimed_ratio = reclaimed_ratio;
355 max_reclaimed_ratio_region = i;
356 max_dead_to_right = dead_to_right;
357 max_live_to_right = live_to_right;
358 }
359
360 ParallelCompactData::RegionData* c = summary_data.region(i);
361 log_develop_trace(gc, compaction)(
362 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
363 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
364 i, p2i(c->destination()),
365 c->partial_obj_size(), c->live_obj_size(),
366 c->data_size(), c->source_region(), c->destination_count(),
367 reclaimed_ratio, dead_to_right, live_to_right);
368
369
370 live_to_right -= c->data_size();
371 ++i;
372 }
373
374 // Any remaining regions are empty. Print one more if there is one.
375 if (i < end_region) {
376 ParallelCompactData::RegionData* c = summary_data.region(i);
377 log_develop_trace(gc, compaction)(
378 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
379 i, p2i(c->destination()),
380 c->partial_obj_size(), c->live_obj_size(),
381 c->data_size(), c->source_region(), c->destination_count());
382 }
383
384 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
385 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
386 }
387
388 void
389 print_initial_summary_data(ParallelCompactData& summary_data,
390 SpaceInfo* space_info) {
391 if (!log_develop_is_enabled(Trace, gc, compaction)) {
392 return;
393 }
394
395 unsigned int id = PSParallelCompact::old_space_id;
396 const MutableSpace* space;
397 do {
398 space = space_info[id].space();
399 print_initial_summary_data(summary_data, space);
400 } while (++id < PSParallelCompact::eden_space_id);
401
402 do {
403 space = space_info[id].space();
404 print_generic_summary_data(summary_data, space->bottom(), space->top());
405 } while (++id < PSParallelCompact::last_space_id);
406 }
407 #endif // #ifndef PRODUCT
408
409 #ifdef ASSERT
410 size_t add_obj_count;
411 size_t add_obj_size;
412 size_t mark_bitmap_count;
413 size_t mark_bitmap_size;
414 #endif // #ifdef ASSERT
415
416 ParallelCompactData::ParallelCompactData() :
417 _region_start(NULL),
418 DEBUG_ONLY(_region_end(NULL) COMMA)
419 _region_vspace(NULL),
420 _reserved_byte_size(0),
421 _region_data(NULL),
422 _region_count(0),
423 _block_vspace(NULL),
424 _block_data(NULL),
425 _block_count(0) {}
426
427 bool ParallelCompactData::initialize(MemRegion covered_region)
428 {
429 _region_start = covered_region.start();
430 const size_t region_size = covered_region.word_size();
431 DEBUG_ONLY(_region_end = _region_start + region_size;)
432
433 assert(region_align_down(_region_start) == _region_start,
434 "region start not aligned");
435 assert((region_size & RegionSizeOffsetMask) == 0,
436 "region size not a multiple of RegionSize");
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 == (size_t) os::vm_page_size() ? 0 :
451 MAX2(page_sz, granularity);
452 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
453 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
454 rs.size());
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 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
474 _region_vspace = create_vspace(count, sizeof(RegionData));
475 if (_region_vspace != 0) {
476 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
477 _region_count = count;
478 return true;
479 }
480 return false;
481 }
482
483 bool ParallelCompactData::initialize_block_data()
484 {
485 assert(_region_count != 0, "region data must be initialized first");
486 const size_t count = _region_count << Log2BlocksPerRegion;
487 _block_vspace = create_vspace(count, sizeof(BlockData));
488 if (_block_vspace != 0) {
489 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
490 _block_count = count;
491 return true;
492 }
493 return false;
494 }
495
496 void ParallelCompactData::clear()
497 {
498 memset(_region_data, 0, _region_vspace->committed_size());
499 memset(_block_data, 0, _block_vspace->committed_size());
500 }
501
502 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
503 assert(beg_region <= _region_count, "beg_region out of range");
504 assert(end_region <= _region_count, "end_region out of range");
505 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
506
507 const size_t region_cnt = end_region - beg_region;
508 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
509
510 const size_t beg_block = beg_region * BlocksPerRegion;
511 const size_t block_cnt = region_cnt * BlocksPerRegion;
512 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
513 }
514
515 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
516 {
517 const RegionData* cur_cp = region(region_idx);
518 const RegionData* const end_cp = region(region_count() - 1);
519
520 HeapWord* result = region_to_addr(region_idx);
521 if (cur_cp < end_cp) {
522 do {
523 result += cur_cp->partial_obj_size();
524 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
525 }
526 return result;
527 }
528
529 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
530 {
531 const size_t obj_ofs = pointer_delta(addr, _region_start);
532 const size_t beg_region = obj_ofs >> Log2RegionSize;
533 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
534
535 DEBUG_ONLY(Atomic::inc(&add_obj_count);)
536 DEBUG_ONLY(Atomic::add(len, &add_obj_size);)
537
538 if (beg_region == end_region) {
539 // All in one region.
540 _region_data[beg_region].add_live_obj(len);
541 return;
542 }
543
544 // First region.
545 const size_t beg_ofs = region_offset(addr);
546 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
547
548 Klass* klass = ((oop)addr)->klass();
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(region_offset(beg) == 0, "not RegionSize aligned");
565 assert(region_offset(end) == 0, "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 == NULL ? NULL : *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 != NULL, "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 (region_offset(dest_addr) == 0) {
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) {
780 assert(addr != NULL, "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, 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::_total_invocations = 0;
843 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
844 jlong PSParallelCompact::_time_of_last_gc = 0;
845 CollectorCounters* PSParallelCompact::_counters = NULL;
846 ParMarkBitMap PSParallelCompact::_mark_bitmap;
847 ParallelCompactData PSParallelCompact::_summary_data;
848
849 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
850
851 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
852
853 class PCReferenceProcessor: public ReferenceProcessor {
854 public:
855 PCReferenceProcessor(
856 BoolObjectClosure* is_subject_to_discovery,
857 BoolObjectClosure* is_alive_non_header) :
858 ReferenceProcessor(is_subject_to_discovery,
859 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
860 ParallelGCThreads, // mt processing degree
861 true, // mt discovery
862 ParallelGCThreads, // mt discovery degree
863 true, // atomic_discovery
864 is_alive_non_header) {
865 }
866
867 template<typename T> bool discover(oop obj, ReferenceType type) {
868 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
869 T heap_oop = RawAccess<>::oop_load(referent_addr);
870 oop referent = CompressedOops::decode_not_null(heap_oop);
871 return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
872 && ReferenceProcessor::discover_reference(obj, type);
873 }
874 virtual bool discover_reference(oop obj, ReferenceType type) {
875 if (UseCompressedOops) {
876 return discover<narrowOop>(obj, type);
877 } else {
878 return discover<oop>(obj, type);
879 }
880 }
881 };
882
883 void PSParallelCompact::post_initialize() {
884 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
885 _span_based_discoverer.set_span(heap->reserved_region());
886 _ref_processor =
887 new PCReferenceProcessor(&_span_based_discoverer,
888 &_is_alive_closure); // non-header is alive closure
889
890 _counters = new CollectorCounters("Parallel full collection pauses", 1);
891
892 // Initialize static fields in ParCompactionManager.
893 ParCompactionManager::initialize(mark_bitmap());
894 }
895
896 bool PSParallelCompact::initialize() {
897 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
898 MemRegion mr = heap->reserved_region();
899
900 // Was the old gen get allocated successfully?
901 if (!heap->old_gen()->is_allocated()) {
902 return false;
903 }
904
905 initialize_space_info();
906 initialize_dead_wood_limiter();
907
908 if (!_mark_bitmap.initialize(mr)) {
909 vm_shutdown_during_initialization(
910 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
911 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
912 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
913 return false;
914 }
915
916 if (!_summary_data.initialize(mr)) {
917 vm_shutdown_during_initialization(
918 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
919 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
920 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
921 return false;
922 }
923
924 return true;
925 }
926
927 void PSParallelCompact::initialize_space_info()
928 {
929 memset(&_space_info, 0, sizeof(_space_info));
930
931 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
932 PSYoungGen* young_gen = heap->young_gen();
933
934 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
935 _space_info[eden_space_id].set_space(young_gen->eden_space());
936 _space_info[from_space_id].set_space(young_gen->from_space());
937 _space_info[to_space_id].set_space(young_gen->to_space());
938
939 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
940 }
941
942 void PSParallelCompact::initialize_dead_wood_limiter()
943 {
944 const size_t max = 100;
945 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
946 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
947 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
948 DEBUG_ONLY(_dwl_initialized = true;)
949 _dwl_adjustment = normal_distribution(1.0);
950 }
951
952 void
953 PSParallelCompact::clear_data_covering_space(SpaceId id)
954 {
955 // At this point, top is the value before GC, new_top() is the value that will
956 // be set at the end of GC. The marking bitmap is cleared to top; nothing
957 // should be marked above top. The summary data is cleared to the larger of
958 // top & new_top.
959 MutableSpace* const space = _space_info[id].space();
960 HeapWord* const bot = space->bottom();
961 HeapWord* const top = space->top();
962 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
963
964 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
965 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
966 _mark_bitmap.clear_range(beg_bit, end_bit);
967
968 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
969 const size_t end_region =
970 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
971 _summary_data.clear_range(beg_region, end_region);
972
973 // Clear the data used to 'split' regions.
974 SplitInfo& split_info = _space_info[id].split_info();
975 if (split_info.is_valid()) {
976 split_info.clear();
977 }
978 DEBUG_ONLY(split_info.verify_clear();)
979 }
980
981 void PSParallelCompact::pre_compact()
982 {
983 // Update the from & to space pointers in space_info, since they are swapped
984 // at each young gen gc. Do the update unconditionally (even though a
985 // promotion failure does not swap spaces) because an unknown number of young
986 // collections will have swapped the spaces an unknown number of times.
987 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
988 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
989 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
990 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
991
992 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
993 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
994
995 // Increment the invocation count
996 heap->increment_total_collections(true);
997
998 // We need to track unique mark sweep invocations as well.
999 _total_invocations++;
1000
1001 heap->print_heap_before_gc();
1002 heap->trace_heap_before_gc(&_gc_tracer);
1003
1004 // Fill in TLABs
1005 heap->ensure_parsability(true); // retire TLABs
1006
1007 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1008 HandleMark hm; // Discard invalid handles created during verification
1009 Universe::verify("Before GC");
1010 }
1011
1012 // Verify object start arrays
1013 if (VerifyObjectStartArray &&
1014 VerifyBeforeGC) {
1015 heap->old_gen()->verify_object_start_array();
1016 }
1017
1018 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1019 DEBUG_ONLY(summary_data().verify_clear();)
1020
1021 // Have worker threads release resources the next time they run a task.
1022 gc_task_manager()->release_all_resources();
1023
1024 ParCompactionManager::reset_all_bitmap_query_caches();
1025 }
1026
1027 void PSParallelCompact::post_compact()
1028 {
1029 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1030
1031 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1032 // Clear the marking bitmap, summary data and split info.
1033 clear_data_covering_space(SpaceId(id));
1034 // Update top(). Must be done after clearing the bitmap and summary data.
1035 _space_info[id].publish_new_top();
1036 }
1037
1038 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1039 MutableSpace* const from_space = _space_info[from_space_id].space();
1040 MutableSpace* const to_space = _space_info[to_space_id].space();
1041
1042 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1043 bool eden_empty = eden_space->is_empty();
1044 if (!eden_empty) {
1045 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1046 heap->young_gen(), heap->old_gen());
1047 }
1048
1049 // Update heap occupancy information which is used as input to the soft ref
1050 // clearing policy at the next gc.
1051 Universe::update_heap_info_at_gc();
1052
1053 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1054 to_space->is_empty();
1055
1056 PSCardTable* ct = heap->card_table();
1057 MemRegion old_mr = heap->old_gen()->reserved();
1058 if (young_gen_empty) {
1059 ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1060 } else {
1061 ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1062 }
1063
1064 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1065 ClassLoaderDataGraph::purge();
1066 MetaspaceUtils::verify_metrics();
1067
1068 heap->prune_scavengable_nmethods();
1069 JvmtiExport::gc_epilogue();
1070
1071 #if COMPILER2_OR_JVMCI
1072 DerivedPointerTable::update_pointers();
1073 #endif
1074
1075 if (ZapUnusedHeapArea) {
1076 heap->gen_mangle_unused_area();
1077 }
1078
1079 // Update time of last GC
1080 reset_millis_since_last_gc();
1081 }
1082
1083 HeapWord*
1084 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1085 bool maximum_compaction)
1086 {
1087 const size_t region_size = ParallelCompactData::RegionSize;
1088 const ParallelCompactData& sd = summary_data();
1089
1090 const MutableSpace* const space = _space_info[id].space();
1091 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1092 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1093 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1094
1095 // Skip full regions at the beginning of the space--they are necessarily part
1096 // of the dense prefix.
1097 size_t full_count = 0;
1098 const RegionData* cp;
1099 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1100 ++full_count;
1101 }
1102
1103 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1104 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1105 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1106 if (maximum_compaction || cp == end_cp || interval_ended) {
1107 _maximum_compaction_gc_num = total_invocations();
1108 return sd.region_to_addr(cp);
1109 }
1110
1111 HeapWord* const new_top = _space_info[id].new_top();
1112 const size_t space_live = pointer_delta(new_top, space->bottom());
1113 const size_t space_used = space->used_in_words();
1114 const size_t space_capacity = space->capacity_in_words();
1115
1116 const double cur_density = double(space_live) / space_capacity;
1117 const double deadwood_density =
1118 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1119 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1120
1121 if (TraceParallelOldGCDensePrefix) {
1122 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1123 cur_density, deadwood_density, deadwood_goal);
1124 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1125 "space_cap=" SIZE_FORMAT,
1126 space_live, space_used,
1127 space_capacity);
1128 }
1129
1130 // XXX - Use binary search?
1131 HeapWord* dense_prefix = sd.region_to_addr(cp);
1132 const RegionData* full_cp = cp;
1133 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1134 while (cp < end_cp) {
1135 HeapWord* region_destination = cp->destination();
1136 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1137 if (TraceParallelOldGCDensePrefix && Verbose) {
1138 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1139 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1140 sd.region(cp), p2i(region_destination),
1141 p2i(dense_prefix), cur_deadwood);
1142 }
1143
1144 if (cur_deadwood >= deadwood_goal) {
1145 // Found the region that has the correct amount of deadwood to the left.
1146 // This typically occurs after crossing a fairly sparse set of regions, so
1147 // iterate backwards over those sparse regions, looking for the region
1148 // that has the lowest density of live objects 'to the right.'
1149 size_t space_to_left = sd.region(cp) * region_size;
1150 size_t live_to_left = space_to_left - cur_deadwood;
1151 size_t space_to_right = space_capacity - space_to_left;
1152 size_t live_to_right = space_live - live_to_left;
1153 double density_to_right = double(live_to_right) / space_to_right;
1154 while (cp > full_cp) {
1155 --cp;
1156 const size_t prev_region_live_to_right = live_to_right -
1157 cp->data_size();
1158 const size_t prev_region_space_to_right = space_to_right + region_size;
1159 double prev_region_density_to_right =
1160 double(prev_region_live_to_right) / prev_region_space_to_right;
1161 if (density_to_right <= prev_region_density_to_right) {
1162 return dense_prefix;
1163 }
1164 if (TraceParallelOldGCDensePrefix && Verbose) {
1165 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1166 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1167 prev_region_density_to_right);
1168 }
1169 dense_prefix -= region_size;
1170 live_to_right = prev_region_live_to_right;
1171 space_to_right = prev_region_space_to_right;
1172 density_to_right = prev_region_density_to_right;
1173 }
1174 return dense_prefix;
1175 }
1176
1177 dense_prefix += region_size;
1178 ++cp;
1179 }
1180
1181 return dense_prefix;
1182 }
1183
1184 #ifndef PRODUCT
1185 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1186 const SpaceId id,
1187 const bool maximum_compaction,
1188 HeapWord* const addr)
1189 {
1190 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1191 RegionData* const cp = summary_data().region(region_idx);
1192 const MutableSpace* const space = _space_info[id].space();
1193 HeapWord* const new_top = _space_info[id].new_top();
1194
1195 const size_t space_live = pointer_delta(new_top, space->bottom());
1196 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1197 const size_t space_cap = space->capacity_in_words();
1198 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1199 const size_t live_to_right = new_top - cp->destination();
1200 const size_t dead_to_right = space->top() - addr - live_to_right;
1201
1202 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1203 "spl=" SIZE_FORMAT " "
1204 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1205 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1206 " ratio=%10.8f",
1207 algorithm, p2i(addr), region_idx,
1208 space_live,
1209 dead_to_left, dead_to_left_pct,
1210 dead_to_right, live_to_right,
1211 double(dead_to_right) / live_to_right);
1212 }
1213 #endif // #ifndef PRODUCT
1214
1215 // Return a fraction indicating how much of the generation can be treated as
1216 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1217 // based on the density of live objects in the generation to determine a limit,
1218 // which is then adjusted so the return value is min_percent when the density is
1219 // 1.
1220 //
1221 // The following table shows some return values for a different values of the
1222 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1223 // min_percent is 1.
1224 //
1225 // fraction allowed as dead wood
1226 // -----------------------------------------------------------------
1227 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1228 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1229 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1230 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1231 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1232 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1233 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1234 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1235 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1236 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1237 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1238 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1239 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1240 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1241 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1242 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1243 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1244 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1245 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1246 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1247 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1248 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1249 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1250
1251 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1252 {
1253 assert(_dwl_initialized, "uninitialized");
1254
1255 // The raw limit is the value of the normal distribution at x = density.
1256 const double raw_limit = normal_distribution(density);
1257
1258 // Adjust the raw limit so it becomes the minimum when the density is 1.
1259 //
1260 // First subtract the adjustment value (which is simply the precomputed value
1261 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1262 // Then add the minimum value, so the minimum is returned when the density is
1263 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1264 const double min = double(min_percent) / 100.0;
1265 const double limit = raw_limit - _dwl_adjustment + min;
1266 return MAX2(limit, 0.0);
1267 }
1268
1269 ParallelCompactData::RegionData*
1270 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1271 const RegionData* end)
1272 {
1273 const size_t region_size = ParallelCompactData::RegionSize;
1274 ParallelCompactData& sd = summary_data();
1275 size_t left = sd.region(beg);
1276 size_t right = end > beg ? sd.region(end) - 1 : left;
1277
1278 // Binary search.
1279 while (left < right) {
1280 // Equivalent to (left + right) / 2, but does not overflow.
1281 const size_t middle = left + (right - left) / 2;
1282 RegionData* const middle_ptr = sd.region(middle);
1283 HeapWord* const dest = middle_ptr->destination();
1284 HeapWord* const addr = sd.region_to_addr(middle);
1285 assert(dest != NULL, "sanity");
1286 assert(dest <= addr, "must move left");
1287
1288 if (middle > left && dest < addr) {
1289 right = middle - 1;
1290 } else if (middle < right && middle_ptr->data_size() == region_size) {
1291 left = middle + 1;
1292 } else {
1293 return middle_ptr;
1294 }
1295 }
1296 return sd.region(left);
1297 }
1298
1299 ParallelCompactData::RegionData*
1300 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1301 const RegionData* end,
1302 size_t dead_words)
1303 {
1304 ParallelCompactData& sd = summary_data();
1305 size_t left = sd.region(beg);
1306 size_t right = end > beg ? sd.region(end) - 1 : left;
1307
1308 // Binary search.
1309 while (left < right) {
1310 // Equivalent to (left + right) / 2, but does not overflow.
1311 const size_t middle = left + (right - left) / 2;
1312 RegionData* const middle_ptr = sd.region(middle);
1313 HeapWord* const dest = middle_ptr->destination();
1314 HeapWord* const addr = sd.region_to_addr(middle);
1315 assert(dest != NULL, "sanity");
1316 assert(dest <= addr, "must move left");
1317
1318 const size_t dead_to_left = pointer_delta(addr, dest);
1319 if (middle > left && dead_to_left > dead_words) {
1320 right = middle - 1;
1321 } else if (middle < right && dead_to_left < dead_words) {
1322 left = middle + 1;
1323 } else {
1324 return middle_ptr;
1325 }
1326 }
1327 return sd.region(left);
1328 }
1329
1330 // The result is valid during the summary phase, after the initial summarization
1331 // of each space into itself, and before final summarization.
1332 inline double
1333 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1334 HeapWord* const bottom,
1335 HeapWord* const top,
1336 HeapWord* const new_top)
1337 {
1338 ParallelCompactData& sd = summary_data();
1339
1340 assert(cp != NULL, "sanity");
1341 assert(bottom != NULL, "sanity");
1342 assert(top != NULL, "sanity");
1343 assert(new_top != NULL, "sanity");
1344 assert(top >= new_top, "summary data problem?");
1345 assert(new_top > bottom, "space is empty; should not be here");
1346 assert(new_top >= cp->destination(), "sanity");
1347 assert(top >= sd.region_to_addr(cp), "sanity");
1348
1349 HeapWord* const destination = cp->destination();
1350 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1351 const size_t compacted_region_live = pointer_delta(new_top, destination);
1352 const size_t compacted_region_used = pointer_delta(top,
1353 sd.region_to_addr(cp));
1354 const size_t reclaimable = compacted_region_used - compacted_region_live;
1355
1356 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1357 return double(reclaimable) / divisor;
1358 }
1359
1360 // Return the address of the end of the dense prefix, a.k.a. the start of the
1361 // compacted region. The address is always on a region boundary.
1362 //
1363 // Completely full regions at the left are skipped, since no compaction can
1364 // occur in those regions. Then the maximum amount of dead wood to allow is
1365 // computed, based on the density (amount live / capacity) of the generation;
1366 // the region with approximately that amount of dead space to the left is
1367 // identified as the limit region. Regions between the last completely full
1368 // region and the limit region are scanned and the one that has the best
1369 // (maximum) reclaimed_ratio() is selected.
1370 HeapWord*
1371 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1372 bool maximum_compaction)
1373 {
1374 const size_t region_size = ParallelCompactData::RegionSize;
1375 const ParallelCompactData& sd = summary_data();
1376
1377 const MutableSpace* const space = _space_info[id].space();
1378 HeapWord* const top = space->top();
1379 HeapWord* const top_aligned_up = sd.region_align_up(top);
1380 HeapWord* const new_top = _space_info[id].new_top();
1381 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1382 HeapWord* const bottom = space->bottom();
1383 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1384 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1385 const RegionData* const new_top_cp =
1386 sd.addr_to_region_ptr(new_top_aligned_up);
1387
1388 // Skip full regions at the beginning of the space--they are necessarily part
1389 // of the dense prefix.
1390 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1391 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1392 space->is_empty(), "no dead space allowed to the left");
1393 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1394 "region must have dead space");
1395
1396 // The gc number is saved whenever a maximum compaction is done, and used to
1397 // determine when the maximum compaction interval has expired. This avoids
1398 // successive max compactions for different reasons.
1399 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1400 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1401 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1402 total_invocations() == HeapFirstMaximumCompactionCount;
1403 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1404 _maximum_compaction_gc_num = total_invocations();
1405 return sd.region_to_addr(full_cp);
1406 }
1407
1408 const size_t space_live = pointer_delta(new_top, bottom);
1409 const size_t space_used = space->used_in_words();
1410 const size_t space_capacity = space->capacity_in_words();
1411
1412 const double density = double(space_live) / double(space_capacity);
1413 const size_t min_percent_free = MarkSweepDeadRatio;
1414 const double limiter = dead_wood_limiter(density, min_percent_free);
1415 const size_t dead_wood_max = space_used - space_live;
1416 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1417 dead_wood_max);
1418
1419 if (TraceParallelOldGCDensePrefix) {
1420 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1421 "space_cap=" SIZE_FORMAT,
1422 space_live, space_used,
1423 space_capacity);
1424 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1425 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1426 density, min_percent_free, limiter,
1427 dead_wood_max, dead_wood_limit);
1428 }
1429
1430 // Locate the region with the desired amount of dead space to the left.
1431 const RegionData* const limit_cp =
1432 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1433
1434 // Scan from the first region with dead space to the limit region and find the
1435 // one with the best (largest) reclaimed ratio.
1436 double best_ratio = 0.0;
1437 const RegionData* best_cp = full_cp;
1438 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1439 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1440 if (tmp_ratio > best_ratio) {
1441 best_cp = cp;
1442 best_ratio = tmp_ratio;
1443 }
1444 }
1445
1446 return sd.region_to_addr(best_cp);
1447 }
1448
1449 void PSParallelCompact::summarize_spaces_quick()
1450 {
1451 for (unsigned int i = 0; i < last_space_id; ++i) {
1452 const MutableSpace* space = _space_info[i].space();
1453 HeapWord** nta = _space_info[i].new_top_addr();
1454 bool result = _summary_data.summarize(_space_info[i].split_info(),
1455 space->bottom(), space->top(), NULL,
1456 space->bottom(), space->end(), nta);
1457 assert(result, "space must fit into itself");
1458 _space_info[i].set_dense_prefix(space->bottom());
1459 }
1460 }
1461
1462 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1463 {
1464 HeapWord* const dense_prefix_end = dense_prefix(id);
1465 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1466 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1467 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1468 // Only enough dead space is filled so that any remaining dead space to the
1469 // left is larger than the minimum filler object. (The remainder is filled
1470 // during the copy/update phase.)
1471 //
1472 // The size of the dead space to the right of the boundary is not a
1473 // concern, since compaction will be able to use whatever space is
1474 // available.
1475 //
1476 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1477 // surrounds the space to be filled with an object.
1478 //
1479 // In the 32-bit VM, each bit represents two 32-bit words:
1480 // +---+
1481 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1482 // end_bits: ... x x x | 0 | || 0 x x ...
1483 // +---+
1484 //
1485 // In the 64-bit VM, each bit represents one 64-bit word:
1486 // +------------+
1487 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1488 // end_bits: ... x x 1 | 0 || 0 | x x ...
1489 // +------------+
1490 // +-------+
1491 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1492 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1493 // +-------+
1494 // +-----------+
1495 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1496 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1497 // +-----------+
1498 // +-------+
1499 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1500 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1501 // +-------+
1502
1503 // Initially assume case a, c or e will apply.
1504 size_t obj_len = CollectedHeap::min_fill_size();
1505 HeapWord* obj_beg = dense_prefix_end - obj_len;
1506
1507 #ifdef _LP64
1508 if (MinObjAlignment > 1) { // object alignment > heap word size
1509 // Cases a, c or e.
1510 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1511 // Case b above.
1512 obj_beg = dense_prefix_end - 1;
1513 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1514 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1515 // Case d above.
1516 obj_beg = dense_prefix_end - 3;
1517 obj_len = 3;
1518 }
1519 #endif // #ifdef _LP64
1520
1521 CollectedHeap::fill_with_object(obj_beg, obj_len);
1522 _mark_bitmap.mark_obj(obj_beg, obj_len);
1523 _summary_data.add_obj(obj_beg, obj_len);
1524 assert(start_array(id) != NULL, "sanity");
1525 start_array(id)->allocate_block(obj_beg);
1526 }
1527 }
1528
1529 void
1530 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1531 {
1532 assert(id < last_space_id, "id out of range");
1533 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1534 "should have been reset in summarize_spaces_quick()");
1535
1536 const MutableSpace* space = _space_info[id].space();
1537 if (_space_info[id].new_top() != space->bottom()) {
1538 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1539 _space_info[id].set_dense_prefix(dense_prefix_end);
1540
1541 #ifndef PRODUCT
1542 if (TraceParallelOldGCDensePrefix) {
1543 print_dense_prefix_stats("ratio", id, maximum_compaction,
1544 dense_prefix_end);
1545 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1546 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1547 }
1548 #endif // #ifndef PRODUCT
1549
1550 // Recompute the summary data, taking into account the dense prefix. If
1551 // every last byte will be reclaimed, then the existing summary data which
1552 // compacts everything can be left in place.
1553 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1554 // If dead space crosses the dense prefix boundary, it is (at least
1555 // partially) filled with a dummy object, marked live and added to the
1556 // summary data. This simplifies the copy/update phase and must be done
1557 // before the final locations of objects are determined, to prevent
1558 // leaving a fragment of dead space that is too small to fill.
1559 fill_dense_prefix_end(id);
1560
1561 // Compute the destination of each Region, and thus each object.
1562 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1563 _summary_data.summarize(_space_info[id].split_info(),
1564 dense_prefix_end, space->top(), NULL,
1565 dense_prefix_end, space->end(),
1566 _space_info[id].new_top_addr());
1567 }
1568 }
1569
1570 if (log_develop_is_enabled(Trace, gc, compaction)) {
1571 const size_t region_size = ParallelCompactData::RegionSize;
1572 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1573 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1574 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1575 HeapWord* const new_top = _space_info[id].new_top();
1576 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1577 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1578 log_develop_trace(gc, compaction)(
1579 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1580 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1581 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1582 id, space->capacity_in_words(), p2i(dense_prefix_end),
1583 dp_region, dp_words / region_size,
1584 cr_words / region_size, p2i(new_top));
1585 }
1586 }
1587
1588 #ifndef PRODUCT
1589 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1590 HeapWord* dst_beg, HeapWord* dst_end,
1591 SpaceId src_space_id,
1592 HeapWord* src_beg, HeapWord* src_end)
1593 {
1594 log_develop_trace(gc, compaction)(
1595 "Summarizing %d [%s] into %d [%s]: "
1596 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1597 SIZE_FORMAT "-" SIZE_FORMAT " "
1598 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1599 SIZE_FORMAT "-" SIZE_FORMAT,
1600 src_space_id, space_names[src_space_id],
1601 dst_space_id, space_names[dst_space_id],
1602 p2i(src_beg), p2i(src_end),
1603 _summary_data.addr_to_region_idx(src_beg),
1604 _summary_data.addr_to_region_idx(src_end),
1605 p2i(dst_beg), p2i(dst_end),
1606 _summary_data.addr_to_region_idx(dst_beg),
1607 _summary_data.addr_to_region_idx(dst_end));
1608 }
1609 #endif // #ifndef PRODUCT
1610
1611 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1612 bool maximum_compaction)
1613 {
1614 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1615
1616 #ifdef ASSERT
1617 if (TraceParallelOldGCMarkingPhase) {
1618 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1619 "add_obj_bytes=" SIZE_FORMAT,
1620 add_obj_count, add_obj_size * HeapWordSize);
1621 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1622 "mark_bitmap_bytes=" SIZE_FORMAT,
1623 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1624 }
1625 #endif // #ifdef ASSERT
1626
1627 // Quick summarization of each space into itself, to see how much is live.
1628 summarize_spaces_quick();
1629
1630 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1631 NOT_PRODUCT(print_region_ranges());
1632 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1633
1634 // The amount of live data that will end up in old space (assuming it fits).
1635 size_t old_space_total_live = 0;
1636 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1637 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1638 _space_info[id].space()->bottom());
1639 }
1640
1641 MutableSpace* const old_space = _space_info[old_space_id].space();
1642 const size_t old_capacity = old_space->capacity_in_words();
1643 if (old_space_total_live > old_capacity) {
1644 // XXX - should also try to expand
1645 maximum_compaction = true;
1646 }
1647
1648 // Old generations.
1649 summarize_space(old_space_id, maximum_compaction);
1650
1651 // Summarize the remaining spaces in the young gen. The initial target space
1652 // is the old gen. If a space does not fit entirely into the target, then the
1653 // remainder is compacted into the space itself and that space becomes the new
1654 // target.
1655 SpaceId dst_space_id = old_space_id;
1656 HeapWord* dst_space_end = old_space->end();
1657 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1658 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1659 const MutableSpace* space = _space_info[id].space();
1660 const size_t live = pointer_delta(_space_info[id].new_top(),
1661 space->bottom());
1662 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1663
1664 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1665 SpaceId(id), space->bottom(), space->top());)
1666 if (live > 0 && live <= available) {
1667 // All the live data will fit.
1668 bool done = _summary_data.summarize(_space_info[id].split_info(),
1669 space->bottom(), space->top(),
1670 NULL,
1671 *new_top_addr, dst_space_end,
1672 new_top_addr);
1673 assert(done, "space must fit into old gen");
1674
1675 // Reset the new_top value for the space.
1676 _space_info[id].set_new_top(space->bottom());
1677 } else if (live > 0) {
1678 // Attempt to fit part of the source space into the target space.
1679 HeapWord* next_src_addr = NULL;
1680 bool done = _summary_data.summarize(_space_info[id].split_info(),
1681 space->bottom(), space->top(),
1682 &next_src_addr,
1683 *new_top_addr, dst_space_end,
1684 new_top_addr);
1685 assert(!done, "space should not fit into old gen");
1686 assert(next_src_addr != NULL, "sanity");
1687
1688 // The source space becomes the new target, so the remainder is compacted
1689 // within the space itself.
1690 dst_space_id = SpaceId(id);
1691 dst_space_end = space->end();
1692 new_top_addr = _space_info[id].new_top_addr();
1693 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1694 space->bottom(), dst_space_end,
1695 SpaceId(id), next_src_addr, space->top());)
1696 done = _summary_data.summarize(_space_info[id].split_info(),
1697 next_src_addr, space->top(),
1698 NULL,
1699 space->bottom(), dst_space_end,
1700 new_top_addr);
1701 assert(done, "space must fit when compacted into itself");
1702 assert(*new_top_addr <= space->top(), "usage should not grow");
1703 }
1704 }
1705
1706 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1707 NOT_PRODUCT(print_region_ranges());
1708 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1709 }
1710
1711 // This method should contain all heap-specific policy for invoking a full
1712 // collection. invoke_no_policy() will only attempt to compact the heap; it
1713 // will do nothing further. If we need to bail out for policy reasons, scavenge
1714 // before full gc, or any other specialized behavior, it needs to be added here.
1715 //
1716 // Note that this method should only be called from the vm_thread while at a
1717 // safepoint.
1718 //
1719 // Note that the all_soft_refs_clear flag in the soft ref policy
1720 // may be true because this method can be called without intervening
1721 // activity. For example when the heap space is tight and full measure
1722 // are being taken to free space.
1723 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1724 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1725 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1726 "should be in vm thread");
1727
1728 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1729 GCCause::Cause gc_cause = heap->gc_cause();
1730 assert(!heap->is_gc_active(), "not reentrant");
1731
1732 PSAdaptiveSizePolicy* policy = heap->size_policy();
1733 IsGCActiveMark mark;
1734
1735 if (ScavengeBeforeFullGC) {
1736 PSScavenge::invoke_no_policy();
1737 }
1738
1739 const bool clear_all_soft_refs =
1740 heap->soft_ref_policy()->should_clear_all_soft_refs();
1741
1742 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1743 maximum_heap_compaction);
1744 }
1745
1746 // This method contains no policy. You should probably
1747 // be calling invoke() instead.
1748 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1749 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1750 assert(ref_processor() != NULL, "Sanity");
1751
1752 if (GCLocker::check_active_before_gc()) {
1753 return false;
1754 }
1755
1756 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1757
1758 GCIdMark gc_id_mark;
1759 _gc_timer.register_gc_start();
1760 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1761
1762 TimeStamp marking_start;
1763 TimeStamp compaction_start;
1764 TimeStamp collection_exit;
1765
1766 GCCause::Cause gc_cause = heap->gc_cause();
1767 PSYoungGen* young_gen = heap->young_gen();
1768 PSOldGen* old_gen = heap->old_gen();
1769 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1770
1771 // The scope of casr should end after code that can change
1772 // SoftRefPolicy::_should_clear_all_soft_refs.
1773 ClearedAllSoftRefs casr(maximum_heap_compaction,
1774 heap->soft_ref_policy());
1775
1776 if (ZapUnusedHeapArea) {
1777 // Save information needed to minimize mangling
1778 heap->record_gen_tops_before_GC();
1779 }
1780
1781 // Make sure data structures are sane, make the heap parsable, and do other
1782 // miscellaneous bookkeeping.
1783 pre_compact();
1784
1785 PreGCValues pre_gc_values(heap);
1786
1787 // Get the compaction manager reserved for the VM thread.
1788 ParCompactionManager* const vmthread_cm =
1789 ParCompactionManager::manager_array(gc_task_manager()->workers());
1790
1791 {
1792 ResourceMark rm;
1793 HandleMark hm;
1794
1795 ParallelScavengeHeap::heap()->workers().update_active_workers(WorkerPolicy::calc_active_workers(
1796 ParallelScavengeHeap::heap()->workers().total_workers(),
1797 ParallelScavengeHeap::heap()->workers().active_workers(),
1798 Threads::number_of_non_daemon_threads()));
1799
1800 // Set the number of GC threads to be used in this collection
1801 gc_task_manager()->set_active_gang();
1802 gc_task_manager()->task_idle_workers();
1803
1804 GCTraceCPUTime tcpu;
1805 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1806
1807 heap->pre_full_gc_dump(&_gc_timer);
1808
1809 TraceCollectorStats tcs(counters());
1810 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1811
1812 if (log_is_enabled(Debug, gc, heap, exit)) {
1813 accumulated_time()->start();
1814 }
1815
1816 // Let the size policy know we're starting
1817 size_policy->major_collection_begin();
1818
1819 #if COMPILER2_OR_JVMCI
1820 DerivedPointerTable::clear();
1821 #endif
1822
1823 ref_processor()->enable_discovery();
1824 ref_processor()->setup_policy(maximum_heap_compaction);
1825
1826 bool marked_for_unloading = false;
1827
1828 marking_start.update();
1829 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
1830
1831 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1832 && GCCause::is_user_requested_gc(gc_cause);
1833 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1834
1835 #if COMPILER2_OR_JVMCI
1836 assert(DerivedPointerTable::is_active(), "Sanity");
1837 DerivedPointerTable::set_active(false);
1838 #endif
1839
1840 // adjust_roots() updates Universe::_intArrayKlassObj which is
1841 // needed by the compaction for filling holes in the dense prefix.
1842 adjust_roots(vmthread_cm);
1843
1844 compaction_start.update();
1845 compact();
1846
1847 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1848 // done before resizing.
1849 post_compact();
1850
1851 // Let the size policy know we're done
1852 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1853
1854 if (UseAdaptiveSizePolicy) {
1855 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1856 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1857 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1858
1859 // Don't check if the size_policy is ready here. Let
1860 // the size_policy check that internally.
1861 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1862 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1863 // Swap the survivor spaces if from_space is empty. The
1864 // resize_young_gen() called below is normally used after
1865 // a successful young GC and swapping of survivor spaces;
1866 // otherwise, it will fail to resize the young gen with
1867 // the current implementation.
1868 if (young_gen->from_space()->is_empty()) {
1869 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1870 young_gen->swap_spaces();
1871 }
1872
1873 // Calculate optimal free space amounts
1874 assert(young_gen->max_size() >
1875 young_gen->from_space()->capacity_in_bytes() +
1876 young_gen->to_space()->capacity_in_bytes(),
1877 "Sizes of space in young gen are out-of-bounds");
1878
1879 size_t young_live = young_gen->used_in_bytes();
1880 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1881 size_t old_live = old_gen->used_in_bytes();
1882 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1883 size_t max_old_gen_size = old_gen->max_gen_size();
1884 size_t max_eden_size = young_gen->max_size() -
1885 young_gen->from_space()->capacity_in_bytes() -
1886 young_gen->to_space()->capacity_in_bytes();
1887
1888 // Used for diagnostics
1889 size_policy->clear_generation_free_space_flags();
1890
1891 size_policy->compute_generations_free_space(young_live,
1892 eden_live,
1893 old_live,
1894 cur_eden,
1895 max_old_gen_size,
1896 max_eden_size,
1897 true /* full gc*/);
1898
1899 size_policy->check_gc_overhead_limit(eden_live,
1900 max_old_gen_size,
1901 max_eden_size,
1902 true /* full gc*/,
1903 gc_cause,
1904 heap->soft_ref_policy());
1905
1906 size_policy->decay_supplemental_growth(true /* full gc*/);
1907
1908 heap->resize_old_gen(
1909 size_policy->calculated_old_free_size_in_bytes());
1910
1911 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1912 size_policy->calculated_survivor_size_in_bytes());
1913 }
1914
1915 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1916 }
1917
1918 if (UsePerfData) {
1919 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1920 counters->update_counters();
1921 counters->update_old_capacity(old_gen->capacity_in_bytes());
1922 counters->update_young_capacity(young_gen->capacity_in_bytes());
1923 }
1924
1925 heap->resize_all_tlabs();
1926
1927 // Resize the metaspace capacity after a collection
1928 MetaspaceGC::compute_new_size();
1929
1930 if (log_is_enabled(Debug, gc, heap, exit)) {
1931 accumulated_time()->stop();
1932 }
1933
1934 young_gen->print_used_change(pre_gc_values.young_gen_used());
1935 old_gen->print_used_change(pre_gc_values.old_gen_used());
1936 MetaspaceUtils::print_metaspace_change(pre_gc_values.metadata_used());
1937
1938 // Track memory usage and detect low memory
1939 MemoryService::track_memory_usage();
1940 heap->update_counters();
1941 gc_task_manager()->release_idle_workers();
1942
1943 heap->post_full_gc_dump(&_gc_timer);
1944 }
1945
1946 #ifdef ASSERT
1947 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1948 ParCompactionManager* const cm =
1949 ParCompactionManager::manager_array(int(i));
1950 assert(cm->marking_stack()->is_empty(), "should be empty");
1951 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i);
1952 }
1953 #endif // ASSERT
1954
1955 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1956 HandleMark hm; // Discard invalid handles created during verification
1957 Universe::verify("After GC");
1958 }
1959
1960 // Re-verify object start arrays
1961 if (VerifyObjectStartArray &&
1962 VerifyAfterGC) {
1963 old_gen->verify_object_start_array();
1964 }
1965
1966 if (ZapUnusedHeapArea) {
1967 old_gen->object_space()->check_mangled_unused_area_complete();
1968 }
1969
1970 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1971
1972 collection_exit.update();
1973
1974 heap->print_heap_after_gc();
1975 heap->trace_heap_after_gc(&_gc_tracer);
1976
1977 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1978 marking_start.ticks(), compaction_start.ticks(),
1979 collection_exit.ticks());
1980 gc_task_manager()->print_task_time_stamps();
1981
1982 #ifdef TRACESPINNING
1983 ParallelTaskTerminator::print_termination_counts();
1984 #endif
1985
1986 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1987
1988 _gc_timer.register_gc_end();
1989
1990 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1991 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1992
1993 return true;
1994 }
1995
1996 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1997 PSYoungGen* young_gen,
1998 PSOldGen* old_gen) {
1999 MutableSpace* const eden_space = young_gen->eden_space();
2000 assert(!eden_space->is_empty(), "eden must be non-empty");
2001 assert(young_gen->virtual_space()->alignment() ==
2002 old_gen->virtual_space()->alignment(), "alignments do not match");
2003
2004 // We also return false when it's a heterogenous heap because old generation cannot absorb data from eden
2005 // when it is allocated on different memory (example, nv-dimm) than young.
2006 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) ||
2007 ParallelArguments::is_heterogeneous_heap()) {
2008 return false;
2009 }
2010
2011 // Both generations must be completely committed.
2012 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2013 return false;
2014 }
2015 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2016 return false;
2017 }
2018
2019 // Figure out how much to take from eden. Include the average amount promoted
2020 // in the total; otherwise the next young gen GC will simply bail out to a
2021 // full GC.
2022 const size_t alignment = old_gen->virtual_space()->alignment();
2023 const size_t eden_used = eden_space->used_in_bytes();
2024 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2025 const size_t absorb_size = align_up(eden_used + promoted, alignment);
2026 const size_t eden_capacity = eden_space->capacity_in_bytes();
2027
2028 if (absorb_size >= eden_capacity) {
2029 return false; // Must leave some space in eden.
2030 }
2031
2032 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2033 if (new_young_size < young_gen->min_gen_size()) {
2034 return false; // Respect young gen minimum size.
2035 }
2036
2037 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: "
2038 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2039 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2040 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2041 absorb_size / K,
2042 eden_capacity / K, (eden_capacity - absorb_size) / K,
2043 young_gen->from_space()->used_in_bytes() / K,
2044 young_gen->to_space()->used_in_bytes() / K,
2045 young_gen->capacity_in_bytes() / K, new_young_size / K);
2046
2047 // Fill the unused part of the old gen.
2048 MutableSpace* const old_space = old_gen->object_space();
2049 HeapWord* const unused_start = old_space->top();
2050 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2051
2052 if (unused_words > 0) {
2053 if (unused_words < CollectedHeap::min_fill_size()) {
2054 return false; // If the old gen cannot be filled, must give up.
2055 }
2056 CollectedHeap::fill_with_objects(unused_start, unused_words);
2057 }
2058
2059 // Take the live data from eden and set both top and end in the old gen to
2060 // eden top. (Need to set end because reset_after_change() mangles the region
2061 // from end to virtual_space->high() in debug builds).
2062 HeapWord* const new_top = eden_space->top();
2063 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2064 absorb_size);
2065 young_gen->reset_after_change();
2066 old_space->set_top(new_top);
2067 old_space->set_end(new_top);
2068 old_gen->reset_after_change();
2069
2070 // Update the object start array for the filler object and the data from eden.
2071 ObjectStartArray* const start_array = old_gen->start_array();
2072 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2073 start_array->allocate_block(p);
2074 }
2075
2076 // Could update the promoted average here, but it is not typically updated at
2077 // full GCs and the value to use is unclear. Something like
2078 //
2079 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2080
2081 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2082 return true;
2083 }
2084
2085 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2086 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2087 "shouldn't return NULL");
2088 return ParallelScavengeHeap::gc_task_manager();
2089 }
2090
2091 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
2092 private:
2093 uint _worker_id;
2094
2095 public:
2096 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
2097 void do_thread(Thread* thread) {
2098 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2099
2100 ResourceMark rm;
2101
2102 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
2103
2104 PCMarkAndPushClosure mark_and_push_closure(cm);
2105 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
2106
2107 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
2108
2109 // Do the real work
2110 cm->follow_marking_stacks();
2111 }
2112 };
2113
2114 void mark_from_roots_task(Parallel::RootType::Value root_type, uint which) {
2115 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2116
2117 ParCompactionManager* cm =
2118 ParCompactionManager::gc_thread_compaction_manager(which);
2119 PCMarkAndPushClosure mark_and_push_closure(cm);
2120
2121 switch (root_type) {
2122 case Parallel::RootType::universe:
2123 Universe::oops_do(&mark_and_push_closure);
2124 break;
2125
2126 case Parallel::RootType::jni_handles:
2127 JNIHandles::oops_do(&mark_and_push_closure);
2128 break;
2129
2130 case Parallel::RootType::object_synchronizer:
2131 ObjectSynchronizer::oops_do(&mark_and_push_closure);
2132 break;
2133
2134 case Parallel::RootType::management:
2135 Management::oops_do(&mark_and_push_closure);
2136 break;
2137
2138 case Parallel::RootType::jvmti:
2139 JvmtiExport::oops_do(&mark_and_push_closure);
2140 break;
2141
2142 case Parallel::RootType::system_dictionary:
2143 SystemDictionary::oops_do(&mark_and_push_closure);
2144 break;
2145
2146 case Parallel::RootType::class_loader_data:
2147 {
2148 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
2149 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
2150 }
2151 break;
2152
2153 case Parallel::RootType::code_cache:
2154 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
2155 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
2156 AOTLoader::oops_do(&mark_and_push_closure);
2157 break;
2158
2159 #if INCLUDE_JVMCI
2160 case Parallel::RootType::jvmci:
2161 JVMCI::oops_do(&mark_and_push_closure);
2162 break;
2163 #endif
2164
2165 case Parallel::RootType::sentinel:
2166 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
2167 fatal("Bad enumeration value: %u", root_type);
2168 break;
2169 }
2170
2171 // Do the real work
2172 cm->follow_marking_stacks();
2173 }
2174
2175 void steal_marking_task(ParallelTaskTerminator& terminator, uint worker_id) {
2176 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2177
2178 ParCompactionManager* cm =
2179 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2180
2181 oop obj = NULL;
2182 ObjArrayTask task;
2183 do {
2184 while (ParCompactionManager::steal_objarray(worker_id, task)) {
2185 cm->follow_array((objArrayOop)task.obj(), task.index());
2186 cm->follow_marking_stacks();
2187 }
2188 while (ParCompactionManager::steal(worker_id, obj)) {
2189 cm->follow_contents(obj);
2190 cm->follow_marking_stacks();
2191 }
2192 } while (!terminator.offer_termination());
2193 }
2194
2195 class MarkFromRootsTask : public AbstractGangTask {
2196 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2197 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2198 EnumClaimer<Parallel::RootType::Value> _enum_claimer;
2199 TaskTerminator _terminator;
2200 uint _active_workers;
2201
2202 public:
2203 MarkFromRootsTask(uint active_workers)
2204 : AbstractGangTask("MarkFromRootsTask"),
2205 _strong_roots_scope(active_workers),
2206 _enum_claimer(Parallel::RootType::sentinel),
2207 _terminator(active_workers, ParCompactionManager::stack_array()),
2208 _active_workers(active_workers) {}
2209
2210 virtual void work(uint worker_id) {
2211 for (Parallel::RootType::Value root_type; _enum_claimer.try_claim(root_type); /* empty */) {
2212 mark_from_roots_task(root_type, worker_id);
2213 }
2214
2215 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2216 Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2217
2218 if (_active_workers > 1) {
2219 steal_marking_task(*_terminator.terminator(), worker_id);
2220 }
2221 }
2222 };
2223
2224 class PCRefProcTask : public AbstractGangTask {
2225 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2226 ProcessTask& _task;
2227 uint _ergo_workers;
2228 TaskTerminator _terminator;
2229
2230 public:
2231 PCRefProcTask(ProcessTask& task, uint ergo_workers)
2232 : AbstractGangTask("PCRefProcTask"),
2233 _task(task),
2234 _ergo_workers(ergo_workers),
2235 _terminator(_ergo_workers, ParCompactionManager::stack_array()) {
2236 }
2237
2238 virtual void work(uint worker_id) {
2239 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2240 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2241
2242 ParCompactionManager* cm =
2243 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2244 PCMarkAndPushClosure mark_and_push_closure(cm);
2245 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2246 _task.work(worker_id, *PSParallelCompact::is_alive_closure(),
2247 mark_and_push_closure, follow_stack_closure);
2248
2249 steal_marking_task(*_terminator.terminator(), worker_id);
2250 }
2251 };
2252
2253 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2254 void execute(ProcessTask& process_task, uint ergo_workers) {
2255 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers,
2256 "Ergonomically chosen workers (%u) must be equal to active workers (%u)",
2257 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers());
2258
2259 PCRefProcTask task(process_task, ergo_workers);
2260 ParallelScavengeHeap::heap()->workers().run_task(&task);
2261 }
2262 };
2263
2264 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2265 bool maximum_heap_compaction,
2266 ParallelOldTracer *gc_tracer) {
2267 // Recursively traverse all live objects and mark them
2268 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2269
2270 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2271 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2272 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2273 TaskQueueSetSuper* qset = ParCompactionManager::stack_array();
2274 TaskTerminator terminator(active_gc_threads, qset);
2275
2276 PCMarkAndPushClosure mark_and_push_closure(cm);
2277 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2278
2279 // Need new claim bits before marking starts.
2280 ClassLoaderDataGraph::clear_claimed_marks();
2281
2282 {
2283 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2284
2285 MarkFromRootsTask task(active_gc_threads);
2286 ParallelScavengeHeap::heap()->workers().run_task(&task);
2287 }
2288
2289 // Process reference objects found during marking
2290 {
2291 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2292
2293 ReferenceProcessorStats stats;
2294 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2295
2296 if (ref_processor()->processing_is_mt()) {
2297 ref_processor()->set_active_mt_degree(active_gc_threads);
2298
2299 RefProcTaskExecutor task_executor;
2300 stats = ref_processor()->process_discovered_references(
2301 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2302 &task_executor, &pt);
2303 } else {
2304 stats = ref_processor()->process_discovered_references(
2305 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2306 &pt);
2307 }
2308
2309 gc_tracer->report_gc_reference_stats(stats);
2310 pt.print_all_references();
2311 }
2312
2313 // This is the point where the entire marking should have completed.
2314 assert(cm->marking_stacks_empty(), "Marking should have completed");
2315
2316 {
2317 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2318 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl);
2319 }
2320
2321 {
2322 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2323
2324 // Follow system dictionary roots and unload classes.
2325 bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2326
2327 // Unload nmethods.
2328 CodeCache::do_unloading(is_alive_closure(), purged_class);
2329
2330 // Prune dead klasses from subklass/sibling/implementor lists.
2331 Klass::clean_weak_klass_links(purged_class);
2332
2333 // Clean JVMCI metadata handles.
2334 JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2335 }
2336
2337 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2338 }
2339
2340 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2341 // Adjust the pointers to reflect the new locations
2342 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2343
2344 // Need new claim bits when tracing through and adjusting pointers.
2345 ClassLoaderDataGraph::clear_claimed_marks();
2346
2347 PCAdjustPointerClosure oop_closure(cm);
2348
2349 // General strong roots.
2350 Universe::oops_do(&oop_closure);
2351 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2352 Threads::oops_do(&oop_closure, NULL);
2353 ObjectSynchronizer::oops_do(&oop_closure);
2354 Management::oops_do(&oop_closure);
2355 JvmtiExport::oops_do(&oop_closure);
2356 SystemDictionary::oops_do(&oop_closure);
2357 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong);
2358 ClassLoaderDataGraph::cld_do(&cld_closure);
2359
2360 // Now adjust pointers in remaining weak roots. (All of which should
2361 // have been cleared if they pointed to non-surviving objects.)
2362 WeakProcessor::oops_do(&oop_closure);
2363
2364 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2365 CodeCache::blobs_do(&adjust_from_blobs);
2366 AOT_ONLY(AOTLoader::oops_do(&oop_closure);)
2367
2368 JVMCI_ONLY(JVMCI::oops_do(&oop_closure);)
2369
2370 ref_processor()->weak_oops_do(&oop_closure);
2371 // Roots were visited so references into the young gen in roots
2372 // may have been scanned. Process them also.
2373 // Should the reference processor have a span that excludes
2374 // young gen objects?
2375 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2376 }
2377
2378 // Helper class to print 8 region numbers per line and then print the total at the end.
2379 class FillableRegionLogger : public StackObj {
2380 private:
2381 Log(gc, compaction) log;
2382 static const int LineLength = 8;
2383 size_t _regions[LineLength];
2384 int _next_index;
2385 bool _enabled;
2386 size_t _total_regions;
2387 public:
2388 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2389 ~FillableRegionLogger() {
2390 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2391 }
2392
2393 void print_line() {
2394 if (!_enabled || _next_index == 0) {
2395 return;
2396 }
2397 FormatBuffer<> line("Fillable: ");
2398 for (int i = 0; i < _next_index; i++) {
2399 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2400 }
2401 log.trace("%s", line.buffer());
2402 _next_index = 0;
2403 }
2404
2405 void handle(size_t region) {
2406 if (!_enabled) {
2407 return;
2408 }
2409 _regions[_next_index++] = region;
2410 if (_next_index == LineLength) {
2411 print_line();
2412 }
2413 _total_regions++;
2414 }
2415 };
2416
2417 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2418 {
2419 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2420
2421 // Find the threads that are active
2422 unsigned int which = 0;
2423
2424 // Find all regions that are available (can be filled immediately) and
2425 // distribute them to the thread stacks. The iteration is done in reverse
2426 // order (high to low) so the regions will be removed in ascending order.
2427
2428 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2429
2430 // id + 1 is used to test termination so unsigned can
2431 // be used with an old_space_id == 0.
2432 FillableRegionLogger region_logger;
2433 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2434 SpaceInfo* const space_info = _space_info + id;
2435 MutableSpace* const space = space_info->space();
2436 HeapWord* const new_top = space_info->new_top();
2437
2438 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2439 const size_t end_region =
2440 sd.addr_to_region_idx(sd.region_align_up(new_top));
2441
2442 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2443 if (sd.region(cur)->claim_unsafe()) {
2444 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2445 cm->region_stack()->push(cur);
2446 region_logger.handle(cur);
2447 // Assign regions to tasks in round-robin fashion.
2448 if (++which == parallel_gc_threads) {
2449 which = 0;
2450 }
2451 }
2452 }
2453 region_logger.print_line();
2454 }
2455 }
2456
2457 class TaskQueue : StackObj {
2458 volatile uint _counter;
2459 uint _size;
2460 uint _insert_index;
2461 PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2462 public:
2463 TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2464 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2465 guarantee(_backing_array != NULL, "alloc failure");
2466 }
2467 ~TaskQueue() {
2468 assert(_counter >= _insert_index, "not all queue elements were claimed");
2469 FREE_C_HEAP_ARRAY(T, _backing_array);
2470 }
2471
2472 void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2473 assert(_insert_index < _size, "to small backing array");
2474 _backing_array[_insert_index++] = value;
2475 }
2476
2477 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2478 uint claimed = Atomic::add(1u, &_counter) - 1; // -1 is so that we start with zero
2479 if (claimed < _insert_index) {
2480 reference = _backing_array[claimed];
2481 return true;
2482 } else {
2483 return false;
2484 }
2485 }
2486 };
2487
2488 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2489
2490 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2491 uint parallel_gc_threads) {
2492 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2493
2494 ParallelCompactData& sd = PSParallelCompact::summary_data();
2495
2496 // Iterate over all the spaces adding tasks for updating
2497 // regions in the dense prefix. Assume that 1 gc thread
2498 // will work on opening the gaps and the remaining gc threads
2499 // will work on the dense prefix.
2500 unsigned int space_id;
2501 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2502 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2503 const MutableSpace* const space = _space_info[space_id].space();
2504
2505 if (dense_prefix_end == space->bottom()) {
2506 // There is no dense prefix for this space.
2507 continue;
2508 }
2509
2510 // The dense prefix is before this region.
2511 size_t region_index_end_dense_prefix =
2512 sd.addr_to_region_idx(dense_prefix_end);
2513 RegionData* const dense_prefix_cp =
2514 sd.region(region_index_end_dense_prefix);
2515 assert(dense_prefix_end == space->end() ||
2516 dense_prefix_cp->available() ||
2517 dense_prefix_cp->claimed(),
2518 "The region after the dense prefix should always be ready to fill");
2519
2520 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2521
2522 // Is there dense prefix work?
2523 size_t total_dense_prefix_regions =
2524 region_index_end_dense_prefix - region_index_start;
2525 // How many regions of the dense prefix should be given to
2526 // each thread?
2527 if (total_dense_prefix_regions > 0) {
2528 uint tasks_for_dense_prefix = 1;
2529 if (total_dense_prefix_regions <=
2530 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2531 // Don't over partition. This assumes that
2532 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2533 // so there are not many regions to process.
2534 tasks_for_dense_prefix = parallel_gc_threads;
2535 } else {
2536 // Over partition
2537 tasks_for_dense_prefix = parallel_gc_threads *
2538 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2539 }
2540 size_t regions_per_thread = total_dense_prefix_regions /
2541 tasks_for_dense_prefix;
2542 // Give each thread at least 1 region.
2543 if (regions_per_thread == 0) {
2544 regions_per_thread = 1;
2545 }
2546
2547 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2548 if (region_index_start >= region_index_end_dense_prefix) {
2549 break;
2550 }
2551 // region_index_end is not processed
2552 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2553 region_index_end_dense_prefix);
2554 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2555 region_index_start,
2556 region_index_end));
2557 region_index_start = region_index_end;
2558 }
2559 }
2560 // This gets any part of the dense prefix that did not
2561 // fit evenly.
2562 if (region_index_start < region_index_end_dense_prefix) {
2563 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2564 region_index_start,
2565 region_index_end_dense_prefix));
2566 }
2567 }
2568 }
2569
2570 #ifdef ASSERT
2571 // Write a histogram of the number of times the block table was filled for a
2572 // region.
2573 void PSParallelCompact::write_block_fill_histogram()
2574 {
2575 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2576 return;
2577 }
2578
2579 Log(gc, compaction) log;
2580 ResourceMark rm;
2581 LogStream ls(log.trace());
2582 outputStream* out = &ls;
2583
2584 typedef ParallelCompactData::RegionData rd_t;
2585 ParallelCompactData& sd = summary_data();
2586
2587 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2588 MutableSpace* const spc = _space_info[id].space();
2589 if (spc->bottom() != spc->top()) {
2590 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2591 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2592 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2593
2594 size_t histo[5] = { 0, 0, 0, 0, 0 };
2595 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2596 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2597
2598 for (const rd_t* cur = beg; cur < end; ++cur) {
2599 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2600 }
2601 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2602 for (size_t i = 0; i < histo_len; ++i) {
2603 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2604 histo[i], 100.0 * histo[i] / region_cnt);
2605 }
2606 out->cr();
2607 }
2608 }
2609 }
2610 #endif // #ifdef ASSERT
2611
2612 //
2613 // CompactionWithStealingTask
2614 //
2615
2616 void compaction_with_stealing_task(ParallelTaskTerminator* terminator, uint which) {
2617 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2618
2619 ParCompactionManager* cm =
2620 ParCompactionManager::gc_thread_compaction_manager(which);
2621
2622 // Drain the stacks that have been preloaded with regions
2623 // that are ready to fill.
2624
2625 cm->drain_region_stacks();
2626
2627 guarantee(cm->region_stack()->is_empty(), "Not empty");
2628
2629 size_t region_index = 0;
2630
2631 while (true) {
2632 if (ParCompactionManager::steal(which, region_index)) {
2633 PSParallelCompact::fill_and_update_region(cm, region_index);
2634 cm->drain_region_stacks();
2635 } else {
2636 if (terminator->offer_termination()) {
2637 break;
2638 }
2639 // Go around again.
2640 }
2641 }
2642 return;
2643 }
2644
2645 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
2646 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2647 TaskQueue& _tq;
2648 TaskTerminator _terminator;
2649 uint _active_workers;
2650
2651 public:
2652 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2653 AbstractGangTask("UpdateDensePrefixAndCompactionTask"),
2654 _tq(tq),
2655 _terminator(active_workers, ParCompactionManager::region_array()),
2656 _active_workers(active_workers) {
2657 }
2658 virtual void work(uint worker_id) {
2659 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2660
2661 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2662 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2663 task._space_id,
2664 task._region_index_start,
2665 task._region_index_end);
2666 }
2667
2668 // Once a thread has drained it's stack, it should try to steal regions from
2669 // other threads.
2670 compaction_with_stealing_task(_terminator.terminator(), worker_id);
2671 }
2672 };
2673
2674 void PSParallelCompact::compact() {
2675 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2676
2677 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2678 PSOldGen* old_gen = heap->old_gen();
2679 old_gen->start_array()->reset();
2680 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2681
2682 TaskQueue task_queue(last_space_id*(active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2683 prepare_region_draining_tasks(active_gc_threads);
2684 enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2685
2686 {
2687 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2688
2689 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2690 ParallelScavengeHeap::heap()->workers().run_task(&task);
2691
2692 #ifdef ASSERT
2693 // Verify that all regions have been processed before the deferred updates.
2694 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2695 verify_complete(SpaceId(id));
2696 }
2697 #endif
2698 }
2699
2700 {
2701 // Update the deferred objects, if any. Any compaction manager can be used.
2702 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2703 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2704 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2705 update_deferred_objects(cm, SpaceId(id));
2706 }
2707 }
2708
2709 DEBUG_ONLY(write_block_fill_histogram());
2710 }
2711
2712 #ifdef ASSERT
2713 void PSParallelCompact::verify_complete(SpaceId space_id) {
2714 // All Regions between space bottom() to new_top() should be marked as filled
2715 // and all Regions between new_top() and top() should be available (i.e.,
2716 // should have been emptied).
2717 ParallelCompactData& sd = summary_data();
2718 SpaceInfo si = _space_info[space_id];
2719 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2720 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2721 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2722 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2723 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2724
2725 bool issued_a_warning = false;
2726
2727 size_t cur_region;
2728 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2729 const RegionData* const c = sd.region(cur_region);
2730 if (!c->completed()) {
2731 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2732 cur_region, c->destination_count());
2733 issued_a_warning = true;
2734 }
2735 }
2736
2737 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2738 const RegionData* const c = sd.region(cur_region);
2739 if (!c->available()) {
2740 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2741 cur_region, c->destination_count());
2742 issued_a_warning = true;
2743 }
2744 }
2745
2746 if (issued_a_warning) {
2747 print_region_ranges();
2748 }
2749 }
2750 #endif // #ifdef ASSERT
2751
2752 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2753 _start_array->allocate_block(addr);
2754 compaction_manager()->update_contents(oop(addr));
2755 }
2756
2757 // Update interior oops in the ranges of regions [beg_region, end_region).
2758 void
2759 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2760 SpaceId space_id,
2761 size_t beg_region,
2762 size_t end_region) {
2763 ParallelCompactData& sd = summary_data();
2764 ParMarkBitMap* const mbm = mark_bitmap();
2765
2766 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2767 HeapWord* const end_addr = sd.region_to_addr(end_region);
2768 assert(beg_region <= end_region, "bad region range");
2769 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2770
2771 #ifdef ASSERT
2772 // Claim the regions to avoid triggering an assert when they are marked as
2773 // filled.
2774 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2775 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2776 }
2777 #endif // #ifdef ASSERT
2778
2779 if (beg_addr != space(space_id)->bottom()) {
2780 // Find the first live object or block of dead space that *starts* in this
2781 // range of regions. If a partial object crosses onto the region, skip it;
2782 // it will be marked for 'deferred update' when the object head is
2783 // processed. If dead space crosses onto the region, it is also skipped; it
2784 // will be filled when the prior region is processed. If neither of those
2785 // apply, the first word in the region is the start of a live object or dead
2786 // space.
2787 assert(beg_addr > space(space_id)->bottom(), "sanity");
2788 const RegionData* const cp = sd.region(beg_region);
2789 if (cp->partial_obj_size() != 0) {
2790 beg_addr = sd.partial_obj_end(beg_region);
2791 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2792 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2793 }
2794 }
2795
2796 if (beg_addr < end_addr) {
2797 // A live object or block of dead space starts in this range of Regions.
2798 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2799
2800 // Create closures and iterate.
2801 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2802 FillClosure fill_closure(cm, space_id);
2803 ParMarkBitMap::IterationStatus status;
2804 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2805 dense_prefix_end);
2806 if (status == ParMarkBitMap::incomplete) {
2807 update_closure.do_addr(update_closure.source());
2808 }
2809 }
2810
2811 // Mark the regions as filled.
2812 RegionData* const beg_cp = sd.region(beg_region);
2813 RegionData* const end_cp = sd.region(end_region);
2814 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2815 cp->set_completed();
2816 }
2817 }
2818
2819 // Return the SpaceId for the space containing addr. If addr is not in the
2820 // heap, last_space_id is returned. In debug mode it expects the address to be
2821 // in the heap and asserts such.
2822 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2823 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2824
2825 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2826 if (_space_info[id].space()->contains(addr)) {
2827 return SpaceId(id);
2828 }
2829 }
2830
2831 assert(false, "no space contains the addr");
2832 return last_space_id;
2833 }
2834
2835 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2836 SpaceId id) {
2837 assert(id < last_space_id, "bad space id");
2838
2839 ParallelCompactData& sd = summary_data();
2840 const SpaceInfo* const space_info = _space_info + id;
2841 ObjectStartArray* const start_array = space_info->start_array();
2842
2843 const MutableSpace* const space = space_info->space();
2844 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2845 HeapWord* const beg_addr = space_info->dense_prefix();
2846 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2847
2848 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2849 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2850 const RegionData* cur_region;
2851 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2852 HeapWord* const addr = cur_region->deferred_obj_addr();
2853 if (addr != NULL) {
2854 if (start_array != NULL) {
2855 start_array->allocate_block(addr);
2856 }
2857 cm->update_contents(oop(addr));
2858 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2859 }
2860 }
2861 }
2862
2863 // Skip over count live words starting from beg, and return the address of the
2864 // next live word. Unless marked, the word corresponding to beg is assumed to
2865 // be dead. Callers must either ensure beg does not correspond to the middle of
2866 // an object, or account for those live words in some other way. Callers must
2867 // also ensure that there are enough live words in the range [beg, end) to skip.
2868 HeapWord*
2869 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2870 {
2871 assert(count > 0, "sanity");
2872
2873 ParMarkBitMap* m = mark_bitmap();
2874 idx_t bits_to_skip = m->words_to_bits(count);
2875 idx_t cur_beg = m->addr_to_bit(beg);
2876 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2877
2878 do {
2879 cur_beg = m->find_obj_beg(cur_beg, search_end);
2880 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2881 const size_t obj_bits = cur_end - cur_beg + 1;
2882 if (obj_bits > bits_to_skip) {
2883 return m->bit_to_addr(cur_beg + bits_to_skip);
2884 }
2885 bits_to_skip -= obj_bits;
2886 cur_beg = cur_end + 1;
2887 } while (bits_to_skip > 0);
2888
2889 // Skipping the desired number of words landed just past the end of an object.
2890 // Find the start of the next object.
2891 cur_beg = m->find_obj_beg(cur_beg, search_end);
2892 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2893 return m->bit_to_addr(cur_beg);
2894 }
2895
2896 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2897 SpaceId src_space_id,
2898 size_t src_region_idx)
2899 {
2900 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2901
2902 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2903 if (split_info.dest_region_addr() == dest_addr) {
2904 // The partial object ending at the split point contains the first word to
2905 // be copied to dest_addr.
2906 return split_info.first_src_addr();
2907 }
2908
2909 const ParallelCompactData& sd = summary_data();
2910 ParMarkBitMap* const bitmap = mark_bitmap();
2911 const size_t RegionSize = ParallelCompactData::RegionSize;
2912
2913 assert(sd.is_region_aligned(dest_addr), "not aligned");
2914 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2915 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2916 HeapWord* const src_region_destination = src_region_ptr->destination();
2917
2918 assert(dest_addr >= src_region_destination, "wrong src region");
2919 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2920
2921 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2922 HeapWord* const src_region_end = src_region_beg + RegionSize;
2923
2924 HeapWord* addr = src_region_beg;
2925 if (dest_addr == src_region_destination) {
2926 // Return the first live word in the source region.
2927 if (partial_obj_size == 0) {
2928 addr = bitmap->find_obj_beg(addr, src_region_end);
2929 assert(addr < src_region_end, "no objects start in src region");
2930 }
2931 return addr;
2932 }
2933
2934 // Must skip some live data.
2935 size_t words_to_skip = dest_addr - src_region_destination;
2936 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2937
2938 if (partial_obj_size >= words_to_skip) {
2939 // All the live words to skip are part of the partial object.
2940 addr += words_to_skip;
2941 if (partial_obj_size == words_to_skip) {
2942 // Find the first live word past the partial object.
2943 addr = bitmap->find_obj_beg(addr, src_region_end);
2944 assert(addr < src_region_end, "wrong src region");
2945 }
2946 return addr;
2947 }
2948
2949 // Skip over the partial object (if any).
2950 if (partial_obj_size != 0) {
2951 words_to_skip -= partial_obj_size;
2952 addr += partial_obj_size;
2953 }
2954
2955 // Skip over live words due to objects that start in the region.
2956 addr = skip_live_words(addr, src_region_end, words_to_skip);
2957 assert(addr < src_region_end, "wrong src region");
2958 return addr;
2959 }
2960
2961 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2962 SpaceId src_space_id,
2963 size_t beg_region,
2964 HeapWord* end_addr)
2965 {
2966 ParallelCompactData& sd = summary_data();
2967
2968 #ifdef ASSERT
2969 MutableSpace* const src_space = _space_info[src_space_id].space();
2970 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2971 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2972 "src_space_id does not match beg_addr");
2973 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2974 "src_space_id does not match end_addr");
2975 #endif // #ifdef ASSERT
2976
2977 RegionData* const beg = sd.region(beg_region);
2978 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2979
2980 // Regions up to new_top() are enqueued if they become available.
2981 HeapWord* const new_top = _space_info[src_space_id].new_top();
2982 RegionData* const enqueue_end =
2983 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2984
2985 for (RegionData* cur = beg; cur < end; ++cur) {
2986 assert(cur->data_size() > 0, "region must have live data");
2987 cur->decrement_destination_count();
2988 if (cur < enqueue_end && cur->available() && cur->claim()) {
2989 cm->push_region(sd.region(cur));
2990 }
2991 }
2992 }
2993
2994 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2995 SpaceId& src_space_id,
2996 HeapWord*& src_space_top,
2997 HeapWord* end_addr)
2998 {
2999 typedef ParallelCompactData::RegionData RegionData;
3000
3001 ParallelCompactData& sd = PSParallelCompact::summary_data();
3002 const size_t region_size = ParallelCompactData::RegionSize;
3003
3004 size_t src_region_idx = 0;
3005
3006 // Skip empty regions (if any) up to the top of the space.
3007 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3008 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3009 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3010 const RegionData* const top_region_ptr =
3011 sd.addr_to_region_ptr(top_aligned_up);
3012 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3013 ++src_region_ptr;
3014 }
3015
3016 if (src_region_ptr < top_region_ptr) {
3017 // The next source region is in the current space. Update src_region_idx
3018 // and the source address to match src_region_ptr.
3019 src_region_idx = sd.region(src_region_ptr);
3020 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3021 if (src_region_addr > closure.source()) {
3022 closure.set_source(src_region_addr);
3023 }
3024 return src_region_idx;
3025 }
3026
3027 // Switch to a new source space and find the first non-empty region.
3028 unsigned int space_id = src_space_id + 1;
3029 assert(space_id < last_space_id, "not enough spaces");
3030
3031 HeapWord* const destination = closure.destination();
3032
3033 do {
3034 MutableSpace* space = _space_info[space_id].space();
3035 HeapWord* const bottom = space->bottom();
3036 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3037
3038 // Iterate over the spaces that do not compact into themselves.
3039 if (bottom_cp->destination() != bottom) {
3040 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3041 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3042
3043 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3044 if (src_cp->live_obj_size() > 0) {
3045 // Found it.
3046 assert(src_cp->destination() == destination,
3047 "first live obj in the space must match the destination");
3048 assert(src_cp->partial_obj_size() == 0,
3049 "a space cannot begin with a partial obj");
3050
3051 src_space_id = SpaceId(space_id);
3052 src_space_top = space->top();
3053 const size_t src_region_idx = sd.region(src_cp);
3054 closure.set_source(sd.region_to_addr(src_region_idx));
3055 return src_region_idx;
3056 } else {
3057 assert(src_cp->data_size() == 0, "sanity");
3058 }
3059 }
3060 }
3061 } while (++space_id < last_space_id);
3062
3063 assert(false, "no source region was found");
3064 return 0;
3065 }
3066
3067 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3068 {
3069 typedef ParMarkBitMap::IterationStatus IterationStatus;
3070 const size_t RegionSize = ParallelCompactData::RegionSize;
3071 ParMarkBitMap* const bitmap = mark_bitmap();
3072 ParallelCompactData& sd = summary_data();
3073 RegionData* const region_ptr = sd.region(region_idx);
3074
3075 // Get the items needed to construct the closure.
3076 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3077 SpaceId dest_space_id = space_id(dest_addr);
3078 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3079 HeapWord* new_top = _space_info[dest_space_id].new_top();
3080 assert(dest_addr < new_top, "sanity");
3081 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3082
3083 // Get the source region and related info.
3084 size_t src_region_idx = region_ptr->source_region();
3085 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3086 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3087
3088 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3089 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3090
3091 // Adjust src_region_idx to prepare for decrementing destination counts (the
3092 // destination count is not decremented when a region is copied to itself).
3093 if (src_region_idx == region_idx) {
3094 src_region_idx += 1;
3095 }
3096
3097 if (bitmap->is_unmarked(closure.source())) {
3098 // The first source word is in the middle of an object; copy the remainder
3099 // of the object or as much as will fit. The fact that pointer updates were
3100 // deferred will be noted when the object header is processed.
3101 HeapWord* const old_src_addr = closure.source();
3102 closure.copy_partial_obj();
3103 if (closure.is_full()) {
3104 decrement_destination_counts(cm, src_space_id, src_region_idx,
3105 closure.source());
3106 region_ptr->set_deferred_obj_addr(NULL);
3107 region_ptr->set_completed();
3108 return;
3109 }
3110
3111 HeapWord* const end_addr = sd.region_align_down(closure.source());
3112 if (sd.region_align_down(old_src_addr) != end_addr) {
3113 // The partial object was copied from more than one source region.
3114 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3115
3116 // Move to the next source region, possibly switching spaces as well. All
3117 // args except end_addr may be modified.
3118 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3119 end_addr);
3120 }
3121 }
3122
3123 do {
3124 HeapWord* const cur_addr = closure.source();
3125 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3126 src_space_top);
3127 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3128
3129 if (status == ParMarkBitMap::incomplete) {
3130 // The last obj that starts in the source region does not end in the
3131 // region.
3132 assert(closure.source() < end_addr, "sanity");
3133 HeapWord* const obj_beg = closure.source();
3134 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3135 src_space_top);
3136 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3137 if (obj_end < range_end) {
3138 // The end was found; the entire object will fit.
3139 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3140 assert(status != ParMarkBitMap::would_overflow, "sanity");
3141 } else {
3142 // The end was not found; the object will not fit.
3143 assert(range_end < src_space_top, "obj cannot cross space boundary");
3144 status = ParMarkBitMap::would_overflow;
3145 }
3146 }
3147
3148 if (status == ParMarkBitMap::would_overflow) {
3149 // The last object did not fit. Note that interior oop updates were
3150 // deferred, then copy enough of the object to fill the region.
3151 region_ptr->set_deferred_obj_addr(closure.destination());
3152 status = closure.copy_until_full(); // copies from closure.source()
3153
3154 decrement_destination_counts(cm, src_space_id, src_region_idx,
3155 closure.source());
3156 region_ptr->set_completed();
3157 return;
3158 }
3159
3160 if (status == ParMarkBitMap::full) {
3161 decrement_destination_counts(cm, src_space_id, src_region_idx,
3162 closure.source());
3163 region_ptr->set_deferred_obj_addr(NULL);
3164 region_ptr->set_completed();
3165 return;
3166 }
3167
3168 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3169
3170 // Move to the next source region, possibly switching spaces as well. All
3171 // args except end_addr may be modified.
3172 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3173 end_addr);
3174 } while (true);
3175 }
3176
3177 void PSParallelCompact::fill_blocks(size_t region_idx)
3178 {
3179 // Fill in the block table elements for the specified region. Each block
3180 // table element holds the number of live words in the region that are to the
3181 // left of the first object that starts in the block. Thus only blocks in
3182 // which an object starts need to be filled.
3183 //
3184 // The algorithm scans the section of the bitmap that corresponds to the
3185 // region, keeping a running total of the live words. When an object start is
3186 // found, if it's the first to start in the block that contains it, the
3187 // current total is written to the block table element.
3188 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3189 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3190 const size_t RegionSize = ParallelCompactData::RegionSize;
3191
3192 ParallelCompactData& sd = summary_data();
3193 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3194 if (partial_obj_size >= RegionSize) {
3195 return; // No objects start in this region.
3196 }
3197
3198 // Ensure the first loop iteration decides that the block has changed.
3199 size_t cur_block = sd.block_count();
3200
3201 const ParMarkBitMap* const bitmap = mark_bitmap();
3202
3203 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3204 assert((size_t)1 << Log2BitsPerBlock ==
3205 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3206
3207 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3208 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3209 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3210 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3211 while (beg_bit < range_end) {
3212 const size_t new_block = beg_bit >> Log2BitsPerBlock;
3213 if (new_block != cur_block) {
3214 cur_block = new_block;
3215 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3216 }
3217
3218 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3219 if (end_bit < range_end - 1) {
3220 live_bits += end_bit - beg_bit + 1;
3221 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3222 } else {
3223 return;
3224 }
3225 }
3226 }
3227
3228 void
3229 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3230 const MutableSpace* sp = space(space_id);
3231 if (sp->is_empty()) {
3232 return;
3233 }
3234
3235 ParallelCompactData& sd = PSParallelCompact::summary_data();
3236 ParMarkBitMap* const bitmap = mark_bitmap();
3237 HeapWord* const dp_addr = dense_prefix(space_id);
3238 HeapWord* beg_addr = sp->bottom();
3239 HeapWord* end_addr = sp->top();
3240
3241 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3242
3243 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3244 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3245 if (beg_region < dp_region) {
3246 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3247 }
3248
3249 // The destination of the first live object that starts in the region is one
3250 // past the end of the partial object entering the region (if any).
3251 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3252 HeapWord* const new_top = _space_info[space_id].new_top();
3253 assert(new_top >= dest_addr, "bad new_top value");
3254 const size_t words = pointer_delta(new_top, dest_addr);
3255
3256 if (words > 0) {
3257 ObjectStartArray* start_array = _space_info[space_id].start_array();
3258 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3259
3260 ParMarkBitMap::IterationStatus status;
3261 status = bitmap->iterate(&closure, dest_addr, end_addr);
3262 assert(status == ParMarkBitMap::full, "iteration not complete");
3263 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3264 "live objects skipped because closure is full");
3265 }
3266 }
3267
3268 jlong PSParallelCompact::millis_since_last_gc() {
3269 // We need a monotonically non-decreasing time in ms but
3270 // os::javaTimeMillis() does not guarantee monotonicity.
3271 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3272 jlong ret_val = now - _time_of_last_gc;
3273 // XXX See note in genCollectedHeap::millis_since_last_gc().
3274 if (ret_val < 0) {
3275 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);)
3276 return 0;
3277 }
3278 return ret_val;
3279 }
3280
3281 void PSParallelCompact::reset_millis_since_last_gc() {
3282 // We need a monotonically non-decreasing time in ms but
3283 // os::javaTimeMillis() does not guarantee monotonicity.
3284 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3285 }
3286
3287 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3288 {
3289 if (source() != destination()) {
3290 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3291 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3292 }
3293 update_state(words_remaining());
3294 assert(is_full(), "sanity");
3295 return ParMarkBitMap::full;
3296 }
3297
3298 void MoveAndUpdateClosure::copy_partial_obj()
3299 {
3300 size_t words = words_remaining();
3301
3302 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3303 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3304 if (end_addr < range_end) {
3305 words = bitmap()->obj_size(source(), end_addr);
3306 }
3307
3308 // This test is necessary; if omitted, the pointer updates to a partial object
3309 // that crosses the dense prefix boundary could be overwritten.
3310 if (source() != destination()) {
3311 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3312 Copy::aligned_conjoint_words(source(), destination(), words);
3313 }
3314 update_state(words);
3315 }
3316
3317 ParMarkBitMapClosure::IterationStatus
3318 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3319 assert(destination() != NULL, "sanity");
3320 assert(bitmap()->obj_size(addr) == words, "bad size");
3321
3322 _source = addr;
3323 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3324 destination(), "wrong destination");
3325
3326 if (words > words_remaining()) {
3327 return ParMarkBitMap::would_overflow;
3328 }
3329
3330 // The start_array must be updated even if the object is not moving.
3331 if (_start_array != NULL) {
3332 _start_array->allocate_block(destination());
3333 }
3334
3335 if (destination() != source()) {
3336 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3337 Copy::aligned_conjoint_words(source(), destination(), words);
3338 }
3339
3340 oop moved_oop = (oop) destination();
3341 compaction_manager()->update_contents(moved_oop);
3342 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3343
3344 update_state(words);
3345 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3346 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3347 }
3348
3349 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3350 ParCompactionManager* cm,
3351 PSParallelCompact::SpaceId space_id) :
3352 ParMarkBitMapClosure(mbm, cm),
3353 _space_id(space_id),
3354 _start_array(PSParallelCompact::start_array(space_id))
3355 {
3356 }
3357
3358 // Updates the references in the object to their new values.
3359 ParMarkBitMapClosure::IterationStatus
3360 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3361 do_addr(addr);
3362 return ParMarkBitMap::incomplete;
3363 }
3364
3365 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3366 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3367 _start_array(PSParallelCompact::start_array(space_id))
3368 {
3369 assert(space_id == PSParallelCompact::old_space_id,
3370 "cannot use FillClosure in the young gen");
3371 }
3372
3373 ParMarkBitMapClosure::IterationStatus
3374 FillClosure::do_addr(HeapWord* addr, size_t size) {
3375 CollectedHeap::fill_with_objects(addr, size);
3376 HeapWord* const end = addr + size;
3377 do {
3378 _start_array->allocate_block(addr);
3379 addr += oop(addr)->size();
3380 } while (addr < end);
3381 return ParMarkBitMap::incomplete;
3382 }