1 /*
2 * Copyright (c) 2005, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "aot/aotLoader.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "gc/parallel/gcTaskManager.hpp"
32 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
33 #include "gc/parallel/parMarkBitMap.inline.hpp"
34 #include "gc/parallel/pcTasks.hpp"
35 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
36 #include "gc/parallel/psCompactionManager.inline.hpp"
37 #include "gc/parallel/psMarkSweep.hpp"
38 #include "gc/parallel/psMarkSweepDecorator.hpp"
39 #include "gc/parallel/psOldGen.hpp"
40 #include "gc/parallel/psParallelCompact.inline.hpp"
41 #include "gc/parallel/psPromotionManager.inline.hpp"
42 #include "gc/parallel/psScavenge.hpp"
43 #include "gc/parallel/psYoungGen.hpp"
44 #include "gc/shared/gcCause.hpp"
45 #include "gc/shared/gcHeapSummary.hpp"
46 #include "gc/shared/gcId.hpp"
47 #include "gc/shared/gcLocker.inline.hpp"
48 #include "gc/shared/gcTimer.hpp"
49 #include "gc/shared/gcTrace.hpp"
50 #include "gc/shared/gcTraceTime.inline.hpp"
51 #include "gc/shared/isGCActiveMark.hpp"
52 #include "gc/shared/referencePolicy.hpp"
53 #include "gc/shared/referenceProcessor.hpp"
54 #include "gc/shared/spaceDecorator.hpp"
55 #include "logging/log.hpp"
56 #include "memory/resourceArea.hpp"
57 #include "oops/instanceKlass.inline.hpp"
58 #include "oops/instanceMirrorKlass.inline.hpp"
59 #include "oops/methodData.hpp"
60 #include "oops/objArrayKlass.inline.hpp"
61 #include "oops/oop.inline.hpp"
62 #include "runtime/atomic.hpp"
63 #include "runtime/fprofiler.hpp"
64 #include "runtime/safepoint.hpp"
65 #include "runtime/vmThread.hpp"
66 #include "services/management.hpp"
67 #include "services/memTracker.hpp"
68 #include "services/memoryService.hpp"
69 #include "utilities/align.hpp"
70 #include "utilities/debug.hpp"
71 #include "utilities/events.hpp"
72 #include "utilities/formatBuffer.hpp"
73 #include "utilities/stack.inline.hpp"
74
75 #include <math.h>
76
77 // All sizes are in HeapWords.
78 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
79 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
80 const size_t ParallelCompactData::RegionSizeBytes =
81 RegionSize << LogHeapWordSize;
82 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
83 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
84 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
85
86 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
87 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
88 const size_t ParallelCompactData::BlockSizeBytes =
89 BlockSize << LogHeapWordSize;
90 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
91 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
92 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
93
94 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
95 const size_t ParallelCompactData::Log2BlocksPerRegion =
96 Log2RegionSize - Log2BlockSize;
97
98 const ParallelCompactData::RegionData::region_sz_t
99 ParallelCompactData::RegionData::dc_shift = 27;
100
101 const ParallelCompactData::RegionData::region_sz_t
102 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
103
104 const ParallelCompactData::RegionData::region_sz_t
105 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
106
107 const ParallelCompactData::RegionData::region_sz_t
108 ParallelCompactData::RegionData::los_mask = ~dc_mask;
109
110 const ParallelCompactData::RegionData::region_sz_t
111 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
112
113 const ParallelCompactData::RegionData::region_sz_t
114 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
115
116 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
117
118 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
119
120 double PSParallelCompact::_dwl_mean;
121 double PSParallelCompact::_dwl_std_dev;
122 double PSParallelCompact::_dwl_first_term;
123 double PSParallelCompact::_dwl_adjustment;
124 #ifdef ASSERT
125 bool PSParallelCompact::_dwl_initialized = false;
126 #endif // #ifdef ASSERT
127
128 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
129 HeapWord* destination)
130 {
131 assert(src_region_idx != 0, "invalid src_region_idx");
132 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
133 assert(destination != NULL, "invalid destination argument");
134
135 _src_region_idx = src_region_idx;
136 _partial_obj_size = partial_obj_size;
137 _destination = destination;
138
139 // These fields may not be updated below, so make sure they're clear.
140 assert(_dest_region_addr == NULL, "should have been cleared");
141 assert(_first_src_addr == NULL, "should have been cleared");
142
143 // Determine the number of destination regions for the partial object.
144 HeapWord* const last_word = destination + partial_obj_size - 1;
145 const ParallelCompactData& sd = PSParallelCompact::summary_data();
146 HeapWord* const beg_region_addr = sd.region_align_down(destination);
147 HeapWord* const end_region_addr = sd.region_align_down(last_word);
148
149 if (beg_region_addr == end_region_addr) {
150 // One destination region.
151 _destination_count = 1;
152 if (end_region_addr == destination) {
153 // The destination falls on a region boundary, thus the first word of the
154 // partial object will be the first word copied to the destination region.
155 _dest_region_addr = end_region_addr;
156 _first_src_addr = sd.region_to_addr(src_region_idx);
157 }
158 } else {
159 // Two destination regions. When copied, the partial object will cross a
160 // destination region boundary, so a word somewhere within the partial
161 // object will be the first word copied to the second destination region.
162 _destination_count = 2;
163 _dest_region_addr = end_region_addr;
164 const size_t ofs = pointer_delta(end_region_addr, destination);
165 assert(ofs < _partial_obj_size, "sanity");
166 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
167 }
168 }
169
170 void SplitInfo::clear()
171 {
172 _src_region_idx = 0;
173 _partial_obj_size = 0;
174 _destination = NULL;
175 _destination_count = 0;
176 _dest_region_addr = NULL;
177 _first_src_addr = NULL;
178 assert(!is_valid(), "sanity");
179 }
180
181 #ifdef ASSERT
182 void SplitInfo::verify_clear()
183 {
184 assert(_src_region_idx == 0, "not clear");
185 assert(_partial_obj_size == 0, "not clear");
186 assert(_destination == NULL, "not clear");
187 assert(_destination_count == 0, "not clear");
188 assert(_dest_region_addr == NULL, "not clear");
189 assert(_first_src_addr == NULL, "not clear");
190 }
191 #endif // #ifdef ASSERT
192
193
194 void PSParallelCompact::print_on_error(outputStream* st) {
195 _mark_bitmap.print_on_error(st);
196 }
197
198 #ifndef PRODUCT
199 const char* PSParallelCompact::space_names[] = {
200 "old ", "eden", "from", "to "
201 };
202
203 void PSParallelCompact::print_region_ranges() {
204 if (!log_develop_is_enabled(Trace, gc, compaction)) {
205 return;
206 }
207 Log(gc, compaction) log;
208 ResourceMark rm;
209 Universe::print_on(log.trace_stream());
210 log.trace("space bottom top end new_top");
211 log.trace("------ ---------- ---------- ---------- ----------");
212
213 for (unsigned int id = 0; id < last_space_id; ++id) {
214 const MutableSpace* space = _space_info[id].space();
215 log.trace("%u %s "
216 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
217 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
218 id, space_names[id],
219 summary_data().addr_to_region_idx(space->bottom()),
220 summary_data().addr_to_region_idx(space->top()),
221 summary_data().addr_to_region_idx(space->end()),
222 summary_data().addr_to_region_idx(_space_info[id].new_top()));
223 }
224 }
225
226 void
227 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
228 {
229 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
230 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
231
232 ParallelCompactData& sd = PSParallelCompact::summary_data();
233 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
234 log_develop_trace(gc, compaction)(
235 REGION_IDX_FORMAT " " PTR_FORMAT " "
236 REGION_IDX_FORMAT " " PTR_FORMAT " "
237 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
238 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
239 i, p2i(c->data_location()), dci, p2i(c->destination()),
240 c->partial_obj_size(), c->live_obj_size(),
241 c->data_size(), c->source_region(), c->destination_count());
242
243 #undef REGION_IDX_FORMAT
244 #undef REGION_DATA_FORMAT
245 }
246
247 void
248 print_generic_summary_data(ParallelCompactData& summary_data,
249 HeapWord* const beg_addr,
250 HeapWord* const end_addr)
251 {
252 size_t total_words = 0;
253 size_t i = summary_data.addr_to_region_idx(beg_addr);
254 const size_t last = summary_data.addr_to_region_idx(end_addr);
255 HeapWord* pdest = 0;
256
257 while (i < last) {
258 ParallelCompactData::RegionData* c = summary_data.region(i);
259 if (c->data_size() != 0 || c->destination() != pdest) {
260 print_generic_summary_region(i, c);
261 total_words += c->data_size();
262 pdest = c->destination();
263 }
264 ++i;
265 }
266
267 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
268 }
269
270 void
271 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
272 HeapWord* const beg_addr,
273 HeapWord* const end_addr) {
274 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
275 }
276
277 void
278 print_generic_summary_data(ParallelCompactData& summary_data,
279 SpaceInfo* space_info)
280 {
281 if (!log_develop_is_enabled(Trace, gc, compaction)) {
282 return;
283 }
284
285 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
286 const MutableSpace* space = space_info[id].space();
287 print_generic_summary_data(summary_data, space->bottom(),
288 MAX2(space->top(), space_info[id].new_top()));
289 }
290 }
291
292 void
293 print_initial_summary_data(ParallelCompactData& summary_data,
294 const MutableSpace* space) {
295 if (space->top() == space->bottom()) {
296 return;
297 }
298
299 const size_t region_size = ParallelCompactData::RegionSize;
300 typedef ParallelCompactData::RegionData RegionData;
301 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
302 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
303 const RegionData* c = summary_data.region(end_region - 1);
304 HeapWord* end_addr = c->destination() + c->data_size();
305 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
306
307 // Print (and count) the full regions at the beginning of the space.
308 size_t full_region_count = 0;
309 size_t i = summary_data.addr_to_region_idx(space->bottom());
310 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
311 ParallelCompactData::RegionData* c = summary_data.region(i);
312 log_develop_trace(gc, compaction)(
313 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
314 i, p2i(c->destination()),
315 c->partial_obj_size(), c->live_obj_size(),
316 c->data_size(), c->source_region(), c->destination_count());
317 ++full_region_count;
318 ++i;
319 }
320
321 size_t live_to_right = live_in_space - full_region_count * region_size;
322
323 double max_reclaimed_ratio = 0.0;
324 size_t max_reclaimed_ratio_region = 0;
325 size_t max_dead_to_right = 0;
326 size_t max_live_to_right = 0;
327
328 // Print the 'reclaimed ratio' for regions while there is something live in
329 // the region or to the right of it. The remaining regions are empty (and
330 // uninteresting), and computing the ratio will result in division by 0.
331 while (i < end_region && live_to_right > 0) {
332 c = summary_data.region(i);
333 HeapWord* const region_addr = summary_data.region_to_addr(i);
334 const size_t used_to_right = pointer_delta(space->top(), region_addr);
335 const size_t dead_to_right = used_to_right - live_to_right;
336 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
337
338 if (reclaimed_ratio > max_reclaimed_ratio) {
339 max_reclaimed_ratio = reclaimed_ratio;
340 max_reclaimed_ratio_region = i;
341 max_dead_to_right = dead_to_right;
342 max_live_to_right = live_to_right;
343 }
344
345 ParallelCompactData::RegionData* c = summary_data.region(i);
346 log_develop_trace(gc, compaction)(
347 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
348 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
349 i, p2i(c->destination()),
350 c->partial_obj_size(), c->live_obj_size(),
351 c->data_size(), c->source_region(), c->destination_count(),
352 reclaimed_ratio, dead_to_right, live_to_right);
353
354
355 live_to_right -= c->data_size();
356 ++i;
357 }
358
359 // Any remaining regions are empty. Print one more if there is one.
360 if (i < end_region) {
361 ParallelCompactData::RegionData* c = summary_data.region(i);
362 log_develop_trace(gc, compaction)(
363 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
364 i, p2i(c->destination()),
365 c->partial_obj_size(), c->live_obj_size(),
366 c->data_size(), c->source_region(), c->destination_count());
367 }
368
369 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
370 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
371 }
372
373 void
374 print_initial_summary_data(ParallelCompactData& summary_data,
375 SpaceInfo* space_info) {
376 if (!log_develop_is_enabled(Trace, gc, compaction)) {
377 return;
378 }
379
380 unsigned int id = PSParallelCompact::old_space_id;
381 const MutableSpace* space;
382 do {
383 space = space_info[id].space();
384 print_initial_summary_data(summary_data, space);
385 } while (++id < PSParallelCompact::eden_space_id);
386
387 do {
388 space = space_info[id].space();
389 print_generic_summary_data(summary_data, space->bottom(), space->top());
390 } while (++id < PSParallelCompact::last_space_id);
391 }
392 #endif // #ifndef PRODUCT
393
394 #ifdef ASSERT
395 size_t add_obj_count;
396 size_t add_obj_size;
397 size_t mark_bitmap_count;
398 size_t mark_bitmap_size;
399 #endif // #ifdef ASSERT
400
401 ParallelCompactData::ParallelCompactData()
402 {
403 _region_start = 0;
404
405 _region_vspace = 0;
406 _reserved_byte_size = 0;
407 _region_data = 0;
408 _region_count = 0;
409
410 _block_vspace = 0;
411 _block_data = 0;
412 _block_count = 0;
413 }
414
415 bool ParallelCompactData::initialize(MemRegion covered_region)
416 {
417 _region_start = covered_region.start();
418 const size_t region_size = covered_region.word_size();
419 DEBUG_ONLY(_region_end = _region_start + region_size;)
420
421 assert(region_align_down(_region_start) == _region_start,
422 "region start not aligned");
423 assert((region_size & RegionSizeOffsetMask) == 0,
424 "region size not a multiple of RegionSize");
425
426 bool result = initialize_region_data(region_size) && initialize_block_data();
427 return result;
428 }
429
430 PSVirtualSpace*
431 ParallelCompactData::create_vspace(size_t count, size_t element_size)
432 {
433 const size_t raw_bytes = count * element_size;
434 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
435 const size_t granularity = os::vm_allocation_granularity();
436 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
437
438 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
439 MAX2(page_sz, granularity);
440 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
441 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
442 rs.size());
443
444 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
445
446 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
447 if (vspace != 0) {
448 if (vspace->expand_by(_reserved_byte_size)) {
449 return vspace;
450 }
451 delete vspace;
452 // Release memory reserved in the space.
453 rs.release();
454 }
455
456 return 0;
457 }
458
459 bool ParallelCompactData::initialize_region_data(size_t region_size)
460 {
461 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
462 _region_vspace = create_vspace(count, sizeof(RegionData));
463 if (_region_vspace != 0) {
464 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
465 _region_count = count;
466 return true;
467 }
468 return false;
469 }
470
471 bool ParallelCompactData::initialize_block_data()
472 {
473 assert(_region_count != 0, "region data must be initialized first");
474 const size_t count = _region_count << Log2BlocksPerRegion;
475 _block_vspace = create_vspace(count, sizeof(BlockData));
476 if (_block_vspace != 0) {
477 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
478 _block_count = count;
479 return true;
480 }
481 return false;
482 }
483
484 void ParallelCompactData::clear()
485 {
486 memset(_region_data, 0, _region_vspace->committed_size());
487 memset(_block_data, 0, _block_vspace->committed_size());
488 }
489
490 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
491 assert(beg_region <= _region_count, "beg_region out of range");
492 assert(end_region <= _region_count, "end_region out of range");
493 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
494
495 const size_t region_cnt = end_region - beg_region;
496 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
497
498 const size_t beg_block = beg_region * BlocksPerRegion;
499 const size_t block_cnt = region_cnt * BlocksPerRegion;
500 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
501 }
502
503 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
504 {
505 const RegionData* cur_cp = region(region_idx);
506 const RegionData* const end_cp = region(region_count() - 1);
507
508 HeapWord* result = region_to_addr(region_idx);
509 if (cur_cp < end_cp) {
510 do {
511 result += cur_cp->partial_obj_size();
512 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
513 }
514 return result;
515 }
516
517 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
518 {
519 const size_t obj_ofs = pointer_delta(addr, _region_start);
520 const size_t beg_region = obj_ofs >> Log2RegionSize;
521 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
522
523 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
524 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
525
526 if (beg_region == end_region) {
527 // All in one region.
528 _region_data[beg_region].add_live_obj(len);
529 return;
530 }
531
532 // First region.
533 const size_t beg_ofs = region_offset(addr);
534 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
535
536 Klass* klass = ((oop)addr)->klass();
537 // Middle regions--completely spanned by this object.
538 for (size_t region = beg_region + 1; region < end_region; ++region) {
539 _region_data[region].set_partial_obj_size(RegionSize);
540 _region_data[region].set_partial_obj_addr(addr);
541 }
542
543 // Last region.
544 const size_t end_ofs = region_offset(addr + len - 1);
545 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
546 _region_data[end_region].set_partial_obj_addr(addr);
547 }
548
549 void
550 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
551 {
552 assert(region_offset(beg) == 0, "not RegionSize aligned");
553 assert(region_offset(end) == 0, "not RegionSize aligned");
554
555 size_t cur_region = addr_to_region_idx(beg);
556 const size_t end_region = addr_to_region_idx(end);
557 HeapWord* addr = beg;
558 while (cur_region < end_region) {
559 _region_data[cur_region].set_destination(addr);
560 _region_data[cur_region].set_destination_count(0);
561 _region_data[cur_region].set_source_region(cur_region);
562 _region_data[cur_region].set_data_location(addr);
563
564 // Update live_obj_size so the region appears completely full.
565 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
566 _region_data[cur_region].set_live_obj_size(live_size);
567
568 ++cur_region;
569 addr += RegionSize;
570 }
571 }
572
573 // Find the point at which a space can be split and, if necessary, record the
574 // split point.
575 //
576 // If the current src region (which overflowed the destination space) doesn't
577 // have a partial object, the split point is at the beginning of the current src
578 // region (an "easy" split, no extra bookkeeping required).
579 //
580 // If the current src region has a partial object, the split point is in the
581 // region where that partial object starts (call it the split_region). If
582 // split_region has a partial object, then the split point is just after that
583 // partial object (a "hard" split where we have to record the split data and
584 // zero the partial_obj_size field). With a "hard" split, we know that the
585 // partial_obj ends within split_region because the partial object that caused
586 // the overflow starts in split_region. If split_region doesn't have a partial
587 // obj, then the split is at the beginning of split_region (another "easy"
588 // split).
589 HeapWord*
590 ParallelCompactData::summarize_split_space(size_t src_region,
591 SplitInfo& split_info,
592 HeapWord* destination,
593 HeapWord* target_end,
594 HeapWord** target_next)
595 {
596 assert(destination <= target_end, "sanity");
597 assert(destination + _region_data[src_region].data_size() > target_end,
598 "region should not fit into target space");
599 assert(is_region_aligned(target_end), "sanity");
600
601 size_t split_region = src_region;
602 HeapWord* split_destination = destination;
603 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
604
605 if (destination + partial_obj_size > target_end) {
606 // The split point is just after the partial object (if any) in the
607 // src_region that contains the start of the object that overflowed the
608 // destination space.
609 //
610 // Find the start of the "overflow" object and set split_region to the
611 // region containing it.
612 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
613 split_region = addr_to_region_idx(overflow_obj);
614
615 // Clear the source_region field of all destination regions whose first word
616 // came from data after the split point (a non-null source_region field
617 // implies a region must be filled).
618 //
619 // An alternative to the simple loop below: clear during post_compact(),
620 // which uses memcpy instead of individual stores, and is easy to
621 // parallelize. (The downside is that it clears the entire RegionData
622 // object as opposed to just one field.)
623 //
624 // post_compact() would have to clear the summary data up to the highest
625 // address that was written during the summary phase, which would be
626 //
627 // max(top, max(new_top, clear_top))
628 //
629 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
630 // to target_end.
631 const RegionData* const sr = region(split_region);
632 const size_t beg_idx =
633 addr_to_region_idx(region_align_up(sr->destination() +
634 sr->partial_obj_size()));
635 const size_t end_idx = addr_to_region_idx(target_end);
636
637 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
638 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
639 _region_data[idx].set_source_region(0);
640 }
641
642 // Set split_destination and partial_obj_size to reflect the split region.
643 split_destination = sr->destination();
644 partial_obj_size = sr->partial_obj_size();
645 }
646
647 // The split is recorded only if a partial object extends onto the region.
648 if (partial_obj_size != 0) {
649 _region_data[split_region].set_partial_obj_size(0);
650 split_info.record(split_region, partial_obj_size, split_destination);
651 }
652
653 // Setup the continuation addresses.
654 *target_next = split_destination + partial_obj_size;
655 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
656
657 if (log_develop_is_enabled(Trace, gc, compaction)) {
658 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
659 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
660 split_type, p2i(source_next), split_region, partial_obj_size);
661 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
662 split_type, p2i(split_destination),
663 addr_to_region_idx(split_destination),
664 p2i(*target_next));
665
666 if (partial_obj_size != 0) {
667 HeapWord* const po_beg = split_info.destination();
668 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
669 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
670 split_type,
671 p2i(po_beg), addr_to_region_idx(po_beg),
672 p2i(po_end), addr_to_region_idx(po_end));
673 }
674 }
675
676 return source_next;
677 }
678
679 bool ParallelCompactData::summarize(SplitInfo& split_info,
680 HeapWord* source_beg, HeapWord* source_end,
681 HeapWord** source_next,
682 HeapWord* target_beg, HeapWord* target_end,
683 HeapWord** target_next)
684 {
685 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
686 log_develop_trace(gc, compaction)(
687 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
688 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
689 p2i(source_beg), p2i(source_end), p2i(source_next_val),
690 p2i(target_beg), p2i(target_end), p2i(*target_next));
691
692 size_t cur_region = addr_to_region_idx(source_beg);
693 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
694
695 HeapWord *dest_addr = target_beg;
696 while (cur_region < end_region) {
697 // The destination must be set even if the region has no data.
698 _region_data[cur_region].set_destination(dest_addr);
699
700 size_t words = _region_data[cur_region].data_size();
701 if (words > 0) {
702 // If cur_region does not fit entirely into the target space, find a point
703 // at which the source space can be 'split' so that part is copied to the
704 // target space and the rest is copied elsewhere.
705 if (dest_addr + words > target_end) {
706 assert(source_next != NULL, "source_next is NULL when splitting");
707 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
708 target_end, target_next);
709 return false;
710 }
711
712 // Compute the destination_count for cur_region, and if necessary, update
713 // source_region for a destination region. The source_region field is
714 // updated if cur_region is the first (left-most) region to be copied to a
715 // destination region.
716 //
717 // The destination_count calculation is a bit subtle. A region that has
718 // data that compacts into itself does not count itself as a destination.
719 // This maintains the invariant that a zero count means the region is
720 // available and can be claimed and then filled.
721 uint destination_count = 0;
722 if (split_info.is_split(cur_region)) {
723 // The current region has been split: the partial object will be copied
724 // to one destination space and the remaining data will be copied to
725 // another destination space. Adjust the initial destination_count and,
726 // if necessary, set the source_region field if the partial object will
727 // cross a destination region boundary.
728 destination_count = split_info.destination_count();
729 if (destination_count == 2) {
730 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
731 _region_data[dest_idx].set_source_region(cur_region);
732 }
733 }
734
735 HeapWord* const last_addr = dest_addr + words - 1;
736 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
737 const size_t dest_region_2 = addr_to_region_idx(last_addr);
738
739 // Initially assume that the destination regions will be the same and
740 // adjust the value below if necessary. Under this assumption, if
741 // cur_region == dest_region_2, then cur_region will be compacted
742 // completely into itself.
743 destination_count += cur_region == dest_region_2 ? 0 : 1;
744 if (dest_region_1 != dest_region_2) {
745 // Destination regions differ; adjust destination_count.
746 destination_count += 1;
747 // Data from cur_region will be copied to the start of dest_region_2.
748 _region_data[dest_region_2].set_source_region(cur_region);
749 } else if (region_offset(dest_addr) == 0) {
750 // Data from cur_region will be copied to the start of the destination
751 // region.
752 _region_data[dest_region_1].set_source_region(cur_region);
753 }
754
755 _region_data[cur_region].set_destination_count(destination_count);
756 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
757 dest_addr += words;
758 }
759
760 ++cur_region;
761 }
762
763 *target_next = dest_addr;
764 return true;
765 }
766
767 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
768 assert(addr != NULL, "Should detect NULL oop earlier");
769 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
770 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
771
772 // Region covering the object.
773 RegionData* const region_ptr = addr_to_region_ptr(addr);
774 HeapWord* result = region_ptr->destination();
775
776 // If the entire Region is live, the new location is region->destination + the
777 // offset of the object within in the Region.
778
779 // Run some performance tests to determine if this special case pays off. It
780 // is worth it for pointers into the dense prefix. If the optimization to
781 // avoid pointer updates in regions that only point to the dense prefix is
782 // ever implemented, this should be revisited.
783 if (region_ptr->data_size() == RegionSize) {
784 result += region_offset(addr);
785 return result;
786 }
787
788 // Otherwise, the new location is region->destination + block offset + the
789 // number of live words in the Block that are (a) to the left of addr and (b)
790 // due to objects that start in the Block.
791
792 // Fill in the block table if necessary. This is unsynchronized, so multiple
793 // threads may fill the block table for a region (harmless, since it is
794 // idempotent).
795 if (!region_ptr->blocks_filled()) {
796 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
797 region_ptr->set_blocks_filled();
798 }
799
800 HeapWord* const search_start = block_align_down(addr);
801 const size_t block_offset = addr_to_block_ptr(addr)->offset();
802
803 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
804 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr));
805 result += block_offset + live;
806 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
807 return result;
808 }
809
810 #ifdef ASSERT
811 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
812 {
813 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
814 const size_t* const end = (const size_t*)vspace->committed_high_addr();
815 for (const size_t* p = beg; p < end; ++p) {
816 assert(*p == 0, "not zero");
817 }
818 }
819
820 void ParallelCompactData::verify_clear()
821 {
822 verify_clear(_region_vspace);
823 verify_clear(_block_vspace);
824 }
825 #endif // #ifdef ASSERT
826
827 STWGCTimer PSParallelCompact::_gc_timer;
828 ParallelOldTracer PSParallelCompact::_gc_tracer;
829 elapsedTimer PSParallelCompact::_accumulated_time;
830 unsigned int PSParallelCompact::_total_invocations = 0;
831 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
832 jlong PSParallelCompact::_time_of_last_gc = 0;
833 CollectorCounters* PSParallelCompact::_counters = NULL;
834 ParMarkBitMap PSParallelCompact::_mark_bitmap;
835 ParallelCompactData PSParallelCompact::_summary_data;
836
837 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
838
839 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
840
841 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
842 PSParallelCompact::AdjustPointerClosure closure(_cm);
843 klass->oops_do(&closure);
844 }
845
846 void PSParallelCompact::post_initialize() {
847 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
848 MemRegion mr = heap->reserved_region();
849 _ref_processor =
850 new ReferenceProcessor(mr, // span
851 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
852 ParallelGCThreads, // mt processing degree
853 true, // mt discovery
854 ParallelGCThreads, // mt discovery degree
855 true, // atomic_discovery
856 &_is_alive_closure); // non-header is alive closure
857 _counters = new CollectorCounters("PSParallelCompact", 1);
858
859 // Initialize static fields in ParCompactionManager.
860 ParCompactionManager::initialize(mark_bitmap());
861 }
862
863 bool PSParallelCompact::initialize() {
864 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
865 MemRegion mr = heap->reserved_region();
866
867 // Was the old gen get allocated successfully?
868 if (!heap->old_gen()->is_allocated()) {
869 return false;
870 }
871
872 initialize_space_info();
873 initialize_dead_wood_limiter();
874
875 if (!_mark_bitmap.initialize(mr)) {
876 vm_shutdown_during_initialization(
877 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
878 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
879 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
880 return false;
881 }
882
883 if (!_summary_data.initialize(mr)) {
884 vm_shutdown_during_initialization(
885 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
886 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
887 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
888 return false;
889 }
890
891 return true;
892 }
893
894 void PSParallelCompact::initialize_space_info()
895 {
896 memset(&_space_info, 0, sizeof(_space_info));
897
898 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
899 PSYoungGen* young_gen = heap->young_gen();
900
901 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
902 _space_info[eden_space_id].set_space(young_gen->eden_space());
903 _space_info[from_space_id].set_space(young_gen->from_space());
904 _space_info[to_space_id].set_space(young_gen->to_space());
905
906 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
907 }
908
909 void PSParallelCompact::initialize_dead_wood_limiter()
910 {
911 const size_t max = 100;
912 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
913 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
914 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
915 DEBUG_ONLY(_dwl_initialized = true;)
916 _dwl_adjustment = normal_distribution(1.0);
917 }
918
919 void
920 PSParallelCompact::clear_data_covering_space(SpaceId id)
921 {
922 // At this point, top is the value before GC, new_top() is the value that will
923 // be set at the end of GC. The marking bitmap is cleared to top; nothing
924 // should be marked above top. The summary data is cleared to the larger of
925 // top & new_top.
926 MutableSpace* const space = _space_info[id].space();
927 HeapWord* const bot = space->bottom();
928 HeapWord* const top = space->top();
929 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
930
931 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
932 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
933 _mark_bitmap.clear_range(beg_bit, end_bit);
934
935 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
936 const size_t end_region =
937 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
938 _summary_data.clear_range(beg_region, end_region);
939
940 // Clear the data used to 'split' regions.
941 SplitInfo& split_info = _space_info[id].split_info();
942 if (split_info.is_valid()) {
943 split_info.clear();
944 }
945 DEBUG_ONLY(split_info.verify_clear();)
946 }
947
948 void PSParallelCompact::pre_compact()
949 {
950 // Update the from & to space pointers in space_info, since they are swapped
951 // at each young gen gc. Do the update unconditionally (even though a
952 // promotion failure does not swap spaces) because an unknown number of young
953 // collections will have swapped the spaces an unknown number of times.
954 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
955 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
956 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
957 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
958
959 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
960 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
961
962 // Increment the invocation count
963 heap->increment_total_collections(true);
964
965 // We need to track unique mark sweep invocations as well.
966 _total_invocations++;
967
968 heap->print_heap_before_gc();
969 heap->trace_heap_before_gc(&_gc_tracer);
970
971 // Fill in TLABs
972 heap->accumulate_statistics_all_tlabs();
973 heap->ensure_parsability(true); // retire TLABs
974
975 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
976 HandleMark hm; // Discard invalid handles created during verification
977 Universe::verify("Before GC");
978 }
979
980 // Verify object start arrays
981 if (VerifyObjectStartArray &&
982 VerifyBeforeGC) {
983 heap->old_gen()->verify_object_start_array();
984 }
985
986 DEBUG_ONLY(mark_bitmap()->verify_clear();)
987 DEBUG_ONLY(summary_data().verify_clear();)
988
989 // Have worker threads release resources the next time they run a task.
990 gc_task_manager()->release_all_resources();
991
992 ParCompactionManager::reset_all_bitmap_query_caches();
993 }
994
995 void PSParallelCompact::post_compact()
996 {
997 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
998
999 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1000 // Clear the marking bitmap, summary data and split info.
1001 clear_data_covering_space(SpaceId(id));
1002 // Update top(). Must be done after clearing the bitmap and summary data.
1003 _space_info[id].publish_new_top();
1004 }
1005
1006 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1007 MutableSpace* const from_space = _space_info[from_space_id].space();
1008 MutableSpace* const to_space = _space_info[to_space_id].space();
1009
1010 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1011 bool eden_empty = eden_space->is_empty();
1012 if (!eden_empty) {
1013 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1014 heap->young_gen(), heap->old_gen());
1015 }
1016
1017 // Update heap occupancy information which is used as input to the soft ref
1018 // clearing policy at the next gc.
1019 Universe::update_heap_info_at_gc();
1020
1021 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1022 to_space->is_empty();
1023
1024 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set());
1025 MemRegion old_mr = heap->old_gen()->reserved();
1026 if (young_gen_empty) {
1027 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1028 } else {
1029 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1030 }
1031
1032 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1033 ClassLoaderDataGraph::purge();
1034 MetaspaceAux::verify_metrics();
1035
1036 CodeCache::gc_epilogue();
1037 JvmtiExport::gc_epilogue();
1038
1039 #if defined(COMPILER2) || INCLUDE_JVMCI
1040 DerivedPointerTable::update_pointers();
1041 #endif
1042
1043 ref_processor()->enqueue_discovered_references(NULL);
1044
1045 if (ZapUnusedHeapArea) {
1046 heap->gen_mangle_unused_area();
1047 }
1048
1049 // Update time of last GC
1050 reset_millis_since_last_gc();
1051 }
1052
1053 HeapWord*
1054 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1055 bool maximum_compaction)
1056 {
1057 const size_t region_size = ParallelCompactData::RegionSize;
1058 const ParallelCompactData& sd = summary_data();
1059
1060 const MutableSpace* const space = _space_info[id].space();
1061 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1062 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1063 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1064
1065 // Skip full regions at the beginning of the space--they are necessarily part
1066 // of the dense prefix.
1067 size_t full_count = 0;
1068 const RegionData* cp;
1069 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1070 ++full_count;
1071 }
1072
1073 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1074 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1075 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1076 if (maximum_compaction || cp == end_cp || interval_ended) {
1077 _maximum_compaction_gc_num = total_invocations();
1078 return sd.region_to_addr(cp);
1079 }
1080
1081 HeapWord* const new_top = _space_info[id].new_top();
1082 const size_t space_live = pointer_delta(new_top, space->bottom());
1083 const size_t space_used = space->used_in_words();
1084 const size_t space_capacity = space->capacity_in_words();
1085
1086 const double cur_density = double(space_live) / space_capacity;
1087 const double deadwood_density =
1088 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1089 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1090
1091 if (TraceParallelOldGCDensePrefix) {
1092 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1093 cur_density, deadwood_density, deadwood_goal);
1094 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1095 "space_cap=" SIZE_FORMAT,
1096 space_live, space_used,
1097 space_capacity);
1098 }
1099
1100 // XXX - Use binary search?
1101 HeapWord* dense_prefix = sd.region_to_addr(cp);
1102 const RegionData* full_cp = cp;
1103 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1104 while (cp < end_cp) {
1105 HeapWord* region_destination = cp->destination();
1106 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1107 if (TraceParallelOldGCDensePrefix && Verbose) {
1108 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1109 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1110 sd.region(cp), p2i(region_destination),
1111 p2i(dense_prefix), cur_deadwood);
1112 }
1113
1114 if (cur_deadwood >= deadwood_goal) {
1115 // Found the region that has the correct amount of deadwood to the left.
1116 // This typically occurs after crossing a fairly sparse set of regions, so
1117 // iterate backwards over those sparse regions, looking for the region
1118 // that has the lowest density of live objects 'to the right.'
1119 size_t space_to_left = sd.region(cp) * region_size;
1120 size_t live_to_left = space_to_left - cur_deadwood;
1121 size_t space_to_right = space_capacity - space_to_left;
1122 size_t live_to_right = space_live - live_to_left;
1123 double density_to_right = double(live_to_right) / space_to_right;
1124 while (cp > full_cp) {
1125 --cp;
1126 const size_t prev_region_live_to_right = live_to_right -
1127 cp->data_size();
1128 const size_t prev_region_space_to_right = space_to_right + region_size;
1129 double prev_region_density_to_right =
1130 double(prev_region_live_to_right) / prev_region_space_to_right;
1131 if (density_to_right <= prev_region_density_to_right) {
1132 return dense_prefix;
1133 }
1134 if (TraceParallelOldGCDensePrefix && Verbose) {
1135 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1136 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1137 prev_region_density_to_right);
1138 }
1139 dense_prefix -= region_size;
1140 live_to_right = prev_region_live_to_right;
1141 space_to_right = prev_region_space_to_right;
1142 density_to_right = prev_region_density_to_right;
1143 }
1144 return dense_prefix;
1145 }
1146
1147 dense_prefix += region_size;
1148 ++cp;
1149 }
1150
1151 return dense_prefix;
1152 }
1153
1154 #ifndef PRODUCT
1155 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1156 const SpaceId id,
1157 const bool maximum_compaction,
1158 HeapWord* const addr)
1159 {
1160 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1161 RegionData* const cp = summary_data().region(region_idx);
1162 const MutableSpace* const space = _space_info[id].space();
1163 HeapWord* const new_top = _space_info[id].new_top();
1164
1165 const size_t space_live = pointer_delta(new_top, space->bottom());
1166 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1167 const size_t space_cap = space->capacity_in_words();
1168 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1169 const size_t live_to_right = new_top - cp->destination();
1170 const size_t dead_to_right = space->top() - addr - live_to_right;
1171
1172 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1173 "spl=" SIZE_FORMAT " "
1174 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1175 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1176 " ratio=%10.8f",
1177 algorithm, p2i(addr), region_idx,
1178 space_live,
1179 dead_to_left, dead_to_left_pct,
1180 dead_to_right, live_to_right,
1181 double(dead_to_right) / live_to_right);
1182 }
1183 #endif // #ifndef PRODUCT
1184
1185 // Return a fraction indicating how much of the generation can be treated as
1186 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1187 // based on the density of live objects in the generation to determine a limit,
1188 // which is then adjusted so the return value is min_percent when the density is
1189 // 1.
1190 //
1191 // The following table shows some return values for a different values of the
1192 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1193 // min_percent is 1.
1194 //
1195 // fraction allowed as dead wood
1196 // -----------------------------------------------------------------
1197 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1198 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1199 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1200 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1201 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1202 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1203 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1204 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1205 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1206 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1207 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1208 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1209 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1210 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1211 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1212 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1213 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1214 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1215 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1216 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1217 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1218 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1219 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1220
1221 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1222 {
1223 assert(_dwl_initialized, "uninitialized");
1224
1225 // The raw limit is the value of the normal distribution at x = density.
1226 const double raw_limit = normal_distribution(density);
1227
1228 // Adjust the raw limit so it becomes the minimum when the density is 1.
1229 //
1230 // First subtract the adjustment value (which is simply the precomputed value
1231 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1232 // Then add the minimum value, so the minimum is returned when the density is
1233 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1234 const double min = double(min_percent) / 100.0;
1235 const double limit = raw_limit - _dwl_adjustment + min;
1236 return MAX2(limit, 0.0);
1237 }
1238
1239 ParallelCompactData::RegionData*
1240 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1241 const RegionData* end)
1242 {
1243 const size_t region_size = ParallelCompactData::RegionSize;
1244 ParallelCompactData& sd = summary_data();
1245 size_t left = sd.region(beg);
1246 size_t right = end > beg ? sd.region(end) - 1 : left;
1247
1248 // Binary search.
1249 while (left < right) {
1250 // Equivalent to (left + right) / 2, but does not overflow.
1251 const size_t middle = left + (right - left) / 2;
1252 RegionData* const middle_ptr = sd.region(middle);
1253 HeapWord* const dest = middle_ptr->destination();
1254 HeapWord* const addr = sd.region_to_addr(middle);
1255 assert(dest != NULL, "sanity");
1256 assert(dest <= addr, "must move left");
1257
1258 if (middle > left && dest < addr) {
1259 right = middle - 1;
1260 } else if (middle < right && middle_ptr->data_size() == region_size) {
1261 left = middle + 1;
1262 } else {
1263 return middle_ptr;
1264 }
1265 }
1266 return sd.region(left);
1267 }
1268
1269 ParallelCompactData::RegionData*
1270 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1271 const RegionData* end,
1272 size_t dead_words)
1273 {
1274 ParallelCompactData& sd = summary_data();
1275 size_t left = sd.region(beg);
1276 size_t right = end > beg ? sd.region(end) - 1 : left;
1277
1278 // Binary search.
1279 while (left < right) {
1280 // Equivalent to (left + right) / 2, but does not overflow.
1281 const size_t middle = left + (right - left) / 2;
1282 RegionData* const middle_ptr = sd.region(middle);
1283 HeapWord* const dest = middle_ptr->destination();
1284 HeapWord* const addr = sd.region_to_addr(middle);
1285 assert(dest != NULL, "sanity");
1286 assert(dest <= addr, "must move left");
1287
1288 const size_t dead_to_left = pointer_delta(addr, dest);
1289 if (middle > left && dead_to_left > dead_words) {
1290 right = middle - 1;
1291 } else if (middle < right && dead_to_left < dead_words) {
1292 left = middle + 1;
1293 } else {
1294 return middle_ptr;
1295 }
1296 }
1297 return sd.region(left);
1298 }
1299
1300 // The result is valid during the summary phase, after the initial summarization
1301 // of each space into itself, and before final summarization.
1302 inline double
1303 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1304 HeapWord* const bottom,
1305 HeapWord* const top,
1306 HeapWord* const new_top)
1307 {
1308 ParallelCompactData& sd = summary_data();
1309
1310 assert(cp != NULL, "sanity");
1311 assert(bottom != NULL, "sanity");
1312 assert(top != NULL, "sanity");
1313 assert(new_top != NULL, "sanity");
1314 assert(top >= new_top, "summary data problem?");
1315 assert(new_top > bottom, "space is empty; should not be here");
1316 assert(new_top >= cp->destination(), "sanity");
1317 assert(top >= sd.region_to_addr(cp), "sanity");
1318
1319 HeapWord* const destination = cp->destination();
1320 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1321 const size_t compacted_region_live = pointer_delta(new_top, destination);
1322 const size_t compacted_region_used = pointer_delta(top,
1323 sd.region_to_addr(cp));
1324 const size_t reclaimable = compacted_region_used - compacted_region_live;
1325
1326 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1327 return double(reclaimable) / divisor;
1328 }
1329
1330 // Return the address of the end of the dense prefix, a.k.a. the start of the
1331 // compacted region. The address is always on a region boundary.
1332 //
1333 // Completely full regions at the left are skipped, since no compaction can
1334 // occur in those regions. Then the maximum amount of dead wood to allow is
1335 // computed, based on the density (amount live / capacity) of the generation;
1336 // the region with approximately that amount of dead space to the left is
1337 // identified as the limit region. Regions between the last completely full
1338 // region and the limit region are scanned and the one that has the best
1339 // (maximum) reclaimed_ratio() is selected.
1340 HeapWord*
1341 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1342 bool maximum_compaction)
1343 {
1344 const size_t region_size = ParallelCompactData::RegionSize;
1345 const ParallelCompactData& sd = summary_data();
1346
1347 const MutableSpace* const space = _space_info[id].space();
1348 HeapWord* const top = space->top();
1349 HeapWord* const top_aligned_up = sd.region_align_up(top);
1350 HeapWord* const new_top = _space_info[id].new_top();
1351 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1352 HeapWord* const bottom = space->bottom();
1353 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1354 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1355 const RegionData* const new_top_cp =
1356 sd.addr_to_region_ptr(new_top_aligned_up);
1357
1358 // Skip full regions at the beginning of the space--they are necessarily part
1359 // of the dense prefix.
1360 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1361 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1362 space->is_empty(), "no dead space allowed to the left");
1363 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1364 "region must have dead space");
1365
1366 // The gc number is saved whenever a maximum compaction is done, and used to
1367 // determine when the maximum compaction interval has expired. This avoids
1368 // successive max compactions for different reasons.
1369 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1370 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1371 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1372 total_invocations() == HeapFirstMaximumCompactionCount;
1373 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1374 _maximum_compaction_gc_num = total_invocations();
1375 return sd.region_to_addr(full_cp);
1376 }
1377
1378 const size_t space_live = pointer_delta(new_top, bottom);
1379 const size_t space_used = space->used_in_words();
1380 const size_t space_capacity = space->capacity_in_words();
1381
1382 const double density = double(space_live) / double(space_capacity);
1383 const size_t min_percent_free = MarkSweepDeadRatio;
1384 const double limiter = dead_wood_limiter(density, min_percent_free);
1385 const size_t dead_wood_max = space_used - space_live;
1386 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1387 dead_wood_max);
1388
1389 if (TraceParallelOldGCDensePrefix) {
1390 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1391 "space_cap=" SIZE_FORMAT,
1392 space_live, space_used,
1393 space_capacity);
1394 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1395 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1396 density, min_percent_free, limiter,
1397 dead_wood_max, dead_wood_limit);
1398 }
1399
1400 // Locate the region with the desired amount of dead space to the left.
1401 const RegionData* const limit_cp =
1402 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1403
1404 // Scan from the first region with dead space to the limit region and find the
1405 // one with the best (largest) reclaimed ratio.
1406 double best_ratio = 0.0;
1407 const RegionData* best_cp = full_cp;
1408 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1409 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1410 if (tmp_ratio > best_ratio) {
1411 best_cp = cp;
1412 best_ratio = tmp_ratio;
1413 }
1414 }
1415
1416 return sd.region_to_addr(best_cp);
1417 }
1418
1419 void PSParallelCompact::summarize_spaces_quick()
1420 {
1421 for (unsigned int i = 0; i < last_space_id; ++i) {
1422 const MutableSpace* space = _space_info[i].space();
1423 HeapWord** nta = _space_info[i].new_top_addr();
1424 bool result = _summary_data.summarize(_space_info[i].split_info(),
1425 space->bottom(), space->top(), NULL,
1426 space->bottom(), space->end(), nta);
1427 assert(result, "space must fit into itself");
1428 _space_info[i].set_dense_prefix(space->bottom());
1429 }
1430 }
1431
1432 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1433 {
1434 HeapWord* const dense_prefix_end = dense_prefix(id);
1435 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1436 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1437 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1438 // Only enough dead space is filled so that any remaining dead space to the
1439 // left is larger than the minimum filler object. (The remainder is filled
1440 // during the copy/update phase.)
1441 //
1442 // The size of the dead space to the right of the boundary is not a
1443 // concern, since compaction will be able to use whatever space is
1444 // available.
1445 //
1446 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1447 // surrounds the space to be filled with an object.
1448 //
1449 // In the 32-bit VM, each bit represents two 32-bit words:
1450 // +---+
1451 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1452 // end_bits: ... x x x | 0 | || 0 x x ...
1453 // +---+
1454 //
1455 // In the 64-bit VM, each bit represents one 64-bit word:
1456 // +------------+
1457 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1458 // end_bits: ... x x 1 | 0 || 0 | x x ...
1459 // +------------+
1460 // +-------+
1461 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1462 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1463 // +-------+
1464 // +-----------+
1465 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1466 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1467 // +-----------+
1468 // +-------+
1469 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1470 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1471 // +-------+
1472
1473 // Initially assume case a, c or e will apply.
1474 size_t obj_len = CollectedHeap::min_fill_size();
1475 HeapWord* obj_beg = dense_prefix_end - obj_len;
1476
1477 #ifdef _LP64
1478 if (MinObjAlignment > 1) { // object alignment > heap word size
1479 // Cases a, c or e.
1480 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1481 // Case b above.
1482 obj_beg = dense_prefix_end - 1;
1483 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1484 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1485 // Case d above.
1486 obj_beg = dense_prefix_end - 3;
1487 obj_len = 3;
1488 }
1489 #endif // #ifdef _LP64
1490
1491 CollectedHeap::fill_with_object(obj_beg, obj_len);
1492 _mark_bitmap.mark_obj(obj_beg, obj_len);
1493 _summary_data.add_obj(obj_beg, obj_len);
1494 assert(start_array(id) != NULL, "sanity");
1495 start_array(id)->allocate_block(obj_beg);
1496 }
1497 }
1498
1499 void
1500 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1501 {
1502 assert(id < last_space_id, "id out of range");
1503 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1504 "should have been reset in summarize_spaces_quick()");
1505
1506 const MutableSpace* space = _space_info[id].space();
1507 if (_space_info[id].new_top() != space->bottom()) {
1508 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1509 _space_info[id].set_dense_prefix(dense_prefix_end);
1510
1511 #ifndef PRODUCT
1512 if (TraceParallelOldGCDensePrefix) {
1513 print_dense_prefix_stats("ratio", id, maximum_compaction,
1514 dense_prefix_end);
1515 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1516 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1517 }
1518 #endif // #ifndef PRODUCT
1519
1520 // Recompute the summary data, taking into account the dense prefix. If
1521 // every last byte will be reclaimed, then the existing summary data which
1522 // compacts everything can be left in place.
1523 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1524 // If dead space crosses the dense prefix boundary, it is (at least
1525 // partially) filled with a dummy object, marked live and added to the
1526 // summary data. This simplifies the copy/update phase and must be done
1527 // before the final locations of objects are determined, to prevent
1528 // leaving a fragment of dead space that is too small to fill.
1529 fill_dense_prefix_end(id);
1530
1531 // Compute the destination of each Region, and thus each object.
1532 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1533 _summary_data.summarize(_space_info[id].split_info(),
1534 dense_prefix_end, space->top(), NULL,
1535 dense_prefix_end, space->end(),
1536 _space_info[id].new_top_addr());
1537 }
1538 }
1539
1540 if (log_develop_is_enabled(Trace, gc, compaction)) {
1541 const size_t region_size = ParallelCompactData::RegionSize;
1542 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1543 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1544 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1545 HeapWord* const new_top = _space_info[id].new_top();
1546 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1547 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1548 log_develop_trace(gc, compaction)(
1549 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1550 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1551 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1552 id, space->capacity_in_words(), p2i(dense_prefix_end),
1553 dp_region, dp_words / region_size,
1554 cr_words / region_size, p2i(new_top));
1555 }
1556 }
1557
1558 #ifndef PRODUCT
1559 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1560 HeapWord* dst_beg, HeapWord* dst_end,
1561 SpaceId src_space_id,
1562 HeapWord* src_beg, HeapWord* src_end)
1563 {
1564 log_develop_trace(gc, compaction)(
1565 "Summarizing %d [%s] into %d [%s]: "
1566 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1567 SIZE_FORMAT "-" SIZE_FORMAT " "
1568 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1569 SIZE_FORMAT "-" SIZE_FORMAT,
1570 src_space_id, space_names[src_space_id],
1571 dst_space_id, space_names[dst_space_id],
1572 p2i(src_beg), p2i(src_end),
1573 _summary_data.addr_to_region_idx(src_beg),
1574 _summary_data.addr_to_region_idx(src_end),
1575 p2i(dst_beg), p2i(dst_end),
1576 _summary_data.addr_to_region_idx(dst_beg),
1577 _summary_data.addr_to_region_idx(dst_end));
1578 }
1579 #endif // #ifndef PRODUCT
1580
1581 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1582 bool maximum_compaction)
1583 {
1584 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1585
1586 #ifdef ASSERT
1587 if (TraceParallelOldGCMarkingPhase) {
1588 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1589 "add_obj_bytes=" SIZE_FORMAT,
1590 add_obj_count, add_obj_size * HeapWordSize);
1591 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1592 "mark_bitmap_bytes=" SIZE_FORMAT,
1593 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1594 }
1595 #endif // #ifdef ASSERT
1596
1597 // Quick summarization of each space into itself, to see how much is live.
1598 summarize_spaces_quick();
1599
1600 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1601 NOT_PRODUCT(print_region_ranges());
1602 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1603
1604 // The amount of live data that will end up in old space (assuming it fits).
1605 size_t old_space_total_live = 0;
1606 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1607 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1608 _space_info[id].space()->bottom());
1609 }
1610
1611 MutableSpace* const old_space = _space_info[old_space_id].space();
1612 const size_t old_capacity = old_space->capacity_in_words();
1613 if (old_space_total_live > old_capacity) {
1614 // XXX - should also try to expand
1615 maximum_compaction = true;
1616 }
1617
1618 // Old generations.
1619 summarize_space(old_space_id, maximum_compaction);
1620
1621 // Summarize the remaining spaces in the young gen. The initial target space
1622 // is the old gen. If a space does not fit entirely into the target, then the
1623 // remainder is compacted into the space itself and that space becomes the new
1624 // target.
1625 SpaceId dst_space_id = old_space_id;
1626 HeapWord* dst_space_end = old_space->end();
1627 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1628 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1629 const MutableSpace* space = _space_info[id].space();
1630 const size_t live = pointer_delta(_space_info[id].new_top(),
1631 space->bottom());
1632 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1633
1634 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1635 SpaceId(id), space->bottom(), space->top());)
1636 if (live > 0 && live <= available) {
1637 // All the live data will fit.
1638 bool done = _summary_data.summarize(_space_info[id].split_info(),
1639 space->bottom(), space->top(),
1640 NULL,
1641 *new_top_addr, dst_space_end,
1642 new_top_addr);
1643 assert(done, "space must fit into old gen");
1644
1645 // Reset the new_top value for the space.
1646 _space_info[id].set_new_top(space->bottom());
1647 } else if (live > 0) {
1648 // Attempt to fit part of the source space into the target space.
1649 HeapWord* next_src_addr = NULL;
1650 bool done = _summary_data.summarize(_space_info[id].split_info(),
1651 space->bottom(), space->top(),
1652 &next_src_addr,
1653 *new_top_addr, dst_space_end,
1654 new_top_addr);
1655 assert(!done, "space should not fit into old gen");
1656 assert(next_src_addr != NULL, "sanity");
1657
1658 // The source space becomes the new target, so the remainder is compacted
1659 // within the space itself.
1660 dst_space_id = SpaceId(id);
1661 dst_space_end = space->end();
1662 new_top_addr = _space_info[id].new_top_addr();
1663 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1664 space->bottom(), dst_space_end,
1665 SpaceId(id), next_src_addr, space->top());)
1666 done = _summary_data.summarize(_space_info[id].split_info(),
1667 next_src_addr, space->top(),
1668 NULL,
1669 space->bottom(), dst_space_end,
1670 new_top_addr);
1671 assert(done, "space must fit when compacted into itself");
1672 assert(*new_top_addr <= space->top(), "usage should not grow");
1673 }
1674 }
1675
1676 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1677 NOT_PRODUCT(print_region_ranges());
1678 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1679 }
1680
1681 // This method should contain all heap-specific policy for invoking a full
1682 // collection. invoke_no_policy() will only attempt to compact the heap; it
1683 // will do nothing further. If we need to bail out for policy reasons, scavenge
1684 // before full gc, or any other specialized behavior, it needs to be added here.
1685 //
1686 // Note that this method should only be called from the vm_thread while at a
1687 // safepoint.
1688 //
1689 // Note that the all_soft_refs_clear flag in the collector policy
1690 // may be true because this method can be called without intervening
1691 // activity. For example when the heap space is tight and full measure
1692 // are being taken to free space.
1693 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1694 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1695 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1696 "should be in vm thread");
1697
1698 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1699 GCCause::Cause gc_cause = heap->gc_cause();
1700 assert(!heap->is_gc_active(), "not reentrant");
1701
1702 PSAdaptiveSizePolicy* policy = heap->size_policy();
1703 IsGCActiveMark mark;
1704
1705 if (ScavengeBeforeFullGC) {
1706 PSScavenge::invoke_no_policy();
1707 }
1708
1709 const bool clear_all_soft_refs =
1710 heap->collector_policy()->should_clear_all_soft_refs();
1711
1712 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1713 maximum_heap_compaction);
1714 }
1715
1716 // This method contains no policy. You should probably
1717 // be calling invoke() instead.
1718 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1719 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1720 assert(ref_processor() != NULL, "Sanity");
1721
1722 if (GCLocker::check_active_before_gc()) {
1723 return false;
1724 }
1725
1726 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1727
1728 GCIdMark gc_id_mark;
1729 _gc_timer.register_gc_start();
1730 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1731
1732 TimeStamp marking_start;
1733 TimeStamp compaction_start;
1734 TimeStamp collection_exit;
1735
1736 GCCause::Cause gc_cause = heap->gc_cause();
1737 PSYoungGen* young_gen = heap->young_gen();
1738 PSOldGen* old_gen = heap->old_gen();
1739 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1740
1741 // The scope of casr should end after code that can change
1742 // CollectorPolicy::_should_clear_all_soft_refs.
1743 ClearedAllSoftRefs casr(maximum_heap_compaction,
1744 heap->collector_policy());
1745
1746 if (ZapUnusedHeapArea) {
1747 // Save information needed to minimize mangling
1748 heap->record_gen_tops_before_GC();
1749 }
1750
1751 // Make sure data structures are sane, make the heap parsable, and do other
1752 // miscellaneous bookkeeping.
1753 pre_compact();
1754
1755 PreGCValues pre_gc_values(heap);
1756
1757 // Get the compaction manager reserved for the VM thread.
1758 ParCompactionManager* const vmthread_cm =
1759 ParCompactionManager::manager_array(gc_task_manager()->workers());
1760
1761 {
1762 ResourceMark rm;
1763 HandleMark hm;
1764
1765 // Set the number of GC threads to be used in this collection
1766 gc_task_manager()->set_active_gang();
1767 gc_task_manager()->task_idle_workers();
1768
1769 GCTraceCPUTime tcpu;
1770 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1771
1772 heap->pre_full_gc_dump(&_gc_timer);
1773
1774 TraceCollectorStats tcs(counters());
1775 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
1776
1777 if (TraceOldGenTime) accumulated_time()->start();
1778
1779 // Let the size policy know we're starting
1780 size_policy->major_collection_begin();
1781
1782 CodeCache::gc_prologue();
1783
1784 #if defined(COMPILER2) || INCLUDE_JVMCI
1785 DerivedPointerTable::clear();
1786 #endif
1787
1788 ref_processor()->enable_discovery();
1789 ref_processor()->setup_policy(maximum_heap_compaction);
1790
1791 bool marked_for_unloading = false;
1792
1793 marking_start.update();
1794 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
1795
1796 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1797 && GCCause::is_user_requested_gc(gc_cause);
1798 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1799
1800 #if defined(COMPILER2) || INCLUDE_JVMCI
1801 assert(DerivedPointerTable::is_active(), "Sanity");
1802 DerivedPointerTable::set_active(false);
1803 #endif
1804
1805 // adjust_roots() updates Universe::_intArrayKlassObj which is
1806 // needed by the compaction for filling holes in the dense prefix.
1807 adjust_roots(vmthread_cm);
1808
1809 compaction_start.update();
1810 compact();
1811
1812 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1813 // done before resizing.
1814 post_compact();
1815
1816 // Let the size policy know we're done
1817 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1818
1819 if (UseAdaptiveSizePolicy) {
1820 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1821 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1822 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1823
1824 // Don't check if the size_policy is ready here. Let
1825 // the size_policy check that internally.
1826 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1827 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1828 // Swap the survivor spaces if from_space is empty. The
1829 // resize_young_gen() called below is normally used after
1830 // a successful young GC and swapping of survivor spaces;
1831 // otherwise, it will fail to resize the young gen with
1832 // the current implementation.
1833 if (young_gen->from_space()->is_empty()) {
1834 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1835 young_gen->swap_spaces();
1836 }
1837
1838 // Calculate optimal free space amounts
1839 assert(young_gen->max_size() >
1840 young_gen->from_space()->capacity_in_bytes() +
1841 young_gen->to_space()->capacity_in_bytes(),
1842 "Sizes of space in young gen are out-of-bounds");
1843
1844 size_t young_live = young_gen->used_in_bytes();
1845 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1846 size_t old_live = old_gen->used_in_bytes();
1847 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1848 size_t max_old_gen_size = old_gen->max_gen_size();
1849 size_t max_eden_size = young_gen->max_size() -
1850 young_gen->from_space()->capacity_in_bytes() -
1851 young_gen->to_space()->capacity_in_bytes();
1852
1853 // Used for diagnostics
1854 size_policy->clear_generation_free_space_flags();
1855
1856 size_policy->compute_generations_free_space(young_live,
1857 eden_live,
1858 old_live,
1859 cur_eden,
1860 max_old_gen_size,
1861 max_eden_size,
1862 true /* full gc*/);
1863
1864 size_policy->check_gc_overhead_limit(young_live,
1865 eden_live,
1866 max_old_gen_size,
1867 max_eden_size,
1868 true /* full gc*/,
1869 gc_cause,
1870 heap->collector_policy());
1871
1872 size_policy->decay_supplemental_growth(true /* full gc*/);
1873
1874 heap->resize_old_gen(
1875 size_policy->calculated_old_free_size_in_bytes());
1876
1877 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1878 size_policy->calculated_survivor_size_in_bytes());
1879 }
1880
1881 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1882 }
1883
1884 if (UsePerfData) {
1885 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1886 counters->update_counters();
1887 counters->update_old_capacity(old_gen->capacity_in_bytes());
1888 counters->update_young_capacity(young_gen->capacity_in_bytes());
1889 }
1890
1891 heap->resize_all_tlabs();
1892
1893 // Resize the metaspace capacity after a collection
1894 MetaspaceGC::compute_new_size();
1895
1896 if (TraceOldGenTime) {
1897 accumulated_time()->stop();
1898 }
1899
1900 young_gen->print_used_change(pre_gc_values.young_gen_used());
1901 old_gen->print_used_change(pre_gc_values.old_gen_used());
1902 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
1903
1904 // Track memory usage and detect low memory
1905 MemoryService::track_memory_usage();
1906 heap->update_counters();
1907 gc_task_manager()->release_idle_workers();
1908
1909 heap->post_full_gc_dump(&_gc_timer);
1910 }
1911
1912 #ifdef ASSERT
1913 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1914 ParCompactionManager* const cm =
1915 ParCompactionManager::manager_array(int(i));
1916 assert(cm->marking_stack()->is_empty(), "should be empty");
1917 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i);
1918 }
1919 #endif // ASSERT
1920
1921 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1922 HandleMark hm; // Discard invalid handles created during verification
1923 Universe::verify("After GC");
1924 }
1925
1926 // Re-verify object start arrays
1927 if (VerifyObjectStartArray &&
1928 VerifyAfterGC) {
1929 old_gen->verify_object_start_array();
1930 }
1931
1932 if (ZapUnusedHeapArea) {
1933 old_gen->object_space()->check_mangled_unused_area_complete();
1934 }
1935
1936 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1937
1938 collection_exit.update();
1939
1940 heap->print_heap_after_gc();
1941 heap->trace_heap_after_gc(&_gc_tracer);
1942
1943 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1944 marking_start.ticks(), compaction_start.ticks(),
1945 collection_exit.ticks());
1946 gc_task_manager()->print_task_time_stamps();
1947
1948 #ifdef TRACESPINNING
1949 ParallelTaskTerminator::print_termination_counts();
1950 #endif
1951
1952 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1953
1954 _gc_timer.register_gc_end();
1955
1956 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1957 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1958
1959 return true;
1960 }
1961
1962 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1963 PSYoungGen* young_gen,
1964 PSOldGen* old_gen) {
1965 MutableSpace* const eden_space = young_gen->eden_space();
1966 assert(!eden_space->is_empty(), "eden must be non-empty");
1967 assert(young_gen->virtual_space()->alignment() ==
1968 old_gen->virtual_space()->alignment(), "alignments do not match");
1969
1970 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
1971 return false;
1972 }
1973
1974 // Both generations must be completely committed.
1975 if (young_gen->virtual_space()->uncommitted_size() != 0) {
1976 return false;
1977 }
1978 if (old_gen->virtual_space()->uncommitted_size() != 0) {
1979 return false;
1980 }
1981
1982 // Figure out how much to take from eden. Include the average amount promoted
1983 // in the total; otherwise the next young gen GC will simply bail out to a
1984 // full GC.
1985 const size_t alignment = old_gen->virtual_space()->alignment();
1986 const size_t eden_used = eden_space->used_in_bytes();
1987 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
1988 const size_t absorb_size = align_up(eden_used + promoted, alignment);
1989 const size_t eden_capacity = eden_space->capacity_in_bytes();
1990
1991 if (absorb_size >= eden_capacity) {
1992 return false; // Must leave some space in eden.
1993 }
1994
1995 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
1996 if (new_young_size < young_gen->min_gen_size()) {
1997 return false; // Respect young gen minimum size.
1998 }
1999
2000 log_trace(heap, ergo)(" absorbing " SIZE_FORMAT "K: "
2001 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2002 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2003 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2004 absorb_size / K,
2005 eden_capacity / K, (eden_capacity - absorb_size) / K,
2006 young_gen->from_space()->used_in_bytes() / K,
2007 young_gen->to_space()->used_in_bytes() / K,
2008 young_gen->capacity_in_bytes() / K, new_young_size / K);
2009
2010 // Fill the unused part of the old gen.
2011 MutableSpace* const old_space = old_gen->object_space();
2012 HeapWord* const unused_start = old_space->top();
2013 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2014
2015 if (unused_words > 0) {
2016 if (unused_words < CollectedHeap::min_fill_size()) {
2017 return false; // If the old gen cannot be filled, must give up.
2018 }
2019 CollectedHeap::fill_with_objects(unused_start, unused_words);
2020 }
2021
2022 // Take the live data from eden and set both top and end in the old gen to
2023 // eden top. (Need to set end because reset_after_change() mangles the region
2024 // from end to virtual_space->high() in debug builds).
2025 HeapWord* const new_top = eden_space->top();
2026 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2027 absorb_size);
2028 young_gen->reset_after_change();
2029 old_space->set_top(new_top);
2030 old_space->set_end(new_top);
2031 old_gen->reset_after_change();
2032
2033 // Update the object start array for the filler object and the data from eden.
2034 ObjectStartArray* const start_array = old_gen->start_array();
2035 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2036 start_array->allocate_block(p);
2037 }
2038
2039 // Could update the promoted average here, but it is not typically updated at
2040 // full GCs and the value to use is unclear. Something like
2041 //
2042 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2043
2044 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2045 return true;
2046 }
2047
2048 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2049 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2050 "shouldn't return NULL");
2051 return ParallelScavengeHeap::gc_task_manager();
2052 }
2053
2054 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2055 bool maximum_heap_compaction,
2056 ParallelOldTracer *gc_tracer) {
2057 // Recursively traverse all live objects and mark them
2058 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2059
2060 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2061 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2062 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2063 TaskQueueSetSuper* qset = ParCompactionManager::stack_array();
2064 ParallelTaskTerminator terminator(active_gc_threads, qset);
2065
2066 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm);
2067 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2068
2069 // Need new claim bits before marking starts.
2070 ClassLoaderDataGraph::clear_claimed_marks();
2071
2072 {
2073 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2074
2075 ParallelScavengeHeap::ParStrongRootsScope psrs;
2076
2077 GCTaskQueue* q = GCTaskQueue::create();
2078
2079 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2080 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2081 // We scan the thread roots in parallel
2082 Threads::create_thread_roots_marking_tasks(q);
2083 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2084 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2085 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2086 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2087 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2088 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2089 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2090
2091 if (active_gc_threads > 1) {
2092 for (uint j = 0; j < active_gc_threads; j++) {
2093 q->enqueue(new StealMarkingTask(&terminator));
2094 }
2095 }
2096
2097 gc_task_manager()->execute_and_wait(q);
2098 }
2099
2100 // Process reference objects found during marking
2101 {
2102 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2103
2104 ReferenceProcessorStats stats;
2105 if (ref_processor()->processing_is_mt()) {
2106 RefProcTaskExecutor task_executor;
2107 stats = ref_processor()->process_discovered_references(
2108 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2109 &task_executor, &_gc_timer);
2110 } else {
2111 stats = ref_processor()->process_discovered_references(
2112 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2113 &_gc_timer);
2114 }
2115
2116 gc_tracer->report_gc_reference_stats(stats);
2117 }
2118
2119 // This is the point where the entire marking should have completed.
2120 assert(cm->marking_stacks_empty(), "Marking should have completed");
2121
2122 {
2123 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2124
2125 // Follow system dictionary roots and unload classes.
2126 bool purged_class = SystemDictionary::do_unloading(is_alive_closure(), &_gc_timer);
2127
2128 // Unload nmethods.
2129 CodeCache::do_unloading(is_alive_closure(), purged_class);
2130
2131 // Prune dead klasses from subklass/sibling/implementor lists.
2132 Klass::clean_weak_klass_links(is_alive_closure());
2133 }
2134
2135 {
2136 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer);
2137 // Delete entries for dead interned strings.
2138 StringTable::unlink(is_alive_closure());
2139 }
2140
2141 {
2142 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer);
2143 // Clean up unreferenced symbols in symbol table.
2144 SymbolTable::unlink();
2145 }
2146
2147 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2148 }
2149
2150 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2151 // Adjust the pointers to reflect the new locations
2152 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2153
2154 // Need new claim bits when tracing through and adjusting pointers.
2155 ClassLoaderDataGraph::clear_claimed_marks();
2156
2157 PSParallelCompact::AdjustPointerClosure oop_closure(cm);
2158 PSParallelCompact::AdjustKlassClosure klass_closure(cm);
2159
2160 // General strong roots.
2161 Universe::oops_do(&oop_closure);
2162 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2163 Threads::oops_do(&oop_closure, NULL);
2164 ObjectSynchronizer::oops_do(&oop_closure);
2165 FlatProfiler::oops_do(&oop_closure);
2166 Management::oops_do(&oop_closure);
2167 JvmtiExport::oops_do(&oop_closure);
2168 SystemDictionary::oops_do(&oop_closure);
2169 ClassLoaderDataGraph::oops_do(&oop_closure, &klass_closure, true);
2170
2171 // Now adjust pointers in remaining weak roots. (All of which should
2172 // have been cleared if they pointed to non-surviving objects.)
2173 // Global (weak) JNI handles
2174 JNIHandles::weak_oops_do(&oop_closure);
2175
2176 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2177 CodeCache::blobs_do(&adjust_from_blobs);
2178 AOTLoader::oops_do(&oop_closure);
2179 StringTable::oops_do(&oop_closure);
2180 ref_processor()->weak_oops_do(&oop_closure);
2181 // Roots were visited so references into the young gen in roots
2182 // may have been scanned. Process them also.
2183 // Should the reference processor have a span that excludes
2184 // young gen objects?
2185 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2186 }
2187
2188 // Helper class to print 8 region numbers per line and then print the total at the end.
2189 class FillableRegionLogger : public StackObj {
2190 private:
2191 Log(gc, compaction) log;
2192 static const int LineLength = 8;
2193 size_t _regions[LineLength];
2194 int _next_index;
2195 bool _enabled;
2196 size_t _total_regions;
2197 public:
2198 FillableRegionLogger() : _next_index(0), _total_regions(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)) { }
2199 ~FillableRegionLogger() {
2200 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2201 }
2202
2203 void print_line() {
2204 if (!_enabled || _next_index == 0) {
2205 return;
2206 }
2207 FormatBuffer<> line("Fillable: ");
2208 for (int i = 0; i < _next_index; i++) {
2209 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2210 }
2211 log.trace("%s", line.buffer());
2212 _next_index = 0;
2213 }
2214
2215 void handle(size_t region) {
2216 if (!_enabled) {
2217 return;
2218 }
2219 _regions[_next_index++] = region;
2220 if (_next_index == LineLength) {
2221 print_line();
2222 }
2223 _total_regions++;
2224 }
2225 };
2226
2227 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q,
2228 uint parallel_gc_threads)
2229 {
2230 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2231
2232 // Find the threads that are active
2233 unsigned int which = 0;
2234
2235 // Find all regions that are available (can be filled immediately) and
2236 // distribute them to the thread stacks. The iteration is done in reverse
2237 // order (high to low) so the regions will be removed in ascending order.
2238
2239 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2240
2241 which = 0;
2242 // id + 1 is used to test termination so unsigned can
2243 // be used with an old_space_id == 0.
2244 FillableRegionLogger region_logger;
2245 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2246 SpaceInfo* const space_info = _space_info + id;
2247 MutableSpace* const space = space_info->space();
2248 HeapWord* const new_top = space_info->new_top();
2249
2250 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2251 const size_t end_region =
2252 sd.addr_to_region_idx(sd.region_align_up(new_top));
2253
2254 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2255 if (sd.region(cur)->claim_unsafe()) {
2256 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2257 cm->region_stack()->push(cur);
2258 region_logger.handle(cur);
2259 // Assign regions to tasks in round-robin fashion.
2260 if (++which == parallel_gc_threads) {
2261 which = 0;
2262 }
2263 }
2264 }
2265 region_logger.print_line();
2266 }
2267 }
2268
2269 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2270
2271 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2272 uint parallel_gc_threads) {
2273 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2274
2275 ParallelCompactData& sd = PSParallelCompact::summary_data();
2276
2277 // Iterate over all the spaces adding tasks for updating
2278 // regions in the dense prefix. Assume that 1 gc thread
2279 // will work on opening the gaps and the remaining gc threads
2280 // will work on the dense prefix.
2281 unsigned int space_id;
2282 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2283 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2284 const MutableSpace* const space = _space_info[space_id].space();
2285
2286 if (dense_prefix_end == space->bottom()) {
2287 // There is no dense prefix for this space.
2288 continue;
2289 }
2290
2291 // The dense prefix is before this region.
2292 size_t region_index_end_dense_prefix =
2293 sd.addr_to_region_idx(dense_prefix_end);
2294 RegionData* const dense_prefix_cp =
2295 sd.region(region_index_end_dense_prefix);
2296 assert(dense_prefix_end == space->end() ||
2297 dense_prefix_cp->available() ||
2298 dense_prefix_cp->claimed(),
2299 "The region after the dense prefix should always be ready to fill");
2300
2301 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2302
2303 // Is there dense prefix work?
2304 size_t total_dense_prefix_regions =
2305 region_index_end_dense_prefix - region_index_start;
2306 // How many regions of the dense prefix should be given to
2307 // each thread?
2308 if (total_dense_prefix_regions > 0) {
2309 uint tasks_for_dense_prefix = 1;
2310 if (total_dense_prefix_regions <=
2311 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2312 // Don't over partition. This assumes that
2313 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2314 // so there are not many regions to process.
2315 tasks_for_dense_prefix = parallel_gc_threads;
2316 } else {
2317 // Over partition
2318 tasks_for_dense_prefix = parallel_gc_threads *
2319 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2320 }
2321 size_t regions_per_thread = total_dense_prefix_regions /
2322 tasks_for_dense_prefix;
2323 // Give each thread at least 1 region.
2324 if (regions_per_thread == 0) {
2325 regions_per_thread = 1;
2326 }
2327
2328 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2329 if (region_index_start >= region_index_end_dense_prefix) {
2330 break;
2331 }
2332 // region_index_end is not processed
2333 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2334 region_index_end_dense_prefix);
2335 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2336 region_index_start,
2337 region_index_end));
2338 region_index_start = region_index_end;
2339 }
2340 }
2341 // This gets any part of the dense prefix that did not
2342 // fit evenly.
2343 if (region_index_start < region_index_end_dense_prefix) {
2344 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2345 region_index_start,
2346 region_index_end_dense_prefix));
2347 }
2348 }
2349 }
2350
2351 void PSParallelCompact::enqueue_region_stealing_tasks(
2352 GCTaskQueue* q,
2353 ParallelTaskTerminator* terminator_ptr,
2354 uint parallel_gc_threads) {
2355 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer);
2356
2357 // Once a thread has drained it's stack, it should try to steal regions from
2358 // other threads.
2359 for (uint j = 0; j < parallel_gc_threads; j++) {
2360 q->enqueue(new CompactionWithStealingTask(terminator_ptr));
2361 }
2362 }
2363
2364 #ifdef ASSERT
2365 // Write a histogram of the number of times the block table was filled for a
2366 // region.
2367 void PSParallelCompact::write_block_fill_histogram()
2368 {
2369 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2370 return;
2371 }
2372
2373 Log(gc, compaction) log;
2374 ResourceMark rm;
2375 outputStream* out = log.trace_stream();
2376
2377 typedef ParallelCompactData::RegionData rd_t;
2378 ParallelCompactData& sd = summary_data();
2379
2380 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2381 MutableSpace* const spc = _space_info[id].space();
2382 if (spc->bottom() != spc->top()) {
2383 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2384 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2385 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2386
2387 size_t histo[5] = { 0, 0, 0, 0, 0 };
2388 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2389 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2390
2391 for (const rd_t* cur = beg; cur < end; ++cur) {
2392 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2393 }
2394 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2395 for (size_t i = 0; i < histo_len; ++i) {
2396 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2397 histo[i], 100.0 * histo[i] / region_cnt);
2398 }
2399 out->cr();
2400 }
2401 }
2402 }
2403 #endif // #ifdef ASSERT
2404
2405 void PSParallelCompact::compact() {
2406 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2407
2408 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2409 PSOldGen* old_gen = heap->old_gen();
2410 old_gen->start_array()->reset();
2411 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2412 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2413 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2414 ParallelTaskTerminator terminator(active_gc_threads, qset);
2415
2416 GCTaskQueue* q = GCTaskQueue::create();
2417 prepare_region_draining_tasks(q, active_gc_threads);
2418 enqueue_dense_prefix_tasks(q, active_gc_threads);
2419 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2420
2421 {
2422 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2423
2424 gc_task_manager()->execute_and_wait(q);
2425
2426 #ifdef ASSERT
2427 // Verify that all regions have been processed before the deferred updates.
2428 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2429 verify_complete(SpaceId(id));
2430 }
2431 #endif
2432 }
2433
2434 {
2435 // Update the deferred objects, if any. Any compaction manager can be used.
2436 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2437 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2438 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2439 update_deferred_objects(cm, SpaceId(id));
2440 }
2441 }
2442
2443 DEBUG_ONLY(write_block_fill_histogram());
2444 }
2445
2446 #ifdef ASSERT
2447 void PSParallelCompact::verify_complete(SpaceId space_id) {
2448 // All Regions between space bottom() to new_top() should be marked as filled
2449 // and all Regions between new_top() and top() should be available (i.e.,
2450 // should have been emptied).
2451 ParallelCompactData& sd = summary_data();
2452 SpaceInfo si = _space_info[space_id];
2453 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2454 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2455 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2456 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2457 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2458
2459 bool issued_a_warning = false;
2460
2461 size_t cur_region;
2462 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2463 const RegionData* const c = sd.region(cur_region);
2464 if (!c->completed()) {
2465 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2466 cur_region, c->destination_count());
2467 issued_a_warning = true;
2468 }
2469 }
2470
2471 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2472 const RegionData* const c = sd.region(cur_region);
2473 if (!c->available()) {
2474 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2475 cur_region, c->destination_count());
2476 issued_a_warning = true;
2477 }
2478 }
2479
2480 if (issued_a_warning) {
2481 print_region_ranges();
2482 }
2483 }
2484 #endif // #ifdef ASSERT
2485
2486 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2487 _start_array->allocate_block(addr);
2488 compaction_manager()->update_contents(oop(addr));
2489 }
2490
2491 // Update interior oops in the ranges of regions [beg_region, end_region).
2492 void
2493 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2494 SpaceId space_id,
2495 size_t beg_region,
2496 size_t end_region) {
2497 ParallelCompactData& sd = summary_data();
2498 ParMarkBitMap* const mbm = mark_bitmap();
2499
2500 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2501 HeapWord* const end_addr = sd.region_to_addr(end_region);
2502 assert(beg_region <= end_region, "bad region range");
2503 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2504
2505 #ifdef ASSERT
2506 // Claim the regions to avoid triggering an assert when they are marked as
2507 // filled.
2508 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2509 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2510 }
2511 #endif // #ifdef ASSERT
2512
2513 if (beg_addr != space(space_id)->bottom()) {
2514 // Find the first live object or block of dead space that *starts* in this
2515 // range of regions. If a partial object crosses onto the region, skip it;
2516 // it will be marked for 'deferred update' when the object head is
2517 // processed. If dead space crosses onto the region, it is also skipped; it
2518 // will be filled when the prior region is processed. If neither of those
2519 // apply, the first word in the region is the start of a live object or dead
2520 // space.
2521 assert(beg_addr > space(space_id)->bottom(), "sanity");
2522 const RegionData* const cp = sd.region(beg_region);
2523 if (cp->partial_obj_size() != 0) {
2524 beg_addr = sd.partial_obj_end(beg_region);
2525 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2526 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2527 }
2528 }
2529
2530 if (beg_addr < end_addr) {
2531 // A live object or block of dead space starts in this range of Regions.
2532 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2533
2534 // Create closures and iterate.
2535 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2536 FillClosure fill_closure(cm, space_id);
2537 ParMarkBitMap::IterationStatus status;
2538 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2539 dense_prefix_end);
2540 if (status == ParMarkBitMap::incomplete) {
2541 update_closure.do_addr(update_closure.source());
2542 }
2543 }
2544
2545 // Mark the regions as filled.
2546 RegionData* const beg_cp = sd.region(beg_region);
2547 RegionData* const end_cp = sd.region(end_region);
2548 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2549 cp->set_completed();
2550 }
2551 }
2552
2553 // Return the SpaceId for the space containing addr. If addr is not in the
2554 // heap, last_space_id is returned. In debug mode it expects the address to be
2555 // in the heap and asserts such.
2556 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2557 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2558
2559 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2560 if (_space_info[id].space()->contains(addr)) {
2561 return SpaceId(id);
2562 }
2563 }
2564
2565 assert(false, "no space contains the addr");
2566 return last_space_id;
2567 }
2568
2569 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2570 SpaceId id) {
2571 assert(id < last_space_id, "bad space id");
2572
2573 ParallelCompactData& sd = summary_data();
2574 const SpaceInfo* const space_info = _space_info + id;
2575 ObjectStartArray* const start_array = space_info->start_array();
2576
2577 const MutableSpace* const space = space_info->space();
2578 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2579 HeapWord* const beg_addr = space_info->dense_prefix();
2580 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2581
2582 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2583 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2584 const RegionData* cur_region;
2585 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2586 HeapWord* const addr = cur_region->deferred_obj_addr();
2587 if (addr != NULL) {
2588 if (start_array != NULL) {
2589 start_array->allocate_block(addr);
2590 }
2591 cm->update_contents(oop(addr));
2592 assert(oop(addr)->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2593 }
2594 }
2595 }
2596
2597 // Skip over count live words starting from beg, and return the address of the
2598 // next live word. Unless marked, the word corresponding to beg is assumed to
2599 // be dead. Callers must either ensure beg does not correspond to the middle of
2600 // an object, or account for those live words in some other way. Callers must
2601 // also ensure that there are enough live words in the range [beg, end) to skip.
2602 HeapWord*
2603 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2604 {
2605 assert(count > 0, "sanity");
2606
2607 ParMarkBitMap* m = mark_bitmap();
2608 idx_t bits_to_skip = m->words_to_bits(count);
2609 idx_t cur_beg = m->addr_to_bit(beg);
2610 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2611
2612 do {
2613 cur_beg = m->find_obj_beg(cur_beg, search_end);
2614 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2615 const size_t obj_bits = cur_end - cur_beg + 1;
2616 if (obj_bits > bits_to_skip) {
2617 return m->bit_to_addr(cur_beg + bits_to_skip);
2618 }
2619 bits_to_skip -= obj_bits;
2620 cur_beg = cur_end + 1;
2621 } while (bits_to_skip > 0);
2622
2623 // Skipping the desired number of words landed just past the end of an object.
2624 // Find the start of the next object.
2625 cur_beg = m->find_obj_beg(cur_beg, search_end);
2626 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2627 return m->bit_to_addr(cur_beg);
2628 }
2629
2630 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2631 SpaceId src_space_id,
2632 size_t src_region_idx)
2633 {
2634 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2635
2636 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2637 if (split_info.dest_region_addr() == dest_addr) {
2638 // The partial object ending at the split point contains the first word to
2639 // be copied to dest_addr.
2640 return split_info.first_src_addr();
2641 }
2642
2643 const ParallelCompactData& sd = summary_data();
2644 ParMarkBitMap* const bitmap = mark_bitmap();
2645 const size_t RegionSize = ParallelCompactData::RegionSize;
2646
2647 assert(sd.is_region_aligned(dest_addr), "not aligned");
2648 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2649 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2650 HeapWord* const src_region_destination = src_region_ptr->destination();
2651
2652 assert(dest_addr >= src_region_destination, "wrong src region");
2653 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2654
2655 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2656 HeapWord* const src_region_end = src_region_beg + RegionSize;
2657
2658 HeapWord* addr = src_region_beg;
2659 if (dest_addr == src_region_destination) {
2660 // Return the first live word in the source region.
2661 if (partial_obj_size == 0) {
2662 addr = bitmap->find_obj_beg(addr, src_region_end);
2663 assert(addr < src_region_end, "no objects start in src region");
2664 }
2665 return addr;
2666 }
2667
2668 // Must skip some live data.
2669 size_t words_to_skip = dest_addr - src_region_destination;
2670 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2671
2672 if (partial_obj_size >= words_to_skip) {
2673 // All the live words to skip are part of the partial object.
2674 addr += words_to_skip;
2675 if (partial_obj_size == words_to_skip) {
2676 // Find the first live word past the partial object.
2677 addr = bitmap->find_obj_beg(addr, src_region_end);
2678 assert(addr < src_region_end, "wrong src region");
2679 }
2680 return addr;
2681 }
2682
2683 // Skip over the partial object (if any).
2684 if (partial_obj_size != 0) {
2685 words_to_skip -= partial_obj_size;
2686 addr += partial_obj_size;
2687 }
2688
2689 // Skip over live words due to objects that start in the region.
2690 addr = skip_live_words(addr, src_region_end, words_to_skip);
2691 assert(addr < src_region_end, "wrong src region");
2692 return addr;
2693 }
2694
2695 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2696 SpaceId src_space_id,
2697 size_t beg_region,
2698 HeapWord* end_addr)
2699 {
2700 ParallelCompactData& sd = summary_data();
2701
2702 #ifdef ASSERT
2703 MutableSpace* const src_space = _space_info[src_space_id].space();
2704 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2705 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2706 "src_space_id does not match beg_addr");
2707 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2708 "src_space_id does not match end_addr");
2709 #endif // #ifdef ASSERT
2710
2711 RegionData* const beg = sd.region(beg_region);
2712 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2713
2714 // Regions up to new_top() are enqueued if they become available.
2715 HeapWord* const new_top = _space_info[src_space_id].new_top();
2716 RegionData* const enqueue_end =
2717 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2718
2719 for (RegionData* cur = beg; cur < end; ++cur) {
2720 assert(cur->data_size() > 0, "region must have live data");
2721 cur->decrement_destination_count();
2722 if (cur < enqueue_end && cur->available() && cur->claim()) {
2723 cm->push_region(sd.region(cur));
2724 }
2725 }
2726 }
2727
2728 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2729 SpaceId& src_space_id,
2730 HeapWord*& src_space_top,
2731 HeapWord* end_addr)
2732 {
2733 typedef ParallelCompactData::RegionData RegionData;
2734
2735 ParallelCompactData& sd = PSParallelCompact::summary_data();
2736 const size_t region_size = ParallelCompactData::RegionSize;
2737
2738 size_t src_region_idx = 0;
2739
2740 // Skip empty regions (if any) up to the top of the space.
2741 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2742 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2743 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2744 const RegionData* const top_region_ptr =
2745 sd.addr_to_region_ptr(top_aligned_up);
2746 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2747 ++src_region_ptr;
2748 }
2749
2750 if (src_region_ptr < top_region_ptr) {
2751 // The next source region is in the current space. Update src_region_idx
2752 // and the source address to match src_region_ptr.
2753 src_region_idx = sd.region(src_region_ptr);
2754 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2755 if (src_region_addr > closure.source()) {
2756 closure.set_source(src_region_addr);
2757 }
2758 return src_region_idx;
2759 }
2760
2761 // Switch to a new source space and find the first non-empty region.
2762 unsigned int space_id = src_space_id + 1;
2763 assert(space_id < last_space_id, "not enough spaces");
2764
2765 HeapWord* const destination = closure.destination();
2766
2767 do {
2768 MutableSpace* space = _space_info[space_id].space();
2769 HeapWord* const bottom = space->bottom();
2770 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2771
2772 // Iterate over the spaces that do not compact into themselves.
2773 if (bottom_cp->destination() != bottom) {
2774 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2775 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2776
2777 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2778 if (src_cp->live_obj_size() > 0) {
2779 // Found it.
2780 assert(src_cp->destination() == destination,
2781 "first live obj in the space must match the destination");
2782 assert(src_cp->partial_obj_size() == 0,
2783 "a space cannot begin with a partial obj");
2784
2785 src_space_id = SpaceId(space_id);
2786 src_space_top = space->top();
2787 const size_t src_region_idx = sd.region(src_cp);
2788 closure.set_source(sd.region_to_addr(src_region_idx));
2789 return src_region_idx;
2790 } else {
2791 assert(src_cp->data_size() == 0, "sanity");
2792 }
2793 }
2794 }
2795 } while (++space_id < last_space_id);
2796
2797 assert(false, "no source region was found");
2798 return 0;
2799 }
2800
2801 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
2802 {
2803 typedef ParMarkBitMap::IterationStatus IterationStatus;
2804 const size_t RegionSize = ParallelCompactData::RegionSize;
2805 ParMarkBitMap* const bitmap = mark_bitmap();
2806 ParallelCompactData& sd = summary_data();
2807 RegionData* const region_ptr = sd.region(region_idx);
2808
2809 // Get the items needed to construct the closure.
2810 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2811 SpaceId dest_space_id = space_id(dest_addr);
2812 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
2813 HeapWord* new_top = _space_info[dest_space_id].new_top();
2814 assert(dest_addr < new_top, "sanity");
2815 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
2816
2817 // Get the source region and related info.
2818 size_t src_region_idx = region_ptr->source_region();
2819 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2820 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2821
2822 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2823 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2824
2825 // Adjust src_region_idx to prepare for decrementing destination counts (the
2826 // destination count is not decremented when a region is copied to itself).
2827 if (src_region_idx == region_idx) {
2828 src_region_idx += 1;
2829 }
2830
2831 if (bitmap->is_unmarked(closure.source())) {
2832 // The first source word is in the middle of an object; copy the remainder
2833 // of the object or as much as will fit. The fact that pointer updates were
2834 // deferred will be noted when the object header is processed.
2835 HeapWord* const old_src_addr = closure.source();
2836 closure.copy_partial_obj();
2837 if (closure.is_full()) {
2838 decrement_destination_counts(cm, src_space_id, src_region_idx,
2839 closure.source());
2840 region_ptr->set_deferred_obj_addr(NULL);
2841 region_ptr->set_completed();
2842 return;
2843 }
2844
2845 HeapWord* const end_addr = sd.region_align_down(closure.source());
2846 if (sd.region_align_down(old_src_addr) != end_addr) {
2847 // The partial object was copied from more than one source region.
2848 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2849
2850 // Move to the next source region, possibly switching spaces as well. All
2851 // args except end_addr may be modified.
2852 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2853 end_addr);
2854 }
2855 }
2856
2857 do {
2858 HeapWord* const cur_addr = closure.source();
2859 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2860 src_space_top);
2861 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2862
2863 if (status == ParMarkBitMap::incomplete) {
2864 // The last obj that starts in the source region does not end in the
2865 // region.
2866 assert(closure.source() < end_addr, "sanity");
2867 HeapWord* const obj_beg = closure.source();
2868 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2869 src_space_top);
2870 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2871 if (obj_end < range_end) {
2872 // The end was found; the entire object will fit.
2873 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2874 assert(status != ParMarkBitMap::would_overflow, "sanity");
2875 } else {
2876 // The end was not found; the object will not fit.
2877 assert(range_end < src_space_top, "obj cannot cross space boundary");
2878 status = ParMarkBitMap::would_overflow;
2879 }
2880 }
2881
2882 if (status == ParMarkBitMap::would_overflow) {
2883 // The last object did not fit. Note that interior oop updates were
2884 // deferred, then copy enough of the object to fill the region.
2885 region_ptr->set_deferred_obj_addr(closure.destination());
2886 status = closure.copy_until_full(); // copies from closure.source()
2887
2888 decrement_destination_counts(cm, src_space_id, src_region_idx,
2889 closure.source());
2890 region_ptr->set_completed();
2891 return;
2892 }
2893
2894 if (status == ParMarkBitMap::full) {
2895 decrement_destination_counts(cm, src_space_id, src_region_idx,
2896 closure.source());
2897 region_ptr->set_deferred_obj_addr(NULL);
2898 region_ptr->set_completed();
2899 return;
2900 }
2901
2902 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2903
2904 // Move to the next source region, possibly switching spaces as well. All
2905 // args except end_addr may be modified.
2906 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2907 end_addr);
2908 } while (true);
2909 }
2910
2911 void PSParallelCompact::fill_blocks(size_t region_idx)
2912 {
2913 // Fill in the block table elements for the specified region. Each block
2914 // table element holds the number of live words in the region that are to the
2915 // left of the first object that starts in the block. Thus only blocks in
2916 // which an object starts need to be filled.
2917 //
2918 // The algorithm scans the section of the bitmap that corresponds to the
2919 // region, keeping a running total of the live words. When an object start is
2920 // found, if it's the first to start in the block that contains it, the
2921 // current total is written to the block table element.
2922 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
2923 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
2924 const size_t RegionSize = ParallelCompactData::RegionSize;
2925
2926 ParallelCompactData& sd = summary_data();
2927 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
2928 if (partial_obj_size >= RegionSize) {
2929 return; // No objects start in this region.
2930 }
2931
2932 // Ensure the first loop iteration decides that the block has changed.
2933 size_t cur_block = sd.block_count();
2934
2935 const ParMarkBitMap* const bitmap = mark_bitmap();
2936
2937 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
2938 assert((size_t)1 << Log2BitsPerBlock ==
2939 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
2940
2941 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
2942 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
2943 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
2944 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
2945 while (beg_bit < range_end) {
2946 const size_t new_block = beg_bit >> Log2BitsPerBlock;
2947 if (new_block != cur_block) {
2948 cur_block = new_block;
2949 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
2950 }
2951
2952 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
2953 if (end_bit < range_end - 1) {
2954 live_bits += end_bit - beg_bit + 1;
2955 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
2956 } else {
2957 return;
2958 }
2959 }
2960 }
2961
2962 void
2963 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
2964 const MutableSpace* sp = space(space_id);
2965 if (sp->is_empty()) {
2966 return;
2967 }
2968
2969 ParallelCompactData& sd = PSParallelCompact::summary_data();
2970 ParMarkBitMap* const bitmap = mark_bitmap();
2971 HeapWord* const dp_addr = dense_prefix(space_id);
2972 HeapWord* beg_addr = sp->bottom();
2973 HeapWord* end_addr = sp->top();
2974
2975 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
2976
2977 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
2978 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
2979 if (beg_region < dp_region) {
2980 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
2981 }
2982
2983 // The destination of the first live object that starts in the region is one
2984 // past the end of the partial object entering the region (if any).
2985 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
2986 HeapWord* const new_top = _space_info[space_id].new_top();
2987 assert(new_top >= dest_addr, "bad new_top value");
2988 const size_t words = pointer_delta(new_top, dest_addr);
2989
2990 if (words > 0) {
2991 ObjectStartArray* start_array = _space_info[space_id].start_array();
2992 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2993
2994 ParMarkBitMap::IterationStatus status;
2995 status = bitmap->iterate(&closure, dest_addr, end_addr);
2996 assert(status == ParMarkBitMap::full, "iteration not complete");
2997 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
2998 "live objects skipped because closure is full");
2999 }
3000 }
3001
3002 jlong PSParallelCompact::millis_since_last_gc() {
3003 // We need a monotonically non-decreasing time in ms but
3004 // os::javaTimeMillis() does not guarantee monotonicity.
3005 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3006 jlong ret_val = now - _time_of_last_gc;
3007 // XXX See note in genCollectedHeap::millis_since_last_gc().
3008 if (ret_val < 0) {
3009 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);)
3010 return 0;
3011 }
3012 return ret_val;
3013 }
3014
3015 void PSParallelCompact::reset_millis_since_last_gc() {
3016 // We need a monotonically non-decreasing time in ms but
3017 // os::javaTimeMillis() does not guarantee monotonicity.
3018 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3019 }
3020
3021 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3022 {
3023 if (source() != destination()) {
3024 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3025 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3026 }
3027 update_state(words_remaining());
3028 assert(is_full(), "sanity");
3029 return ParMarkBitMap::full;
3030 }
3031
3032 void MoveAndUpdateClosure::copy_partial_obj()
3033 {
3034 size_t words = words_remaining();
3035
3036 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3037 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3038 if (end_addr < range_end) {
3039 words = bitmap()->obj_size(source(), end_addr);
3040 }
3041
3042 // This test is necessary; if omitted, the pointer updates to a partial object
3043 // that crosses the dense prefix boundary could be overwritten.
3044 if (source() != destination()) {
3045 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3046 Copy::aligned_conjoint_words(source(), destination(), words);
3047 }
3048 update_state(words);
3049 }
3050
3051 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3052 PSParallelCompact::AdjustPointerClosure closure(cm);
3053 oop_oop_iterate_oop_maps<true>(obj, &closure);
3054 }
3055
3056 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3057 InstanceKlass::oop_pc_update_pointers(obj, cm);
3058
3059 PSParallelCompact::AdjustPointerClosure closure(cm);
3060 oop_oop_iterate_statics<true>(obj, &closure);
3061 }
3062
3063 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3064 InstanceKlass::oop_pc_update_pointers(obj, cm);
3065 }
3066
3067 #ifdef ASSERT
3068 template <class T> static void trace_reference_gc(const char *s, oop obj,
3069 T* referent_addr,
3070 T* next_addr,
3071 T* discovered_addr) {
3072 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj));
3073 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3074 p2i(referent_addr), referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL);
3075 log_develop_trace(gc, ref)(" next_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3076 p2i(next_addr), next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL);
3077 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3078 p2i(discovered_addr), discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL);
3079 }
3080 #endif
3081
3082 template <class T>
3083 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) {
3084 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj);
3085 PSParallelCompact::adjust_pointer(referent_addr, cm);
3086 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj);
3087 PSParallelCompact::adjust_pointer(next_addr, cm);
3088 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj);
3089 PSParallelCompact::adjust_pointer(discovered_addr, cm);
3090 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj,
3091 referent_addr, next_addr, discovered_addr);)
3092 }
3093
3094 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3095 InstanceKlass::oop_pc_update_pointers(obj, cm);
3096
3097 if (UseCompressedOops) {
3098 oop_pc_update_pointers_specialized<narrowOop>(obj, cm);
3099 } else {
3100 oop_pc_update_pointers_specialized<oop>(obj, cm);
3101 }
3102 }
3103
3104 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3105 assert(obj->is_objArray(), "obj must be obj array");
3106 PSParallelCompact::AdjustPointerClosure closure(cm);
3107 oop_oop_iterate_elements<true>(objArrayOop(obj), &closure);
3108 }
3109
3110 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3111 assert(obj->is_typeArray(),"must be a type array");
3112 }
3113
3114 ParMarkBitMapClosure::IterationStatus
3115 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3116 assert(destination() != NULL, "sanity");
3117 assert(bitmap()->obj_size(addr) == words, "bad size");
3118
3119 _source = addr;
3120 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3121 destination(), "wrong destination");
3122
3123 if (words > words_remaining()) {
3124 return ParMarkBitMap::would_overflow;
3125 }
3126
3127 // The start_array must be updated even if the object is not moving.
3128 if (_start_array != NULL) {
3129 _start_array->allocate_block(destination());
3130 }
3131
3132 if (destination() != source()) {
3133 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3134 Copy::aligned_conjoint_words(source(), destination(), words);
3135 }
3136
3137 oop moved_oop = (oop) destination();
3138 compaction_manager()->update_contents(moved_oop);
3139 assert(moved_oop->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3140
3141 update_state(words);
3142 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3143 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3144 }
3145
3146 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3147 ParCompactionManager* cm,
3148 PSParallelCompact::SpaceId space_id) :
3149 ParMarkBitMapClosure(mbm, cm),
3150 _space_id(space_id),
3151 _start_array(PSParallelCompact::start_array(space_id))
3152 {
3153 }
3154
3155 // Updates the references in the object to their new values.
3156 ParMarkBitMapClosure::IterationStatus
3157 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3158 do_addr(addr);
3159 return ParMarkBitMap::incomplete;
3160 }
3161
3162 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3163 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3164 _start_array(PSParallelCompact::start_array(space_id))
3165 {
3166 assert(space_id == PSParallelCompact::old_space_id,
3167 "cannot use FillClosure in the young gen");
3168 }
3169
3170 ParMarkBitMapClosure::IterationStatus
3171 FillClosure::do_addr(HeapWord* addr, size_t size) {
3172 CollectedHeap::fill_with_objects(addr, size);
3173 HeapWord* const end = addr + size;
3174 do {
3175 _start_array->allocate_block(addr);
3176 addr += oop(addr)->size();
3177 } while (addr < end);
3178 return ParMarkBitMap::incomplete;
3179 }