1 /*
2 * Copyright (c) 2005, 2021, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderDataGraph.hpp"
27 #include "classfile/javaClasses.inline.hpp"
28 #include "classfile/stringTable.hpp"
29 #include "classfile/symbolTable.hpp"
30 #include "classfile/systemDictionary.hpp"
31 #include "code/codeCache.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/parallel/parallelArguments.hpp"
34 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
35 #include "gc/parallel/parMarkBitMap.inline.hpp"
36 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
37 #include "gc/parallel/psCompactionManager.inline.hpp"
38 #include "gc/parallel/psOldGen.hpp"
39 #include "gc/parallel/psParallelCompact.inline.hpp"
40 #include "gc/parallel/psPromotionManager.inline.hpp"
41 #include "gc/parallel/psRootType.hpp"
42 #include "gc/parallel/psScavenge.hpp"
43 #include "gc/parallel/psStringDedup.hpp"
44 #include "gc/parallel/psYoungGen.hpp"
45 #include "gc/shared/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/oopStorage.inline.hpp"
54 #include "gc/shared/oopStorageSet.inline.hpp"
55 #include "gc/shared/oopStorageSetParState.inline.hpp"
56 #include "gc/shared/referencePolicy.hpp"
57 #include "gc/shared/referenceProcessor.hpp"
58 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
59 #include "gc/shared/spaceDecorator.inline.hpp"
60 #include "gc/shared/taskTerminator.hpp"
61 #include "gc/shared/weakProcessor.inline.hpp"
62 #include "gc/shared/workerPolicy.hpp"
63 #include "gc/shared/workgroup.hpp"
64 #include "logging/log.hpp"
65 #include "memory/iterator.inline.hpp"
66 #include "memory/metaspaceUtils.hpp"
67 #include "memory/resourceArea.hpp"
68 #include "memory/universe.hpp"
69 #include "oops/access.inline.hpp"
70 #include "oops/flatArrayKlass.inline.hpp"
71 #include "oops/instanceClassLoaderKlass.inline.hpp"
72 #include "oops/instanceKlass.inline.hpp"
73 #include "oops/instanceMirrorKlass.inline.hpp"
74 #include "oops/methodData.hpp"
75 #include "oops/objArrayKlass.inline.hpp"
76 #include "oops/oop.inline.hpp"
77 #include "runtime/atomic.hpp"
78 #include "runtime/handles.inline.hpp"
79 #include "runtime/java.hpp"
80 #include "runtime/safepoint.hpp"
81 #include "runtime/vmThread.hpp"
82 #include "services/memTracker.hpp"
83 #include "services/memoryService.hpp"
84 #include "utilities/align.hpp"
85 #include "utilities/debug.hpp"
86 #include "utilities/events.hpp"
87 #include "utilities/formatBuffer.hpp"
88 #include "utilities/macros.hpp"
89 #include "utilities/stack.inline.hpp"
90 #if INCLUDE_JVMCI
91 #include "jvmci/jvmci.hpp"
92 #endif
93
94 #include <math.h>
95
96 // All sizes are in HeapWords.
97 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
98 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
99 const size_t ParallelCompactData::RegionSizeBytes =
100 RegionSize << LogHeapWordSize;
101 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
102 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
103 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
104
105 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
106 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
107 const size_t ParallelCompactData::BlockSizeBytes =
108 BlockSize << LogHeapWordSize;
109 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
110 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
111 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
112
113 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
114 const size_t ParallelCompactData::Log2BlocksPerRegion =
115 Log2RegionSize - Log2BlockSize;
116
117 const ParallelCompactData::RegionData::region_sz_t
118 ParallelCompactData::RegionData::dc_shift = 27;
119
120 const ParallelCompactData::RegionData::region_sz_t
121 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
122
123 const ParallelCompactData::RegionData::region_sz_t
124 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
125
126 const ParallelCompactData::RegionData::region_sz_t
127 ParallelCompactData::RegionData::los_mask = ~dc_mask;
128
129 const ParallelCompactData::RegionData::region_sz_t
130 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
131
132 const ParallelCompactData::RegionData::region_sz_t
133 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
134
135 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
136
137 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
138 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
139
140 double PSParallelCompact::_dwl_mean;
141 double PSParallelCompact::_dwl_std_dev;
142 double PSParallelCompact::_dwl_first_term;
143 double PSParallelCompact::_dwl_adjustment;
144 #ifdef ASSERT
145 bool PSParallelCompact::_dwl_initialized = false;
146 #endif // #ifdef ASSERT
147
148 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
149 HeapWord* destination)
150 {
151 assert(src_region_idx != 0, "invalid src_region_idx");
152 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
153 assert(destination != NULL, "invalid destination argument");
154
155 _src_region_idx = src_region_idx;
156 _partial_obj_size = partial_obj_size;
157 _destination = destination;
158
159 // These fields may not be updated below, so make sure they're clear.
160 assert(_dest_region_addr == NULL, "should have been cleared");
161 assert(_first_src_addr == NULL, "should have been cleared");
162
163 // Determine the number of destination regions for the partial object.
164 HeapWord* const last_word = destination + partial_obj_size - 1;
165 const ParallelCompactData& sd = PSParallelCompact::summary_data();
166 HeapWord* const beg_region_addr = sd.region_align_down(destination);
167 HeapWord* const end_region_addr = sd.region_align_down(last_word);
168
169 if (beg_region_addr == end_region_addr) {
170 // One destination region.
171 _destination_count = 1;
172 if (end_region_addr == destination) {
173 // The destination falls on a region boundary, thus the first word of the
174 // partial object will be the first word copied to the destination region.
175 _dest_region_addr = end_region_addr;
176 _first_src_addr = sd.region_to_addr(src_region_idx);
177 }
178 } else {
179 // Two destination regions. When copied, the partial object will cross a
180 // destination region boundary, so a word somewhere within the partial
181 // object will be the first word copied to the second destination region.
182 _destination_count = 2;
183 _dest_region_addr = end_region_addr;
184 const size_t ofs = pointer_delta(end_region_addr, destination);
185 assert(ofs < _partial_obj_size, "sanity");
186 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
187 }
188 }
189
190 void SplitInfo::clear()
191 {
192 _src_region_idx = 0;
193 _partial_obj_size = 0;
194 _destination = NULL;
195 _destination_count = 0;
196 _dest_region_addr = NULL;
197 _first_src_addr = NULL;
198 assert(!is_valid(), "sanity");
199 }
200
201 #ifdef ASSERT
202 void SplitInfo::verify_clear()
203 {
204 assert(_src_region_idx == 0, "not clear");
205 assert(_partial_obj_size == 0, "not clear");
206 assert(_destination == NULL, "not clear");
207 assert(_destination_count == 0, "not clear");
208 assert(_dest_region_addr == NULL, "not clear");
209 assert(_first_src_addr == NULL, "not clear");
210 }
211 #endif // #ifdef ASSERT
212
213
214 void PSParallelCompact::print_on_error(outputStream* st) {
215 _mark_bitmap.print_on_error(st);
216 }
217
218 #ifndef PRODUCT
219 const char* PSParallelCompact::space_names[] = {
220 "old ", "eden", "from", "to "
221 };
222
223 void PSParallelCompact::print_region_ranges() {
224 if (!log_develop_is_enabled(Trace, gc, compaction)) {
225 return;
226 }
227 Log(gc, compaction) log;
228 ResourceMark rm;
229 LogStream ls(log.trace());
230 Universe::print_on(&ls);
231 log.trace("space bottom top end new_top");
232 log.trace("------ ---------- ---------- ---------- ----------");
233
234 for (unsigned int id = 0; id < last_space_id; ++id) {
235 const MutableSpace* space = _space_info[id].space();
236 log.trace("%u %s "
237 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
238 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
239 id, space_names[id],
240 summary_data().addr_to_region_idx(space->bottom()),
241 summary_data().addr_to_region_idx(space->top()),
242 summary_data().addr_to_region_idx(space->end()),
243 summary_data().addr_to_region_idx(_space_info[id].new_top()));
244 }
245 }
246
247 void
248 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
249 {
250 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
251 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
252
253 ParallelCompactData& sd = PSParallelCompact::summary_data();
254 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
255 log_develop_trace(gc, compaction)(
256 REGION_IDX_FORMAT " " PTR_FORMAT " "
257 REGION_IDX_FORMAT " " PTR_FORMAT " "
258 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
259 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
260 i, p2i(c->data_location()), dci, p2i(c->destination()),
261 c->partial_obj_size(), c->live_obj_size(),
262 c->data_size(), c->source_region(), c->destination_count());
263
264 #undef REGION_IDX_FORMAT
265 #undef REGION_DATA_FORMAT
266 }
267
268 void
269 print_generic_summary_data(ParallelCompactData& summary_data,
270 HeapWord* const beg_addr,
271 HeapWord* const end_addr)
272 {
273 size_t total_words = 0;
274 size_t i = summary_data.addr_to_region_idx(beg_addr);
275 const size_t last = summary_data.addr_to_region_idx(end_addr);
276 HeapWord* pdest = 0;
277
278 while (i < last) {
279 ParallelCompactData::RegionData* c = summary_data.region(i);
280 if (c->data_size() != 0 || c->destination() != pdest) {
281 print_generic_summary_region(i, c);
282 total_words += c->data_size();
283 pdest = c->destination();
284 }
285 ++i;
286 }
287
288 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
289 }
290
291 void
292 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
293 HeapWord* const beg_addr,
294 HeapWord* const end_addr) {
295 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
296 }
297
298 void
299 print_generic_summary_data(ParallelCompactData& summary_data,
300 SpaceInfo* space_info)
301 {
302 if (!log_develop_is_enabled(Trace, gc, compaction)) {
303 return;
304 }
305
306 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
307 const MutableSpace* space = space_info[id].space();
308 print_generic_summary_data(summary_data, space->bottom(),
309 MAX2(space->top(), space_info[id].new_top()));
310 }
311 }
312
313 void
314 print_initial_summary_data(ParallelCompactData& summary_data,
315 const MutableSpace* space) {
316 if (space->top() == space->bottom()) {
317 return;
318 }
319
320 const size_t region_size = ParallelCompactData::RegionSize;
321 typedef ParallelCompactData::RegionData RegionData;
322 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
323 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
324 const RegionData* c = summary_data.region(end_region - 1);
325 HeapWord* end_addr = c->destination() + c->data_size();
326 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
327
328 // Print (and count) the full regions at the beginning of the space.
329 size_t full_region_count = 0;
330 size_t i = summary_data.addr_to_region_idx(space->bottom());
331 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
332 ParallelCompactData::RegionData* c = summary_data.region(i);
333 log_develop_trace(gc, compaction)(
334 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
335 i, p2i(c->destination()),
336 c->partial_obj_size(), c->live_obj_size(),
337 c->data_size(), c->source_region(), c->destination_count());
338 ++full_region_count;
339 ++i;
340 }
341
342 size_t live_to_right = live_in_space - full_region_count * region_size;
343
344 double max_reclaimed_ratio = 0.0;
345 size_t max_reclaimed_ratio_region = 0;
346 size_t max_dead_to_right = 0;
347 size_t max_live_to_right = 0;
348
349 // Print the 'reclaimed ratio' for regions while there is something live in
350 // the region or to the right of it. The remaining regions are empty (and
351 // uninteresting), and computing the ratio will result in division by 0.
352 while (i < end_region && live_to_right > 0) {
353 c = summary_data.region(i);
354 HeapWord* const region_addr = summary_data.region_to_addr(i);
355 const size_t used_to_right = pointer_delta(space->top(), region_addr);
356 const size_t dead_to_right = used_to_right - live_to_right;
357 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
358
359 if (reclaimed_ratio > max_reclaimed_ratio) {
360 max_reclaimed_ratio = reclaimed_ratio;
361 max_reclaimed_ratio_region = i;
362 max_dead_to_right = dead_to_right;
363 max_live_to_right = live_to_right;
364 }
365
366 ParallelCompactData::RegionData* c = summary_data.region(i);
367 log_develop_trace(gc, compaction)(
368 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
369 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
370 i, p2i(c->destination()),
371 c->partial_obj_size(), c->live_obj_size(),
372 c->data_size(), c->source_region(), c->destination_count(),
373 reclaimed_ratio, dead_to_right, live_to_right);
374
375
376 live_to_right -= c->data_size();
377 ++i;
378 }
379
380 // Any remaining regions are empty. Print one more if there is one.
381 if (i < end_region) {
382 ParallelCompactData::RegionData* c = summary_data.region(i);
383 log_develop_trace(gc, compaction)(
384 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
385 i, p2i(c->destination()),
386 c->partial_obj_size(), c->live_obj_size(),
387 c->data_size(), c->source_region(), c->destination_count());
388 }
389
390 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
391 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
392 }
393
394 void
395 print_initial_summary_data(ParallelCompactData& summary_data,
396 SpaceInfo* space_info) {
397 if (!log_develop_is_enabled(Trace, gc, compaction)) {
398 return;
399 }
400
401 unsigned int id = PSParallelCompact::old_space_id;
402 const MutableSpace* space;
403 do {
404 space = space_info[id].space();
405 print_initial_summary_data(summary_data, space);
406 } while (++id < PSParallelCompact::eden_space_id);
407
408 do {
409 space = space_info[id].space();
410 print_generic_summary_data(summary_data, space->bottom(), space->top());
411 } while (++id < PSParallelCompact::last_space_id);
412 }
413 #endif // #ifndef PRODUCT
414
415 ParallelCompactData::ParallelCompactData() :
416 _region_start(NULL),
417 DEBUG_ONLY(_region_end(NULL) COMMA)
418 _region_vspace(NULL),
419 _reserved_byte_size(0),
420 _region_data(NULL),
421 _region_count(0),
422 _block_vspace(NULL),
423 _block_data(NULL),
424 _block_count(0) {}
425
426 bool ParallelCompactData::initialize(MemRegion covered_region)
427 {
428 _region_start = covered_region.start();
429 const size_t region_size = covered_region.word_size();
430 DEBUG_ONLY(_region_end = _region_start + region_size;)
431
432 assert(region_align_down(_region_start) == _region_start,
433 "region start not aligned");
434 assert((region_size & RegionSizeOffsetMask) == 0,
435 "region size not a multiple of RegionSize");
436
437 bool result = initialize_region_data(region_size) && initialize_block_data();
438 return result;
439 }
440
441 PSVirtualSpace*
442 ParallelCompactData::create_vspace(size_t count, size_t element_size)
443 {
444 const size_t raw_bytes = count * element_size;
445 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
446 const size_t granularity = os::vm_allocation_granularity();
447 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
448
449 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
450 MAX2(page_sz, granularity);
451 ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
452 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
453 rs.size());
454
455 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
456
457 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
458 if (vspace != 0) {
459 if (vspace->expand_by(_reserved_byte_size)) {
460 return vspace;
461 }
462 delete vspace;
463 // Release memory reserved in the space.
464 rs.release();
465 }
466
467 return 0;
468 }
469
470 bool ParallelCompactData::initialize_region_data(size_t region_size)
471 {
472 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
473 _region_vspace = create_vspace(count, sizeof(RegionData));
474 if (_region_vspace != 0) {
475 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
476 _region_count = count;
477 return true;
478 }
479 return false;
480 }
481
482 bool ParallelCompactData::initialize_block_data()
483 {
484 assert(_region_count != 0, "region data must be initialized first");
485 const size_t count = _region_count << Log2BlocksPerRegion;
486 _block_vspace = create_vspace(count, sizeof(BlockData));
487 if (_block_vspace != 0) {
488 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
489 _block_count = count;
490 return true;
491 }
492 return false;
493 }
494
495 void ParallelCompactData::clear()
496 {
497 memset(_region_data, 0, _region_vspace->committed_size());
498 memset(_block_data, 0, _block_vspace->committed_size());
499 }
500
501 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
502 assert(beg_region <= _region_count, "beg_region out of range");
503 assert(end_region <= _region_count, "end_region out of range");
504 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
505
506 const size_t region_cnt = end_region - beg_region;
507 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
508
509 const size_t beg_block = beg_region * BlocksPerRegion;
510 const size_t block_cnt = region_cnt * BlocksPerRegion;
511 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
512 }
513
514 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
515 {
516 const RegionData* cur_cp = region(region_idx);
517 const RegionData* const end_cp = region(region_count() - 1);
518
519 HeapWord* result = region_to_addr(region_idx);
520 if (cur_cp < end_cp) {
521 do {
522 result += cur_cp->partial_obj_size();
523 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
524 }
525 return result;
526 }
527
528 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
529 {
530 const size_t obj_ofs = pointer_delta(addr, _region_start);
531 const size_t beg_region = obj_ofs >> Log2RegionSize;
532 // end_region is inclusive
533 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
534
535 if (beg_region == end_region) {
536 // All in one region.
537 _region_data[beg_region].add_live_obj(len);
538 return;
539 }
540
541 // First region.
542 const size_t beg_ofs = region_offset(addr);
543 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
544
545 // Middle regions--completely spanned by this object.
546 for (size_t region = beg_region + 1; region < end_region; ++region) {
547 _region_data[region].set_partial_obj_size(RegionSize);
548 _region_data[region].set_partial_obj_addr(addr);
549 }
550
551 // Last region.
552 const size_t end_ofs = region_offset(addr + len - 1);
553 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
554 _region_data[end_region].set_partial_obj_addr(addr);
555 }
556
557 void
558 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
559 {
560 assert(is_region_aligned(beg), "not RegionSize aligned");
561 assert(is_region_aligned(end), "not RegionSize aligned");
562
563 size_t cur_region = addr_to_region_idx(beg);
564 const size_t end_region = addr_to_region_idx(end);
565 HeapWord* addr = beg;
566 while (cur_region < end_region) {
567 _region_data[cur_region].set_destination(addr);
568 _region_data[cur_region].set_destination_count(0);
569 _region_data[cur_region].set_source_region(cur_region);
570 _region_data[cur_region].set_data_location(addr);
571
572 // Update live_obj_size so the region appears completely full.
573 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
574 _region_data[cur_region].set_live_obj_size(live_size);
575
576 ++cur_region;
577 addr += RegionSize;
578 }
579 }
580
581 // Find the point at which a space can be split and, if necessary, record the
582 // split point.
583 //
584 // If the current src region (which overflowed the destination space) doesn't
585 // have a partial object, the split point is at the beginning of the current src
586 // region (an "easy" split, no extra bookkeeping required).
587 //
588 // If the current src region has a partial object, the split point is in the
589 // region where that partial object starts (call it the split_region). If
590 // split_region has a partial object, then the split point is just after that
591 // partial object (a "hard" split where we have to record the split data and
592 // zero the partial_obj_size field). With a "hard" split, we know that the
593 // partial_obj ends within split_region because the partial object that caused
594 // the overflow starts in split_region. If split_region doesn't have a partial
595 // obj, then the split is at the beginning of split_region (another "easy"
596 // split).
597 HeapWord*
598 ParallelCompactData::summarize_split_space(size_t src_region,
599 SplitInfo& split_info,
600 HeapWord* destination,
601 HeapWord* target_end,
602 HeapWord** target_next)
603 {
604 assert(destination <= target_end, "sanity");
605 assert(destination + _region_data[src_region].data_size() > target_end,
606 "region should not fit into target space");
607 assert(is_region_aligned(target_end), "sanity");
608
609 size_t split_region = src_region;
610 HeapWord* split_destination = destination;
611 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
612
613 if (destination + partial_obj_size > target_end) {
614 // The split point is just after the partial object (if any) in the
615 // src_region that contains the start of the object that overflowed the
616 // destination space.
617 //
618 // Find the start of the "overflow" object and set split_region to the
619 // region containing it.
620 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
621 split_region = addr_to_region_idx(overflow_obj);
622
623 // Clear the source_region field of all destination regions whose first word
624 // came from data after the split point (a non-null source_region field
625 // implies a region must be filled).
626 //
627 // An alternative to the simple loop below: clear during post_compact(),
628 // which uses memcpy instead of individual stores, and is easy to
629 // parallelize. (The downside is that it clears the entire RegionData
630 // object as opposed to just one field.)
631 //
632 // post_compact() would have to clear the summary data up to the highest
633 // address that was written during the summary phase, which would be
634 //
635 // max(top, max(new_top, clear_top))
636 //
637 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
638 // to target_end.
639 const RegionData* const sr = region(split_region);
640 const size_t beg_idx =
641 addr_to_region_idx(region_align_up(sr->destination() +
642 sr->partial_obj_size()));
643 const size_t end_idx = addr_to_region_idx(target_end);
644
645 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
646 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
647 _region_data[idx].set_source_region(0);
648 }
649
650 // Set split_destination and partial_obj_size to reflect the split region.
651 split_destination = sr->destination();
652 partial_obj_size = sr->partial_obj_size();
653 }
654
655 // The split is recorded only if a partial object extends onto the region.
656 if (partial_obj_size != 0) {
657 _region_data[split_region].set_partial_obj_size(0);
658 split_info.record(split_region, partial_obj_size, split_destination);
659 }
660
661 // Setup the continuation addresses.
662 *target_next = split_destination + partial_obj_size;
663 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
664
665 if (log_develop_is_enabled(Trace, gc, compaction)) {
666 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
667 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
668 split_type, p2i(source_next), split_region, partial_obj_size);
669 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
670 split_type, p2i(split_destination),
671 addr_to_region_idx(split_destination),
672 p2i(*target_next));
673
674 if (partial_obj_size != 0) {
675 HeapWord* const po_beg = split_info.destination();
676 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
677 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
678 split_type,
679 p2i(po_beg), addr_to_region_idx(po_beg),
680 p2i(po_end), addr_to_region_idx(po_end));
681 }
682 }
683
684 return source_next;
685 }
686
687 bool ParallelCompactData::summarize(SplitInfo& split_info,
688 HeapWord* source_beg, HeapWord* source_end,
689 HeapWord** source_next,
690 HeapWord* target_beg, HeapWord* target_end,
691 HeapWord** target_next)
692 {
693 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
694 log_develop_trace(gc, compaction)(
695 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
696 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
697 p2i(source_beg), p2i(source_end), p2i(source_next_val),
698 p2i(target_beg), p2i(target_end), p2i(*target_next));
699
700 size_t cur_region = addr_to_region_idx(source_beg);
701 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
702
703 HeapWord *dest_addr = target_beg;
704 while (cur_region < end_region) {
705 // The destination must be set even if the region has no data.
706 _region_data[cur_region].set_destination(dest_addr);
707
708 size_t words = _region_data[cur_region].data_size();
709 if (words > 0) {
710 // If cur_region does not fit entirely into the target space, find a point
711 // at which the source space can be 'split' so that part is copied to the
712 // target space and the rest is copied elsewhere.
713 if (dest_addr + words > target_end) {
714 assert(source_next != NULL, "source_next is NULL when splitting");
715 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
716 target_end, target_next);
717 return false;
718 }
719
720 // Compute the destination_count for cur_region, and if necessary, update
721 // source_region for a destination region. The source_region field is
722 // updated if cur_region is the first (left-most) region to be copied to a
723 // destination region.
724 //
725 // The destination_count calculation is a bit subtle. A region that has
726 // data that compacts into itself does not count itself as a destination.
727 // This maintains the invariant that a zero count means the region is
728 // available and can be claimed and then filled.
729 uint destination_count = 0;
730 if (split_info.is_split(cur_region)) {
731 // The current region has been split: the partial object will be copied
732 // to one destination space and the remaining data will be copied to
733 // another destination space. Adjust the initial destination_count and,
734 // if necessary, set the source_region field if the partial object will
735 // cross a destination region boundary.
736 destination_count = split_info.destination_count();
737 if (destination_count == 2) {
738 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
739 _region_data[dest_idx].set_source_region(cur_region);
740 }
741 }
742
743 HeapWord* const last_addr = dest_addr + words - 1;
744 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
745 const size_t dest_region_2 = addr_to_region_idx(last_addr);
746
747 // Initially assume that the destination regions will be the same and
748 // adjust the value below if necessary. Under this assumption, if
749 // cur_region == dest_region_2, then cur_region will be compacted
750 // completely into itself.
751 destination_count += cur_region == dest_region_2 ? 0 : 1;
752 if (dest_region_1 != dest_region_2) {
753 // Destination regions differ; adjust destination_count.
754 destination_count += 1;
755 // Data from cur_region will be copied to the start of dest_region_2.
756 _region_data[dest_region_2].set_source_region(cur_region);
757 } else if (is_region_aligned(dest_addr)) {
758 // Data from cur_region will be copied to the start of the destination
759 // region.
760 _region_data[dest_region_1].set_source_region(cur_region);
761 }
762
763 _region_data[cur_region].set_destination_count(destination_count);
764 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
765 dest_addr += words;
766 }
767
768 ++cur_region;
769 }
770
771 *target_next = dest_addr;
772 return true;
773 }
774
775 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const {
776 assert(addr != NULL, "Should detect NULL oop earlier");
777 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
778 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
779
780 // Region covering the object.
781 RegionData* const region_ptr = addr_to_region_ptr(addr);
782 HeapWord* result = region_ptr->destination();
783
784 // If the entire Region is live, the new location is region->destination + the
785 // offset of the object within in the Region.
786
787 // Run some performance tests to determine if this special case pays off. It
788 // is worth it for pointers into the dense prefix. If the optimization to
789 // avoid pointer updates in regions that only point to the dense prefix is
790 // ever implemented, this should be revisited.
791 if (region_ptr->data_size() == RegionSize) {
792 result += region_offset(addr);
793 return result;
794 }
795
796 // Otherwise, the new location is region->destination + block offset + the
797 // number of live words in the Block that are (a) to the left of addr and (b)
798 // due to objects that start in the Block.
799
800 // Fill in the block table if necessary. This is unsynchronized, so multiple
801 // threads may fill the block table for a region (harmless, since it is
802 // idempotent).
803 if (!region_ptr->blocks_filled()) {
804 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
805 region_ptr->set_blocks_filled();
806 }
807
808 HeapWord* const search_start = block_align_down(addr);
809 const size_t block_offset = addr_to_block_ptr(addr)->offset();
810
811 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
812 const size_t live = bitmap->live_words_in_range(cm, search_start, cast_to_oop(addr));
813 result += block_offset + live;
814 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
815 return result;
816 }
817
818 #ifdef ASSERT
819 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
820 {
821 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
822 const size_t* const end = (const size_t*)vspace->committed_high_addr();
823 for (const size_t* p = beg; p < end; ++p) {
824 assert(*p == 0, "not zero");
825 }
826 }
827
828 void ParallelCompactData::verify_clear()
829 {
830 verify_clear(_region_vspace);
831 verify_clear(_block_vspace);
832 }
833 #endif // #ifdef ASSERT
834
835 STWGCTimer PSParallelCompact::_gc_timer;
836 ParallelOldTracer PSParallelCompact::_gc_tracer;
837 elapsedTimer PSParallelCompact::_accumulated_time;
838 unsigned int PSParallelCompact::_total_invocations = 0;
839 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
840 CollectorCounters* PSParallelCompact::_counters = NULL;
841 ParMarkBitMap PSParallelCompact::_mark_bitmap;
842 ParallelCompactData PSParallelCompact::_summary_data;
843
844 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
845
846 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
847
848 class PCReferenceProcessor: public ReferenceProcessor {
849 public:
850 PCReferenceProcessor(
851 BoolObjectClosure* is_subject_to_discovery,
852 BoolObjectClosure* is_alive_non_header) :
853 ReferenceProcessor(is_subject_to_discovery,
854 ParallelGCThreads, // mt processing degree
855 true, // mt discovery
856 ParallelGCThreads, // mt discovery degree
857 true, // atomic_discovery
858 is_alive_non_header) {
859 }
860
861 template<typename T> bool discover(oop obj, ReferenceType type) {
862 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
863 T heap_oop = RawAccess<>::oop_load(referent_addr);
864 oop referent = CompressedOops::decode_not_null(heap_oop);
865 return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
866 && ReferenceProcessor::discover_reference(obj, type);
867 }
868 virtual bool discover_reference(oop obj, ReferenceType type) {
869 if (UseCompressedOops) {
870 return discover<narrowOop>(obj, type);
871 } else {
872 return discover<oop>(obj, type);
873 }
874 }
875 };
876
877 void PSParallelCompact::post_initialize() {
878 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
879 _span_based_discoverer.set_span(heap->reserved_region());
880 _ref_processor =
881 new PCReferenceProcessor(&_span_based_discoverer,
882 &_is_alive_closure); // non-header is alive closure
883
884 _counters = new CollectorCounters("Parallel full collection pauses", 1);
885
886 // Initialize static fields in ParCompactionManager.
887 ParCompactionManager::initialize(mark_bitmap());
888 }
889
890 bool PSParallelCompact::initialize() {
891 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
892 MemRegion mr = heap->reserved_region();
893
894 // Was the old gen get allocated successfully?
895 if (!heap->old_gen()->is_allocated()) {
896 return false;
897 }
898
899 initialize_space_info();
900 initialize_dead_wood_limiter();
901
902 if (!_mark_bitmap.initialize(mr)) {
903 vm_shutdown_during_initialization(
904 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
905 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
906 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
907 return false;
908 }
909
910 if (!_summary_data.initialize(mr)) {
911 vm_shutdown_during_initialization(
912 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
913 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
914 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
915 return false;
916 }
917
918 return true;
919 }
920
921 void PSParallelCompact::initialize_space_info()
922 {
923 memset(&_space_info, 0, sizeof(_space_info));
924
925 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
926 PSYoungGen* young_gen = heap->young_gen();
927
928 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
929 _space_info[eden_space_id].set_space(young_gen->eden_space());
930 _space_info[from_space_id].set_space(young_gen->from_space());
931 _space_info[to_space_id].set_space(young_gen->to_space());
932
933 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
934 }
935
936 void PSParallelCompact::initialize_dead_wood_limiter()
937 {
938 const size_t max = 100;
939 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
940 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
941 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
942 DEBUG_ONLY(_dwl_initialized = true;)
943 _dwl_adjustment = normal_distribution(1.0);
944 }
945
946 void
947 PSParallelCompact::clear_data_covering_space(SpaceId id)
948 {
949 // At this point, top is the value before GC, new_top() is the value that will
950 // be set at the end of GC. The marking bitmap is cleared to top; nothing
951 // should be marked above top. The summary data is cleared to the larger of
952 // top & new_top.
953 MutableSpace* const space = _space_info[id].space();
954 HeapWord* const bot = space->bottom();
955 HeapWord* const top = space->top();
956 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
957
958 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
959 const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
960 _mark_bitmap.clear_range(beg_bit, end_bit);
961
962 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
963 const size_t end_region =
964 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
965 _summary_data.clear_range(beg_region, end_region);
966
967 // Clear the data used to 'split' regions.
968 SplitInfo& split_info = _space_info[id].split_info();
969 if (split_info.is_valid()) {
970 split_info.clear();
971 }
972 DEBUG_ONLY(split_info.verify_clear();)
973 }
974
975 void PSParallelCompact::pre_compact()
976 {
977 // Update the from & to space pointers in space_info, since they are swapped
978 // at each young gen gc. Do the update unconditionally (even though a
979 // promotion failure does not swap spaces) because an unknown number of young
980 // collections will have swapped the spaces an unknown number of times.
981 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
982 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
983 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
984 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
985
986 // Increment the invocation count
987 heap->increment_total_collections(true);
988
989 // We need to track unique mark sweep invocations as well.
990 _total_invocations++;
991
992 heap->print_heap_before_gc();
993 heap->trace_heap_before_gc(&_gc_tracer);
994
995 // Fill in TLABs
996 heap->ensure_parsability(true); // retire TLABs
997
998 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
999 Universe::verify("Before GC");
1000 }
1001
1002 // Verify object start arrays
1003 if (VerifyObjectStartArray &&
1004 VerifyBeforeGC) {
1005 heap->old_gen()->verify_object_start_array();
1006 }
1007
1008 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1009 DEBUG_ONLY(summary_data().verify_clear();)
1010
1011 ParCompactionManager::reset_all_bitmap_query_caches();
1012 }
1013
1014 void PSParallelCompact::post_compact()
1015 {
1016 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1017 ParCompactionManager::remove_all_shadow_regions();
1018
1019 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1020 // Clear the marking bitmap, summary data and split info.
1021 clear_data_covering_space(SpaceId(id));
1022 // Update top(). Must be done after clearing the bitmap and summary data.
1023 _space_info[id].publish_new_top();
1024 }
1025
1026 ParCompactionManager::flush_all_string_dedup_requests();
1027
1028 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1029 MutableSpace* const from_space = _space_info[from_space_id].space();
1030 MutableSpace* const to_space = _space_info[to_space_id].space();
1031
1032 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1033 bool eden_empty = eden_space->is_empty();
1034
1035 // Update heap occupancy information which is used as input to the soft ref
1036 // clearing policy at the next gc.
1037 Universe::heap()->update_capacity_and_used_at_gc();
1038
1039 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1040 to_space->is_empty();
1041
1042 PSCardTable* ct = heap->card_table();
1043 MemRegion old_mr = heap->old_gen()->reserved();
1044 if (young_gen_empty) {
1045 ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1046 } else {
1047 ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1048 }
1049
1050 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1051 ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1052 DEBUG_ONLY(MetaspaceUtils::verify();)
1053
1054 heap->prune_scavengable_nmethods();
1055
1056 #if COMPILER2_OR_JVMCI
1057 DerivedPointerTable::update_pointers();
1058 #endif
1059
1060 if (ZapUnusedHeapArea) {
1061 heap->gen_mangle_unused_area();
1062 }
1063
1064 // Signal that we have completed a visit to all live objects.
1065 Universe::heap()->record_whole_heap_examined_timestamp();
1066 }
1067
1068 HeapWord*
1069 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1070 bool maximum_compaction)
1071 {
1072 const size_t region_size = ParallelCompactData::RegionSize;
1073 const ParallelCompactData& sd = summary_data();
1074
1075 const MutableSpace* const space = _space_info[id].space();
1076 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1077 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1078 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1079
1080 // Skip full regions at the beginning of the space--they are necessarily part
1081 // of the dense prefix.
1082 size_t full_count = 0;
1083 const RegionData* cp;
1084 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1085 ++full_count;
1086 }
1087
1088 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1089 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1090 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1091 if (maximum_compaction || cp == end_cp || interval_ended) {
1092 _maximum_compaction_gc_num = total_invocations();
1093 return sd.region_to_addr(cp);
1094 }
1095
1096 HeapWord* const new_top = _space_info[id].new_top();
1097 const size_t space_live = pointer_delta(new_top, space->bottom());
1098 const size_t space_used = space->used_in_words();
1099 const size_t space_capacity = space->capacity_in_words();
1100
1101 const double cur_density = double(space_live) / space_capacity;
1102 const double deadwood_density =
1103 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1104 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1105
1106 log_develop_debug(gc, compaction)(
1107 "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1108 cur_density, deadwood_density, deadwood_goal);
1109 log_develop_debug(gc, compaction)(
1110 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1111 "space_cap=" SIZE_FORMAT,
1112 space_live, space_used,
1113 space_capacity);
1114
1115 // XXX - Use binary search?
1116 HeapWord* dense_prefix = sd.region_to_addr(cp);
1117 const RegionData* full_cp = cp;
1118 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1119 while (cp < end_cp) {
1120 HeapWord* region_destination = cp->destination();
1121 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1122
1123 log_develop_trace(gc, compaction)(
1124 "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1125 "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1126 sd.region(cp), p2i(region_destination),
1127 p2i(dense_prefix), cur_deadwood);
1128
1129 if (cur_deadwood >= deadwood_goal) {
1130 // Found the region that has the correct amount of deadwood to the left.
1131 // This typically occurs after crossing a fairly sparse set of regions, so
1132 // iterate backwards over those sparse regions, looking for the region
1133 // that has the lowest density of live objects 'to the right.'
1134 size_t space_to_left = sd.region(cp) * region_size;
1135 size_t live_to_left = space_to_left - cur_deadwood;
1136 size_t space_to_right = space_capacity - space_to_left;
1137 size_t live_to_right = space_live - live_to_left;
1138 double density_to_right = double(live_to_right) / space_to_right;
1139 while (cp > full_cp) {
1140 --cp;
1141 const size_t prev_region_live_to_right = live_to_right -
1142 cp->data_size();
1143 const size_t prev_region_space_to_right = space_to_right + region_size;
1144 double prev_region_density_to_right =
1145 double(prev_region_live_to_right) / prev_region_space_to_right;
1146 if (density_to_right <= prev_region_density_to_right) {
1147 return dense_prefix;
1148 }
1149
1150 log_develop_trace(gc, compaction)(
1151 "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1152 "pc_d2r=%10.8f",
1153 sd.region(cp), density_to_right,
1154 prev_region_density_to_right);
1155
1156 dense_prefix -= region_size;
1157 live_to_right = prev_region_live_to_right;
1158 space_to_right = prev_region_space_to_right;
1159 density_to_right = prev_region_density_to_right;
1160 }
1161 return dense_prefix;
1162 }
1163
1164 dense_prefix += region_size;
1165 ++cp;
1166 }
1167
1168 return dense_prefix;
1169 }
1170
1171 #ifndef PRODUCT
1172 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1173 const SpaceId id,
1174 const bool maximum_compaction,
1175 HeapWord* const addr)
1176 {
1177 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1178 RegionData* const cp = summary_data().region(region_idx);
1179 const MutableSpace* const space = _space_info[id].space();
1180 HeapWord* const new_top = _space_info[id].new_top();
1181
1182 const size_t space_live = pointer_delta(new_top, space->bottom());
1183 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1184 const size_t space_cap = space->capacity_in_words();
1185 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1186 const size_t live_to_right = new_top - cp->destination();
1187 const size_t dead_to_right = space->top() - addr - live_to_right;
1188
1189 log_develop_debug(gc, compaction)(
1190 "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1191 "spl=" SIZE_FORMAT " "
1192 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1193 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1194 "ratio=%10.8f",
1195 algorithm, p2i(addr), region_idx,
1196 space_live,
1197 dead_to_left, dead_to_left_pct,
1198 dead_to_right, live_to_right,
1199 double(dead_to_right) / live_to_right);
1200 }
1201 #endif // #ifndef PRODUCT
1202
1203 // Return a fraction indicating how much of the generation can be treated as
1204 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1205 // based on the density of live objects in the generation to determine a limit,
1206 // which is then adjusted so the return value is min_percent when the density is
1207 // 1.
1208 //
1209 // The following table shows some return values for a different values of the
1210 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1211 // min_percent is 1.
1212 //
1213 // fraction allowed as dead wood
1214 // -----------------------------------------------------------------
1215 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1216 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1217 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1218 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1219 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1220 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1221 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1222 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1223 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1224 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1225 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1226 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1227 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1228 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1229 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1230 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1231 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1232 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1233 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1234 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1235 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1236 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1237 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1238
1239 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1240 {
1241 assert(_dwl_initialized, "uninitialized");
1242
1243 // The raw limit is the value of the normal distribution at x = density.
1244 const double raw_limit = normal_distribution(density);
1245
1246 // Adjust the raw limit so it becomes the minimum when the density is 1.
1247 //
1248 // First subtract the adjustment value (which is simply the precomputed value
1249 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1250 // Then add the minimum value, so the minimum is returned when the density is
1251 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1252 const double min = double(min_percent) / 100.0;
1253 const double limit = raw_limit - _dwl_adjustment + min;
1254 return MAX2(limit, 0.0);
1255 }
1256
1257 ParallelCompactData::RegionData*
1258 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1259 const RegionData* end)
1260 {
1261 const size_t region_size = ParallelCompactData::RegionSize;
1262 ParallelCompactData& sd = summary_data();
1263 size_t left = sd.region(beg);
1264 size_t right = end > beg ? sd.region(end) - 1 : left;
1265
1266 // Binary search.
1267 while (left < right) {
1268 // Equivalent to (left + right) / 2, but does not overflow.
1269 const size_t middle = left + (right - left) / 2;
1270 RegionData* const middle_ptr = sd.region(middle);
1271 HeapWord* const dest = middle_ptr->destination();
1272 HeapWord* const addr = sd.region_to_addr(middle);
1273 assert(dest != NULL, "sanity");
1274 assert(dest <= addr, "must move left");
1275
1276 if (middle > left && dest < addr) {
1277 right = middle - 1;
1278 } else if (middle < right && middle_ptr->data_size() == region_size) {
1279 left = middle + 1;
1280 } else {
1281 return middle_ptr;
1282 }
1283 }
1284 return sd.region(left);
1285 }
1286
1287 ParallelCompactData::RegionData*
1288 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1289 const RegionData* end,
1290 size_t dead_words)
1291 {
1292 ParallelCompactData& sd = summary_data();
1293 size_t left = sd.region(beg);
1294 size_t right = end > beg ? sd.region(end) - 1 : left;
1295
1296 // Binary search.
1297 while (left < right) {
1298 // Equivalent to (left + right) / 2, but does not overflow.
1299 const size_t middle = left + (right - left) / 2;
1300 RegionData* const middle_ptr = sd.region(middle);
1301 HeapWord* const dest = middle_ptr->destination();
1302 HeapWord* const addr = sd.region_to_addr(middle);
1303 assert(dest != NULL, "sanity");
1304 assert(dest <= addr, "must move left");
1305
1306 const size_t dead_to_left = pointer_delta(addr, dest);
1307 if (middle > left && dead_to_left > dead_words) {
1308 right = middle - 1;
1309 } else if (middle < right && dead_to_left < dead_words) {
1310 left = middle + 1;
1311 } else {
1312 return middle_ptr;
1313 }
1314 }
1315 return sd.region(left);
1316 }
1317
1318 // The result is valid during the summary phase, after the initial summarization
1319 // of each space into itself, and before final summarization.
1320 inline double
1321 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1322 HeapWord* const bottom,
1323 HeapWord* const top,
1324 HeapWord* const new_top)
1325 {
1326 ParallelCompactData& sd = summary_data();
1327
1328 assert(cp != NULL, "sanity");
1329 assert(bottom != NULL, "sanity");
1330 assert(top != NULL, "sanity");
1331 assert(new_top != NULL, "sanity");
1332 assert(top >= new_top, "summary data problem?");
1333 assert(new_top > bottom, "space is empty; should not be here");
1334 assert(new_top >= cp->destination(), "sanity");
1335 assert(top >= sd.region_to_addr(cp), "sanity");
1336
1337 HeapWord* const destination = cp->destination();
1338 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1339 const size_t compacted_region_live = pointer_delta(new_top, destination);
1340 const size_t compacted_region_used = pointer_delta(top,
1341 sd.region_to_addr(cp));
1342 const size_t reclaimable = compacted_region_used - compacted_region_live;
1343
1344 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1345 return double(reclaimable) / divisor;
1346 }
1347
1348 // Return the address of the end of the dense prefix, a.k.a. the start of the
1349 // compacted region. The address is always on a region boundary.
1350 //
1351 // Completely full regions at the left are skipped, since no compaction can
1352 // occur in those regions. Then the maximum amount of dead wood to allow is
1353 // computed, based on the density (amount live / capacity) of the generation;
1354 // the region with approximately that amount of dead space to the left is
1355 // identified as the limit region. Regions between the last completely full
1356 // region and the limit region are scanned and the one that has the best
1357 // (maximum) reclaimed_ratio() is selected.
1358 HeapWord*
1359 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1360 bool maximum_compaction)
1361 {
1362 const size_t region_size = ParallelCompactData::RegionSize;
1363 const ParallelCompactData& sd = summary_data();
1364
1365 const MutableSpace* const space = _space_info[id].space();
1366 HeapWord* const top = space->top();
1367 HeapWord* const top_aligned_up = sd.region_align_up(top);
1368 HeapWord* const new_top = _space_info[id].new_top();
1369 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1370 HeapWord* const bottom = space->bottom();
1371 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1372 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1373 const RegionData* const new_top_cp =
1374 sd.addr_to_region_ptr(new_top_aligned_up);
1375
1376 // Skip full regions at the beginning of the space--they are necessarily part
1377 // of the dense prefix.
1378 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1379 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1380 space->is_empty(), "no dead space allowed to the left");
1381 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1382 "region must have dead space");
1383
1384 // The gc number is saved whenever a maximum compaction is done, and used to
1385 // determine when the maximum compaction interval has expired. This avoids
1386 // successive max compactions for different reasons.
1387 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1388 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1389 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1390 total_invocations() == HeapFirstMaximumCompactionCount;
1391 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1392 _maximum_compaction_gc_num = total_invocations();
1393 return sd.region_to_addr(full_cp);
1394 }
1395
1396 const size_t space_live = pointer_delta(new_top, bottom);
1397 const size_t space_used = space->used_in_words();
1398 const size_t space_capacity = space->capacity_in_words();
1399
1400 const double density = double(space_live) / double(space_capacity);
1401 const size_t min_percent_free = MarkSweepDeadRatio;
1402 const double limiter = dead_wood_limiter(density, min_percent_free);
1403 const size_t dead_wood_max = space_used - space_live;
1404 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1405 dead_wood_max);
1406
1407 log_develop_debug(gc, compaction)(
1408 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1409 "space_cap=" SIZE_FORMAT,
1410 space_live, space_used,
1411 space_capacity);
1412 log_develop_debug(gc, compaction)(
1413 "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1414 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1415 density, min_percent_free, limiter,
1416 dead_wood_max, dead_wood_limit);
1417
1418 // Locate the region with the desired amount of dead space to the left.
1419 const RegionData* const limit_cp =
1420 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1421
1422 // Scan from the first region with dead space to the limit region and find the
1423 // one with the best (largest) reclaimed ratio.
1424 double best_ratio = 0.0;
1425 const RegionData* best_cp = full_cp;
1426 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1427 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1428 if (tmp_ratio > best_ratio) {
1429 best_cp = cp;
1430 best_ratio = tmp_ratio;
1431 }
1432 }
1433
1434 return sd.region_to_addr(best_cp);
1435 }
1436
1437 void PSParallelCompact::summarize_spaces_quick()
1438 {
1439 for (unsigned int i = 0; i < last_space_id; ++i) {
1440 const MutableSpace* space = _space_info[i].space();
1441 HeapWord** nta = _space_info[i].new_top_addr();
1442 bool result = _summary_data.summarize(_space_info[i].split_info(),
1443 space->bottom(), space->top(), NULL,
1444 space->bottom(), space->end(), nta);
1445 assert(result, "space must fit into itself");
1446 _space_info[i].set_dense_prefix(space->bottom());
1447 }
1448 }
1449
1450 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1451 {
1452 HeapWord* const dense_prefix_end = dense_prefix(id);
1453 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1454 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1455 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1456 // Only enough dead space is filled so that any remaining dead space to the
1457 // left is larger than the minimum filler object. (The remainder is filled
1458 // during the copy/update phase.)
1459 //
1460 // The size of the dead space to the right of the boundary is not a
1461 // concern, since compaction will be able to use whatever space is
1462 // available.
1463 //
1464 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1465 // surrounds the space to be filled with an object.
1466 //
1467 // In the 32-bit VM, each bit represents two 32-bit words:
1468 // +---+
1469 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1470 // end_bits: ... x x x | 0 | || 0 x x ...
1471 // +---+
1472 //
1473 // In the 64-bit VM, each bit represents one 64-bit word:
1474 // +------------+
1475 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1476 // end_bits: ... x x 1 | 0 || 0 | x x ...
1477 // +------------+
1478 // +-------+
1479 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1480 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1481 // +-------+
1482 // +-----------+
1483 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1484 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1485 // +-----------+
1486 // +-------+
1487 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1488 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1489 // +-------+
1490
1491 // Initially assume case a, c or e will apply.
1492 size_t obj_len = CollectedHeap::min_fill_size();
1493 HeapWord* obj_beg = dense_prefix_end - obj_len;
1494
1495 #ifdef _LP64
1496 if (MinObjAlignment > 1) { // object alignment > heap word size
1497 // Cases a, c or e.
1498 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1499 // Case b above.
1500 obj_beg = dense_prefix_end - 1;
1501 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1502 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1503 // Case d above.
1504 obj_beg = dense_prefix_end - 3;
1505 obj_len = 3;
1506 }
1507 #endif // #ifdef _LP64
1508
1509 CollectedHeap::fill_with_object(obj_beg, obj_len);
1510 _mark_bitmap.mark_obj(obj_beg, obj_len);
1511 _summary_data.add_obj(obj_beg, obj_len);
1512 assert(start_array(id) != NULL, "sanity");
1513 start_array(id)->allocate_block(obj_beg);
1514 }
1515 }
1516
1517 void
1518 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1519 {
1520 assert(id < last_space_id, "id out of range");
1521 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1522 "should have been reset in summarize_spaces_quick()");
1523
1524 const MutableSpace* space = _space_info[id].space();
1525 if (_space_info[id].new_top() != space->bottom()) {
1526 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1527 _space_info[id].set_dense_prefix(dense_prefix_end);
1528
1529 #ifndef PRODUCT
1530 if (log_is_enabled(Debug, gc, compaction)) {
1531 print_dense_prefix_stats("ratio", id, maximum_compaction,
1532 dense_prefix_end);
1533 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1534 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1535 }
1536 #endif // #ifndef PRODUCT
1537
1538 // Recompute the summary data, taking into account the dense prefix. If
1539 // every last byte will be reclaimed, then the existing summary data which
1540 // compacts everything can be left in place.
1541 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1542 // If dead space crosses the dense prefix boundary, it is (at least
1543 // partially) filled with a dummy object, marked live and added to the
1544 // summary data. This simplifies the copy/update phase and must be done
1545 // before the final locations of objects are determined, to prevent
1546 // leaving a fragment of dead space that is too small to fill.
1547 fill_dense_prefix_end(id);
1548
1549 // Compute the destination of each Region, and thus each object.
1550 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1551 _summary_data.summarize(_space_info[id].split_info(),
1552 dense_prefix_end, space->top(), NULL,
1553 dense_prefix_end, space->end(),
1554 _space_info[id].new_top_addr());
1555 }
1556 }
1557
1558 if (log_develop_is_enabled(Trace, gc, compaction)) {
1559 const size_t region_size = ParallelCompactData::RegionSize;
1560 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1561 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1562 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1563 HeapWord* const new_top = _space_info[id].new_top();
1564 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1565 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1566 log_develop_trace(gc, compaction)(
1567 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1568 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1569 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1570 id, space->capacity_in_words(), p2i(dense_prefix_end),
1571 dp_region, dp_words / region_size,
1572 cr_words / region_size, p2i(new_top));
1573 }
1574 }
1575
1576 #ifndef PRODUCT
1577 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1578 HeapWord* dst_beg, HeapWord* dst_end,
1579 SpaceId src_space_id,
1580 HeapWord* src_beg, HeapWord* src_end)
1581 {
1582 log_develop_trace(gc, compaction)(
1583 "Summarizing %d [%s] into %d [%s]: "
1584 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1585 SIZE_FORMAT "-" SIZE_FORMAT " "
1586 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1587 SIZE_FORMAT "-" SIZE_FORMAT,
1588 src_space_id, space_names[src_space_id],
1589 dst_space_id, space_names[dst_space_id],
1590 p2i(src_beg), p2i(src_end),
1591 _summary_data.addr_to_region_idx(src_beg),
1592 _summary_data.addr_to_region_idx(src_end),
1593 p2i(dst_beg), p2i(dst_end),
1594 _summary_data.addr_to_region_idx(dst_beg),
1595 _summary_data.addr_to_region_idx(dst_end));
1596 }
1597 #endif // #ifndef PRODUCT
1598
1599 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1600 bool maximum_compaction)
1601 {
1602 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1603
1604 // Quick summarization of each space into itself, to see how much is live.
1605 summarize_spaces_quick();
1606
1607 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1608 NOT_PRODUCT(print_region_ranges());
1609 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1610
1611 // The amount of live data that will end up in old space (assuming it fits).
1612 size_t old_space_total_live = 0;
1613 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1614 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1615 _space_info[id].space()->bottom());
1616 }
1617
1618 MutableSpace* const old_space = _space_info[old_space_id].space();
1619 const size_t old_capacity = old_space->capacity_in_words();
1620 if (old_space_total_live > old_capacity) {
1621 // XXX - should also try to expand
1622 maximum_compaction = true;
1623 }
1624
1625 // Old generations.
1626 summarize_space(old_space_id, maximum_compaction);
1627
1628 // Summarize the remaining spaces in the young gen. The initial target space
1629 // is the old gen. If a space does not fit entirely into the target, then the
1630 // remainder is compacted into the space itself and that space becomes the new
1631 // target.
1632 SpaceId dst_space_id = old_space_id;
1633 HeapWord* dst_space_end = old_space->end();
1634 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1635 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1636 const MutableSpace* space = _space_info[id].space();
1637 const size_t live = pointer_delta(_space_info[id].new_top(),
1638 space->bottom());
1639 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1640
1641 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1642 SpaceId(id), space->bottom(), space->top());)
1643 if (live > 0 && live <= available) {
1644 // All the live data will fit.
1645 bool done = _summary_data.summarize(_space_info[id].split_info(),
1646 space->bottom(), space->top(),
1647 NULL,
1648 *new_top_addr, dst_space_end,
1649 new_top_addr);
1650 assert(done, "space must fit into old gen");
1651
1652 // Reset the new_top value for the space.
1653 _space_info[id].set_new_top(space->bottom());
1654 } else if (live > 0) {
1655 // Attempt to fit part of the source space into the target space.
1656 HeapWord* next_src_addr = NULL;
1657 bool done = _summary_data.summarize(_space_info[id].split_info(),
1658 space->bottom(), space->top(),
1659 &next_src_addr,
1660 *new_top_addr, dst_space_end,
1661 new_top_addr);
1662 assert(!done, "space should not fit into old gen");
1663 assert(next_src_addr != NULL, "sanity");
1664
1665 // The source space becomes the new target, so the remainder is compacted
1666 // within the space itself.
1667 dst_space_id = SpaceId(id);
1668 dst_space_end = space->end();
1669 new_top_addr = _space_info[id].new_top_addr();
1670 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1671 space->bottom(), dst_space_end,
1672 SpaceId(id), next_src_addr, space->top());)
1673 done = _summary_data.summarize(_space_info[id].split_info(),
1674 next_src_addr, space->top(),
1675 NULL,
1676 space->bottom(), dst_space_end,
1677 new_top_addr);
1678 assert(done, "space must fit when compacted into itself");
1679 assert(*new_top_addr <= space->top(), "usage should not grow");
1680 }
1681 }
1682
1683 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1684 NOT_PRODUCT(print_region_ranges());
1685 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1686 }
1687
1688 // This method should contain all heap-specific policy for invoking a full
1689 // collection. invoke_no_policy() will only attempt to compact the heap; it
1690 // will do nothing further. If we need to bail out for policy reasons, scavenge
1691 // before full gc, or any other specialized behavior, it needs to be added here.
1692 //
1693 // Note that this method should only be called from the vm_thread while at a
1694 // safepoint.
1695 //
1696 // Note that the all_soft_refs_clear flag in the soft ref policy
1697 // may be true because this method can be called without intervening
1698 // activity. For example when the heap space is tight and full measure
1699 // are being taken to free space.
1700 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1701 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1702 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1703 "should be in vm thread");
1704
1705 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1706 GCCause::Cause gc_cause = heap->gc_cause();
1707 assert(!heap->is_gc_active(), "not reentrant");
1708
1709 PSAdaptiveSizePolicy* policy = heap->size_policy();
1710 IsGCActiveMark mark;
1711
1712 if (ScavengeBeforeFullGC) {
1713 PSScavenge::invoke_no_policy();
1714 }
1715
1716 const bool clear_all_soft_refs =
1717 heap->soft_ref_policy()->should_clear_all_soft_refs();
1718
1719 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1720 maximum_heap_compaction);
1721 }
1722
1723 // This method contains no policy. You should probably
1724 // be calling invoke() instead.
1725 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1726 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1727 assert(ref_processor() != NULL, "Sanity");
1728
1729 if (GCLocker::check_active_before_gc()) {
1730 return false;
1731 }
1732
1733 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1734
1735 GCIdMark gc_id_mark;
1736 _gc_timer.register_gc_start();
1737 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1738
1739 TimeStamp marking_start;
1740 TimeStamp compaction_start;
1741 TimeStamp collection_exit;
1742
1743 GCCause::Cause gc_cause = heap->gc_cause();
1744 PSYoungGen* young_gen = heap->young_gen();
1745 PSOldGen* old_gen = heap->old_gen();
1746 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1747
1748 // The scope of casr should end after code that can change
1749 // SoftRefPolicy::_should_clear_all_soft_refs.
1750 ClearedAllSoftRefs casr(maximum_heap_compaction,
1751 heap->soft_ref_policy());
1752
1753 if (ZapUnusedHeapArea) {
1754 // Save information needed to minimize mangling
1755 heap->record_gen_tops_before_GC();
1756 }
1757
1758 // Make sure data structures are sane, make the heap parsable, and do other
1759 // miscellaneous bookkeeping.
1760 pre_compact();
1761
1762 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1763
1764 // Get the compaction manager reserved for the VM thread.
1765 ParCompactionManager* const vmthread_cm = ParCompactionManager::get_vmthread_cm();
1766
1767 {
1768 const uint active_workers =
1769 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(),
1770 ParallelScavengeHeap::heap()->workers().active_workers(),
1771 Threads::number_of_non_daemon_threads());
1772 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers);
1773
1774 GCTraceCPUTime tcpu;
1775 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1776
1777 heap->pre_full_gc_dump(&_gc_timer);
1778
1779 TraceCollectorStats tcs(counters());
1780 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1781
1782 if (log_is_enabled(Debug, gc, heap, exit)) {
1783 accumulated_time()->start();
1784 }
1785
1786 // Let the size policy know we're starting
1787 size_policy->major_collection_begin();
1788
1789 #if COMPILER2_OR_JVMCI
1790 DerivedPointerTable::clear();
1791 #endif
1792
1793 ref_processor()->start_discovery(maximum_heap_compaction);
1794
1795 marking_start.update();
1796 marking_phase(vmthread_cm, &_gc_tracer);
1797
1798 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1799 && GCCause::is_user_requested_gc(gc_cause);
1800 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1801
1802 #if COMPILER2_OR_JVMCI
1803 assert(DerivedPointerTable::is_active(), "Sanity");
1804 DerivedPointerTable::set_active(false);
1805 #endif
1806
1807 // adjust_roots() updates Universe::_intArrayKlassObj which is
1808 // needed by the compaction for filling holes in the dense prefix.
1809 adjust_roots();
1810
1811 compaction_start.update();
1812 compact();
1813
1814 ParCompactionManager::verify_all_region_stack_empty();
1815
1816 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1817 // done before resizing.
1818 post_compact();
1819
1820 // Let the size policy know we're done
1821 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1822
1823 if (UseAdaptiveSizePolicy) {
1824 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1825 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1826 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1827
1828 // Don't check if the size_policy is ready here. Let
1829 // the size_policy check that internally.
1830 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1831 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1832 // Swap the survivor spaces if from_space is empty. The
1833 // resize_young_gen() called below is normally used after
1834 // a successful young GC and swapping of survivor spaces;
1835 // otherwise, it will fail to resize the young gen with
1836 // the current implementation.
1837 if (young_gen->from_space()->is_empty()) {
1838 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1839 young_gen->swap_spaces();
1840 }
1841
1842 // Calculate optimal free space amounts
1843 assert(young_gen->max_gen_size() >
1844 young_gen->from_space()->capacity_in_bytes() +
1845 young_gen->to_space()->capacity_in_bytes(),
1846 "Sizes of space in young gen are out-of-bounds");
1847
1848 size_t young_live = young_gen->used_in_bytes();
1849 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1850 size_t old_live = old_gen->used_in_bytes();
1851 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1852 size_t max_old_gen_size = old_gen->max_gen_size();
1853 size_t max_eden_size = young_gen->max_gen_size() -
1854 young_gen->from_space()->capacity_in_bytes() -
1855 young_gen->to_space()->capacity_in_bytes();
1856
1857 // Used for diagnostics
1858 size_policy->clear_generation_free_space_flags();
1859
1860 size_policy->compute_generations_free_space(young_live,
1861 eden_live,
1862 old_live,
1863 cur_eden,
1864 max_old_gen_size,
1865 max_eden_size,
1866 true /* full gc*/);
1867
1868 size_policy->check_gc_overhead_limit(eden_live,
1869 max_old_gen_size,
1870 max_eden_size,
1871 true /* full gc*/,
1872 gc_cause,
1873 heap->soft_ref_policy());
1874
1875 size_policy->decay_supplemental_growth(true /* full gc*/);
1876
1877 heap->resize_old_gen(
1878 size_policy->calculated_old_free_size_in_bytes());
1879
1880 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1881 size_policy->calculated_survivor_size_in_bytes());
1882 }
1883
1884 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1885 }
1886
1887 if (UsePerfData) {
1888 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1889 counters->update_counters();
1890 counters->update_old_capacity(old_gen->capacity_in_bytes());
1891 counters->update_young_capacity(young_gen->capacity_in_bytes());
1892 }
1893
1894 heap->resize_all_tlabs();
1895
1896 // Resize the metaspace capacity after a collection
1897 MetaspaceGC::compute_new_size();
1898
1899 if (log_is_enabled(Debug, gc, heap, exit)) {
1900 accumulated_time()->stop();
1901 }
1902
1903 heap->print_heap_change(pre_gc_values);
1904
1905 // Track memory usage and detect low memory
1906 MemoryService::track_memory_usage();
1907 heap->update_counters();
1908
1909 heap->post_full_gc_dump(&_gc_timer);
1910 }
1911
1912 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1913 Universe::verify("After GC");
1914 }
1915
1916 // Re-verify object start arrays
1917 if (VerifyObjectStartArray &&
1918 VerifyAfterGC) {
1919 old_gen->verify_object_start_array();
1920 }
1921
1922 if (ZapUnusedHeapArea) {
1923 old_gen->object_space()->check_mangled_unused_area_complete();
1924 }
1925
1926 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1927
1928 collection_exit.update();
1929
1930 heap->print_heap_after_gc();
1931 heap->trace_heap_after_gc(&_gc_tracer);
1932
1933 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1934 marking_start.ticks(), compaction_start.ticks(),
1935 collection_exit.ticks());
1936
1937 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1938
1939 _gc_timer.register_gc_end();
1940
1941 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1942 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1943
1944 return true;
1945 }
1946
1947 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1948 private:
1949 uint _worker_id;
1950
1951 public:
1952 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1953 void do_thread(Thread* thread) {
1954 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1955
1956 ResourceMark rm;
1957
1958 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1959
1960 PCMarkAndPushClosure mark_and_push_closure(cm);
1961 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
1962
1963 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1964
1965 // Do the real work
1966 cm->follow_marking_stacks();
1967 }
1968 };
1969
1970 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
1971 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1972
1973 ParCompactionManager* cm =
1974 ParCompactionManager::gc_thread_compaction_manager(worker_id);
1975 PCMarkAndPushClosure mark_and_push_closure(cm);
1976
1977 switch (root_type) {
1978 case ParallelRootType::class_loader_data:
1979 {
1980 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
1981 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1982 }
1983 break;
1984
1985 case ParallelRootType::code_cache:
1986 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
1987 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
1988 break;
1989
1990 case ParallelRootType::sentinel:
1991 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
1992 fatal("Bad enumeration value: %u", root_type);
1993 break;
1994 }
1995
1996 // Do the real work
1997 cm->follow_marking_stacks();
1998 }
1999
2000 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
2001 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2002
2003 ParCompactionManager* cm =
2004 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2005
2006 oop obj = NULL;
2007 ObjArrayTask task;
2008 do {
2009 while (ParCompactionManager::steal_objarray(worker_id, task)) {
2010 cm->follow_array((objArrayOop)task.obj(), task.index());
2011 cm->follow_marking_stacks();
2012 }
2013 while (ParCompactionManager::steal(worker_id, obj)) {
2014 cm->follow_contents(obj);
2015 cm->follow_marking_stacks();
2016 }
2017 } while (!terminator.offer_termination());
2018 }
2019
2020 class MarkFromRootsTask : public AbstractGangTask {
2021 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2022 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2023 SequentialSubTasksDone _subtasks;
2024 TaskTerminator _terminator;
2025 uint _active_workers;
2026
2027 public:
2028 MarkFromRootsTask(uint active_workers) :
2029 AbstractGangTask("MarkFromRootsTask"),
2030 _strong_roots_scope(active_workers),
2031 _subtasks(ParallelRootType::sentinel),
2032 _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2033 _active_workers(active_workers) {
2034 }
2035
2036 virtual void work(uint worker_id) {
2037 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2038 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2039 }
2040
2041 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2042 Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2043
2044 // Mark from OopStorages
2045 {
2046 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2047 PCMarkAndPushClosure closure(cm);
2048 _oop_storage_set_par_state.oops_do(&closure);
2049 // Do the real work
2050 cm->follow_marking_stacks();
2051 }
2052
2053 if (_active_workers > 1) {
2054 steal_marking_work(_terminator, worker_id);
2055 }
2056 }
2057 };
2058
2059 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2060 TaskTerminator _terminator;
2061
2062 public:
2063 ParallelCompactRefProcProxyTask(uint max_workers)
2064 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2065 _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2066
2067 void work(uint worker_id) override {
2068 assert(worker_id < _max_workers, "sanity");
2069 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2070 PCMarkAndPushClosure keep_alive(cm);
2071 BarrierEnqueueDiscoveredFieldClosure enqueue;
2072 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2073 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2074 }
2075
2076 void prepare_run_task_hook() override {
2077 _terminator.reset_for_reuse(_queue_count);
2078 }
2079 };
2080
2081 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2082 ParallelOldTracer *gc_tracer) {
2083 // Recursively traverse all live objects and mark them
2084 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2085
2086 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2087
2088 // Need new claim bits before marking starts.
2089 ClassLoaderDataGraph::clear_claimed_marks();
2090
2091 {
2092 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2093
2094 MarkFromRootsTask task(active_gc_threads);
2095 ParallelScavengeHeap::heap()->workers().run_task(&task);
2096 }
2097
2098 // Process reference objects found during marking
2099 {
2100 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2101
2102 ReferenceProcessorStats stats;
2103 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2104
2105 ref_processor()->set_active_mt_degree(active_gc_threads);
2106 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2107 stats = ref_processor()->process_discovered_references(task, pt);
2108
2109 gc_tracer->report_gc_reference_stats(stats);
2110 pt.print_all_references();
2111 }
2112
2113 // This is the point where the entire marking should have completed.
2114 ParCompactionManager::verify_all_marking_stack_empty();
2115
2116 {
2117 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2118 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2119 is_alive_closure(),
2120 &do_nothing_cl,
2121 1);
2122 }
2123
2124 {
2125 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2126
2127 // Follow system dictionary roots and unload classes.
2128 bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2129
2130 // Unload nmethods.
2131 CodeCache::do_unloading(is_alive_closure(), purged_class);
2132
2133 // Prune dead klasses from subklass/sibling/implementor lists.
2134 Klass::clean_weak_klass_links(purged_class);
2135
2136 // Clean JVMCI metadata handles.
2137 JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2138 }
2139
2140 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2141 }
2142
2143 #ifdef ASSERT
2144 void PCAdjustPointerClosure::verify_cm(ParCompactionManager* cm) {
2145 assert(cm != NULL, "associate ParCompactionManage should not be NULL");
2146 auto vmthread_cm = ParCompactionManager::get_vmthread_cm();
2147 if (Thread::current()->is_VM_thread()) {
2148 assert(cm == vmthread_cm, "VM threads should use ParCompactionManager from get_vmthread_cm()");
2149 } else {
2150 assert(Thread::current()->is_Worker_thread(), "Must be a GC thread");
2151 assert(cm != vmthread_cm, "GC threads should use ParCompactionManager from gc_thread_compaction_manager()");
2152 }
2153 }
2154 #endif
2155
2156 class PSAdjustTask final : public AbstractGangTask {
2157 SubTasksDone _sub_tasks;
2158 WeakProcessor::Task _weak_proc_task;
2159 OopStorageSetStrongParState<false, false> _oop_storage_iter;
2160 uint _nworkers;
2161
2162 enum PSAdjustSubTask {
2163 PSAdjustSubTask_code_cache,
2164 PSAdjustSubTask_old_ref_process,
2165 PSAdjustSubTask_young_ref_process,
2166
2167 PSAdjustSubTask_num_elements
2168 };
2169
2170 public:
2171 PSAdjustTask(uint nworkers) :
2172 AbstractGangTask("PSAdjust task"),
2173 _sub_tasks(PSAdjustSubTask_num_elements),
2174 _weak_proc_task(nworkers),
2175 _nworkers(nworkers) {
2176 // Need new claim bits when tracing through and adjusting pointers.
2177 ClassLoaderDataGraph::clear_claimed_marks();
2178 if (nworkers > 1) {
2179 Threads::change_thread_claim_token();
2180 }
2181 }
2182
2183 ~PSAdjustTask() {
2184 Threads::assert_all_threads_claimed();
2185 }
2186
2187 void work(uint worker_id) {
2188 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2189 PCAdjustPointerClosure adjust(cm);
2190 {
2191 ResourceMark rm;
2192 Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2193 }
2194 _oop_storage_iter.oops_do(&adjust);
2195 {
2196 CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
2197 ClassLoaderDataGraph::cld_do(&cld_closure);
2198 }
2199 {
2200 AlwaysTrueClosure always_alive;
2201 _weak_proc_task.work(worker_id, &always_alive, &adjust);
2202 }
2203 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2204 CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2205 CodeCache::blobs_do(&adjust_code);
2206 }
2207 if (_sub_tasks.try_claim_task(PSAdjustSubTask_old_ref_process)) {
2208 PSParallelCompact::ref_processor()->weak_oops_do(&adjust);
2209 }
2210 if (_sub_tasks.try_claim_task(PSAdjustSubTask_young_ref_process)) {
2211 // Roots were visited so references into the young gen in roots
2212 // may have been scanned. Process them also.
2213 // Should the reference processor have a span that excludes
2214 // young gen objects?
2215 PSScavenge::reference_processor()->weak_oops_do(&adjust);
2216 }
2217 _sub_tasks.all_tasks_claimed();
2218 }
2219 };
2220
2221 void PSParallelCompact::adjust_roots() {
2222 // Adjust the pointers to reflect the new locations
2223 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2224 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2225 PSAdjustTask task(nworkers);
2226 ParallelScavengeHeap::heap()->workers().run_task(&task);
2227 }
2228
2229 // Helper class to print 8 region numbers per line and then print the total at the end.
2230 class FillableRegionLogger : public StackObj {
2231 private:
2232 Log(gc, compaction) log;
2233 static const int LineLength = 8;
2234 size_t _regions[LineLength];
2235 int _next_index;
2236 bool _enabled;
2237 size_t _total_regions;
2238 public:
2239 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2240 ~FillableRegionLogger() {
2241 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2242 }
2243
2244 void print_line() {
2245 if (!_enabled || _next_index == 0) {
2246 return;
2247 }
2248 FormatBuffer<> line("Fillable: ");
2249 for (int i = 0; i < _next_index; i++) {
2250 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2251 }
2252 log.trace("%s", line.buffer());
2253 _next_index = 0;
2254 }
2255
2256 void handle(size_t region) {
2257 if (!_enabled) {
2258 return;
2259 }
2260 _regions[_next_index++] = region;
2261 if (_next_index == LineLength) {
2262 print_line();
2263 }
2264 _total_regions++;
2265 }
2266 };
2267
2268 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2269 {
2270 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2271
2272 // Find the threads that are active
2273 uint worker_id = 0;
2274
2275 // Find all regions that are available (can be filled immediately) and
2276 // distribute them to the thread stacks. The iteration is done in reverse
2277 // order (high to low) so the regions will be removed in ascending order.
2278
2279 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2280
2281 // id + 1 is used to test termination so unsigned can
2282 // be used with an old_space_id == 0.
2283 FillableRegionLogger region_logger;
2284 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2285 SpaceInfo* const space_info = _space_info + id;
2286 HeapWord* const new_top = space_info->new_top();
2287
2288 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2289 const size_t end_region =
2290 sd.addr_to_region_idx(sd.region_align_up(new_top));
2291
2292 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2293 if (sd.region(cur)->claim_unsafe()) {
2294 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2295 bool result = sd.region(cur)->mark_normal();
2296 assert(result, "Must succeed at this point.");
2297 cm->region_stack()->push(cur);
2298 region_logger.handle(cur);
2299 // Assign regions to tasks in round-robin fashion.
2300 if (++worker_id == parallel_gc_threads) {
2301 worker_id = 0;
2302 }
2303 }
2304 }
2305 region_logger.print_line();
2306 }
2307 }
2308
2309 class TaskQueue : StackObj {
2310 volatile uint _counter;
2311 uint _size;
2312 uint _insert_index;
2313 PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2314 public:
2315 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2316 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2317 }
2318 ~TaskQueue() {
2319 assert(_counter >= _insert_index, "not all queue elements were claimed");
2320 FREE_C_HEAP_ARRAY(T, _backing_array);
2321 }
2322
2323 void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2324 assert(_insert_index < _size, "too small backing array");
2325 _backing_array[_insert_index++] = value;
2326 }
2327
2328 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2329 uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2330 if (claimed < _insert_index) {
2331 reference = _backing_array[claimed];
2332 return true;
2333 } else {
2334 return false;
2335 }
2336 }
2337 };
2338
2339 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2340
2341 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2342 uint parallel_gc_threads) {
2343 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2344
2345 ParallelCompactData& sd = PSParallelCompact::summary_data();
2346
2347 // Iterate over all the spaces adding tasks for updating
2348 // regions in the dense prefix. Assume that 1 gc thread
2349 // will work on opening the gaps and the remaining gc threads
2350 // will work on the dense prefix.
2351 unsigned int space_id;
2352 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2353 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2354 const MutableSpace* const space = _space_info[space_id].space();
2355
2356 if (dense_prefix_end == space->bottom()) {
2357 // There is no dense prefix for this space.
2358 continue;
2359 }
2360
2361 // The dense prefix is before this region.
2362 size_t region_index_end_dense_prefix =
2363 sd.addr_to_region_idx(dense_prefix_end);
2364 RegionData* const dense_prefix_cp =
2365 sd.region(region_index_end_dense_prefix);
2366 assert(dense_prefix_end == space->end() ||
2367 dense_prefix_cp->available() ||
2368 dense_prefix_cp->claimed(),
2369 "The region after the dense prefix should always be ready to fill");
2370
2371 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2372
2373 // Is there dense prefix work?
2374 size_t total_dense_prefix_regions =
2375 region_index_end_dense_prefix - region_index_start;
2376 // How many regions of the dense prefix should be given to
2377 // each thread?
2378 if (total_dense_prefix_regions > 0) {
2379 uint tasks_for_dense_prefix = 1;
2380 if (total_dense_prefix_regions <=
2381 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2382 // Don't over partition. This assumes that
2383 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2384 // so there are not many regions to process.
2385 tasks_for_dense_prefix = parallel_gc_threads;
2386 } else {
2387 // Over partition
2388 tasks_for_dense_prefix = parallel_gc_threads *
2389 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2390 }
2391 size_t regions_per_thread = total_dense_prefix_regions /
2392 tasks_for_dense_prefix;
2393 // Give each thread at least 1 region.
2394 if (regions_per_thread == 0) {
2395 regions_per_thread = 1;
2396 }
2397
2398 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2399 if (region_index_start >= region_index_end_dense_prefix) {
2400 break;
2401 }
2402 // region_index_end is not processed
2403 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2404 region_index_end_dense_prefix);
2405 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2406 region_index_start,
2407 region_index_end));
2408 region_index_start = region_index_end;
2409 }
2410 }
2411 // This gets any part of the dense prefix that did not
2412 // fit evenly.
2413 if (region_index_start < region_index_end_dense_prefix) {
2414 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2415 region_index_start,
2416 region_index_end_dense_prefix));
2417 }
2418 }
2419 }
2420
2421 #ifdef ASSERT
2422 // Write a histogram of the number of times the block table was filled for a
2423 // region.
2424 void PSParallelCompact::write_block_fill_histogram()
2425 {
2426 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2427 return;
2428 }
2429
2430 Log(gc, compaction) log;
2431 ResourceMark rm;
2432 LogStream ls(log.trace());
2433 outputStream* out = &ls;
2434
2435 typedef ParallelCompactData::RegionData rd_t;
2436 ParallelCompactData& sd = summary_data();
2437
2438 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2439 MutableSpace* const spc = _space_info[id].space();
2440 if (spc->bottom() != spc->top()) {
2441 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2442 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2443 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2444
2445 size_t histo[5] = { 0, 0, 0, 0, 0 };
2446 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2447 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2448
2449 for (const rd_t* cur = beg; cur < end; ++cur) {
2450 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2451 }
2452 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2453 for (size_t i = 0; i < histo_len; ++i) {
2454 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2455 histo[i], 100.0 * histo[i] / region_cnt);
2456 }
2457 out->cr();
2458 }
2459 }
2460 }
2461 #endif // #ifdef ASSERT
2462
2463 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2464 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2465
2466 ParCompactionManager* cm =
2467 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2468
2469 // Drain the stacks that have been preloaded with regions
2470 // that are ready to fill.
2471
2472 cm->drain_region_stacks();
2473
2474 guarantee(cm->region_stack()->is_empty(), "Not empty");
2475
2476 size_t region_index = 0;
2477
2478 while (true) {
2479 if (ParCompactionManager::steal(worker_id, region_index)) {
2480 PSParallelCompact::fill_and_update_region(cm, region_index);
2481 cm->drain_region_stacks();
2482 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2483 // Fill and update an unavailable region with the help of a shadow region
2484 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2485 cm->drain_region_stacks();
2486 } else {
2487 if (terminator->offer_termination()) {
2488 break;
2489 }
2490 // Go around again.
2491 }
2492 }
2493 }
2494
2495 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
2496 TaskQueue& _tq;
2497 TaskTerminator _terminator;
2498 uint _active_workers;
2499
2500 public:
2501 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2502 AbstractGangTask("UpdateDensePrefixAndCompactionTask"),
2503 _tq(tq),
2504 _terminator(active_workers, ParCompactionManager::region_task_queues()),
2505 _active_workers(active_workers) {
2506 }
2507 virtual void work(uint worker_id) {
2508 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2509
2510 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2511 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2512 task._space_id,
2513 task._region_index_start,
2514 task._region_index_end);
2515 }
2516
2517 // Once a thread has drained it's stack, it should try to steal regions from
2518 // other threads.
2519 compaction_with_stealing_work(&_terminator, worker_id);
2520 }
2521 };
2522
2523 void PSParallelCompact::compact() {
2524 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2525
2526 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2527 PSOldGen* old_gen = heap->old_gen();
2528 old_gen->start_array()->reset();
2529 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2530
2531 // for [0..last_space_id)
2532 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2533 // push
2534 // push
2535 //
2536 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2537 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2538 initialize_shadow_regions(active_gc_threads);
2539 prepare_region_draining_tasks(active_gc_threads);
2540 enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2541
2542 {
2543 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2544
2545 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2546 ParallelScavengeHeap::heap()->workers().run_task(&task);
2547
2548 #ifdef ASSERT
2549 // Verify that all regions have been processed before the deferred updates.
2550 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2551 verify_complete(SpaceId(id));
2552 }
2553 #endif
2554 }
2555
2556 {
2557 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2558 // Update the deferred objects, if any. In principle, any compaction
2559 // manager can be used. However, since the current thread is VM thread, we
2560 // use the rightful one to keep the verification logic happy.
2561 ParCompactionManager* cm = ParCompactionManager::get_vmthread_cm();
2562 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2563 update_deferred_objects(cm, SpaceId(id));
2564 }
2565 }
2566
2567 DEBUG_ONLY(write_block_fill_histogram());
2568 }
2569
2570 #ifdef ASSERT
2571 void PSParallelCompact::verify_complete(SpaceId space_id) {
2572 // All Regions between space bottom() to new_top() should be marked as filled
2573 // and all Regions between new_top() and top() should be available (i.e.,
2574 // should have been emptied).
2575 ParallelCompactData& sd = summary_data();
2576 SpaceInfo si = _space_info[space_id];
2577 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2578 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2579 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2580 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2581 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2582
2583 bool issued_a_warning = false;
2584
2585 size_t cur_region;
2586 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2587 const RegionData* const c = sd.region(cur_region);
2588 if (!c->completed()) {
2589 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2590 cur_region, c->destination_count());
2591 issued_a_warning = true;
2592 }
2593 }
2594
2595 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2596 const RegionData* const c = sd.region(cur_region);
2597 if (!c->available()) {
2598 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2599 cur_region, c->destination_count());
2600 issued_a_warning = true;
2601 }
2602 }
2603
2604 if (issued_a_warning) {
2605 print_region_ranges();
2606 }
2607 }
2608 #endif // #ifdef ASSERT
2609
2610 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2611 _start_array->allocate_block(addr);
2612 compaction_manager()->update_contents(cast_to_oop(addr));
2613 }
2614
2615 // Update interior oops in the ranges of regions [beg_region, end_region).
2616 void
2617 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2618 SpaceId space_id,
2619 size_t beg_region,
2620 size_t end_region) {
2621 ParallelCompactData& sd = summary_data();
2622 ParMarkBitMap* const mbm = mark_bitmap();
2623
2624 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2625 HeapWord* const end_addr = sd.region_to_addr(end_region);
2626 assert(beg_region <= end_region, "bad region range");
2627 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2628
2629 #ifdef ASSERT
2630 // Claim the regions to avoid triggering an assert when they are marked as
2631 // filled.
2632 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2633 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2634 }
2635 #endif // #ifdef ASSERT
2636
2637 if (beg_addr != space(space_id)->bottom()) {
2638 // Find the first live object or block of dead space that *starts* in this
2639 // range of regions. If a partial object crosses onto the region, skip it;
2640 // it will be marked for 'deferred update' when the object head is
2641 // processed. If dead space crosses onto the region, it is also skipped; it
2642 // will be filled when the prior region is processed. If neither of those
2643 // apply, the first word in the region is the start of a live object or dead
2644 // space.
2645 assert(beg_addr > space(space_id)->bottom(), "sanity");
2646 const RegionData* const cp = sd.region(beg_region);
2647 if (cp->partial_obj_size() != 0) {
2648 beg_addr = sd.partial_obj_end(beg_region);
2649 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2650 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2651 }
2652 }
2653
2654 if (beg_addr < end_addr) {
2655 // A live object or block of dead space starts in this range of Regions.
2656 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2657
2658 // Create closures and iterate.
2659 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2660 FillClosure fill_closure(cm, space_id);
2661 ParMarkBitMap::IterationStatus status;
2662 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2663 dense_prefix_end);
2664 if (status == ParMarkBitMap::incomplete) {
2665 update_closure.do_addr(update_closure.source());
2666 }
2667 }
2668
2669 // Mark the regions as filled.
2670 RegionData* const beg_cp = sd.region(beg_region);
2671 RegionData* const end_cp = sd.region(end_region);
2672 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2673 cp->set_completed();
2674 }
2675 }
2676
2677 // Return the SpaceId for the space containing addr. If addr is not in the
2678 // heap, last_space_id is returned. In debug mode it expects the address to be
2679 // in the heap and asserts such.
2680 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2681 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2682
2683 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2684 if (_space_info[id].space()->contains(addr)) {
2685 return SpaceId(id);
2686 }
2687 }
2688
2689 assert(false, "no space contains the addr");
2690 return last_space_id;
2691 }
2692
2693 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2694 SpaceId id) {
2695 assert(id < last_space_id, "bad space id");
2696
2697 ParallelCompactData& sd = summary_data();
2698 const SpaceInfo* const space_info = _space_info + id;
2699 ObjectStartArray* const start_array = space_info->start_array();
2700
2701 const MutableSpace* const space = space_info->space();
2702 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2703 HeapWord* const beg_addr = space_info->dense_prefix();
2704 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2705
2706 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2707 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2708 const RegionData* cur_region;
2709 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2710 HeapWord* const addr = cur_region->deferred_obj_addr();
2711 if (addr != NULL) {
2712 if (start_array != NULL) {
2713 start_array->allocate_block(addr);
2714 }
2715 cm->update_contents(cast_to_oop(addr));
2716 assert(oopDesc::is_oop_or_null(cast_to_oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2717 }
2718 }
2719 }
2720
2721 // Skip over count live words starting from beg, and return the address of the
2722 // next live word. Unless marked, the word corresponding to beg is assumed to
2723 // be dead. Callers must either ensure beg does not correspond to the middle of
2724 // an object, or account for those live words in some other way. Callers must
2725 // also ensure that there are enough live words in the range [beg, end) to skip.
2726 HeapWord*
2727 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2728 {
2729 assert(count > 0, "sanity");
2730
2731 ParMarkBitMap* m = mark_bitmap();
2732 idx_t bits_to_skip = m->words_to_bits(count);
2733 idx_t cur_beg = m->addr_to_bit(beg);
2734 const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2735
2736 do {
2737 cur_beg = m->find_obj_beg(cur_beg, search_end);
2738 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2739 const size_t obj_bits = cur_end - cur_beg + 1;
2740 if (obj_bits > bits_to_skip) {
2741 return m->bit_to_addr(cur_beg + bits_to_skip);
2742 }
2743 bits_to_skip -= obj_bits;
2744 cur_beg = cur_end + 1;
2745 } while (bits_to_skip > 0);
2746
2747 // Skipping the desired number of words landed just past the end of an object.
2748 // Find the start of the next object.
2749 cur_beg = m->find_obj_beg(cur_beg, search_end);
2750 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2751 return m->bit_to_addr(cur_beg);
2752 }
2753
2754 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2755 SpaceId src_space_id,
2756 size_t src_region_idx)
2757 {
2758 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2759
2760 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2761 if (split_info.dest_region_addr() == dest_addr) {
2762 // The partial object ending at the split point contains the first word to
2763 // be copied to dest_addr.
2764 return split_info.first_src_addr();
2765 }
2766
2767 const ParallelCompactData& sd = summary_data();
2768 ParMarkBitMap* const bitmap = mark_bitmap();
2769 const size_t RegionSize = ParallelCompactData::RegionSize;
2770
2771 assert(sd.is_region_aligned(dest_addr), "not aligned");
2772 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2773 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2774 HeapWord* const src_region_destination = src_region_ptr->destination();
2775
2776 assert(dest_addr >= src_region_destination, "wrong src region");
2777 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2778
2779 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2780 HeapWord* const src_region_end = src_region_beg + RegionSize;
2781
2782 HeapWord* addr = src_region_beg;
2783 if (dest_addr == src_region_destination) {
2784 // Return the first live word in the source region.
2785 if (partial_obj_size == 0) {
2786 addr = bitmap->find_obj_beg(addr, src_region_end);
2787 assert(addr < src_region_end, "no objects start in src region");
2788 }
2789 return addr;
2790 }
2791
2792 // Must skip some live data.
2793 size_t words_to_skip = dest_addr - src_region_destination;
2794 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2795
2796 if (partial_obj_size >= words_to_skip) {
2797 // All the live words to skip are part of the partial object.
2798 addr += words_to_skip;
2799 if (partial_obj_size == words_to_skip) {
2800 // Find the first live word past the partial object.
2801 addr = bitmap->find_obj_beg(addr, src_region_end);
2802 assert(addr < src_region_end, "wrong src region");
2803 }
2804 return addr;
2805 }
2806
2807 // Skip over the partial object (if any).
2808 if (partial_obj_size != 0) {
2809 words_to_skip -= partial_obj_size;
2810 addr += partial_obj_size;
2811 }
2812
2813 // Skip over live words due to objects that start in the region.
2814 addr = skip_live_words(addr, src_region_end, words_to_skip);
2815 assert(addr < src_region_end, "wrong src region");
2816 return addr;
2817 }
2818
2819 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2820 SpaceId src_space_id,
2821 size_t beg_region,
2822 HeapWord* end_addr)
2823 {
2824 ParallelCompactData& sd = summary_data();
2825
2826 #ifdef ASSERT
2827 MutableSpace* const src_space = _space_info[src_space_id].space();
2828 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2829 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2830 "src_space_id does not match beg_addr");
2831 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2832 "src_space_id does not match end_addr");
2833 #endif // #ifdef ASSERT
2834
2835 RegionData* const beg = sd.region(beg_region);
2836 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2837
2838 // Regions up to new_top() are enqueued if they become available.
2839 HeapWord* const new_top = _space_info[src_space_id].new_top();
2840 RegionData* const enqueue_end =
2841 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2842
2843 for (RegionData* cur = beg; cur < end; ++cur) {
2844 assert(cur->data_size() > 0, "region must have live data");
2845 cur->decrement_destination_count();
2846 if (cur < enqueue_end && cur->available() && cur->claim()) {
2847 if (cur->mark_normal()) {
2848 cm->push_region(sd.region(cur));
2849 } else if (cur->mark_copied()) {
2850 // Try to copy the content of the shadow region back to its corresponding
2851 // heap region if the shadow region is filled. Otherwise, the GC thread
2852 // fills the shadow region will copy the data back (see
2853 // MoveAndUpdateShadowClosure::complete_region).
2854 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2855 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2856 cur->set_completed();
2857 }
2858 }
2859 }
2860 }
2861
2862 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2863 SpaceId& src_space_id,
2864 HeapWord*& src_space_top,
2865 HeapWord* end_addr)
2866 {
2867 typedef ParallelCompactData::RegionData RegionData;
2868
2869 ParallelCompactData& sd = PSParallelCompact::summary_data();
2870 const size_t region_size = ParallelCompactData::RegionSize;
2871
2872 size_t src_region_idx = 0;
2873
2874 // Skip empty regions (if any) up to the top of the space.
2875 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2876 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2877 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2878 const RegionData* const top_region_ptr =
2879 sd.addr_to_region_ptr(top_aligned_up);
2880 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2881 ++src_region_ptr;
2882 }
2883
2884 if (src_region_ptr < top_region_ptr) {
2885 // The next source region is in the current space. Update src_region_idx
2886 // and the source address to match src_region_ptr.
2887 src_region_idx = sd.region(src_region_ptr);
2888 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2889 if (src_region_addr > closure.source()) {
2890 closure.set_source(src_region_addr);
2891 }
2892 return src_region_idx;
2893 }
2894
2895 // Switch to a new source space and find the first non-empty region.
2896 unsigned int space_id = src_space_id + 1;
2897 assert(space_id < last_space_id, "not enough spaces");
2898
2899 HeapWord* const destination = closure.destination();
2900
2901 do {
2902 MutableSpace* space = _space_info[space_id].space();
2903 HeapWord* const bottom = space->bottom();
2904 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2905
2906 // Iterate over the spaces that do not compact into themselves.
2907 if (bottom_cp->destination() != bottom) {
2908 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2909 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2910
2911 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2912 if (src_cp->live_obj_size() > 0) {
2913 // Found it.
2914 assert(src_cp->destination() == destination,
2915 "first live obj in the space must match the destination");
2916 assert(src_cp->partial_obj_size() == 0,
2917 "a space cannot begin with a partial obj");
2918
2919 src_space_id = SpaceId(space_id);
2920 src_space_top = space->top();
2921 const size_t src_region_idx = sd.region(src_cp);
2922 closure.set_source(sd.region_to_addr(src_region_idx));
2923 return src_region_idx;
2924 } else {
2925 assert(src_cp->data_size() == 0, "sanity");
2926 }
2927 }
2928 }
2929 } while (++space_id < last_space_id);
2930
2931 assert(false, "no source region was found");
2932 return 0;
2933 }
2934
2935 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2936 {
2937 typedef ParMarkBitMap::IterationStatus IterationStatus;
2938 ParMarkBitMap* const bitmap = mark_bitmap();
2939 ParallelCompactData& sd = summary_data();
2940 RegionData* const region_ptr = sd.region(region_idx);
2941
2942 // Get the source region and related info.
2943 size_t src_region_idx = region_ptr->source_region();
2944 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2945 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2946 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2947
2948 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2949
2950 // Adjust src_region_idx to prepare for decrementing destination counts (the
2951 // destination count is not decremented when a region is copied to itself).
2952 if (src_region_idx == region_idx) {
2953 src_region_idx += 1;
2954 }
2955
2956 if (bitmap->is_unmarked(closure.source())) {
2957 // The first source word is in the middle of an object; copy the remainder
2958 // of the object or as much as will fit. The fact that pointer updates were
2959 // deferred will be noted when the object header is processed.
2960 HeapWord* const old_src_addr = closure.source();
2961 closure.copy_partial_obj();
2962 if (closure.is_full()) {
2963 decrement_destination_counts(cm, src_space_id, src_region_idx,
2964 closure.source());
2965 region_ptr->set_deferred_obj_addr(NULL);
2966 closure.complete_region(cm, dest_addr, region_ptr);
2967 return;
2968 }
2969
2970 HeapWord* const end_addr = sd.region_align_down(closure.source());
2971 if (sd.region_align_down(old_src_addr) != end_addr) {
2972 // The partial object was copied from more than one source region.
2973 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2974
2975 // Move to the next source region, possibly switching spaces as well. All
2976 // args except end_addr may be modified.
2977 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2978 end_addr);
2979 }
2980 }
2981
2982 do {
2983 HeapWord* const cur_addr = closure.source();
2984 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2985 src_space_top);
2986 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2987
2988 if (status == ParMarkBitMap::incomplete) {
2989 // The last obj that starts in the source region does not end in the
2990 // region.
2991 assert(closure.source() < end_addr, "sanity");
2992 HeapWord* const obj_beg = closure.source();
2993 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2994 src_space_top);
2995 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2996 if (obj_end < range_end) {
2997 // The end was found; the entire object will fit.
2998 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2999 assert(status != ParMarkBitMap::would_overflow, "sanity");
3000 } else {
3001 // The end was not found; the object will not fit.
3002 assert(range_end < src_space_top, "obj cannot cross space boundary");
3003 status = ParMarkBitMap::would_overflow;
3004 }
3005 }
3006
3007 if (status == ParMarkBitMap::would_overflow) {
3008 // The last object did not fit. Note that interior oop updates were
3009 // deferred, then copy enough of the object to fill the region.
3010 region_ptr->set_deferred_obj_addr(closure.destination());
3011 status = closure.copy_until_full(); // copies from closure.source()
3012
3013 decrement_destination_counts(cm, src_space_id, src_region_idx,
3014 closure.source());
3015 closure.complete_region(cm, dest_addr, region_ptr);
3016 return;
3017 }
3018
3019 if (status == ParMarkBitMap::full) {
3020 decrement_destination_counts(cm, src_space_id, src_region_idx,
3021 closure.source());
3022 region_ptr->set_deferred_obj_addr(NULL);
3023 closure.complete_region(cm, dest_addr, region_ptr);
3024 return;
3025 }
3026
3027 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3028
3029 // Move to the next source region, possibly switching spaces as well. All
3030 // args except end_addr may be modified.
3031 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3032 end_addr);
3033 } while (true);
3034 }
3035
3036 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3037 {
3038 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3039 fill_region(cm, cl, region_idx);
3040 }
3041
3042 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3043 {
3044 // Get a shadow region first
3045 ParallelCompactData& sd = summary_data();
3046 RegionData* const region_ptr = sd.region(region_idx);
3047 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3048 // The InvalidShadow return value indicates the corresponding heap region is available,
3049 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3050 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3051 if (shadow_region == ParCompactionManager::InvalidShadow) {
3052 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3053 region_ptr->shadow_to_normal();
3054 return fill_region(cm, cl, region_idx);
3055 } else {
3056 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3057 return fill_region(cm, cl, region_idx);
3058 }
3059 }
3060
3061 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3062 {
3063 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3064 }
3065
3066 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
3067 {
3068 size_t next = cm->next_shadow_region();
3069 ParallelCompactData& sd = summary_data();
3070 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3071 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3072
3073 while (next < old_new_top) {
3074 if (sd.region(next)->mark_shadow()) {
3075 region_idx = next;
3076 return true;
3077 }
3078 next = cm->move_next_shadow_region_by(active_gc_threads);
3079 }
3080
3081 return false;
3082 }
3083
3084 // The shadow region is an optimization to address region dependencies in full GC. The basic
3085 // idea is making more regions available by temporally storing their live objects in empty
3086 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3087 // GC threads need not wait destination regions to be available before processing sources.
3088 //
3089 // A typical workflow would be:
3090 // After draining its own stack and failing to steal from others, a GC worker would pick an
3091 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3092 // the shadow region by copying live objects from source regions of the unavailable one. Once
3093 // the unavailable region becomes available, the data in the shadow region will be copied back.
3094 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3095 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3096 {
3097 const ParallelCompactData& sd = PSParallelCompact::summary_data();
3098
3099 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3100 SpaceInfo* const space_info = _space_info + id;
3101 MutableSpace* const space = space_info->space();
3102
3103 const size_t beg_region =
3104 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3105 const size_t end_region =
3106 sd.addr_to_region_idx(sd.region_align_down(space->end()));
3107
3108 for (size_t cur = beg_region; cur < end_region; ++cur) {
3109 ParCompactionManager::push_shadow_region(cur);
3110 }
3111 }
3112
3113 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3114 for (uint i = 0; i < parallel_gc_threads; i++) {
3115 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3116 cm->set_next_shadow_region(beg_region + i);
3117 }
3118 }
3119
3120 void PSParallelCompact::fill_blocks(size_t region_idx)
3121 {
3122 // Fill in the block table elements for the specified region. Each block
3123 // table element holds the number of live words in the region that are to the
3124 // left of the first object that starts in the block. Thus only blocks in
3125 // which an object starts need to be filled.
3126 //
3127 // The algorithm scans the section of the bitmap that corresponds to the
3128 // region, keeping a running total of the live words. When an object start is
3129 // found, if it's the first to start in the block that contains it, the
3130 // current total is written to the block table element.
3131 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3132 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3133 const size_t RegionSize = ParallelCompactData::RegionSize;
3134
3135 ParallelCompactData& sd = summary_data();
3136 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3137 if (partial_obj_size >= RegionSize) {
3138 return; // No objects start in this region.
3139 }
3140
3141 // Ensure the first loop iteration decides that the block has changed.
3142 size_t cur_block = sd.block_count();
3143
3144 const ParMarkBitMap* const bitmap = mark_bitmap();
3145
3146 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3147 assert((size_t)1 << Log2BitsPerBlock ==
3148 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3149
3150 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3151 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3152 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3153 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3154 while (beg_bit < range_end) {
3155 const size_t new_block = beg_bit >> Log2BitsPerBlock;
3156 if (new_block != cur_block) {
3157 cur_block = new_block;
3158 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3159 }
3160
3161 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3162 if (end_bit < range_end - 1) {
3163 live_bits += end_bit - beg_bit + 1;
3164 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3165 } else {
3166 return;
3167 }
3168 }
3169 }
3170
3171 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3172 {
3173 if (source() != copy_destination()) {
3174 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3175 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3176 }
3177 update_state(words_remaining());
3178 assert(is_full(), "sanity");
3179 return ParMarkBitMap::full;
3180 }
3181
3182 void MoveAndUpdateClosure::copy_partial_obj()
3183 {
3184 size_t words = words_remaining();
3185
3186 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3187 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3188 if (end_addr < range_end) {
3189 words = bitmap()->obj_size(source(), end_addr);
3190 }
3191
3192 // This test is necessary; if omitted, the pointer updates to a partial object
3193 // that crosses the dense prefix boundary could be overwritten.
3194 if (source() != copy_destination()) {
3195 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3196 Copy::aligned_conjoint_words(source(), copy_destination(), words);
3197 }
3198 update_state(words);
3199 }
3200
3201 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3202 PSParallelCompact::RegionData *region_ptr) {
3203 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3204 region_ptr->set_completed();
3205 }
3206
3207 ParMarkBitMapClosure::IterationStatus
3208 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3209 assert(destination() != NULL, "sanity");
3210 assert(bitmap()->obj_size(addr) == words, "bad size");
3211
3212 _source = addr;
3213 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3214 destination(), "wrong destination");
3215
3216 if (words > words_remaining()) {
3217 return ParMarkBitMap::would_overflow;
3218 }
3219
3220 // The start_array must be updated even if the object is not moving.
3221 if (_start_array != NULL) {
3222 _start_array->allocate_block(destination());
3223 }
3224
3225 if (copy_destination() != source()) {
3226 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3227 Copy::aligned_conjoint_words(source(), copy_destination(), words);
3228 }
3229
3230 oop moved_oop = cast_to_oop(copy_destination());
3231 compaction_manager()->update_contents(moved_oop);
3232 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3233
3234 update_state(words);
3235 assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3236 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3237 }
3238
3239 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3240 PSParallelCompact::RegionData *region_ptr) {
3241 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3242 // Record the shadow region index
3243 region_ptr->set_shadow_region(_shadow);
3244 // Mark the shadow region as filled to indicate the data is ready to be
3245 // copied back
3246 region_ptr->mark_filled();
3247 // Try to copy the content of the shadow region back to its corresponding
3248 // heap region if available; the GC thread that decreases the destination
3249 // count to zero will do the copying otherwise (see
3250 // PSParallelCompact::decrement_destination_counts).
3251 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3252 region_ptr->set_completed();
3253 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3254 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3255 }
3256 }
3257
3258 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3259 ParCompactionManager* cm,
3260 PSParallelCompact::SpaceId space_id) :
3261 ParMarkBitMapClosure(mbm, cm),
3262 _space_id(space_id),
3263 _start_array(PSParallelCompact::start_array(space_id))
3264 {
3265 }
3266
3267 // Updates the references in the object to their new values.
3268 ParMarkBitMapClosure::IterationStatus
3269 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3270 do_addr(addr);
3271 return ParMarkBitMap::incomplete;
3272 }
3273
3274 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3275 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3276 _start_array(PSParallelCompact::start_array(space_id))
3277 {
3278 assert(space_id == PSParallelCompact::old_space_id,
3279 "cannot use FillClosure in the young gen");
3280 }
3281
3282 ParMarkBitMapClosure::IterationStatus
3283 FillClosure::do_addr(HeapWord* addr, size_t size) {
3284 CollectedHeap::fill_with_objects(addr, size);
3285 HeapWord* const end = addr + size;
3286 do {
3287 _start_array->allocate_block(addr);
3288 addr += cast_to_oop(addr)->size();
3289 } while (addr < end);
3290 return ParMarkBitMap::incomplete;
3291 }