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