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