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/safepoint.hpp"
64 #include "runtime/vmThread.hpp"
65 #include "services/management.hpp"
66 #include "services/memTracker.hpp"
67 #include "services/memoryService.hpp"
68 #include "utilities/align.hpp"
69 #include "utilities/debug.hpp"
70 #include "utilities/events.hpp"
71 #include "utilities/formatBuffer.hpp"
72 #include "utilities/stack.inline.hpp"
73
74 #include <math.h>
75
76 // All sizes are in HeapWords.
77 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
78 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
79 const size_t ParallelCompactData::RegionSizeBytes =
80 RegionSize << LogHeapWordSize;
81 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
82 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
83 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
84
85 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
86 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
87 const size_t ParallelCompactData::BlockSizeBytes =
88 BlockSize << LogHeapWordSize;
89 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
90 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
91 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
92
93 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
94 const size_t ParallelCompactData::Log2BlocksPerRegion =
95 Log2RegionSize - Log2BlockSize;
96
97 const ParallelCompactData::RegionData::region_sz_t
98 ParallelCompactData::RegionData::dc_shift = 27;
99
100 const ParallelCompactData::RegionData::region_sz_t
101 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
102
103 const ParallelCompactData::RegionData::region_sz_t
104 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
105
106 const ParallelCompactData::RegionData::region_sz_t
107 ParallelCompactData::RegionData::los_mask = ~dc_mask;
108
109 const ParallelCompactData::RegionData::region_sz_t
110 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
111
112 const ParallelCompactData::RegionData::region_sz_t
113 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
114
115 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
116
117 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
118
119 double PSParallelCompact::_dwl_mean;
120 double PSParallelCompact::_dwl_std_dev;
121 double PSParallelCompact::_dwl_first_term;
122 double PSParallelCompact::_dwl_adjustment;
123 #ifdef ASSERT
124 bool PSParallelCompact::_dwl_initialized = false;
125 #endif // #ifdef ASSERT
126
127 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
128 HeapWord* destination)
129 {
130 assert(src_region_idx != 0, "invalid src_region_idx");
131 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
132 assert(destination != NULL, "invalid destination argument");
133
134 _src_region_idx = src_region_idx;
135 _partial_obj_size = partial_obj_size;
136 _destination = destination;
137
138 // These fields may not be updated below, so make sure they're clear.
139 assert(_dest_region_addr == NULL, "should have been cleared");
140 assert(_first_src_addr == NULL, "should have been cleared");
141
142 // Determine the number of destination regions for the partial object.
143 HeapWord* const last_word = destination + partial_obj_size - 1;
144 const ParallelCompactData& sd = PSParallelCompact::summary_data();
145 HeapWord* const beg_region_addr = sd.region_align_down(destination);
146 HeapWord* const end_region_addr = sd.region_align_down(last_word);
147
148 if (beg_region_addr == end_region_addr) {
149 // One destination region.
150 _destination_count = 1;
151 if (end_region_addr == destination) {
152 // The destination falls on a region boundary, thus the first word of the
153 // partial object will be the first word copied to the destination region.
154 _dest_region_addr = end_region_addr;
155 _first_src_addr = sd.region_to_addr(src_region_idx);
156 }
157 } else {
158 // Two destination regions. When copied, the partial object will cross a
159 // destination region boundary, so a word somewhere within the partial
160 // object will be the first word copied to the second destination region.
161 _destination_count = 2;
162 _dest_region_addr = end_region_addr;
163 const size_t ofs = pointer_delta(end_region_addr, destination);
164 assert(ofs < _partial_obj_size, "sanity");
165 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
166 }
167 }
168
169 void SplitInfo::clear()
170 {
171 _src_region_idx = 0;
172 _partial_obj_size = 0;
173 _destination = NULL;
174 _destination_count = 0;
175 _dest_region_addr = NULL;
176 _first_src_addr = NULL;
177 assert(!is_valid(), "sanity");
178 }
179
180 #ifdef ASSERT
181 void SplitInfo::verify_clear()
182 {
183 assert(_src_region_idx == 0, "not clear");
184 assert(_partial_obj_size == 0, "not clear");
185 assert(_destination == NULL, "not clear");
186 assert(_destination_count == 0, "not clear");
187 assert(_dest_region_addr == NULL, "not clear");
188 assert(_first_src_addr == NULL, "not clear");
189 }
190 #endif // #ifdef ASSERT
191
192
193 void PSParallelCompact::print_on_error(outputStream* st) {
194 _mark_bitmap.print_on_error(st);
195 }
196
197 #ifndef PRODUCT
198 const char* PSParallelCompact::space_names[] = {
199 "old ", "eden", "from", "to "
200 };
201
202 void PSParallelCompact::print_region_ranges() {
203 if (!log_develop_is_enabled(Trace, gc, compaction)) {
204 return;
205 }
206 Log(gc, compaction) log;
207 ResourceMark rm;
208 LogStream ls(log.trace());
209 Universe::print_on(&ls);
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 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->num_q());
1044
1045 ref_processor()->enqueue_discovered_references(NULL, &pt);
1046
1047 pt.print_enqueue_phase();
1048
1049 if (ZapUnusedHeapArea) {
1050 heap->gen_mangle_unused_area();
1051 }
1052
1053 // Update time of last GC
1054 reset_millis_since_last_gc();
1055 }
1056
1057 HeapWord*
1058 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1059 bool maximum_compaction)
1060 {
1061 const size_t region_size = ParallelCompactData::RegionSize;
1062 const ParallelCompactData& sd = summary_data();
1063
1064 const MutableSpace* const space = _space_info[id].space();
1065 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1066 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1067 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1068
1069 // Skip full regions at the beginning of the space--they are necessarily part
1070 // of the dense prefix.
1071 size_t full_count = 0;
1072 const RegionData* cp;
1073 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1074 ++full_count;
1075 }
1076
1077 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1078 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1079 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1080 if (maximum_compaction || cp == end_cp || interval_ended) {
1081 _maximum_compaction_gc_num = total_invocations();
1082 return sd.region_to_addr(cp);
1083 }
1084
1085 HeapWord* const new_top = _space_info[id].new_top();
1086 const size_t space_live = pointer_delta(new_top, space->bottom());
1087 const size_t space_used = space->used_in_words();
1088 const size_t space_capacity = space->capacity_in_words();
1089
1090 const double cur_density = double(space_live) / space_capacity;
1091 const double deadwood_density =
1092 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1093 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1094
1095 if (TraceParallelOldGCDensePrefix) {
1096 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1097 cur_density, deadwood_density, deadwood_goal);
1098 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1099 "space_cap=" SIZE_FORMAT,
1100 space_live, space_used,
1101 space_capacity);
1102 }
1103
1104 // XXX - Use binary search?
1105 HeapWord* dense_prefix = sd.region_to_addr(cp);
1106 const RegionData* full_cp = cp;
1107 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1108 while (cp < end_cp) {
1109 HeapWord* region_destination = cp->destination();
1110 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1111 if (TraceParallelOldGCDensePrefix && Verbose) {
1112 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1113 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1114 sd.region(cp), p2i(region_destination),
1115 p2i(dense_prefix), cur_deadwood);
1116 }
1117
1118 if (cur_deadwood >= deadwood_goal) {
1119 // Found the region that has the correct amount of deadwood to the left.
1120 // This typically occurs after crossing a fairly sparse set of regions, so
1121 // iterate backwards over those sparse regions, looking for the region
1122 // that has the lowest density of live objects 'to the right.'
1123 size_t space_to_left = sd.region(cp) * region_size;
1124 size_t live_to_left = space_to_left - cur_deadwood;
1125 size_t space_to_right = space_capacity - space_to_left;
1126 size_t live_to_right = space_live - live_to_left;
1127 double density_to_right = double(live_to_right) / space_to_right;
1128 while (cp > full_cp) {
1129 --cp;
1130 const size_t prev_region_live_to_right = live_to_right -
1131 cp->data_size();
1132 const size_t prev_region_space_to_right = space_to_right + region_size;
1133 double prev_region_density_to_right =
1134 double(prev_region_live_to_right) / prev_region_space_to_right;
1135 if (density_to_right <= prev_region_density_to_right) {
1136 return dense_prefix;
1137 }
1138 if (TraceParallelOldGCDensePrefix && Verbose) {
1139 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1140 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1141 prev_region_density_to_right);
1142 }
1143 dense_prefix -= region_size;
1144 live_to_right = prev_region_live_to_right;
1145 space_to_right = prev_region_space_to_right;
1146 density_to_right = prev_region_density_to_right;
1147 }
1148 return dense_prefix;
1149 }
1150
1151 dense_prefix += region_size;
1152 ++cp;
1153 }
1154
1155 return dense_prefix;
1156 }
1157
1158 #ifndef PRODUCT
1159 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1160 const SpaceId id,
1161 const bool maximum_compaction,
1162 HeapWord* const addr)
1163 {
1164 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1165 RegionData* const cp = summary_data().region(region_idx);
1166 const MutableSpace* const space = _space_info[id].space();
1167 HeapWord* const new_top = _space_info[id].new_top();
1168
1169 const size_t space_live = pointer_delta(new_top, space->bottom());
1170 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1171 const size_t space_cap = space->capacity_in_words();
1172 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1173 const size_t live_to_right = new_top - cp->destination();
1174 const size_t dead_to_right = space->top() - addr - live_to_right;
1175
1176 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1177 "spl=" SIZE_FORMAT " "
1178 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1179 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1180 " ratio=%10.8f",
1181 algorithm, p2i(addr), region_idx,
1182 space_live,
1183 dead_to_left, dead_to_left_pct,
1184 dead_to_right, live_to_right,
1185 double(dead_to_right) / live_to_right);
1186 }
1187 #endif // #ifndef PRODUCT
1188
1189 // Return a fraction indicating how much of the generation can be treated as
1190 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1191 // based on the density of live objects in the generation to determine a limit,
1192 // which is then adjusted so the return value is min_percent when the density is
1193 // 1.
1194 //
1195 // The following table shows some return values for a different values of the
1196 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1197 // min_percent is 1.
1198 //
1199 // fraction allowed as dead wood
1200 // -----------------------------------------------------------------
1201 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1202 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1203 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1204 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1205 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1206 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1207 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1208 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1209 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1210 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1211 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1212 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1213 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1214 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1215 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1216 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1217 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1218 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1219 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1220 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1221 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1222 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1223 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1224
1225 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1226 {
1227 assert(_dwl_initialized, "uninitialized");
1228
1229 // The raw limit is the value of the normal distribution at x = density.
1230 const double raw_limit = normal_distribution(density);
1231
1232 // Adjust the raw limit so it becomes the minimum when the density is 1.
1233 //
1234 // First subtract the adjustment value (which is simply the precomputed value
1235 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1236 // Then add the minimum value, so the minimum is returned when the density is
1237 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1238 const double min = double(min_percent) / 100.0;
1239 const double limit = raw_limit - _dwl_adjustment + min;
1240 return MAX2(limit, 0.0);
1241 }
1242
1243 ParallelCompactData::RegionData*
1244 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1245 const RegionData* end)
1246 {
1247 const size_t region_size = ParallelCompactData::RegionSize;
1248 ParallelCompactData& sd = summary_data();
1249 size_t left = sd.region(beg);
1250 size_t right = end > beg ? sd.region(end) - 1 : left;
1251
1252 // Binary search.
1253 while (left < right) {
1254 // Equivalent to (left + right) / 2, but does not overflow.
1255 const size_t middle = left + (right - left) / 2;
1256 RegionData* const middle_ptr = sd.region(middle);
1257 HeapWord* const dest = middle_ptr->destination();
1258 HeapWord* const addr = sd.region_to_addr(middle);
1259 assert(dest != NULL, "sanity");
1260 assert(dest <= addr, "must move left");
1261
1262 if (middle > left && dest < addr) {
1263 right = middle - 1;
1264 } else if (middle < right && middle_ptr->data_size() == region_size) {
1265 left = middle + 1;
1266 } else {
1267 return middle_ptr;
1268 }
1269 }
1270 return sd.region(left);
1271 }
1272
1273 ParallelCompactData::RegionData*
1274 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1275 const RegionData* end,
1276 size_t dead_words)
1277 {
1278 ParallelCompactData& sd = summary_data();
1279 size_t left = sd.region(beg);
1280 size_t right = end > beg ? sd.region(end) - 1 : left;
1281
1282 // Binary search.
1283 while (left < right) {
1284 // Equivalent to (left + right) / 2, but does not overflow.
1285 const size_t middle = left + (right - left) / 2;
1286 RegionData* const middle_ptr = sd.region(middle);
1287 HeapWord* const dest = middle_ptr->destination();
1288 HeapWord* const addr = sd.region_to_addr(middle);
1289 assert(dest != NULL, "sanity");
1290 assert(dest <= addr, "must move left");
1291
1292 const size_t dead_to_left = pointer_delta(addr, dest);
1293 if (middle > left && dead_to_left > dead_words) {
1294 right = middle - 1;
1295 } else if (middle < right && dead_to_left < dead_words) {
1296 left = middle + 1;
1297 } else {
1298 return middle_ptr;
1299 }
1300 }
1301 return sd.region(left);
1302 }
1303
1304 // The result is valid during the summary phase, after the initial summarization
1305 // of each space into itself, and before final summarization.
1306 inline double
1307 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1308 HeapWord* const bottom,
1309 HeapWord* const top,
1310 HeapWord* const new_top)
1311 {
1312 ParallelCompactData& sd = summary_data();
1313
1314 assert(cp != NULL, "sanity");
1315 assert(bottom != NULL, "sanity");
1316 assert(top != NULL, "sanity");
1317 assert(new_top != NULL, "sanity");
1318 assert(top >= new_top, "summary data problem?");
1319 assert(new_top > bottom, "space is empty; should not be here");
1320 assert(new_top >= cp->destination(), "sanity");
1321 assert(top >= sd.region_to_addr(cp), "sanity");
1322
1323 HeapWord* const destination = cp->destination();
1324 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1325 const size_t compacted_region_live = pointer_delta(new_top, destination);
1326 const size_t compacted_region_used = pointer_delta(top,
1327 sd.region_to_addr(cp));
1328 const size_t reclaimable = compacted_region_used - compacted_region_live;
1329
1330 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1331 return double(reclaimable) / divisor;
1332 }
1333
1334 // Return the address of the end of the dense prefix, a.k.a. the start of the
1335 // compacted region. The address is always on a region boundary.
1336 //
1337 // Completely full regions at the left are skipped, since no compaction can
1338 // occur in those regions. Then the maximum amount of dead wood to allow is
1339 // computed, based on the density (amount live / capacity) of the generation;
1340 // the region with approximately that amount of dead space to the left is
1341 // identified as the limit region. Regions between the last completely full
1342 // region and the limit region are scanned and the one that has the best
1343 // (maximum) reclaimed_ratio() is selected.
1344 HeapWord*
1345 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1346 bool maximum_compaction)
1347 {
1348 const size_t region_size = ParallelCompactData::RegionSize;
1349 const ParallelCompactData& sd = summary_data();
1350
1351 const MutableSpace* const space = _space_info[id].space();
1352 HeapWord* const top = space->top();
1353 HeapWord* const top_aligned_up = sd.region_align_up(top);
1354 HeapWord* const new_top = _space_info[id].new_top();
1355 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1356 HeapWord* const bottom = space->bottom();
1357 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1358 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1359 const RegionData* const new_top_cp =
1360 sd.addr_to_region_ptr(new_top_aligned_up);
1361
1362 // Skip full regions at the beginning of the space--they are necessarily part
1363 // of the dense prefix.
1364 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1365 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1366 space->is_empty(), "no dead space allowed to the left");
1367 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1368 "region must have dead space");
1369
1370 // The gc number is saved whenever a maximum compaction is done, and used to
1371 // determine when the maximum compaction interval has expired. This avoids
1372 // successive max compactions for different reasons.
1373 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1374 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1375 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1376 total_invocations() == HeapFirstMaximumCompactionCount;
1377 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1378 _maximum_compaction_gc_num = total_invocations();
1379 return sd.region_to_addr(full_cp);
1380 }
1381
1382 const size_t space_live = pointer_delta(new_top, bottom);
1383 const size_t space_used = space->used_in_words();
1384 const size_t space_capacity = space->capacity_in_words();
1385
1386 const double density = double(space_live) / double(space_capacity);
1387 const size_t min_percent_free = MarkSweepDeadRatio;
1388 const double limiter = dead_wood_limiter(density, min_percent_free);
1389 const size_t dead_wood_max = space_used - space_live;
1390 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1391 dead_wood_max);
1392
1393 if (TraceParallelOldGCDensePrefix) {
1394 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1395 "space_cap=" SIZE_FORMAT,
1396 space_live, space_used,
1397 space_capacity);
1398 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1399 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1400 density, min_percent_free, limiter,
1401 dead_wood_max, dead_wood_limit);
1402 }
1403
1404 // Locate the region with the desired amount of dead space to the left.
1405 const RegionData* const limit_cp =
1406 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1407
1408 // Scan from the first region with dead space to the limit region and find the
1409 // one with the best (largest) reclaimed ratio.
1410 double best_ratio = 0.0;
1411 const RegionData* best_cp = full_cp;
1412 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1413 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1414 if (tmp_ratio > best_ratio) {
1415 best_cp = cp;
1416 best_ratio = tmp_ratio;
1417 }
1418 }
1419
1420 return sd.region_to_addr(best_cp);
1421 }
1422
1423 void PSParallelCompact::summarize_spaces_quick()
1424 {
1425 for (unsigned int i = 0; i < last_space_id; ++i) {
1426 const MutableSpace* space = _space_info[i].space();
1427 HeapWord** nta = _space_info[i].new_top_addr();
1428 bool result = _summary_data.summarize(_space_info[i].split_info(),
1429 space->bottom(), space->top(), NULL,
1430 space->bottom(), space->end(), nta);
1431 assert(result, "space must fit into itself");
1432 _space_info[i].set_dense_prefix(space->bottom());
1433 }
1434 }
1435
1436 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1437 {
1438 HeapWord* const dense_prefix_end = dense_prefix(id);
1439 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1440 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1441 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1442 // Only enough dead space is filled so that any remaining dead space to the
1443 // left is larger than the minimum filler object. (The remainder is filled
1444 // during the copy/update phase.)
1445 //
1446 // The size of the dead space to the right of the boundary is not a
1447 // concern, since compaction will be able to use whatever space is
1448 // available.
1449 //
1450 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1451 // surrounds the space to be filled with an object.
1452 //
1453 // In the 32-bit VM, each bit represents two 32-bit words:
1454 // +---+
1455 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1456 // end_bits: ... x x x | 0 | || 0 x x ...
1457 // +---+
1458 //
1459 // In the 64-bit VM, each bit represents one 64-bit word:
1460 // +------------+
1461 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1462 // end_bits: ... x x 1 | 0 || 0 | x x ...
1463 // +------------+
1464 // +-------+
1465 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1466 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1467 // +-------+
1468 // +-----------+
1469 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1470 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1471 // +-----------+
1472 // +-------+
1473 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1474 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1475 // +-------+
1476
1477 // Initially assume case a, c or e will apply.
1478 size_t obj_len = CollectedHeap::min_fill_size();
1479 HeapWord* obj_beg = dense_prefix_end - obj_len;
1480
1481 #ifdef _LP64
1482 if (MinObjAlignment > 1) { // object alignment > heap word size
1483 // Cases a, c or e.
1484 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1485 // Case b above.
1486 obj_beg = dense_prefix_end - 1;
1487 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1488 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1489 // Case d above.
1490 obj_beg = dense_prefix_end - 3;
1491 obj_len = 3;
1492 }
1493 #endif // #ifdef _LP64
1494
1495 CollectedHeap::fill_with_object(obj_beg, obj_len);
1496 _mark_bitmap.mark_obj(obj_beg, obj_len);
1497 _summary_data.add_obj(obj_beg, obj_len);
1498 assert(start_array(id) != NULL, "sanity");
1499 start_array(id)->allocate_block(obj_beg);
1500 }
1501 }
1502
1503 void
1504 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1505 {
1506 assert(id < last_space_id, "id out of range");
1507 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1508 "should have been reset in summarize_spaces_quick()");
1509
1510 const MutableSpace* space = _space_info[id].space();
1511 if (_space_info[id].new_top() != space->bottom()) {
1512 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1513 _space_info[id].set_dense_prefix(dense_prefix_end);
1514
1515 #ifndef PRODUCT
1516 if (TraceParallelOldGCDensePrefix) {
1517 print_dense_prefix_stats("ratio", id, maximum_compaction,
1518 dense_prefix_end);
1519 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1520 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1521 }
1522 #endif // #ifndef PRODUCT
1523
1524 // Recompute the summary data, taking into account the dense prefix. If
1525 // every last byte will be reclaimed, then the existing summary data which
1526 // compacts everything can be left in place.
1527 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1528 // If dead space crosses the dense prefix boundary, it is (at least
1529 // partially) filled with a dummy object, marked live and added to the
1530 // summary data. This simplifies the copy/update phase and must be done
1531 // before the final locations of objects are determined, to prevent
1532 // leaving a fragment of dead space that is too small to fill.
1533 fill_dense_prefix_end(id);
1534
1535 // Compute the destination of each Region, and thus each object.
1536 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1537 _summary_data.summarize(_space_info[id].split_info(),
1538 dense_prefix_end, space->top(), NULL,
1539 dense_prefix_end, space->end(),
1540 _space_info[id].new_top_addr());
1541 }
1542 }
1543
1544 if (log_develop_is_enabled(Trace, gc, compaction)) {
1545 const size_t region_size = ParallelCompactData::RegionSize;
1546 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1547 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1548 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1549 HeapWord* const new_top = _space_info[id].new_top();
1550 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1551 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1552 log_develop_trace(gc, compaction)(
1553 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1554 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1555 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1556 id, space->capacity_in_words(), p2i(dense_prefix_end),
1557 dp_region, dp_words / region_size,
1558 cr_words / region_size, p2i(new_top));
1559 }
1560 }
1561
1562 #ifndef PRODUCT
1563 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1564 HeapWord* dst_beg, HeapWord* dst_end,
1565 SpaceId src_space_id,
1566 HeapWord* src_beg, HeapWord* src_end)
1567 {
1568 log_develop_trace(gc, compaction)(
1569 "Summarizing %d [%s] into %d [%s]: "
1570 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1571 SIZE_FORMAT "-" SIZE_FORMAT " "
1572 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1573 SIZE_FORMAT "-" SIZE_FORMAT,
1574 src_space_id, space_names[src_space_id],
1575 dst_space_id, space_names[dst_space_id],
1576 p2i(src_beg), p2i(src_end),
1577 _summary_data.addr_to_region_idx(src_beg),
1578 _summary_data.addr_to_region_idx(src_end),
1579 p2i(dst_beg), p2i(dst_end),
1580 _summary_data.addr_to_region_idx(dst_beg),
1581 _summary_data.addr_to_region_idx(dst_end));
1582 }
1583 #endif // #ifndef PRODUCT
1584
1585 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1586 bool maximum_compaction)
1587 {
1588 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1589
1590 #ifdef ASSERT
1591 if (TraceParallelOldGCMarkingPhase) {
1592 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1593 "add_obj_bytes=" SIZE_FORMAT,
1594 add_obj_count, add_obj_size * HeapWordSize);
1595 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1596 "mark_bitmap_bytes=" SIZE_FORMAT,
1597 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1598 }
1599 #endif // #ifdef ASSERT
1600
1601 // Quick summarization of each space into itself, to see how much is live.
1602 summarize_spaces_quick();
1603
1604 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1605 NOT_PRODUCT(print_region_ranges());
1606 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1607
1608 // The amount of live data that will end up in old space (assuming it fits).
1609 size_t old_space_total_live = 0;
1610 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1611 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1612 _space_info[id].space()->bottom());
1613 }
1614
1615 MutableSpace* const old_space = _space_info[old_space_id].space();
1616 const size_t old_capacity = old_space->capacity_in_words();
1617 if (old_space_total_live > old_capacity) {
1618 // XXX - should also try to expand
1619 maximum_compaction = true;
1620 }
1621
1622 // Old generations.
1623 summarize_space(old_space_id, maximum_compaction);
1624
1625 // Summarize the remaining spaces in the young gen. The initial target space
1626 // is the old gen. If a space does not fit entirely into the target, then the
1627 // remainder is compacted into the space itself and that space becomes the new
1628 // target.
1629 SpaceId dst_space_id = old_space_id;
1630 HeapWord* dst_space_end = old_space->end();
1631 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1632 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1633 const MutableSpace* space = _space_info[id].space();
1634 const size_t live = pointer_delta(_space_info[id].new_top(),
1635 space->bottom());
1636 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1637
1638 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1639 SpaceId(id), space->bottom(), space->top());)
1640 if (live > 0 && live <= available) {
1641 // All the live data will fit.
1642 bool done = _summary_data.summarize(_space_info[id].split_info(),
1643 space->bottom(), space->top(),
1644 NULL,
1645 *new_top_addr, dst_space_end,
1646 new_top_addr);
1647 assert(done, "space must fit into old gen");
1648
1649 // Reset the new_top value for the space.
1650 _space_info[id].set_new_top(space->bottom());
1651 } else if (live > 0) {
1652 // Attempt to fit part of the source space into the target space.
1653 HeapWord* next_src_addr = NULL;
1654 bool done = _summary_data.summarize(_space_info[id].split_info(),
1655 space->bottom(), space->top(),
1656 &next_src_addr,
1657 *new_top_addr, dst_space_end,
1658 new_top_addr);
1659 assert(!done, "space should not fit into old gen");
1660 assert(next_src_addr != NULL, "sanity");
1661
1662 // The source space becomes the new target, so the remainder is compacted
1663 // within the space itself.
1664 dst_space_id = SpaceId(id);
1665 dst_space_end = space->end();
1666 new_top_addr = _space_info[id].new_top_addr();
1667 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1668 space->bottom(), dst_space_end,
1669 SpaceId(id), next_src_addr, space->top());)
1670 done = _summary_data.summarize(_space_info[id].split_info(),
1671 next_src_addr, space->top(),
1672 NULL,
1673 space->bottom(), dst_space_end,
1674 new_top_addr);
1675 assert(done, "space must fit when compacted into itself");
1676 assert(*new_top_addr <= space->top(), "usage should not grow");
1677 }
1678 }
1679
1680 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1681 NOT_PRODUCT(print_region_ranges());
1682 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1683 }
1684
1685 // This method should contain all heap-specific policy for invoking a full
1686 // collection. invoke_no_policy() will only attempt to compact the heap; it
1687 // will do nothing further. If we need to bail out for policy reasons, scavenge
1688 // before full gc, or any other specialized behavior, it needs to be added here.
1689 //
1690 // Note that this method should only be called from the vm_thread while at a
1691 // safepoint.
1692 //
1693 // Note that the all_soft_refs_clear flag in the collector policy
1694 // may be true because this method can be called without intervening
1695 // activity. For example when the heap space is tight and full measure
1696 // are being taken to free space.
1697 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1698 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1699 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1700 "should be in vm thread");
1701
1702 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1703 GCCause::Cause gc_cause = heap->gc_cause();
1704 assert(!heap->is_gc_active(), "not reentrant");
1705
1706 PSAdaptiveSizePolicy* policy = heap->size_policy();
1707 IsGCActiveMark mark;
1708
1709 if (ScavengeBeforeFullGC) {
1710 PSScavenge::invoke_no_policy();
1711 }
1712
1713 const bool clear_all_soft_refs =
1714 heap->collector_policy()->should_clear_all_soft_refs();
1715
1716 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1717 maximum_heap_compaction);
1718 }
1719
1720 // This method contains no policy. You should probably
1721 // be calling invoke() instead.
1722 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1723 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1724 assert(ref_processor() != NULL, "Sanity");
1725
1726 if (GCLocker::check_active_before_gc()) {
1727 return false;
1728 }
1729
1730 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1731
1732 GCIdMark gc_id_mark;
1733 _gc_timer.register_gc_start();
1734 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1735
1736 TimeStamp marking_start;
1737 TimeStamp compaction_start;
1738 TimeStamp collection_exit;
1739
1740 GCCause::Cause gc_cause = heap->gc_cause();
1741 PSYoungGen* young_gen = heap->young_gen();
1742 PSOldGen* old_gen = heap->old_gen();
1743 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1744
1745 // The scope of casr should end after code that can change
1746 // CollectorPolicy::_should_clear_all_soft_refs.
1747 ClearedAllSoftRefs casr(maximum_heap_compaction,
1748 heap->collector_policy());
1749
1750 if (ZapUnusedHeapArea) {
1751 // Save information needed to minimize mangling
1752 heap->record_gen_tops_before_GC();
1753 }
1754
1755 // Make sure data structures are sane, make the heap parsable, and do other
1756 // miscellaneous bookkeeping.
1757 pre_compact();
1758
1759 PreGCValues pre_gc_values(heap);
1760
1761 // Get the compaction manager reserved for the VM thread.
1762 ParCompactionManager* const vmthread_cm =
1763 ParCompactionManager::manager_array(gc_task_manager()->workers());
1764
1765 {
1766 ResourceMark rm;
1767 HandleMark hm;
1768
1769 // Set the number of GC threads to be used in this collection
1770 gc_task_manager()->set_active_gang();
1771 gc_task_manager()->task_idle_workers();
1772
1773 GCTraceCPUTime tcpu;
1774 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1775
1776 heap->pre_full_gc_dump(&_gc_timer);
1777
1778 TraceCollectorStats tcs(counters());
1779 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
1780
1781 if (TraceOldGenTime) accumulated_time()->start();
1782
1783 // Let the size policy know we're starting
1784 size_policy->major_collection_begin();
1785
1786 CodeCache::gc_prologue();
1787
1788 #if defined(COMPILER2) || INCLUDE_JVMCI
1789 DerivedPointerTable::clear();
1790 #endif
1791
1792 ref_processor()->enable_discovery();
1793 ref_processor()->setup_policy(maximum_heap_compaction);
1794
1795 bool marked_for_unloading = false;
1796
1797 marking_start.update();
1798 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
1799
1800 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1801 && GCCause::is_user_requested_gc(gc_cause);
1802 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1803
1804 #if defined(COMPILER2) || INCLUDE_JVMCI
1805 assert(DerivedPointerTable::is_active(), "Sanity");
1806 DerivedPointerTable::set_active(false);
1807 #endif
1808
1809 // adjust_roots() updates Universe::_intArrayKlassObj which is
1810 // needed by the compaction for filling holes in the dense prefix.
1811 adjust_roots(vmthread_cm);
1812
1813 compaction_start.update();
1814 compact();
1815
1816 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1817 // done before resizing.
1818 post_compact();
1819
1820 // Let the size policy know we're done
1821 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1822
1823 if (UseAdaptiveSizePolicy) {
1824 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1825 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1826 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1827
1828 // Don't check if the size_policy is ready here. Let
1829 // the size_policy check that internally.
1830 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1831 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1832 // Swap the survivor spaces if from_space is empty. The
1833 // resize_young_gen() called below is normally used after
1834 // a successful young GC and swapping of survivor spaces;
1835 // otherwise, it will fail to resize the young gen with
1836 // the current implementation.
1837 if (young_gen->from_space()->is_empty()) {
1838 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1839 young_gen->swap_spaces();
1840 }
1841
1842 // Calculate optimal free space amounts
1843 assert(young_gen->max_size() >
1844 young_gen->from_space()->capacity_in_bytes() +
1845 young_gen->to_space()->capacity_in_bytes(),
1846 "Sizes of space in young gen are out-of-bounds");
1847
1848 size_t young_live = young_gen->used_in_bytes();
1849 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1850 size_t old_live = old_gen->used_in_bytes();
1851 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1852 size_t max_old_gen_size = old_gen->max_gen_size();
1853 size_t max_eden_size = young_gen->max_size() -
1854 young_gen->from_space()->capacity_in_bytes() -
1855 young_gen->to_space()->capacity_in_bytes();
1856
1857 // Used for diagnostics
1858 size_policy->clear_generation_free_space_flags();
1859
1860 size_policy->compute_generations_free_space(young_live,
1861 eden_live,
1862 old_live,
1863 cur_eden,
1864 max_old_gen_size,
1865 max_eden_size,
1866 true /* full gc*/);
1867
1868 size_policy->check_gc_overhead_limit(young_live,
1869 eden_live,
1870 max_old_gen_size,
1871 max_eden_size,
1872 true /* full gc*/,
1873 gc_cause,
1874 heap->collector_policy());
1875
1876 size_policy->decay_supplemental_growth(true /* full gc*/);
1877
1878 heap->resize_old_gen(
1879 size_policy->calculated_old_free_size_in_bytes());
1880
1881 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1882 size_policy->calculated_survivor_size_in_bytes());
1883 }
1884
1885 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1886 }
1887
1888 if (UsePerfData) {
1889 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1890 counters->update_counters();
1891 counters->update_old_capacity(old_gen->capacity_in_bytes());
1892 counters->update_young_capacity(young_gen->capacity_in_bytes());
1893 }
1894
1895 heap->resize_all_tlabs();
1896
1897 // Resize the metaspace capacity after a collection
1898 MetaspaceGC::compute_new_size();
1899
1900 if (TraceOldGenTime) {
1901 accumulated_time()->stop();
1902 }
1903
1904 young_gen->print_used_change(pre_gc_values.young_gen_used());
1905 old_gen->print_used_change(pre_gc_values.old_gen_used());
1906 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
1907
1908 // Track memory usage and detect low memory
1909 MemoryService::track_memory_usage();
1910 heap->update_counters();
1911 gc_task_manager()->release_idle_workers();
1912
1913 heap->post_full_gc_dump(&_gc_timer);
1914 }
1915
1916 #ifdef ASSERT
1917 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1918 ParCompactionManager* const cm =
1919 ParCompactionManager::manager_array(int(i));
1920 assert(cm->marking_stack()->is_empty(), "should be empty");
1921 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i);
1922 }
1923 #endif // ASSERT
1924
1925 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1926 HandleMark hm; // Discard invalid handles created during verification
1927 Universe::verify("After GC");
1928 }
1929
1930 // Re-verify object start arrays
1931 if (VerifyObjectStartArray &&
1932 VerifyAfterGC) {
1933 old_gen->verify_object_start_array();
1934 }
1935
1936 if (ZapUnusedHeapArea) {
1937 old_gen->object_space()->check_mangled_unused_area_complete();
1938 }
1939
1940 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1941
1942 collection_exit.update();
1943
1944 heap->print_heap_after_gc();
1945 heap->trace_heap_after_gc(&_gc_tracer);
1946
1947 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1948 marking_start.ticks(), compaction_start.ticks(),
1949 collection_exit.ticks());
1950 gc_task_manager()->print_task_time_stamps();
1951
1952 #ifdef TRACESPINNING
1953 ParallelTaskTerminator::print_termination_counts();
1954 #endif
1955
1956 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1957
1958 _gc_timer.register_gc_end();
1959
1960 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1961 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1962
1963 return true;
1964 }
1965
1966 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1967 PSYoungGen* young_gen,
1968 PSOldGen* old_gen) {
1969 MutableSpace* const eden_space = young_gen->eden_space();
1970 assert(!eden_space->is_empty(), "eden must be non-empty");
1971 assert(young_gen->virtual_space()->alignment() ==
1972 old_gen->virtual_space()->alignment(), "alignments do not match");
1973
1974 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
1975 return false;
1976 }
1977
1978 // Both generations must be completely committed.
1979 if (young_gen->virtual_space()->uncommitted_size() != 0) {
1980 return false;
1981 }
1982 if (old_gen->virtual_space()->uncommitted_size() != 0) {
1983 return false;
1984 }
1985
1986 // Figure out how much to take from eden. Include the average amount promoted
1987 // in the total; otherwise the next young gen GC will simply bail out to a
1988 // full GC.
1989 const size_t alignment = old_gen->virtual_space()->alignment();
1990 const size_t eden_used = eden_space->used_in_bytes();
1991 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
1992 const size_t absorb_size = align_up(eden_used + promoted, alignment);
1993 const size_t eden_capacity = eden_space->capacity_in_bytes();
1994
1995 if (absorb_size >= eden_capacity) {
1996 return false; // Must leave some space in eden.
1997 }
1998
1999 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2000 if (new_young_size < young_gen->min_gen_size()) {
2001 return false; // Respect young gen minimum size.
2002 }
2003
2004 log_trace(heap, ergo)(" absorbing " SIZE_FORMAT "K: "
2005 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2006 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2007 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2008 absorb_size / K,
2009 eden_capacity / K, (eden_capacity - absorb_size) / K,
2010 young_gen->from_space()->used_in_bytes() / K,
2011 young_gen->to_space()->used_in_bytes() / K,
2012 young_gen->capacity_in_bytes() / K, new_young_size / K);
2013
2014 // Fill the unused part of the old gen.
2015 MutableSpace* const old_space = old_gen->object_space();
2016 HeapWord* const unused_start = old_space->top();
2017 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2018
2019 if (unused_words > 0) {
2020 if (unused_words < CollectedHeap::min_fill_size()) {
2021 return false; // If the old gen cannot be filled, must give up.
2022 }
2023 CollectedHeap::fill_with_objects(unused_start, unused_words);
2024 }
2025
2026 // Take the live data from eden and set both top and end in the old gen to
2027 // eden top. (Need to set end because reset_after_change() mangles the region
2028 // from end to virtual_space->high() in debug builds).
2029 HeapWord* const new_top = eden_space->top();
2030 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2031 absorb_size);
2032 young_gen->reset_after_change();
2033 old_space->set_top(new_top);
2034 old_space->set_end(new_top);
2035 old_gen->reset_after_change();
2036
2037 // Update the object start array for the filler object and the data from eden.
2038 ObjectStartArray* const start_array = old_gen->start_array();
2039 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2040 start_array->allocate_block(p);
2041 }
2042
2043 // Could update the promoted average here, but it is not typically updated at
2044 // full GCs and the value to use is unclear. Something like
2045 //
2046 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2047
2048 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2049 return true;
2050 }
2051
2052 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2053 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2054 "shouldn't return NULL");
2055 return ParallelScavengeHeap::gc_task_manager();
2056 }
2057
2058 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2059 bool maximum_heap_compaction,
2060 ParallelOldTracer *gc_tracer) {
2061 // Recursively traverse all live objects and mark them
2062 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2063
2064 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2065 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2066 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2067 TaskQueueSetSuper* qset = ParCompactionManager::stack_array();
2068 ParallelTaskTerminator terminator(active_gc_threads, qset);
2069
2070 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm);
2071 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2072
2073 // Need new claim bits before marking starts.
2074 ClassLoaderDataGraph::clear_claimed_marks();
2075
2076 {
2077 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2078
2079 ParallelScavengeHeap::ParStrongRootsScope psrs;
2080
2081 GCTaskQueue* q = GCTaskQueue::create();
2082
2083 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2084 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2085 // We scan the thread roots in parallel
2086 Threads::create_thread_roots_marking_tasks(q);
2087 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2088 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2089 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2090 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2091 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2092 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2093
2094 if (active_gc_threads > 1) {
2095 for (uint j = 0; j < active_gc_threads; j++) {
2096 q->enqueue(new StealMarkingTask(&terminator));
2097 }
2098 }
2099
2100 gc_task_manager()->execute_and_wait(q);
2101 }
2102
2103 // Process reference objects found during marking
2104 {
2105 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2106
2107 ReferenceProcessorStats stats;
2108 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->num_q());
2109 if (ref_processor()->processing_is_mt()) {
2110 RefProcTaskExecutor task_executor;
2111 stats = ref_processor()->process_discovered_references(
2112 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2113 &task_executor, &pt);
2114 } else {
2115 stats = ref_processor()->process_discovered_references(
2116 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2117 &pt);
2118 }
2119
2120 gc_tracer->report_gc_reference_stats(stats);
2121 pt.print_all_references();
2122 }
2123
2124 // This is the point where the entire marking should have completed.
2125 assert(cm->marking_stacks_empty(), "Marking should have completed");
2126
2127 {
2128 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2129
2130 // Follow system dictionary roots and unload classes.
2131 bool purged_class = SystemDictionary::do_unloading(is_alive_closure(), &_gc_timer);
2132
2133 // Unload nmethods.
2134 CodeCache::do_unloading(is_alive_closure(), purged_class);
2135
2136 // Prune dead klasses from subklass/sibling/implementor lists.
2137 Klass::clean_weak_klass_links(is_alive_closure());
2138 }
2139
2140 {
2141 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer);
2142 // Delete entries for dead interned strings.
2143 StringTable::unlink(is_alive_closure());
2144 }
2145
2146 {
2147 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer);
2148 // Clean up unreferenced symbols in symbol table.
2149 SymbolTable::unlink();
2150 }
2151
2152 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2153 }
2154
2155 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2156 // Adjust the pointers to reflect the new locations
2157 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2158
2159 // Need new claim bits when tracing through and adjusting pointers.
2160 ClassLoaderDataGraph::clear_claimed_marks();
2161
2162 PSParallelCompact::AdjustPointerClosure oop_closure(cm);
2163 PSParallelCompact::AdjustKlassClosure klass_closure(cm);
2164
2165 // General strong roots.
2166 Universe::oops_do(&oop_closure);
2167 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2168 Threads::oops_do(&oop_closure, NULL);
2169 ObjectSynchronizer::oops_do(&oop_closure);
2170 Management::oops_do(&oop_closure);
2171 JvmtiExport::oops_do(&oop_closure);
2172 SystemDictionary::oops_do(&oop_closure);
2173 ClassLoaderDataGraph::oops_do(&oop_closure, &klass_closure, true);
2174
2175 // Now adjust pointers in remaining weak roots. (All of which should
2176 // have been cleared if they pointed to non-surviving objects.)
2177 // Global (weak) JNI handles
2178 JNIHandles::weak_oops_do(&oop_closure);
2179
2180 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2181 CodeCache::blobs_do(&adjust_from_blobs);
2182 AOTLoader::oops_do(&oop_closure);
2183 StringTable::oops_do(&oop_closure);
2184 ref_processor()->weak_oops_do(&oop_closure);
2185 // Roots were visited so references into the young gen in roots
2186 // may have been scanned. Process them also.
2187 // Should the reference processor have a span that excludes
2188 // young gen objects?
2189 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2190 }
2191
2192 // Helper class to print 8 region numbers per line and then print the total at the end.
2193 class FillableRegionLogger : public StackObj {
2194 private:
2195 Log(gc, compaction) log;
2196 static const int LineLength = 8;
2197 size_t _regions[LineLength];
2198 int _next_index;
2199 bool _enabled;
2200 size_t _total_regions;
2201 public:
2202 FillableRegionLogger() : _next_index(0), _total_regions(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)) { }
2203 ~FillableRegionLogger() {
2204 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2205 }
2206
2207 void print_line() {
2208 if (!_enabled || _next_index == 0) {
2209 return;
2210 }
2211 FormatBuffer<> line("Fillable: ");
2212 for (int i = 0; i < _next_index; i++) {
2213 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2214 }
2215 log.trace("%s", line.buffer());
2216 _next_index = 0;
2217 }
2218
2219 void handle(size_t region) {
2220 if (!_enabled) {
2221 return;
2222 }
2223 _regions[_next_index++] = region;
2224 if (_next_index == LineLength) {
2225 print_line();
2226 }
2227 _total_regions++;
2228 }
2229 };
2230
2231 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q,
2232 uint parallel_gc_threads)
2233 {
2234 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2235
2236 // Find the threads that are active
2237 unsigned int which = 0;
2238
2239 // Find all regions that are available (can be filled immediately) and
2240 // distribute them to the thread stacks. The iteration is done in reverse
2241 // order (high to low) so the regions will be removed in ascending order.
2242
2243 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2244
2245 which = 0;
2246 // id + 1 is used to test termination so unsigned can
2247 // be used with an old_space_id == 0.
2248 FillableRegionLogger region_logger;
2249 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2250 SpaceInfo* const space_info = _space_info + id;
2251 MutableSpace* const space = space_info->space();
2252 HeapWord* const new_top = space_info->new_top();
2253
2254 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2255 const size_t end_region =
2256 sd.addr_to_region_idx(sd.region_align_up(new_top));
2257
2258 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2259 if (sd.region(cur)->claim_unsafe()) {
2260 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2261 cm->region_stack()->push(cur);
2262 region_logger.handle(cur);
2263 // Assign regions to tasks in round-robin fashion.
2264 if (++which == parallel_gc_threads) {
2265 which = 0;
2266 }
2267 }
2268 }
2269 region_logger.print_line();
2270 }
2271 }
2272
2273 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2274
2275 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2276 uint parallel_gc_threads) {
2277 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2278
2279 ParallelCompactData& sd = PSParallelCompact::summary_data();
2280
2281 // Iterate over all the spaces adding tasks for updating
2282 // regions in the dense prefix. Assume that 1 gc thread
2283 // will work on opening the gaps and the remaining gc threads
2284 // will work on the dense prefix.
2285 unsigned int space_id;
2286 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2287 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2288 const MutableSpace* const space = _space_info[space_id].space();
2289
2290 if (dense_prefix_end == space->bottom()) {
2291 // There is no dense prefix for this space.
2292 continue;
2293 }
2294
2295 // The dense prefix is before this region.
2296 size_t region_index_end_dense_prefix =
2297 sd.addr_to_region_idx(dense_prefix_end);
2298 RegionData* const dense_prefix_cp =
2299 sd.region(region_index_end_dense_prefix);
2300 assert(dense_prefix_end == space->end() ||
2301 dense_prefix_cp->available() ||
2302 dense_prefix_cp->claimed(),
2303 "The region after the dense prefix should always be ready to fill");
2304
2305 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2306
2307 // Is there dense prefix work?
2308 size_t total_dense_prefix_regions =
2309 region_index_end_dense_prefix - region_index_start;
2310 // How many regions of the dense prefix should be given to
2311 // each thread?
2312 if (total_dense_prefix_regions > 0) {
2313 uint tasks_for_dense_prefix = 1;
2314 if (total_dense_prefix_regions <=
2315 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2316 // Don't over partition. This assumes that
2317 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2318 // so there are not many regions to process.
2319 tasks_for_dense_prefix = parallel_gc_threads;
2320 } else {
2321 // Over partition
2322 tasks_for_dense_prefix = parallel_gc_threads *
2323 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2324 }
2325 size_t regions_per_thread = total_dense_prefix_regions /
2326 tasks_for_dense_prefix;
2327 // Give each thread at least 1 region.
2328 if (regions_per_thread == 0) {
2329 regions_per_thread = 1;
2330 }
2331
2332 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2333 if (region_index_start >= region_index_end_dense_prefix) {
2334 break;
2335 }
2336 // region_index_end is not processed
2337 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2338 region_index_end_dense_prefix);
2339 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2340 region_index_start,
2341 region_index_end));
2342 region_index_start = region_index_end;
2343 }
2344 }
2345 // This gets any part of the dense prefix that did not
2346 // fit evenly.
2347 if (region_index_start < region_index_end_dense_prefix) {
2348 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2349 region_index_start,
2350 region_index_end_dense_prefix));
2351 }
2352 }
2353 }
2354
2355 void PSParallelCompact::enqueue_region_stealing_tasks(
2356 GCTaskQueue* q,
2357 ParallelTaskTerminator* terminator_ptr,
2358 uint parallel_gc_threads) {
2359 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer);
2360
2361 // Once a thread has drained it's stack, it should try to steal regions from
2362 // other threads.
2363 for (uint j = 0; j < parallel_gc_threads; j++) {
2364 q->enqueue(new CompactionWithStealingTask(terminator_ptr));
2365 }
2366 }
2367
2368 #ifdef ASSERT
2369 // Write a histogram of the number of times the block table was filled for a
2370 // region.
2371 void PSParallelCompact::write_block_fill_histogram()
2372 {
2373 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2374 return;
2375 }
2376
2377 Log(gc, compaction) log;
2378 ResourceMark rm;
2379 LogStream ls(log.trace());
2380 outputStream* out = &ls;
2381
2382 typedef ParallelCompactData::RegionData rd_t;
2383 ParallelCompactData& sd = summary_data();
2384
2385 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2386 MutableSpace* const spc = _space_info[id].space();
2387 if (spc->bottom() != spc->top()) {
2388 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2389 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2390 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2391
2392 size_t histo[5] = { 0, 0, 0, 0, 0 };
2393 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2394 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2395
2396 for (const rd_t* cur = beg; cur < end; ++cur) {
2397 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2398 }
2399 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2400 for (size_t i = 0; i < histo_len; ++i) {
2401 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2402 histo[i], 100.0 * histo[i] / region_cnt);
2403 }
2404 out->cr();
2405 }
2406 }
2407 }
2408 #endif // #ifdef ASSERT
2409
2410 void PSParallelCompact::compact() {
2411 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2412
2413 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2414 PSOldGen* old_gen = heap->old_gen();
2415 old_gen->start_array()->reset();
2416 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2417 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2418 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2419 ParallelTaskTerminator terminator(active_gc_threads, qset);
2420
2421 GCTaskQueue* q = GCTaskQueue::create();
2422 prepare_region_draining_tasks(q, active_gc_threads);
2423 enqueue_dense_prefix_tasks(q, active_gc_threads);
2424 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2425
2426 {
2427 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2428
2429 gc_task_manager()->execute_and_wait(q);
2430
2431 #ifdef ASSERT
2432 // Verify that all regions have been processed before the deferred updates.
2433 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2434 verify_complete(SpaceId(id));
2435 }
2436 #endif
2437 }
2438
2439 {
2440 // Update the deferred objects, if any. Any compaction manager can be used.
2441 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2442 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2443 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2444 update_deferred_objects(cm, SpaceId(id));
2445 }
2446 }
2447
2448 DEBUG_ONLY(write_block_fill_histogram());
2449 }
2450
2451 #ifdef ASSERT
2452 void PSParallelCompact::verify_complete(SpaceId space_id) {
2453 // All Regions between space bottom() to new_top() should be marked as filled
2454 // and all Regions between new_top() and top() should be available (i.e.,
2455 // should have been emptied).
2456 ParallelCompactData& sd = summary_data();
2457 SpaceInfo si = _space_info[space_id];
2458 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2459 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2460 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2461 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2462 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2463
2464 bool issued_a_warning = false;
2465
2466 size_t cur_region;
2467 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2468 const RegionData* const c = sd.region(cur_region);
2469 if (!c->completed()) {
2470 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2471 cur_region, c->destination_count());
2472 issued_a_warning = true;
2473 }
2474 }
2475
2476 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2477 const RegionData* const c = sd.region(cur_region);
2478 if (!c->available()) {
2479 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2480 cur_region, c->destination_count());
2481 issued_a_warning = true;
2482 }
2483 }
2484
2485 if (issued_a_warning) {
2486 print_region_ranges();
2487 }
2488 }
2489 #endif // #ifdef ASSERT
2490
2491 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2492 _start_array->allocate_block(addr);
2493 compaction_manager()->update_contents(oop(addr));
2494 }
2495
2496 // Update interior oops in the ranges of regions [beg_region, end_region).
2497 void
2498 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2499 SpaceId space_id,
2500 size_t beg_region,
2501 size_t end_region) {
2502 ParallelCompactData& sd = summary_data();
2503 ParMarkBitMap* const mbm = mark_bitmap();
2504
2505 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2506 HeapWord* const end_addr = sd.region_to_addr(end_region);
2507 assert(beg_region <= end_region, "bad region range");
2508 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2509
2510 #ifdef ASSERT
2511 // Claim the regions to avoid triggering an assert when they are marked as
2512 // filled.
2513 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2514 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2515 }
2516 #endif // #ifdef ASSERT
2517
2518 if (beg_addr != space(space_id)->bottom()) {
2519 // Find the first live object or block of dead space that *starts* in this
2520 // range of regions. If a partial object crosses onto the region, skip it;
2521 // it will be marked for 'deferred update' when the object head is
2522 // processed. If dead space crosses onto the region, it is also skipped; it
2523 // will be filled when the prior region is processed. If neither of those
2524 // apply, the first word in the region is the start of a live object or dead
2525 // space.
2526 assert(beg_addr > space(space_id)->bottom(), "sanity");
2527 const RegionData* const cp = sd.region(beg_region);
2528 if (cp->partial_obj_size() != 0) {
2529 beg_addr = sd.partial_obj_end(beg_region);
2530 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2531 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2532 }
2533 }
2534
2535 if (beg_addr < end_addr) {
2536 // A live object or block of dead space starts in this range of Regions.
2537 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2538
2539 // Create closures and iterate.
2540 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2541 FillClosure fill_closure(cm, space_id);
2542 ParMarkBitMap::IterationStatus status;
2543 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2544 dense_prefix_end);
2545 if (status == ParMarkBitMap::incomplete) {
2546 update_closure.do_addr(update_closure.source());
2547 }
2548 }
2549
2550 // Mark the regions as filled.
2551 RegionData* const beg_cp = sd.region(beg_region);
2552 RegionData* const end_cp = sd.region(end_region);
2553 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2554 cp->set_completed();
2555 }
2556 }
2557
2558 // Return the SpaceId for the space containing addr. If addr is not in the
2559 // heap, last_space_id is returned. In debug mode it expects the address to be
2560 // in the heap and asserts such.
2561 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2562 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2563
2564 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2565 if (_space_info[id].space()->contains(addr)) {
2566 return SpaceId(id);
2567 }
2568 }
2569
2570 assert(false, "no space contains the addr");
2571 return last_space_id;
2572 }
2573
2574 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2575 SpaceId id) {
2576 assert(id < last_space_id, "bad space id");
2577
2578 ParallelCompactData& sd = summary_data();
2579 const SpaceInfo* const space_info = _space_info + id;
2580 ObjectStartArray* const start_array = space_info->start_array();
2581
2582 const MutableSpace* const space = space_info->space();
2583 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2584 HeapWord* const beg_addr = space_info->dense_prefix();
2585 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2586
2587 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2588 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2589 const RegionData* cur_region;
2590 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2591 HeapWord* const addr = cur_region->deferred_obj_addr();
2592 if (addr != NULL) {
2593 if (start_array != NULL) {
2594 start_array->allocate_block(addr);
2595 }
2596 cm->update_contents(oop(addr));
2597 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2598 }
2599 }
2600 }
2601
2602 // Skip over count live words starting from beg, and return the address of the
2603 // next live word. Unless marked, the word corresponding to beg is assumed to
2604 // be dead. Callers must either ensure beg does not correspond to the middle of
2605 // an object, or account for those live words in some other way. Callers must
2606 // also ensure that there are enough live words in the range [beg, end) to skip.
2607 HeapWord*
2608 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2609 {
2610 assert(count > 0, "sanity");
2611
2612 ParMarkBitMap* m = mark_bitmap();
2613 idx_t bits_to_skip = m->words_to_bits(count);
2614 idx_t cur_beg = m->addr_to_bit(beg);
2615 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2616
2617 do {
2618 cur_beg = m->find_obj_beg(cur_beg, search_end);
2619 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2620 const size_t obj_bits = cur_end - cur_beg + 1;
2621 if (obj_bits > bits_to_skip) {
2622 return m->bit_to_addr(cur_beg + bits_to_skip);
2623 }
2624 bits_to_skip -= obj_bits;
2625 cur_beg = cur_end + 1;
2626 } while (bits_to_skip > 0);
2627
2628 // Skipping the desired number of words landed just past the end of an object.
2629 // Find the start of the next object.
2630 cur_beg = m->find_obj_beg(cur_beg, search_end);
2631 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2632 return m->bit_to_addr(cur_beg);
2633 }
2634
2635 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2636 SpaceId src_space_id,
2637 size_t src_region_idx)
2638 {
2639 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2640
2641 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2642 if (split_info.dest_region_addr() == dest_addr) {
2643 // The partial object ending at the split point contains the first word to
2644 // be copied to dest_addr.
2645 return split_info.first_src_addr();
2646 }
2647
2648 const ParallelCompactData& sd = summary_data();
2649 ParMarkBitMap* const bitmap = mark_bitmap();
2650 const size_t RegionSize = ParallelCompactData::RegionSize;
2651
2652 assert(sd.is_region_aligned(dest_addr), "not aligned");
2653 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2654 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2655 HeapWord* const src_region_destination = src_region_ptr->destination();
2656
2657 assert(dest_addr >= src_region_destination, "wrong src region");
2658 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2659
2660 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2661 HeapWord* const src_region_end = src_region_beg + RegionSize;
2662
2663 HeapWord* addr = src_region_beg;
2664 if (dest_addr == src_region_destination) {
2665 // Return the first live word in the source region.
2666 if (partial_obj_size == 0) {
2667 addr = bitmap->find_obj_beg(addr, src_region_end);
2668 assert(addr < src_region_end, "no objects start in src region");
2669 }
2670 return addr;
2671 }
2672
2673 // Must skip some live data.
2674 size_t words_to_skip = dest_addr - src_region_destination;
2675 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2676
2677 if (partial_obj_size >= words_to_skip) {
2678 // All the live words to skip are part of the partial object.
2679 addr += words_to_skip;
2680 if (partial_obj_size == words_to_skip) {
2681 // Find the first live word past the partial object.
2682 addr = bitmap->find_obj_beg(addr, src_region_end);
2683 assert(addr < src_region_end, "wrong src region");
2684 }
2685 return addr;
2686 }
2687
2688 // Skip over the partial object (if any).
2689 if (partial_obj_size != 0) {
2690 words_to_skip -= partial_obj_size;
2691 addr += partial_obj_size;
2692 }
2693
2694 // Skip over live words due to objects that start in the region.
2695 addr = skip_live_words(addr, src_region_end, words_to_skip);
2696 assert(addr < src_region_end, "wrong src region");
2697 return addr;
2698 }
2699
2700 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2701 SpaceId src_space_id,
2702 size_t beg_region,
2703 HeapWord* end_addr)
2704 {
2705 ParallelCompactData& sd = summary_data();
2706
2707 #ifdef ASSERT
2708 MutableSpace* const src_space = _space_info[src_space_id].space();
2709 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2710 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2711 "src_space_id does not match beg_addr");
2712 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2713 "src_space_id does not match end_addr");
2714 #endif // #ifdef ASSERT
2715
2716 RegionData* const beg = sd.region(beg_region);
2717 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2718
2719 // Regions up to new_top() are enqueued if they become available.
2720 HeapWord* const new_top = _space_info[src_space_id].new_top();
2721 RegionData* const enqueue_end =
2722 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2723
2724 for (RegionData* cur = beg; cur < end; ++cur) {
2725 assert(cur->data_size() > 0, "region must have live data");
2726 cur->decrement_destination_count();
2727 if (cur < enqueue_end && cur->available() && cur->claim()) {
2728 cm->push_region(sd.region(cur));
2729 }
2730 }
2731 }
2732
2733 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2734 SpaceId& src_space_id,
2735 HeapWord*& src_space_top,
2736 HeapWord* end_addr)
2737 {
2738 typedef ParallelCompactData::RegionData RegionData;
2739
2740 ParallelCompactData& sd = PSParallelCompact::summary_data();
2741 const size_t region_size = ParallelCompactData::RegionSize;
2742
2743 size_t src_region_idx = 0;
2744
2745 // Skip empty regions (if any) up to the top of the space.
2746 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2747 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2748 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2749 const RegionData* const top_region_ptr =
2750 sd.addr_to_region_ptr(top_aligned_up);
2751 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2752 ++src_region_ptr;
2753 }
2754
2755 if (src_region_ptr < top_region_ptr) {
2756 // The next source region is in the current space. Update src_region_idx
2757 // and the source address to match src_region_ptr.
2758 src_region_idx = sd.region(src_region_ptr);
2759 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2760 if (src_region_addr > closure.source()) {
2761 closure.set_source(src_region_addr);
2762 }
2763 return src_region_idx;
2764 }
2765
2766 // Switch to a new source space and find the first non-empty region.
2767 unsigned int space_id = src_space_id + 1;
2768 assert(space_id < last_space_id, "not enough spaces");
2769
2770 HeapWord* const destination = closure.destination();
2771
2772 do {
2773 MutableSpace* space = _space_info[space_id].space();
2774 HeapWord* const bottom = space->bottom();
2775 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2776
2777 // Iterate over the spaces that do not compact into themselves.
2778 if (bottom_cp->destination() != bottom) {
2779 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2780 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2781
2782 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2783 if (src_cp->live_obj_size() > 0) {
2784 // Found it.
2785 assert(src_cp->destination() == destination,
2786 "first live obj in the space must match the destination");
2787 assert(src_cp->partial_obj_size() == 0,
2788 "a space cannot begin with a partial obj");
2789
2790 src_space_id = SpaceId(space_id);
2791 src_space_top = space->top();
2792 const size_t src_region_idx = sd.region(src_cp);
2793 closure.set_source(sd.region_to_addr(src_region_idx));
2794 return src_region_idx;
2795 } else {
2796 assert(src_cp->data_size() == 0, "sanity");
2797 }
2798 }
2799 }
2800 } while (++space_id < last_space_id);
2801
2802 assert(false, "no source region was found");
2803 return 0;
2804 }
2805
2806 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
2807 {
2808 typedef ParMarkBitMap::IterationStatus IterationStatus;
2809 const size_t RegionSize = ParallelCompactData::RegionSize;
2810 ParMarkBitMap* const bitmap = mark_bitmap();
2811 ParallelCompactData& sd = summary_data();
2812 RegionData* const region_ptr = sd.region(region_idx);
2813
2814 // Get the items needed to construct the closure.
2815 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2816 SpaceId dest_space_id = space_id(dest_addr);
2817 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
2818 HeapWord* new_top = _space_info[dest_space_id].new_top();
2819 assert(dest_addr < new_top, "sanity");
2820 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
2821
2822 // Get the source region and related info.
2823 size_t src_region_idx = region_ptr->source_region();
2824 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2825 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2826
2827 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2828 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2829
2830 // Adjust src_region_idx to prepare for decrementing destination counts (the
2831 // destination count is not decremented when a region is copied to itself).
2832 if (src_region_idx == region_idx) {
2833 src_region_idx += 1;
2834 }
2835
2836 if (bitmap->is_unmarked(closure.source())) {
2837 // The first source word is in the middle of an object; copy the remainder
2838 // of the object or as much as will fit. The fact that pointer updates were
2839 // deferred will be noted when the object header is processed.
2840 HeapWord* const old_src_addr = closure.source();
2841 closure.copy_partial_obj();
2842 if (closure.is_full()) {
2843 decrement_destination_counts(cm, src_space_id, src_region_idx,
2844 closure.source());
2845 region_ptr->set_deferred_obj_addr(NULL);
2846 region_ptr->set_completed();
2847 return;
2848 }
2849
2850 HeapWord* const end_addr = sd.region_align_down(closure.source());
2851 if (sd.region_align_down(old_src_addr) != end_addr) {
2852 // The partial object was copied from more than one source region.
2853 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2854
2855 // Move to the next source region, possibly switching spaces as well. All
2856 // args except end_addr may be modified.
2857 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2858 end_addr);
2859 }
2860 }
2861
2862 do {
2863 HeapWord* const cur_addr = closure.source();
2864 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2865 src_space_top);
2866 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2867
2868 if (status == ParMarkBitMap::incomplete) {
2869 // The last obj that starts in the source region does not end in the
2870 // region.
2871 assert(closure.source() < end_addr, "sanity");
2872 HeapWord* const obj_beg = closure.source();
2873 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2874 src_space_top);
2875 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2876 if (obj_end < range_end) {
2877 // The end was found; the entire object will fit.
2878 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2879 assert(status != ParMarkBitMap::would_overflow, "sanity");
2880 } else {
2881 // The end was not found; the object will not fit.
2882 assert(range_end < src_space_top, "obj cannot cross space boundary");
2883 status = ParMarkBitMap::would_overflow;
2884 }
2885 }
2886
2887 if (status == ParMarkBitMap::would_overflow) {
2888 // The last object did not fit. Note that interior oop updates were
2889 // deferred, then copy enough of the object to fill the region.
2890 region_ptr->set_deferred_obj_addr(closure.destination());
2891 status = closure.copy_until_full(); // copies from closure.source()
2892
2893 decrement_destination_counts(cm, src_space_id, src_region_idx,
2894 closure.source());
2895 region_ptr->set_completed();
2896 return;
2897 }
2898
2899 if (status == ParMarkBitMap::full) {
2900 decrement_destination_counts(cm, src_space_id, src_region_idx,
2901 closure.source());
2902 region_ptr->set_deferred_obj_addr(NULL);
2903 region_ptr->set_completed();
2904 return;
2905 }
2906
2907 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2908
2909 // Move to the next source region, possibly switching spaces as well. All
2910 // args except end_addr may be modified.
2911 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2912 end_addr);
2913 } while (true);
2914 }
2915
2916 void PSParallelCompact::fill_blocks(size_t region_idx)
2917 {
2918 // Fill in the block table elements for the specified region. Each block
2919 // table element holds the number of live words in the region that are to the
2920 // left of the first object that starts in the block. Thus only blocks in
2921 // which an object starts need to be filled.
2922 //
2923 // The algorithm scans the section of the bitmap that corresponds to the
2924 // region, keeping a running total of the live words. When an object start is
2925 // found, if it's the first to start in the block that contains it, the
2926 // current total is written to the block table element.
2927 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
2928 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
2929 const size_t RegionSize = ParallelCompactData::RegionSize;
2930
2931 ParallelCompactData& sd = summary_data();
2932 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
2933 if (partial_obj_size >= RegionSize) {
2934 return; // No objects start in this region.
2935 }
2936
2937 // Ensure the first loop iteration decides that the block has changed.
2938 size_t cur_block = sd.block_count();
2939
2940 const ParMarkBitMap* const bitmap = mark_bitmap();
2941
2942 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
2943 assert((size_t)1 << Log2BitsPerBlock ==
2944 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
2945
2946 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
2947 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
2948 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
2949 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
2950 while (beg_bit < range_end) {
2951 const size_t new_block = beg_bit >> Log2BitsPerBlock;
2952 if (new_block != cur_block) {
2953 cur_block = new_block;
2954 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
2955 }
2956
2957 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
2958 if (end_bit < range_end - 1) {
2959 live_bits += end_bit - beg_bit + 1;
2960 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
2961 } else {
2962 return;
2963 }
2964 }
2965 }
2966
2967 void
2968 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
2969 const MutableSpace* sp = space(space_id);
2970 if (sp->is_empty()) {
2971 return;
2972 }
2973
2974 ParallelCompactData& sd = PSParallelCompact::summary_data();
2975 ParMarkBitMap* const bitmap = mark_bitmap();
2976 HeapWord* const dp_addr = dense_prefix(space_id);
2977 HeapWord* beg_addr = sp->bottom();
2978 HeapWord* end_addr = sp->top();
2979
2980 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
2981
2982 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
2983 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
2984 if (beg_region < dp_region) {
2985 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
2986 }
2987
2988 // The destination of the first live object that starts in the region is one
2989 // past the end of the partial object entering the region (if any).
2990 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
2991 HeapWord* const new_top = _space_info[space_id].new_top();
2992 assert(new_top >= dest_addr, "bad new_top value");
2993 const size_t words = pointer_delta(new_top, dest_addr);
2994
2995 if (words > 0) {
2996 ObjectStartArray* start_array = _space_info[space_id].start_array();
2997 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2998
2999 ParMarkBitMap::IterationStatus status;
3000 status = bitmap->iterate(&closure, dest_addr, end_addr);
3001 assert(status == ParMarkBitMap::full, "iteration not complete");
3002 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3003 "live objects skipped because closure is full");
3004 }
3005 }
3006
3007 jlong PSParallelCompact::millis_since_last_gc() {
3008 // We need a monotonically non-decreasing time in ms but
3009 // os::javaTimeMillis() does not guarantee monotonicity.
3010 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3011 jlong ret_val = now - _time_of_last_gc;
3012 // XXX See note in genCollectedHeap::millis_since_last_gc().
3013 if (ret_val < 0) {
3014 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);)
3015 return 0;
3016 }
3017 return ret_val;
3018 }
3019
3020 void PSParallelCompact::reset_millis_since_last_gc() {
3021 // We need a monotonically non-decreasing time in ms but
3022 // os::javaTimeMillis() does not guarantee monotonicity.
3023 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3024 }
3025
3026 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3027 {
3028 if (source() != destination()) {
3029 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3030 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3031 }
3032 update_state(words_remaining());
3033 assert(is_full(), "sanity");
3034 return ParMarkBitMap::full;
3035 }
3036
3037 void MoveAndUpdateClosure::copy_partial_obj()
3038 {
3039 size_t words = words_remaining();
3040
3041 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3042 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3043 if (end_addr < range_end) {
3044 words = bitmap()->obj_size(source(), end_addr);
3045 }
3046
3047 // This test is necessary; if omitted, the pointer updates to a partial object
3048 // that crosses the dense prefix boundary could be overwritten.
3049 if (source() != destination()) {
3050 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3051 Copy::aligned_conjoint_words(source(), destination(), words);
3052 }
3053 update_state(words);
3054 }
3055
3056 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3057 PSParallelCompact::AdjustPointerClosure closure(cm);
3058 oop_oop_iterate_oop_maps<true>(obj, &closure);
3059 }
3060
3061 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3062 InstanceKlass::oop_pc_update_pointers(obj, cm);
3063
3064 PSParallelCompact::AdjustPointerClosure closure(cm);
3065 oop_oop_iterate_statics<true>(obj, &closure);
3066 }
3067
3068 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3069 InstanceKlass::oop_pc_update_pointers(obj, cm);
3070 }
3071
3072 #ifdef ASSERT
3073 template <class T> static void trace_reference_gc(const char *s, oop obj,
3074 T* referent_addr,
3075 T* next_addr,
3076 T* discovered_addr) {
3077 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj));
3078 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3079 p2i(referent_addr), referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL);
3080 log_develop_trace(gc, ref)(" next_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3081 p2i(next_addr), next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL);
3082 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3083 p2i(discovered_addr), discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL);
3084 }
3085 #endif
3086
3087 template <class T>
3088 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) {
3089 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj);
3090 PSParallelCompact::adjust_pointer(referent_addr, cm);
3091 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj);
3092 PSParallelCompact::adjust_pointer(next_addr, cm);
3093 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj);
3094 PSParallelCompact::adjust_pointer(discovered_addr, cm);
3095 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj,
3096 referent_addr, next_addr, discovered_addr);)
3097 }
3098
3099 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3100 InstanceKlass::oop_pc_update_pointers(obj, cm);
3101
3102 if (UseCompressedOops) {
3103 oop_pc_update_pointers_specialized<narrowOop>(obj, cm);
3104 } else {
3105 oop_pc_update_pointers_specialized<oop>(obj, cm);
3106 }
3107 }
3108
3109 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3110 assert(obj->is_objArray(), "obj must be obj array");
3111 PSParallelCompact::AdjustPointerClosure closure(cm);
3112 oop_oop_iterate_elements<true>(objArrayOop(obj), &closure);
3113 }
3114
3115 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3116 assert(obj->is_typeArray(),"must be a type array");
3117 }
3118
3119 ParMarkBitMapClosure::IterationStatus
3120 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3121 assert(destination() != NULL, "sanity");
3122 assert(bitmap()->obj_size(addr) == words, "bad size");
3123
3124 _source = addr;
3125 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3126 destination(), "wrong destination");
3127
3128 if (words > words_remaining()) {
3129 return ParMarkBitMap::would_overflow;
3130 }
3131
3132 // The start_array must be updated even if the object is not moving.
3133 if (_start_array != NULL) {
3134 _start_array->allocate_block(destination());
3135 }
3136
3137 if (destination() != source()) {
3138 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3139 Copy::aligned_conjoint_words(source(), destination(), words);
3140 }
3141
3142 oop moved_oop = (oop) destination();
3143 compaction_manager()->update_contents(moved_oop);
3144 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3145
3146 update_state(words);
3147 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3148 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3149 }
3150
3151 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3152 ParCompactionManager* cm,
3153 PSParallelCompact::SpaceId space_id) :
3154 ParMarkBitMapClosure(mbm, cm),
3155 _space_id(space_id),
3156 _start_array(PSParallelCompact::start_array(space_id))
3157 {
3158 }
3159
3160 // Updates the references in the object to their new values.
3161 ParMarkBitMapClosure::IterationStatus
3162 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3163 do_addr(addr);
3164 return ParMarkBitMap::incomplete;
3165 }
3166
3167 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3168 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3169 _start_array(PSParallelCompact::start_array(space_id))
3170 {
3171 assert(space_id == PSParallelCompact::old_space_id,
3172 "cannot use FillClosure in the young gen");
3173 }
3174
3175 ParMarkBitMapClosure::IterationStatus
3176 FillClosure::do_addr(HeapWord* addr, size_t size) {
3177 CollectedHeap::fill_with_objects(addr, size);
3178 HeapWord* const end = addr + size;
3179 do {
3180 _start_array->allocate_block(addr);
3181 addr += oop(addr)->size();
3182 } while (addr < end);
3183 return ParMarkBitMap::incomplete;
3184 }