1 /*
2 * Copyright (c) 2005, 2018, 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.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 "gc/shared/weakProcessor.hpp"
56 #include "logging/log.hpp"
57 #include "memory/resourceArea.hpp"
58 #include "oops/access.inline.hpp"
59 #include "oops/instanceKlass.inline.hpp"
60 #include "oops/instanceMirrorKlass.inline.hpp"
61 #include "oops/methodData.hpp"
62 #include "oops/objArrayKlass.inline.hpp"
63 #include "oops/oop.inline.hpp"
64 #include "runtime/atomic.hpp"
65 #include "runtime/handles.inline.hpp"
66 #include "runtime/safepoint.hpp"
67 #include "runtime/vmThread.hpp"
68 #include "services/management.hpp"
69 #include "services/memTracker.hpp"
70 #include "services/memoryService.hpp"
71 #include "utilities/align.hpp"
72 #include "utilities/debug.hpp"
73 #include "utilities/events.hpp"
74 #include "utilities/formatBuffer.hpp"
75 #include "utilities/stack.inline.hpp"
76
77 #include <math.h>
78
79 // All sizes are in HeapWords.
80 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
81 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
82 const size_t ParallelCompactData::RegionSizeBytes =
83 RegionSize << LogHeapWordSize;
84 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
85 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
86 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
87
88 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
89 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
90 const size_t ParallelCompactData::BlockSizeBytes =
91 BlockSize << LogHeapWordSize;
92 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
93 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
94 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
95
96 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
97 const size_t ParallelCompactData::Log2BlocksPerRegion =
98 Log2RegionSize - Log2BlockSize;
99
100 const ParallelCompactData::RegionData::region_sz_t
101 ParallelCompactData::RegionData::dc_shift = 27;
102
103 const ParallelCompactData::RegionData::region_sz_t
104 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
105
106 const ParallelCompactData::RegionData::region_sz_t
107 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
108
109 const ParallelCompactData::RegionData::region_sz_t
110 ParallelCompactData::RegionData::los_mask = ~dc_mask;
111
112 const ParallelCompactData::RegionData::region_sz_t
113 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
114
115 const ParallelCompactData::RegionData::region_sz_t
116 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
117
118 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
119
120 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
121
122 double PSParallelCompact::_dwl_mean;
123 double PSParallelCompact::_dwl_std_dev;
124 double PSParallelCompact::_dwl_first_term;
125 double PSParallelCompact::_dwl_adjustment;
126 #ifdef ASSERT
127 bool PSParallelCompact::_dwl_initialized = false;
128 #endif // #ifdef ASSERT
129
130 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
131 HeapWord* destination)
132 {
133 assert(src_region_idx != 0, "invalid src_region_idx");
134 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
135 assert(destination != NULL, "invalid destination argument");
136
137 _src_region_idx = src_region_idx;
138 _partial_obj_size = partial_obj_size;
139 _destination = destination;
140
141 // These fields may not be updated below, so make sure they're clear.
142 assert(_dest_region_addr == NULL, "should have been cleared");
143 assert(_first_src_addr == NULL, "should have been cleared");
144
145 // Determine the number of destination regions for the partial object.
146 HeapWord* const last_word = destination + partial_obj_size - 1;
147 const ParallelCompactData& sd = PSParallelCompact::summary_data();
148 HeapWord* const beg_region_addr = sd.region_align_down(destination);
149 HeapWord* const end_region_addr = sd.region_align_down(last_word);
150
151 if (beg_region_addr == end_region_addr) {
152 // One destination region.
153 _destination_count = 1;
154 if (end_region_addr == destination) {
155 // The destination falls on a region boundary, thus the first word of the
156 // partial object will be the first word copied to the destination region.
157 _dest_region_addr = end_region_addr;
158 _first_src_addr = sd.region_to_addr(src_region_idx);
159 }
160 } else {
161 // Two destination regions. When copied, the partial object will cross a
162 // destination region boundary, so a word somewhere within the partial
163 // object will be the first word copied to the second destination region.
164 _destination_count = 2;
165 _dest_region_addr = end_region_addr;
166 const size_t ofs = pointer_delta(end_region_addr, destination);
167 assert(ofs < _partial_obj_size, "sanity");
168 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
169 }
170 }
171
172 void SplitInfo::clear()
173 {
174 _src_region_idx = 0;
175 _partial_obj_size = 0;
176 _destination = NULL;
177 _destination_count = 0;
178 _dest_region_addr = NULL;
179 _first_src_addr = NULL;
180 assert(!is_valid(), "sanity");
181 }
182
183 #ifdef ASSERT
184 void SplitInfo::verify_clear()
185 {
186 assert(_src_region_idx == 0, "not clear");
187 assert(_partial_obj_size == 0, "not clear");
188 assert(_destination == NULL, "not clear");
189 assert(_destination_count == 0, "not clear");
190 assert(_dest_region_addr == NULL, "not clear");
191 assert(_first_src_addr == NULL, "not clear");
192 }
193 #endif // #ifdef ASSERT
194
195
196 void PSParallelCompact::print_on_error(outputStream* st) {
197 _mark_bitmap.print_on_error(st);
198 }
199
200 #ifndef PRODUCT
201 const char* PSParallelCompact::space_names[] = {
202 "old ", "eden", "from", "to "
203 };
204
205 void PSParallelCompact::print_region_ranges() {
206 if (!log_develop_is_enabled(Trace, gc, compaction)) {
207 return;
208 }
209 Log(gc, compaction) log;
210 ResourceMark rm;
211 LogStream ls(log.trace());
212 Universe::print_on(&ls);
213 log.trace("space bottom top end new_top");
214 log.trace("------ ---------- ---------- ---------- ----------");
215
216 for (unsigned int id = 0; id < last_space_id; ++id) {
217 const MutableSpace* space = _space_info[id].space();
218 log.trace("%u %s "
219 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
220 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
221 id, space_names[id],
222 summary_data().addr_to_region_idx(space->bottom()),
223 summary_data().addr_to_region_idx(space->top()),
224 summary_data().addr_to_region_idx(space->end()),
225 summary_data().addr_to_region_idx(_space_info[id].new_top()));
226 }
227 }
228
229 void
230 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
231 {
232 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
233 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
234
235 ParallelCompactData& sd = PSParallelCompact::summary_data();
236 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
237 log_develop_trace(gc, compaction)(
238 REGION_IDX_FORMAT " " PTR_FORMAT " "
239 REGION_IDX_FORMAT " " PTR_FORMAT " "
240 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
241 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
242 i, p2i(c->data_location()), dci, p2i(c->destination()),
243 c->partial_obj_size(), c->live_obj_size(),
244 c->data_size(), c->source_region(), c->destination_count());
245
246 #undef REGION_IDX_FORMAT
247 #undef REGION_DATA_FORMAT
248 }
249
250 void
251 print_generic_summary_data(ParallelCompactData& summary_data,
252 HeapWord* const beg_addr,
253 HeapWord* const end_addr)
254 {
255 size_t total_words = 0;
256 size_t i = summary_data.addr_to_region_idx(beg_addr);
257 const size_t last = summary_data.addr_to_region_idx(end_addr);
258 HeapWord* pdest = 0;
259
260 while (i < last) {
261 ParallelCompactData::RegionData* c = summary_data.region(i);
262 if (c->data_size() != 0 || c->destination() != pdest) {
263 print_generic_summary_region(i, c);
264 total_words += c->data_size();
265 pdest = c->destination();
266 }
267 ++i;
268 }
269
270 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
271 }
272
273 void
274 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
275 HeapWord* const beg_addr,
276 HeapWord* const end_addr) {
277 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
278 }
279
280 void
281 print_generic_summary_data(ParallelCompactData& summary_data,
282 SpaceInfo* space_info)
283 {
284 if (!log_develop_is_enabled(Trace, gc, compaction)) {
285 return;
286 }
287
288 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
289 const MutableSpace* space = space_info[id].space();
290 print_generic_summary_data(summary_data, space->bottom(),
291 MAX2(space->top(), space_info[id].new_top()));
292 }
293 }
294
295 void
296 print_initial_summary_data(ParallelCompactData& summary_data,
297 const MutableSpace* space) {
298 if (space->top() == space->bottom()) {
299 return;
300 }
301
302 const size_t region_size = ParallelCompactData::RegionSize;
303 typedef ParallelCompactData::RegionData RegionData;
304 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
305 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
306 const RegionData* c = summary_data.region(end_region - 1);
307 HeapWord* end_addr = c->destination() + c->data_size();
308 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
309
310 // Print (and count) the full regions at the beginning of the space.
311 size_t full_region_count = 0;
312 size_t i = summary_data.addr_to_region_idx(space->bottom());
313 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
314 ParallelCompactData::RegionData* c = summary_data.region(i);
315 log_develop_trace(gc, compaction)(
316 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
317 i, p2i(c->destination()),
318 c->partial_obj_size(), c->live_obj_size(),
319 c->data_size(), c->source_region(), c->destination_count());
320 ++full_region_count;
321 ++i;
322 }
323
324 size_t live_to_right = live_in_space - full_region_count * region_size;
325
326 double max_reclaimed_ratio = 0.0;
327 size_t max_reclaimed_ratio_region = 0;
328 size_t max_dead_to_right = 0;
329 size_t max_live_to_right = 0;
330
331 // Print the 'reclaimed ratio' for regions while there is something live in
332 // the region or to the right of it. The remaining regions are empty (and
333 // uninteresting), and computing the ratio will result in division by 0.
334 while (i < end_region && live_to_right > 0) {
335 c = summary_data.region(i);
336 HeapWord* const region_addr = summary_data.region_to_addr(i);
337 const size_t used_to_right = pointer_delta(space->top(), region_addr);
338 const size_t dead_to_right = used_to_right - live_to_right;
339 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
340
341 if (reclaimed_ratio > max_reclaimed_ratio) {
342 max_reclaimed_ratio = reclaimed_ratio;
343 max_reclaimed_ratio_region = i;
344 max_dead_to_right = dead_to_right;
345 max_live_to_right = live_to_right;
346 }
347
348 ParallelCompactData::RegionData* c = summary_data.region(i);
349 log_develop_trace(gc, compaction)(
350 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
351 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
352 i, p2i(c->destination()),
353 c->partial_obj_size(), c->live_obj_size(),
354 c->data_size(), c->source_region(), c->destination_count(),
355 reclaimed_ratio, dead_to_right, live_to_right);
356
357
358 live_to_right -= c->data_size();
359 ++i;
360 }
361
362 // Any remaining regions are empty. Print one more if there is one.
363 if (i < end_region) {
364 ParallelCompactData::RegionData* c = summary_data.region(i);
365 log_develop_trace(gc, compaction)(
366 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
367 i, p2i(c->destination()),
368 c->partial_obj_size(), c->live_obj_size(),
369 c->data_size(), c->source_region(), c->destination_count());
370 }
371
372 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
373 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
374 }
375
376 void
377 print_initial_summary_data(ParallelCompactData& summary_data,
378 SpaceInfo* space_info) {
379 if (!log_develop_is_enabled(Trace, gc, compaction)) {
380 return;
381 }
382
383 unsigned int id = PSParallelCompact::old_space_id;
384 const MutableSpace* space;
385 do {
386 space = space_info[id].space();
387 print_initial_summary_data(summary_data, space);
388 } while (++id < PSParallelCompact::eden_space_id);
389
390 do {
391 space = space_info[id].space();
392 print_generic_summary_data(summary_data, space->bottom(), space->top());
393 } while (++id < PSParallelCompact::last_space_id);
394 }
395 #endif // #ifndef PRODUCT
396
397 #ifdef ASSERT
398 size_t add_obj_count;
399 size_t add_obj_size;
400 size_t mark_bitmap_count;
401 size_t mark_bitmap_size;
402 #endif // #ifdef ASSERT
403
404 ParallelCompactData::ParallelCompactData()
405 {
406 _region_start = 0;
407
408 _region_vspace = 0;
409 _reserved_byte_size = 0;
410 _region_data = 0;
411 _region_count = 0;
412
413 _block_vspace = 0;
414 _block_data = 0;
415 _block_count = 0;
416 }
417
418 bool ParallelCompactData::initialize(MemRegion covered_region)
419 {
420 _region_start = covered_region.start();
421 const size_t region_size = covered_region.word_size();
422 DEBUG_ONLY(_region_end = _region_start + region_size;)
423
424 assert(region_align_down(_region_start) == _region_start,
425 "region start not aligned");
426 assert((region_size & RegionSizeOffsetMask) == 0,
427 "region size not a multiple of RegionSize");
428
429 bool result = initialize_region_data(region_size) && initialize_block_data();
430 return result;
431 }
432
433 PSVirtualSpace*
434 ParallelCompactData::create_vspace(size_t count, size_t element_size)
435 {
436 const size_t raw_bytes = count * element_size;
437 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
438 const size_t granularity = os::vm_allocation_granularity();
439 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
440
441 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
442 MAX2(page_sz, granularity);
443 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
444 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
445 rs.size());
446
447 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
448
449 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
450 if (vspace != 0) {
451 if (vspace->expand_by(_reserved_byte_size)) {
452 return vspace;
453 }
454 delete vspace;
455 // Release memory reserved in the space.
456 rs.release();
457 }
458
459 return 0;
460 }
461
462 bool ParallelCompactData::initialize_region_data(size_t region_size)
463 {
464 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
465 _region_vspace = create_vspace(count, sizeof(RegionData));
466 if (_region_vspace != 0) {
467 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
468 _region_count = count;
469 return true;
470 }
471 return false;
472 }
473
474 bool ParallelCompactData::initialize_block_data()
475 {
476 assert(_region_count != 0, "region data must be initialized first");
477 const size_t count = _region_count << Log2BlocksPerRegion;
478 _block_vspace = create_vspace(count, sizeof(BlockData));
479 if (_block_vspace != 0) {
480 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
481 _block_count = count;
482 return true;
483 }
484 return false;
485 }
486
487 void ParallelCompactData::clear()
488 {
489 memset(_region_data, 0, _region_vspace->committed_size());
490 memset(_block_data, 0, _block_vspace->committed_size());
491 }
492
493 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
494 assert(beg_region <= _region_count, "beg_region out of range");
495 assert(end_region <= _region_count, "end_region out of range");
496 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
497
498 const size_t region_cnt = end_region - beg_region;
499 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
500
501 const size_t beg_block = beg_region * BlocksPerRegion;
502 const size_t block_cnt = region_cnt * BlocksPerRegion;
503 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
504 }
505
506 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
507 {
508 const RegionData* cur_cp = region(region_idx);
509 const RegionData* const end_cp = region(region_count() - 1);
510
511 HeapWord* result = region_to_addr(region_idx);
512 if (cur_cp < end_cp) {
513 do {
514 result += cur_cp->partial_obj_size();
515 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
516 }
517 return result;
518 }
519
520 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
521 {
522 const size_t obj_ofs = pointer_delta(addr, _region_start);
523 const size_t beg_region = obj_ofs >> Log2RegionSize;
524 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
525
526 DEBUG_ONLY(Atomic::inc(&add_obj_count);)
527 DEBUG_ONLY(Atomic::add(len, &add_obj_size);)
528
529 if (beg_region == end_region) {
530 // All in one region.
531 _region_data[beg_region].add_live_obj(len);
532 return;
533 }
534
535 // First region.
536 const size_t beg_ofs = region_offset(addr);
537 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
538
539 Klass* klass = ((oop)addr)->klass();
540 // Middle regions--completely spanned by this object.
541 for (size_t region = beg_region + 1; region < end_region; ++region) {
542 _region_data[region].set_partial_obj_size(RegionSize);
543 _region_data[region].set_partial_obj_addr(addr);
544 }
545
546 // Last region.
547 const size_t end_ofs = region_offset(addr + len - 1);
548 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
549 _region_data[end_region].set_partial_obj_addr(addr);
550 }
551
552 void
553 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
554 {
555 assert(region_offset(beg) == 0, "not RegionSize aligned");
556 assert(region_offset(end) == 0, "not RegionSize aligned");
557
558 size_t cur_region = addr_to_region_idx(beg);
559 const size_t end_region = addr_to_region_idx(end);
560 HeapWord* addr = beg;
561 while (cur_region < end_region) {
562 _region_data[cur_region].set_destination(addr);
563 _region_data[cur_region].set_destination_count(0);
564 _region_data[cur_region].set_source_region(cur_region);
565 _region_data[cur_region].set_data_location(addr);
566
567 // Update live_obj_size so the region appears completely full.
568 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
569 _region_data[cur_region].set_live_obj_size(live_size);
570
571 ++cur_region;
572 addr += RegionSize;
573 }
574 }
575
576 // Find the point at which a space can be split and, if necessary, record the
577 // split point.
578 //
579 // If the current src region (which overflowed the destination space) doesn't
580 // have a partial object, the split point is at the beginning of the current src
581 // region (an "easy" split, no extra bookkeeping required).
582 //
583 // If the current src region has a partial object, the split point is in the
584 // region where that partial object starts (call it the split_region). If
585 // split_region has a partial object, then the split point is just after that
586 // partial object (a "hard" split where we have to record the split data and
587 // zero the partial_obj_size field). With a "hard" split, we know that the
588 // partial_obj ends within split_region because the partial object that caused
589 // the overflow starts in split_region. If split_region doesn't have a partial
590 // obj, then the split is at the beginning of split_region (another "easy"
591 // split).
592 HeapWord*
593 ParallelCompactData::summarize_split_space(size_t src_region,
594 SplitInfo& split_info,
595 HeapWord* destination,
596 HeapWord* target_end,
597 HeapWord** target_next)
598 {
599 assert(destination <= target_end, "sanity");
600 assert(destination + _region_data[src_region].data_size() > target_end,
601 "region should not fit into target space");
602 assert(is_region_aligned(target_end), "sanity");
603
604 size_t split_region = src_region;
605 HeapWord* split_destination = destination;
606 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
607
608 if (destination + partial_obj_size > target_end) {
609 // The split point is just after the partial object (if any) in the
610 // src_region that contains the start of the object that overflowed the
611 // destination space.
612 //
613 // Find the start of the "overflow" object and set split_region to the
614 // region containing it.
615 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
616 split_region = addr_to_region_idx(overflow_obj);
617
618 // Clear the source_region field of all destination regions whose first word
619 // came from data after the split point (a non-null source_region field
620 // implies a region must be filled).
621 //
622 // An alternative to the simple loop below: clear during post_compact(),
623 // which uses memcpy instead of individual stores, and is easy to
624 // parallelize. (The downside is that it clears the entire RegionData
625 // object as opposed to just one field.)
626 //
627 // post_compact() would have to clear the summary data up to the highest
628 // address that was written during the summary phase, which would be
629 //
630 // max(top, max(new_top, clear_top))
631 //
632 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
633 // to target_end.
634 const RegionData* const sr = region(split_region);
635 const size_t beg_idx =
636 addr_to_region_idx(region_align_up(sr->destination() +
637 sr->partial_obj_size()));
638 const size_t end_idx = addr_to_region_idx(target_end);
639
640 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
641 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
642 _region_data[idx].set_source_region(0);
643 }
644
645 // Set split_destination and partial_obj_size to reflect the split region.
646 split_destination = sr->destination();
647 partial_obj_size = sr->partial_obj_size();
648 }
649
650 // The split is recorded only if a partial object extends onto the region.
651 if (partial_obj_size != 0) {
652 _region_data[split_region].set_partial_obj_size(0);
653 split_info.record(split_region, partial_obj_size, split_destination);
654 }
655
656 // Setup the continuation addresses.
657 *target_next = split_destination + partial_obj_size;
658 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
659
660 if (log_develop_is_enabled(Trace, gc, compaction)) {
661 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
662 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
663 split_type, p2i(source_next), split_region, partial_obj_size);
664 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
665 split_type, p2i(split_destination),
666 addr_to_region_idx(split_destination),
667 p2i(*target_next));
668
669 if (partial_obj_size != 0) {
670 HeapWord* const po_beg = split_info.destination();
671 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
672 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
673 split_type,
674 p2i(po_beg), addr_to_region_idx(po_beg),
675 p2i(po_end), addr_to_region_idx(po_end));
676 }
677 }
678
679 return source_next;
680 }
681
682 bool ParallelCompactData::summarize(SplitInfo& split_info,
683 HeapWord* source_beg, HeapWord* source_end,
684 HeapWord** source_next,
685 HeapWord* target_beg, HeapWord* target_end,
686 HeapWord** target_next)
687 {
688 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
689 log_develop_trace(gc, compaction)(
690 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
691 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
692 p2i(source_beg), p2i(source_end), p2i(source_next_val),
693 p2i(target_beg), p2i(target_end), p2i(*target_next));
694
695 size_t cur_region = addr_to_region_idx(source_beg);
696 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
697
698 HeapWord *dest_addr = target_beg;
699 while (cur_region < end_region) {
700 // The destination must be set even if the region has no data.
701 _region_data[cur_region].set_destination(dest_addr);
702
703 size_t words = _region_data[cur_region].data_size();
704 if (words > 0) {
705 // If cur_region does not fit entirely into the target space, find a point
706 // at which the source space can be 'split' so that part is copied to the
707 // target space and the rest is copied elsewhere.
708 if (dest_addr + words > target_end) {
709 assert(source_next != NULL, "source_next is NULL when splitting");
710 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
711 target_end, target_next);
712 return false;
713 }
714
715 // Compute the destination_count for cur_region, and if necessary, update
716 // source_region for a destination region. The source_region field is
717 // updated if cur_region is the first (left-most) region to be copied to a
718 // destination region.
719 //
720 // The destination_count calculation is a bit subtle. A region that has
721 // data that compacts into itself does not count itself as a destination.
722 // This maintains the invariant that a zero count means the region is
723 // available and can be claimed and then filled.
724 uint destination_count = 0;
725 if (split_info.is_split(cur_region)) {
726 // The current region has been split: the partial object will be copied
727 // to one destination space and the remaining data will be copied to
728 // another destination space. Adjust the initial destination_count and,
729 // if necessary, set the source_region field if the partial object will
730 // cross a destination region boundary.
731 destination_count = split_info.destination_count();
732 if (destination_count == 2) {
733 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
734 _region_data[dest_idx].set_source_region(cur_region);
735 }
736 }
737
738 HeapWord* const last_addr = dest_addr + words - 1;
739 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
740 const size_t dest_region_2 = addr_to_region_idx(last_addr);
741
742 // Initially assume that the destination regions will be the same and
743 // adjust the value below if necessary. Under this assumption, if
744 // cur_region == dest_region_2, then cur_region will be compacted
745 // completely into itself.
746 destination_count += cur_region == dest_region_2 ? 0 : 1;
747 if (dest_region_1 != dest_region_2) {
748 // Destination regions differ; adjust destination_count.
749 destination_count += 1;
750 // Data from cur_region will be copied to the start of dest_region_2.
751 _region_data[dest_region_2].set_source_region(cur_region);
752 } else if (region_offset(dest_addr) == 0) {
753 // Data from cur_region will be copied to the start of the destination
754 // region.
755 _region_data[dest_region_1].set_source_region(cur_region);
756 }
757
758 _region_data[cur_region].set_destination_count(destination_count);
759 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
760 dest_addr += words;
761 }
762
763 ++cur_region;
764 }
765
766 *target_next = dest_addr;
767 return true;
768 }
769
770 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
771 assert(addr != NULL, "Should detect NULL oop earlier");
772 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
773 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
774
775 // Region covering the object.
776 RegionData* const region_ptr = addr_to_region_ptr(addr);
777 HeapWord* result = region_ptr->destination();
778
779 // If the entire Region is live, the new location is region->destination + the
780 // offset of the object within in the Region.
781
782 // Run some performance tests to determine if this special case pays off. It
783 // is worth it for pointers into the dense prefix. If the optimization to
784 // avoid pointer updates in regions that only point to the dense prefix is
785 // ever implemented, this should be revisited.
786 if (region_ptr->data_size() == RegionSize) {
787 result += region_offset(addr);
788 return result;
789 }
790
791 // Otherwise, the new location is region->destination + block offset + the
792 // number of live words in the Block that are (a) to the left of addr and (b)
793 // due to objects that start in the Block.
794
795 // Fill in the block table if necessary. This is unsynchronized, so multiple
796 // threads may fill the block table for a region (harmless, since it is
797 // idempotent).
798 if (!region_ptr->blocks_filled()) {
799 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
800 region_ptr->set_blocks_filled();
801 }
802
803 HeapWord* const search_start = block_align_down(addr);
804 const size_t block_offset = addr_to_block_ptr(addr)->offset();
805
806 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
807 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr));
808 result += block_offset + live;
809 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
810 return result;
811 }
812
813 #ifdef ASSERT
814 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
815 {
816 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
817 const size_t* const end = (const size_t*)vspace->committed_high_addr();
818 for (const size_t* p = beg; p < end; ++p) {
819 assert(*p == 0, "not zero");
820 }
821 }
822
823 void ParallelCompactData::verify_clear()
824 {
825 verify_clear(_region_vspace);
826 verify_clear(_block_vspace);
827 }
828 #endif // #ifdef ASSERT
829
830 STWGCTimer PSParallelCompact::_gc_timer;
831 ParallelOldTracer PSParallelCompact::_gc_tracer;
832 elapsedTimer PSParallelCompact::_accumulated_time;
833 unsigned int PSParallelCompact::_total_invocations = 0;
834 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
835 jlong PSParallelCompact::_time_of_last_gc = 0;
836 CollectorCounters* PSParallelCompact::_counters = NULL;
837 ParMarkBitMap PSParallelCompact::_mark_bitmap;
838 ParallelCompactData PSParallelCompact::_summary_data;
839
840 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
841
842 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
843
844 void PSParallelCompact::post_initialize() {
845 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
846 MemRegion mr = heap->reserved_region();
847 _ref_processor =
848 new ReferenceProcessor(mr, // span
849 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
850 ParallelGCThreads, // mt processing degree
851 true, // mt discovery
852 ParallelGCThreads, // mt discovery degree
853 true, // atomic_discovery
854 &_is_alive_closure); // non-header is alive closure
855 _counters = new CollectorCounters("PSParallelCompact", 1);
856
857 // Initialize static fields in ParCompactionManager.
858 ParCompactionManager::initialize(mark_bitmap());
859 }
860
861 bool PSParallelCompact::initialize() {
862 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
863 MemRegion mr = heap->reserved_region();
864
865 // Was the old gen get allocated successfully?
866 if (!heap->old_gen()->is_allocated()) {
867 return false;
868 }
869
870 initialize_space_info();
871 initialize_dead_wood_limiter();
872
873 if (!_mark_bitmap.initialize(mr)) {
874 vm_shutdown_during_initialization(
875 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
876 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
877 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
878 return false;
879 }
880
881 if (!_summary_data.initialize(mr)) {
882 vm_shutdown_during_initialization(
883 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
884 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
885 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
886 return false;
887 }
888
889 return true;
890 }
891
892 void PSParallelCompact::initialize_space_info()
893 {
894 memset(&_space_info, 0, sizeof(_space_info));
895
896 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
897 PSYoungGen* young_gen = heap->young_gen();
898
899 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
900 _space_info[eden_space_id].set_space(young_gen->eden_space());
901 _space_info[from_space_id].set_space(young_gen->from_space());
902 _space_info[to_space_id].set_space(young_gen->to_space());
903
904 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
905 }
906
907 void PSParallelCompact::initialize_dead_wood_limiter()
908 {
909 const size_t max = 100;
910 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
911 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
912 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
913 DEBUG_ONLY(_dwl_initialized = true;)
914 _dwl_adjustment = normal_distribution(1.0);
915 }
916
917 void
918 PSParallelCompact::clear_data_covering_space(SpaceId id)
919 {
920 // At this point, top is the value before GC, new_top() is the value that will
921 // be set at the end of GC. The marking bitmap is cleared to top; nothing
922 // should be marked above top. The summary data is cleared to the larger of
923 // top & new_top.
924 MutableSpace* const space = _space_info[id].space();
925 HeapWord* const bot = space->bottom();
926 HeapWord* const top = space->top();
927 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
928
929 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
930 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
931 _mark_bitmap.clear_range(beg_bit, end_bit);
932
933 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
934 const size_t end_region =
935 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
936 _summary_data.clear_range(beg_region, end_region);
937
938 // Clear the data used to 'split' regions.
939 SplitInfo& split_info = _space_info[id].split_info();
940 if (split_info.is_valid()) {
941 split_info.clear();
942 }
943 DEBUG_ONLY(split_info.verify_clear();)
944 }
945
946 void PSParallelCompact::pre_compact()
947 {
948 // Update the from & to space pointers in space_info, since they are swapped
949 // at each young gen gc. Do the update unconditionally (even though a
950 // promotion failure does not swap spaces) because an unknown number of young
951 // collections will have swapped the spaces an unknown number of times.
952 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
953 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
954 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
955 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
956
957 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
958 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
959
960 // Increment the invocation count
961 heap->increment_total_collections(true);
962
963 // We need to track unique mark sweep invocations as well.
964 _total_invocations++;
965
966 heap->print_heap_before_gc();
967 heap->trace_heap_before_gc(&_gc_tracer);
968
969 // Fill in TLABs
970 heap->accumulate_statistics_all_tlabs();
971 heap->ensure_parsability(true); // retire TLABs
972
973 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
974 HandleMark hm; // Discard invalid handles created during verification
975 Universe::verify("Before GC");
976 }
977
978 // Verify object start arrays
979 if (VerifyObjectStartArray &&
980 VerifyBeforeGC) {
981 heap->old_gen()->verify_object_start_array();
982 }
983
984 DEBUG_ONLY(mark_bitmap()->verify_clear();)
985 DEBUG_ONLY(summary_data().verify_clear();)
986
987 // Have worker threads release resources the next time they run a task.
988 gc_task_manager()->release_all_resources();
989
990 ParCompactionManager::reset_all_bitmap_query_caches();
991 }
992
993 void PSParallelCompact::post_compact()
994 {
995 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
996
997 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
998 // Clear the marking bitmap, summary data and split info.
999 clear_data_covering_space(SpaceId(id));
1000 // Update top(). Must be done after clearing the bitmap and summary data.
1001 _space_info[id].publish_new_top();
1002 }
1003
1004 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1005 MutableSpace* const from_space = _space_info[from_space_id].space();
1006 MutableSpace* const to_space = _space_info[to_space_id].space();
1007
1008 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1009 bool eden_empty = eden_space->is_empty();
1010 if (!eden_empty) {
1011 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1012 heap->young_gen(), heap->old_gen());
1013 }
1014
1015 // Update heap occupancy information which is used as input to the soft ref
1016 // clearing policy at the next gc.
1017 Universe::update_heap_info_at_gc();
1018
1019 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1020 to_space->is_empty();
1021
1022 PSCardTable* ct = heap->card_table();
1023 MemRegion old_mr = heap->old_gen()->reserved();
1024 if (young_gen_empty) {
1025 ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1026 } else {
1027 ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1028 }
1029
1030 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1031 ClassLoaderDataGraph::purge();
1032 MetaspaceUtils::verify_metrics();
1033
1034 CodeCache::gc_epilogue();
1035 JvmtiExport::gc_epilogue();
1036
1037 #if COMPILER2_OR_JVMCI
1038 DerivedPointerTable::update_pointers();
1039 #endif
1040
1041 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->num_q());
1042
1043 ref_processor()->enqueue_discovered_references(NULL, &pt);
1044
1045 pt.print_enqueue_phase();
1046
1047 if (ZapUnusedHeapArea) {
1048 heap->gen_mangle_unused_area();
1049 }
1050
1051 // Update time of last GC
1052 reset_millis_since_last_gc();
1053 }
1054
1055 HeapWord*
1056 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1057 bool maximum_compaction)
1058 {
1059 const size_t region_size = ParallelCompactData::RegionSize;
1060 const ParallelCompactData& sd = summary_data();
1061
1062 const MutableSpace* const space = _space_info[id].space();
1063 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1064 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1065 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1066
1067 // Skip full regions at the beginning of the space--they are necessarily part
1068 // of the dense prefix.
1069 size_t full_count = 0;
1070 const RegionData* cp;
1071 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1072 ++full_count;
1073 }
1074
1075 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1076 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1077 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1078 if (maximum_compaction || cp == end_cp || interval_ended) {
1079 _maximum_compaction_gc_num = total_invocations();
1080 return sd.region_to_addr(cp);
1081 }
1082
1083 HeapWord* const new_top = _space_info[id].new_top();
1084 const size_t space_live = pointer_delta(new_top, space->bottom());
1085 const size_t space_used = space->used_in_words();
1086 const size_t space_capacity = space->capacity_in_words();
1087
1088 const double cur_density = double(space_live) / space_capacity;
1089 const double deadwood_density =
1090 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1091 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1092
1093 if (TraceParallelOldGCDensePrefix) {
1094 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1095 cur_density, deadwood_density, deadwood_goal);
1096 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1097 "space_cap=" SIZE_FORMAT,
1098 space_live, space_used,
1099 space_capacity);
1100 }
1101
1102 // XXX - Use binary search?
1103 HeapWord* dense_prefix = sd.region_to_addr(cp);
1104 const RegionData* full_cp = cp;
1105 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1106 while (cp < end_cp) {
1107 HeapWord* region_destination = cp->destination();
1108 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1109 if (TraceParallelOldGCDensePrefix && Verbose) {
1110 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1111 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1112 sd.region(cp), p2i(region_destination),
1113 p2i(dense_prefix), cur_deadwood);
1114 }
1115
1116 if (cur_deadwood >= deadwood_goal) {
1117 // Found the region that has the correct amount of deadwood to the left.
1118 // This typically occurs after crossing a fairly sparse set of regions, so
1119 // iterate backwards over those sparse regions, looking for the region
1120 // that has the lowest density of live objects 'to the right.'
1121 size_t space_to_left = sd.region(cp) * region_size;
1122 size_t live_to_left = space_to_left - cur_deadwood;
1123 size_t space_to_right = space_capacity - space_to_left;
1124 size_t live_to_right = space_live - live_to_left;
1125 double density_to_right = double(live_to_right) / space_to_right;
1126 while (cp > full_cp) {
1127 --cp;
1128 const size_t prev_region_live_to_right = live_to_right -
1129 cp->data_size();
1130 const size_t prev_region_space_to_right = space_to_right + region_size;
1131 double prev_region_density_to_right =
1132 double(prev_region_live_to_right) / prev_region_space_to_right;
1133 if (density_to_right <= prev_region_density_to_right) {
1134 return dense_prefix;
1135 }
1136 if (TraceParallelOldGCDensePrefix && Verbose) {
1137 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1138 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1139 prev_region_density_to_right);
1140 }
1141 dense_prefix -= region_size;
1142 live_to_right = prev_region_live_to_right;
1143 space_to_right = prev_region_space_to_right;
1144 density_to_right = prev_region_density_to_right;
1145 }
1146 return dense_prefix;
1147 }
1148
1149 dense_prefix += region_size;
1150 ++cp;
1151 }
1152
1153 return dense_prefix;
1154 }
1155
1156 #ifndef PRODUCT
1157 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1158 const SpaceId id,
1159 const bool maximum_compaction,
1160 HeapWord* const addr)
1161 {
1162 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1163 RegionData* const cp = summary_data().region(region_idx);
1164 const MutableSpace* const space = _space_info[id].space();
1165 HeapWord* const new_top = _space_info[id].new_top();
1166
1167 const size_t space_live = pointer_delta(new_top, space->bottom());
1168 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1169 const size_t space_cap = space->capacity_in_words();
1170 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1171 const size_t live_to_right = new_top - cp->destination();
1172 const size_t dead_to_right = space->top() - addr - live_to_right;
1173
1174 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1175 "spl=" SIZE_FORMAT " "
1176 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1177 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1178 " ratio=%10.8f",
1179 algorithm, p2i(addr), region_idx,
1180 space_live,
1181 dead_to_left, dead_to_left_pct,
1182 dead_to_right, live_to_right,
1183 double(dead_to_right) / live_to_right);
1184 }
1185 #endif // #ifndef PRODUCT
1186
1187 // Return a fraction indicating how much of the generation can be treated as
1188 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1189 // based on the density of live objects in the generation to determine a limit,
1190 // which is then adjusted so the return value is min_percent when the density is
1191 // 1.
1192 //
1193 // The following table shows some return values for a different values of the
1194 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1195 // min_percent is 1.
1196 //
1197 // fraction allowed as dead wood
1198 // -----------------------------------------------------------------
1199 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1200 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1201 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1202 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1203 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1204 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1205 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1206 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1207 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1208 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1209 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1210 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1211 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1212 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1213 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1214 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1215 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1216 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1217 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1218 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1219 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1220 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1221 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1222
1223 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1224 {
1225 assert(_dwl_initialized, "uninitialized");
1226
1227 // The raw limit is the value of the normal distribution at x = density.
1228 const double raw_limit = normal_distribution(density);
1229
1230 // Adjust the raw limit so it becomes the minimum when the density is 1.
1231 //
1232 // First subtract the adjustment value (which is simply the precomputed value
1233 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1234 // Then add the minimum value, so the minimum is returned when the density is
1235 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1236 const double min = double(min_percent) / 100.0;
1237 const double limit = raw_limit - _dwl_adjustment + min;
1238 return MAX2(limit, 0.0);
1239 }
1240
1241 ParallelCompactData::RegionData*
1242 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1243 const RegionData* end)
1244 {
1245 const size_t region_size = ParallelCompactData::RegionSize;
1246 ParallelCompactData& sd = summary_data();
1247 size_t left = sd.region(beg);
1248 size_t right = end > beg ? sd.region(end) - 1 : left;
1249
1250 // Binary search.
1251 while (left < right) {
1252 // Equivalent to (left + right) / 2, but does not overflow.
1253 const size_t middle = left + (right - left) / 2;
1254 RegionData* const middle_ptr = sd.region(middle);
1255 HeapWord* const dest = middle_ptr->destination();
1256 HeapWord* const addr = sd.region_to_addr(middle);
1257 assert(dest != NULL, "sanity");
1258 assert(dest <= addr, "must move left");
1259
1260 if (middle > left && dest < addr) {
1261 right = middle - 1;
1262 } else if (middle < right && middle_ptr->data_size() == region_size) {
1263 left = middle + 1;
1264 } else {
1265 return middle_ptr;
1266 }
1267 }
1268 return sd.region(left);
1269 }
1270
1271 ParallelCompactData::RegionData*
1272 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1273 const RegionData* end,
1274 size_t dead_words)
1275 {
1276 ParallelCompactData& sd = summary_data();
1277 size_t left = sd.region(beg);
1278 size_t right = end > beg ? sd.region(end) - 1 : left;
1279
1280 // Binary search.
1281 while (left < right) {
1282 // Equivalent to (left + right) / 2, but does not overflow.
1283 const size_t middle = left + (right - left) / 2;
1284 RegionData* const middle_ptr = sd.region(middle);
1285 HeapWord* const dest = middle_ptr->destination();
1286 HeapWord* const addr = sd.region_to_addr(middle);
1287 assert(dest != NULL, "sanity");
1288 assert(dest <= addr, "must move left");
1289
1290 const size_t dead_to_left = pointer_delta(addr, dest);
1291 if (middle > left && dead_to_left > dead_words) {
1292 right = middle - 1;
1293 } else if (middle < right && dead_to_left < dead_words) {
1294 left = middle + 1;
1295 } else {
1296 return middle_ptr;
1297 }
1298 }
1299 return sd.region(left);
1300 }
1301
1302 // The result is valid during the summary phase, after the initial summarization
1303 // of each space into itself, and before final summarization.
1304 inline double
1305 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1306 HeapWord* const bottom,
1307 HeapWord* const top,
1308 HeapWord* const new_top)
1309 {
1310 ParallelCompactData& sd = summary_data();
1311
1312 assert(cp != NULL, "sanity");
1313 assert(bottom != NULL, "sanity");
1314 assert(top != NULL, "sanity");
1315 assert(new_top != NULL, "sanity");
1316 assert(top >= new_top, "summary data problem?");
1317 assert(new_top > bottom, "space is empty; should not be here");
1318 assert(new_top >= cp->destination(), "sanity");
1319 assert(top >= sd.region_to_addr(cp), "sanity");
1320
1321 HeapWord* const destination = cp->destination();
1322 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1323 const size_t compacted_region_live = pointer_delta(new_top, destination);
1324 const size_t compacted_region_used = pointer_delta(top,
1325 sd.region_to_addr(cp));
1326 const size_t reclaimable = compacted_region_used - compacted_region_live;
1327
1328 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1329 return double(reclaimable) / divisor;
1330 }
1331
1332 // Return the address of the end of the dense prefix, a.k.a. the start of the
1333 // compacted region. The address is always on a region boundary.
1334 //
1335 // Completely full regions at the left are skipped, since no compaction can
1336 // occur in those regions. Then the maximum amount of dead wood to allow is
1337 // computed, based on the density (amount live / capacity) of the generation;
1338 // the region with approximately that amount of dead space to the left is
1339 // identified as the limit region. Regions between the last completely full
1340 // region and the limit region are scanned and the one that has the best
1341 // (maximum) reclaimed_ratio() is selected.
1342 HeapWord*
1343 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1344 bool maximum_compaction)
1345 {
1346 const size_t region_size = ParallelCompactData::RegionSize;
1347 const ParallelCompactData& sd = summary_data();
1348
1349 const MutableSpace* const space = _space_info[id].space();
1350 HeapWord* const top = space->top();
1351 HeapWord* const top_aligned_up = sd.region_align_up(top);
1352 HeapWord* const new_top = _space_info[id].new_top();
1353 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1354 HeapWord* const bottom = space->bottom();
1355 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1356 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1357 const RegionData* const new_top_cp =
1358 sd.addr_to_region_ptr(new_top_aligned_up);
1359
1360 // Skip full regions at the beginning of the space--they are necessarily part
1361 // of the dense prefix.
1362 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1363 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1364 space->is_empty(), "no dead space allowed to the left");
1365 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1366 "region must have dead space");
1367
1368 // The gc number is saved whenever a maximum compaction is done, and used to
1369 // determine when the maximum compaction interval has expired. This avoids
1370 // successive max compactions for different reasons.
1371 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1372 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1373 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1374 total_invocations() == HeapFirstMaximumCompactionCount;
1375 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1376 _maximum_compaction_gc_num = total_invocations();
1377 return sd.region_to_addr(full_cp);
1378 }
1379
1380 const size_t space_live = pointer_delta(new_top, bottom);
1381 const size_t space_used = space->used_in_words();
1382 const size_t space_capacity = space->capacity_in_words();
1383
1384 const double density = double(space_live) / double(space_capacity);
1385 const size_t min_percent_free = MarkSweepDeadRatio;
1386 const double limiter = dead_wood_limiter(density, min_percent_free);
1387 const size_t dead_wood_max = space_used - space_live;
1388 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1389 dead_wood_max);
1390
1391 if (TraceParallelOldGCDensePrefix) {
1392 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1393 "space_cap=" SIZE_FORMAT,
1394 space_live, space_used,
1395 space_capacity);
1396 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1397 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1398 density, min_percent_free, limiter,
1399 dead_wood_max, dead_wood_limit);
1400 }
1401
1402 // Locate the region with the desired amount of dead space to the left.
1403 const RegionData* const limit_cp =
1404 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1405
1406 // Scan from the first region with dead space to the limit region and find the
1407 // one with the best (largest) reclaimed ratio.
1408 double best_ratio = 0.0;
1409 const RegionData* best_cp = full_cp;
1410 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1411 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1412 if (tmp_ratio > best_ratio) {
1413 best_cp = cp;
1414 best_ratio = tmp_ratio;
1415 }
1416 }
1417
1418 return sd.region_to_addr(best_cp);
1419 }
1420
1421 void PSParallelCompact::summarize_spaces_quick()
1422 {
1423 for (unsigned int i = 0; i < last_space_id; ++i) {
1424 const MutableSpace* space = _space_info[i].space();
1425 HeapWord** nta = _space_info[i].new_top_addr();
1426 bool result = _summary_data.summarize(_space_info[i].split_info(),
1427 space->bottom(), space->top(), NULL,
1428 space->bottom(), space->end(), nta);
1429 assert(result, "space must fit into itself");
1430 _space_info[i].set_dense_prefix(space->bottom());
1431 }
1432 }
1433
1434 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1435 {
1436 HeapWord* const dense_prefix_end = dense_prefix(id);
1437 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1438 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1439 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1440 // Only enough dead space is filled so that any remaining dead space to the
1441 // left is larger than the minimum filler object. (The remainder is filled
1442 // during the copy/update phase.)
1443 //
1444 // The size of the dead space to the right of the boundary is not a
1445 // concern, since compaction will be able to use whatever space is
1446 // available.
1447 //
1448 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1449 // surrounds the space to be filled with an object.
1450 //
1451 // In the 32-bit VM, each bit represents two 32-bit words:
1452 // +---+
1453 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1454 // end_bits: ... x x x | 0 | || 0 x x ...
1455 // +---+
1456 //
1457 // In the 64-bit VM, each bit represents one 64-bit word:
1458 // +------------+
1459 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1460 // end_bits: ... x x 1 | 0 || 0 | x x ...
1461 // +------------+
1462 // +-------+
1463 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1464 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1465 // +-------+
1466 // +-----------+
1467 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1468 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1469 // +-----------+
1470 // +-------+
1471 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1472 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1473 // +-------+
1474
1475 // Initially assume case a, c or e will apply.
1476 size_t obj_len = CollectedHeap::min_fill_size();
1477 HeapWord* obj_beg = dense_prefix_end - obj_len;
1478
1479 #ifdef _LP64
1480 if (MinObjAlignment > 1) { // object alignment > heap word size
1481 // Cases a, c or e.
1482 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1483 // Case b above.
1484 obj_beg = dense_prefix_end - 1;
1485 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1486 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1487 // Case d above.
1488 obj_beg = dense_prefix_end - 3;
1489 obj_len = 3;
1490 }
1491 #endif // #ifdef _LP64
1492
1493 CollectedHeap::fill_with_object(obj_beg, obj_len);
1494 _mark_bitmap.mark_obj(obj_beg, obj_len);
1495 _summary_data.add_obj(obj_beg, obj_len);
1496 assert(start_array(id) != NULL, "sanity");
1497 start_array(id)->allocate_block(obj_beg);
1498 }
1499 }
1500
1501 void
1502 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1503 {
1504 assert(id < last_space_id, "id out of range");
1505 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1506 "should have been reset in summarize_spaces_quick()");
1507
1508 const MutableSpace* space = _space_info[id].space();
1509 if (_space_info[id].new_top() != space->bottom()) {
1510 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1511 _space_info[id].set_dense_prefix(dense_prefix_end);
1512
1513 #ifndef PRODUCT
1514 if (TraceParallelOldGCDensePrefix) {
1515 print_dense_prefix_stats("ratio", id, maximum_compaction,
1516 dense_prefix_end);
1517 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1518 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1519 }
1520 #endif // #ifndef PRODUCT
1521
1522 // Recompute the summary data, taking into account the dense prefix. If
1523 // every last byte will be reclaimed, then the existing summary data which
1524 // compacts everything can be left in place.
1525 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1526 // If dead space crosses the dense prefix boundary, it is (at least
1527 // partially) filled with a dummy object, marked live and added to the
1528 // summary data. This simplifies the copy/update phase and must be done
1529 // before the final locations of objects are determined, to prevent
1530 // leaving a fragment of dead space that is too small to fill.
1531 fill_dense_prefix_end(id);
1532
1533 // Compute the destination of each Region, and thus each object.
1534 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1535 _summary_data.summarize(_space_info[id].split_info(),
1536 dense_prefix_end, space->top(), NULL,
1537 dense_prefix_end, space->end(),
1538 _space_info[id].new_top_addr());
1539 }
1540 }
1541
1542 if (log_develop_is_enabled(Trace, gc, compaction)) {
1543 const size_t region_size = ParallelCompactData::RegionSize;
1544 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1545 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1546 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1547 HeapWord* const new_top = _space_info[id].new_top();
1548 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1549 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1550 log_develop_trace(gc, compaction)(
1551 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1552 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1553 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1554 id, space->capacity_in_words(), p2i(dense_prefix_end),
1555 dp_region, dp_words / region_size,
1556 cr_words / region_size, p2i(new_top));
1557 }
1558 }
1559
1560 #ifndef PRODUCT
1561 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1562 HeapWord* dst_beg, HeapWord* dst_end,
1563 SpaceId src_space_id,
1564 HeapWord* src_beg, HeapWord* src_end)
1565 {
1566 log_develop_trace(gc, compaction)(
1567 "Summarizing %d [%s] into %d [%s]: "
1568 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1569 SIZE_FORMAT "-" SIZE_FORMAT " "
1570 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1571 SIZE_FORMAT "-" SIZE_FORMAT,
1572 src_space_id, space_names[src_space_id],
1573 dst_space_id, space_names[dst_space_id],
1574 p2i(src_beg), p2i(src_end),
1575 _summary_data.addr_to_region_idx(src_beg),
1576 _summary_data.addr_to_region_idx(src_end),
1577 p2i(dst_beg), p2i(dst_end),
1578 _summary_data.addr_to_region_idx(dst_beg),
1579 _summary_data.addr_to_region_idx(dst_end));
1580 }
1581 #endif // #ifndef PRODUCT
1582
1583 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1584 bool maximum_compaction)
1585 {
1586 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1587
1588 #ifdef ASSERT
1589 if (TraceParallelOldGCMarkingPhase) {
1590 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1591 "add_obj_bytes=" SIZE_FORMAT,
1592 add_obj_count, add_obj_size * HeapWordSize);
1593 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1594 "mark_bitmap_bytes=" SIZE_FORMAT,
1595 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1596 }
1597 #endif // #ifdef ASSERT
1598
1599 // Quick summarization of each space into itself, to see how much is live.
1600 summarize_spaces_quick();
1601
1602 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1603 NOT_PRODUCT(print_region_ranges());
1604 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1605
1606 // The amount of live data that will end up in old space (assuming it fits).
1607 size_t old_space_total_live = 0;
1608 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1609 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1610 _space_info[id].space()->bottom());
1611 }
1612
1613 MutableSpace* const old_space = _space_info[old_space_id].space();
1614 const size_t old_capacity = old_space->capacity_in_words();
1615 if (old_space_total_live > old_capacity) {
1616 // XXX - should also try to expand
1617 maximum_compaction = true;
1618 }
1619
1620 // Old generations.
1621 summarize_space(old_space_id, maximum_compaction);
1622
1623 // Summarize the remaining spaces in the young gen. The initial target space
1624 // is the old gen. If a space does not fit entirely into the target, then the
1625 // remainder is compacted into the space itself and that space becomes the new
1626 // target.
1627 SpaceId dst_space_id = old_space_id;
1628 HeapWord* dst_space_end = old_space->end();
1629 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1630 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1631 const MutableSpace* space = _space_info[id].space();
1632 const size_t live = pointer_delta(_space_info[id].new_top(),
1633 space->bottom());
1634 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1635
1636 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1637 SpaceId(id), space->bottom(), space->top());)
1638 if (live > 0 && live <= available) {
1639 // All the live data will fit.
1640 bool done = _summary_data.summarize(_space_info[id].split_info(),
1641 space->bottom(), space->top(),
1642 NULL,
1643 *new_top_addr, dst_space_end,
1644 new_top_addr);
1645 assert(done, "space must fit into old gen");
1646
1647 // Reset the new_top value for the space.
1648 _space_info[id].set_new_top(space->bottom());
1649 } else if (live > 0) {
1650 // Attempt to fit part of the source space into the target space.
1651 HeapWord* next_src_addr = NULL;
1652 bool done = _summary_data.summarize(_space_info[id].split_info(),
1653 space->bottom(), space->top(),
1654 &next_src_addr,
1655 *new_top_addr, dst_space_end,
1656 new_top_addr);
1657 assert(!done, "space should not fit into old gen");
1658 assert(next_src_addr != NULL, "sanity");
1659
1660 // The source space becomes the new target, so the remainder is compacted
1661 // within the space itself.
1662 dst_space_id = SpaceId(id);
1663 dst_space_end = space->end();
1664 new_top_addr = _space_info[id].new_top_addr();
1665 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1666 space->bottom(), dst_space_end,
1667 SpaceId(id), next_src_addr, space->top());)
1668 done = _summary_data.summarize(_space_info[id].split_info(),
1669 next_src_addr, space->top(),
1670 NULL,
1671 space->bottom(), dst_space_end,
1672 new_top_addr);
1673 assert(done, "space must fit when compacted into itself");
1674 assert(*new_top_addr <= space->top(), "usage should not grow");
1675 }
1676 }
1677
1678 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1679 NOT_PRODUCT(print_region_ranges());
1680 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1681 }
1682
1683 // This method should contain all heap-specific policy for invoking a full
1684 // collection. invoke_no_policy() will only attempt to compact the heap; it
1685 // will do nothing further. If we need to bail out for policy reasons, scavenge
1686 // before full gc, or any other specialized behavior, it needs to be added here.
1687 //
1688 // Note that this method should only be called from the vm_thread while at a
1689 // safepoint.
1690 //
1691 // Note that the all_soft_refs_clear flag in the collector policy
1692 // may be true because this method can be called without intervening
1693 // activity. For example when the heap space is tight and full measure
1694 // are being taken to free space.
1695 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1696 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1697 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1698 "should be in vm thread");
1699
1700 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1701 GCCause::Cause gc_cause = heap->gc_cause();
1702 assert(!heap->is_gc_active(), "not reentrant");
1703
1704 PSAdaptiveSizePolicy* policy = heap->size_policy();
1705 IsGCActiveMark mark;
1706
1707 if (ScavengeBeforeFullGC) {
1708 PSScavenge::invoke_no_policy();
1709 }
1710
1711 const bool clear_all_soft_refs =
1712 heap->soft_ref_policy()->should_clear_all_soft_refs();
1713
1714 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1715 maximum_heap_compaction);
1716 }
1717
1718 // This method contains no policy. You should probably
1719 // be calling invoke() instead.
1720 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1721 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1722 assert(ref_processor() != NULL, "Sanity");
1723
1724 if (GCLocker::check_active_before_gc()) {
1725 return false;
1726 }
1727
1728 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1729
1730 GCIdMark gc_id_mark;
1731 _gc_timer.register_gc_start();
1732 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1733
1734 TimeStamp marking_start;
1735 TimeStamp compaction_start;
1736 TimeStamp collection_exit;
1737
1738 GCCause::Cause gc_cause = heap->gc_cause();
1739 PSYoungGen* young_gen = heap->young_gen();
1740 PSOldGen* old_gen = heap->old_gen();
1741 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1742
1743 // The scope of casr should end after code that can change
1744 // CollectorPolicy::_should_clear_all_soft_refs.
1745 ClearedAllSoftRefs casr(maximum_heap_compaction,
1746 heap->soft_ref_policy());
1747
1748 if (ZapUnusedHeapArea) {
1749 // Save information needed to minimize mangling
1750 heap->record_gen_tops_before_GC();
1751 }
1752
1753 // Make sure data structures are sane, make the heap parsable, and do other
1754 // miscellaneous bookkeeping.
1755 pre_compact();
1756
1757 PreGCValues pre_gc_values(heap);
1758
1759 // Get the compaction manager reserved for the VM thread.
1760 ParCompactionManager* const vmthread_cm =
1761 ParCompactionManager::manager_array(gc_task_manager()->workers());
1762
1763 {
1764 ResourceMark rm;
1765 HandleMark hm;
1766
1767 // Set the number of GC threads to be used in this collection
1768 gc_task_manager()->set_active_gang();
1769 gc_task_manager()->task_idle_workers();
1770
1771 GCTraceCPUTime tcpu;
1772 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1773
1774 heap->pre_full_gc_dump(&_gc_timer);
1775
1776 TraceCollectorStats tcs(counters());
1777 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1778
1779 if (log_is_enabled(Debug, gc, heap, exit)) {
1780 accumulated_time()->start();
1781 }
1782
1783 // Let the size policy know we're starting
1784 size_policy->major_collection_begin();
1785
1786 CodeCache::gc_prologue();
1787
1788 #if COMPILER2_OR_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 COMPILER2_OR_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->soft_ref_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 (log_is_enabled(Debug, gc, heap, exit)) {
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 MetaspaceUtils::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("Weak Processing", &_gc_timer);
2129 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl);
2130 }
2131
2132 {
2133 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2134
2135 // Follow system dictionary roots and unload classes.
2136 bool purged_class = SystemDictionary::do_unloading(is_alive_closure(), &_gc_timer);
2137
2138 // Unload nmethods.
2139 CodeCache::do_unloading(is_alive_closure(), purged_class);
2140
2141 // Prune dead klasses from subklass/sibling/implementor lists.
2142 Klass::clean_weak_klass_links();
2143 }
2144
2145 {
2146 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer);
2147 // Delete entries for dead interned strings.
2148 StringTable::unlink(is_alive_closure());
2149 }
2150
2151 {
2152 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer);
2153 // Clean up unreferenced symbols in symbol table.
2154 SymbolTable::unlink();
2155 }
2156
2157 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2158 }
2159
2160 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2161 // Adjust the pointers to reflect the new locations
2162 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2163
2164 // Need new claim bits when tracing through and adjusting pointers.
2165 ClassLoaderDataGraph::clear_claimed_marks();
2166
2167 PSParallelCompact::AdjustPointerClosure oop_closure(cm);
2168
2169 // General strong roots.
2170 Universe::oops_do(&oop_closure);
2171 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2172 Threads::oops_do(&oop_closure, NULL);
2173 ObjectSynchronizer::oops_do(&oop_closure);
2174 Management::oops_do(&oop_closure);
2175 JvmtiExport::oops_do(&oop_closure);
2176 SystemDictionary::oops_do(&oop_closure);
2177 ClassLoaderDataGraph::oops_do(&oop_closure, true);
2178
2179 // Now adjust pointers in remaining weak roots. (All of which should
2180 // have been cleared if they pointed to non-surviving objects.)
2181 WeakProcessor::oops_do(&oop_closure);
2182
2183 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2184 CodeCache::blobs_do(&adjust_from_blobs);
2185 AOTLoader::oops_do(&oop_closure);
2186 StringTable::oops_do(&oop_closure);
2187 ref_processor()->weak_oops_do(&oop_closure);
2188 // Roots were visited so references into the young gen in roots
2189 // may have been scanned. Process them also.
2190 // Should the reference processor have a span that excludes
2191 // young gen objects?
2192 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2193 }
2194
2195 // Helper class to print 8 region numbers per line and then print the total at the end.
2196 class FillableRegionLogger : public StackObj {
2197 private:
2198 Log(gc, compaction) log;
2199 static const int LineLength = 8;
2200 size_t _regions[LineLength];
2201 int _next_index;
2202 bool _enabled;
2203 size_t _total_regions;
2204 public:
2205 FillableRegionLogger() : _next_index(0), _total_regions(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)) { }
2206 ~FillableRegionLogger() {
2207 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2208 }
2209
2210 void print_line() {
2211 if (!_enabled || _next_index == 0) {
2212 return;
2213 }
2214 FormatBuffer<> line("Fillable: ");
2215 for (int i = 0; i < _next_index; i++) {
2216 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2217 }
2218 log.trace("%s", line.buffer());
2219 _next_index = 0;
2220 }
2221
2222 void handle(size_t region) {
2223 if (!_enabled) {
2224 return;
2225 }
2226 _regions[_next_index++] = region;
2227 if (_next_index == LineLength) {
2228 print_line();
2229 }
2230 _total_regions++;
2231 }
2232 };
2233
2234 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q,
2235 uint parallel_gc_threads)
2236 {
2237 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2238
2239 // Find the threads that are active
2240 unsigned int which = 0;
2241
2242 // Find all regions that are available (can be filled immediately) and
2243 // distribute them to the thread stacks. The iteration is done in reverse
2244 // order (high to low) so the regions will be removed in ascending order.
2245
2246 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2247
2248 which = 0;
2249 // id + 1 is used to test termination so unsigned can
2250 // be used with an old_space_id == 0.
2251 FillableRegionLogger region_logger;
2252 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2253 SpaceInfo* const space_info = _space_info + id;
2254 MutableSpace* const space = space_info->space();
2255 HeapWord* const new_top = space_info->new_top();
2256
2257 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2258 const size_t end_region =
2259 sd.addr_to_region_idx(sd.region_align_up(new_top));
2260
2261 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2262 if (sd.region(cur)->claim_unsafe()) {
2263 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2264 cm->region_stack()->push(cur);
2265 region_logger.handle(cur);
2266 // Assign regions to tasks in round-robin fashion.
2267 if (++which == parallel_gc_threads) {
2268 which = 0;
2269 }
2270 }
2271 }
2272 region_logger.print_line();
2273 }
2274 }
2275
2276 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2277
2278 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2279 uint parallel_gc_threads) {
2280 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2281
2282 ParallelCompactData& sd = PSParallelCompact::summary_data();
2283
2284 // Iterate over all the spaces adding tasks for updating
2285 // regions in the dense prefix. Assume that 1 gc thread
2286 // will work on opening the gaps and the remaining gc threads
2287 // will work on the dense prefix.
2288 unsigned int space_id;
2289 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2290 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2291 const MutableSpace* const space = _space_info[space_id].space();
2292
2293 if (dense_prefix_end == space->bottom()) {
2294 // There is no dense prefix for this space.
2295 continue;
2296 }
2297
2298 // The dense prefix is before this region.
2299 size_t region_index_end_dense_prefix =
2300 sd.addr_to_region_idx(dense_prefix_end);
2301 RegionData* const dense_prefix_cp =
2302 sd.region(region_index_end_dense_prefix);
2303 assert(dense_prefix_end == space->end() ||
2304 dense_prefix_cp->available() ||
2305 dense_prefix_cp->claimed(),
2306 "The region after the dense prefix should always be ready to fill");
2307
2308 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2309
2310 // Is there dense prefix work?
2311 size_t total_dense_prefix_regions =
2312 region_index_end_dense_prefix - region_index_start;
2313 // How many regions of the dense prefix should be given to
2314 // each thread?
2315 if (total_dense_prefix_regions > 0) {
2316 uint tasks_for_dense_prefix = 1;
2317 if (total_dense_prefix_regions <=
2318 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2319 // Don't over partition. This assumes that
2320 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2321 // so there are not many regions to process.
2322 tasks_for_dense_prefix = parallel_gc_threads;
2323 } else {
2324 // Over partition
2325 tasks_for_dense_prefix = parallel_gc_threads *
2326 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2327 }
2328 size_t regions_per_thread = total_dense_prefix_regions /
2329 tasks_for_dense_prefix;
2330 // Give each thread at least 1 region.
2331 if (regions_per_thread == 0) {
2332 regions_per_thread = 1;
2333 }
2334
2335 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2336 if (region_index_start >= region_index_end_dense_prefix) {
2337 break;
2338 }
2339 // region_index_end is not processed
2340 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2341 region_index_end_dense_prefix);
2342 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2343 region_index_start,
2344 region_index_end));
2345 region_index_start = region_index_end;
2346 }
2347 }
2348 // This gets any part of the dense prefix that did not
2349 // fit evenly.
2350 if (region_index_start < region_index_end_dense_prefix) {
2351 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2352 region_index_start,
2353 region_index_end_dense_prefix));
2354 }
2355 }
2356 }
2357
2358 void PSParallelCompact::enqueue_region_stealing_tasks(
2359 GCTaskQueue* q,
2360 ParallelTaskTerminator* terminator_ptr,
2361 uint parallel_gc_threads) {
2362 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer);
2363
2364 // Once a thread has drained it's stack, it should try to steal regions from
2365 // other threads.
2366 for (uint j = 0; j < parallel_gc_threads; j++) {
2367 q->enqueue(new CompactionWithStealingTask(terminator_ptr));
2368 }
2369 }
2370
2371 #ifdef ASSERT
2372 // Write a histogram of the number of times the block table was filled for a
2373 // region.
2374 void PSParallelCompact::write_block_fill_histogram()
2375 {
2376 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2377 return;
2378 }
2379
2380 Log(gc, compaction) log;
2381 ResourceMark rm;
2382 LogStream ls(log.trace());
2383 outputStream* out = &ls;
2384
2385 typedef ParallelCompactData::RegionData rd_t;
2386 ParallelCompactData& sd = summary_data();
2387
2388 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2389 MutableSpace* const spc = _space_info[id].space();
2390 if (spc->bottom() != spc->top()) {
2391 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2392 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2393 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2394
2395 size_t histo[5] = { 0, 0, 0, 0, 0 };
2396 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2397 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2398
2399 for (const rd_t* cur = beg; cur < end; ++cur) {
2400 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2401 }
2402 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2403 for (size_t i = 0; i < histo_len; ++i) {
2404 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2405 histo[i], 100.0 * histo[i] / region_cnt);
2406 }
2407 out->cr();
2408 }
2409 }
2410 }
2411 #endif // #ifdef ASSERT
2412
2413 void PSParallelCompact::compact() {
2414 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2415
2416 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2417 PSOldGen* old_gen = heap->old_gen();
2418 old_gen->start_array()->reset();
2419 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2420 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2421 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2422 ParallelTaskTerminator terminator(active_gc_threads, qset);
2423
2424 GCTaskQueue* q = GCTaskQueue::create();
2425 prepare_region_draining_tasks(q, active_gc_threads);
2426 enqueue_dense_prefix_tasks(q, active_gc_threads);
2427 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2428
2429 {
2430 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2431
2432 gc_task_manager()->execute_and_wait(q);
2433
2434 #ifdef ASSERT
2435 // Verify that all regions have been processed before the deferred updates.
2436 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2437 verify_complete(SpaceId(id));
2438 }
2439 #endif
2440 }
2441
2442 {
2443 // Update the deferred objects, if any. Any compaction manager can be used.
2444 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2445 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2446 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2447 update_deferred_objects(cm, SpaceId(id));
2448 }
2449 }
2450
2451 DEBUG_ONLY(write_block_fill_histogram());
2452 }
2453
2454 #ifdef ASSERT
2455 void PSParallelCompact::verify_complete(SpaceId space_id) {
2456 // All Regions between space bottom() to new_top() should be marked as filled
2457 // and all Regions between new_top() and top() should be available (i.e.,
2458 // should have been emptied).
2459 ParallelCompactData& sd = summary_data();
2460 SpaceInfo si = _space_info[space_id];
2461 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2462 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2463 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2464 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2465 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2466
2467 bool issued_a_warning = false;
2468
2469 size_t cur_region;
2470 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2471 const RegionData* const c = sd.region(cur_region);
2472 if (!c->completed()) {
2473 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2474 cur_region, c->destination_count());
2475 issued_a_warning = true;
2476 }
2477 }
2478
2479 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2480 const RegionData* const c = sd.region(cur_region);
2481 if (!c->available()) {
2482 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2483 cur_region, c->destination_count());
2484 issued_a_warning = true;
2485 }
2486 }
2487
2488 if (issued_a_warning) {
2489 print_region_ranges();
2490 }
2491 }
2492 #endif // #ifdef ASSERT
2493
2494 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2495 _start_array->allocate_block(addr);
2496 compaction_manager()->update_contents(oop(addr));
2497 }
2498
2499 // Update interior oops in the ranges of regions [beg_region, end_region).
2500 void
2501 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2502 SpaceId space_id,
2503 size_t beg_region,
2504 size_t end_region) {
2505 ParallelCompactData& sd = summary_data();
2506 ParMarkBitMap* const mbm = mark_bitmap();
2507
2508 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2509 HeapWord* const end_addr = sd.region_to_addr(end_region);
2510 assert(beg_region <= end_region, "bad region range");
2511 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2512
2513 #ifdef ASSERT
2514 // Claim the regions to avoid triggering an assert when they are marked as
2515 // filled.
2516 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2517 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2518 }
2519 #endif // #ifdef ASSERT
2520
2521 if (beg_addr != space(space_id)->bottom()) {
2522 // Find the first live object or block of dead space that *starts* in this
2523 // range of regions. If a partial object crosses onto the region, skip it;
2524 // it will be marked for 'deferred update' when the object head is
2525 // processed. If dead space crosses onto the region, it is also skipped; it
2526 // will be filled when the prior region is processed. If neither of those
2527 // apply, the first word in the region is the start of a live object or dead
2528 // space.
2529 assert(beg_addr > space(space_id)->bottom(), "sanity");
2530 const RegionData* const cp = sd.region(beg_region);
2531 if (cp->partial_obj_size() != 0) {
2532 beg_addr = sd.partial_obj_end(beg_region);
2533 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2534 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2535 }
2536 }
2537
2538 if (beg_addr < end_addr) {
2539 // A live object or block of dead space starts in this range of Regions.
2540 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2541
2542 // Create closures and iterate.
2543 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2544 FillClosure fill_closure(cm, space_id);
2545 ParMarkBitMap::IterationStatus status;
2546 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2547 dense_prefix_end);
2548 if (status == ParMarkBitMap::incomplete) {
2549 update_closure.do_addr(update_closure.source());
2550 }
2551 }
2552
2553 // Mark the regions as filled.
2554 RegionData* const beg_cp = sd.region(beg_region);
2555 RegionData* const end_cp = sd.region(end_region);
2556 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2557 cp->set_completed();
2558 }
2559 }
2560
2561 // Return the SpaceId for the space containing addr. If addr is not in the
2562 // heap, last_space_id is returned. In debug mode it expects the address to be
2563 // in the heap and asserts such.
2564 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2565 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2566
2567 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2568 if (_space_info[id].space()->contains(addr)) {
2569 return SpaceId(id);
2570 }
2571 }
2572
2573 assert(false, "no space contains the addr");
2574 return last_space_id;
2575 }
2576
2577 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2578 SpaceId id) {
2579 assert(id < last_space_id, "bad space id");
2580
2581 ParallelCompactData& sd = summary_data();
2582 const SpaceInfo* const space_info = _space_info + id;
2583 ObjectStartArray* const start_array = space_info->start_array();
2584
2585 const MutableSpace* const space = space_info->space();
2586 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2587 HeapWord* const beg_addr = space_info->dense_prefix();
2588 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2589
2590 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2591 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2592 const RegionData* cur_region;
2593 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2594 HeapWord* const addr = cur_region->deferred_obj_addr();
2595 if (addr != NULL) {
2596 if (start_array != NULL) {
2597 start_array->allocate_block(addr);
2598 }
2599 cm->update_contents(oop(addr));
2600 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2601 }
2602 }
2603 }
2604
2605 // Skip over count live words starting from beg, and return the address of the
2606 // next live word. Unless marked, the word corresponding to beg is assumed to
2607 // be dead. Callers must either ensure beg does not correspond to the middle of
2608 // an object, or account for those live words in some other way. Callers must
2609 // also ensure that there are enough live words in the range [beg, end) to skip.
2610 HeapWord*
2611 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2612 {
2613 assert(count > 0, "sanity");
2614
2615 ParMarkBitMap* m = mark_bitmap();
2616 idx_t bits_to_skip = m->words_to_bits(count);
2617 idx_t cur_beg = m->addr_to_bit(beg);
2618 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2619
2620 do {
2621 cur_beg = m->find_obj_beg(cur_beg, search_end);
2622 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2623 const size_t obj_bits = cur_end - cur_beg + 1;
2624 if (obj_bits > bits_to_skip) {
2625 return m->bit_to_addr(cur_beg + bits_to_skip);
2626 }
2627 bits_to_skip -= obj_bits;
2628 cur_beg = cur_end + 1;
2629 } while (bits_to_skip > 0);
2630
2631 // Skipping the desired number of words landed just past the end of an object.
2632 // Find the start of the next object.
2633 cur_beg = m->find_obj_beg(cur_beg, search_end);
2634 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2635 return m->bit_to_addr(cur_beg);
2636 }
2637
2638 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2639 SpaceId src_space_id,
2640 size_t src_region_idx)
2641 {
2642 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2643
2644 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2645 if (split_info.dest_region_addr() == dest_addr) {
2646 // The partial object ending at the split point contains the first word to
2647 // be copied to dest_addr.
2648 return split_info.first_src_addr();
2649 }
2650
2651 const ParallelCompactData& sd = summary_data();
2652 ParMarkBitMap* const bitmap = mark_bitmap();
2653 const size_t RegionSize = ParallelCompactData::RegionSize;
2654
2655 assert(sd.is_region_aligned(dest_addr), "not aligned");
2656 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2657 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2658 HeapWord* const src_region_destination = src_region_ptr->destination();
2659
2660 assert(dest_addr >= src_region_destination, "wrong src region");
2661 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2662
2663 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2664 HeapWord* const src_region_end = src_region_beg + RegionSize;
2665
2666 HeapWord* addr = src_region_beg;
2667 if (dest_addr == src_region_destination) {
2668 // Return the first live word in the source region.
2669 if (partial_obj_size == 0) {
2670 addr = bitmap->find_obj_beg(addr, src_region_end);
2671 assert(addr < src_region_end, "no objects start in src region");
2672 }
2673 return addr;
2674 }
2675
2676 // Must skip some live data.
2677 size_t words_to_skip = dest_addr - src_region_destination;
2678 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2679
2680 if (partial_obj_size >= words_to_skip) {
2681 // All the live words to skip are part of the partial object.
2682 addr += words_to_skip;
2683 if (partial_obj_size == words_to_skip) {
2684 // Find the first live word past the partial object.
2685 addr = bitmap->find_obj_beg(addr, src_region_end);
2686 assert(addr < src_region_end, "wrong src region");
2687 }
2688 return addr;
2689 }
2690
2691 // Skip over the partial object (if any).
2692 if (partial_obj_size != 0) {
2693 words_to_skip -= partial_obj_size;
2694 addr += partial_obj_size;
2695 }
2696
2697 // Skip over live words due to objects that start in the region.
2698 addr = skip_live_words(addr, src_region_end, words_to_skip);
2699 assert(addr < src_region_end, "wrong src region");
2700 return addr;
2701 }
2702
2703 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2704 SpaceId src_space_id,
2705 size_t beg_region,
2706 HeapWord* end_addr)
2707 {
2708 ParallelCompactData& sd = summary_data();
2709
2710 #ifdef ASSERT
2711 MutableSpace* const src_space = _space_info[src_space_id].space();
2712 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2713 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2714 "src_space_id does not match beg_addr");
2715 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2716 "src_space_id does not match end_addr");
2717 #endif // #ifdef ASSERT
2718
2719 RegionData* const beg = sd.region(beg_region);
2720 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2721
2722 // Regions up to new_top() are enqueued if they become available.
2723 HeapWord* const new_top = _space_info[src_space_id].new_top();
2724 RegionData* const enqueue_end =
2725 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2726
2727 for (RegionData* cur = beg; cur < end; ++cur) {
2728 assert(cur->data_size() > 0, "region must have live data");
2729 cur->decrement_destination_count();
2730 if (cur < enqueue_end && cur->available() && cur->claim()) {
2731 cm->push_region(sd.region(cur));
2732 }
2733 }
2734 }
2735
2736 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2737 SpaceId& src_space_id,
2738 HeapWord*& src_space_top,
2739 HeapWord* end_addr)
2740 {
2741 typedef ParallelCompactData::RegionData RegionData;
2742
2743 ParallelCompactData& sd = PSParallelCompact::summary_data();
2744 const size_t region_size = ParallelCompactData::RegionSize;
2745
2746 size_t src_region_idx = 0;
2747
2748 // Skip empty regions (if any) up to the top of the space.
2749 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2750 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2751 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2752 const RegionData* const top_region_ptr =
2753 sd.addr_to_region_ptr(top_aligned_up);
2754 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2755 ++src_region_ptr;
2756 }
2757
2758 if (src_region_ptr < top_region_ptr) {
2759 // The next source region is in the current space. Update src_region_idx
2760 // and the source address to match src_region_ptr.
2761 src_region_idx = sd.region(src_region_ptr);
2762 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2763 if (src_region_addr > closure.source()) {
2764 closure.set_source(src_region_addr);
2765 }
2766 return src_region_idx;
2767 }
2768
2769 // Switch to a new source space and find the first non-empty region.
2770 unsigned int space_id = src_space_id + 1;
2771 assert(space_id < last_space_id, "not enough spaces");
2772
2773 HeapWord* const destination = closure.destination();
2774
2775 do {
2776 MutableSpace* space = _space_info[space_id].space();
2777 HeapWord* const bottom = space->bottom();
2778 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2779
2780 // Iterate over the spaces that do not compact into themselves.
2781 if (bottom_cp->destination() != bottom) {
2782 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2783 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2784
2785 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2786 if (src_cp->live_obj_size() > 0) {
2787 // Found it.
2788 assert(src_cp->destination() == destination,
2789 "first live obj in the space must match the destination");
2790 assert(src_cp->partial_obj_size() == 0,
2791 "a space cannot begin with a partial obj");
2792
2793 src_space_id = SpaceId(space_id);
2794 src_space_top = space->top();
2795 const size_t src_region_idx = sd.region(src_cp);
2796 closure.set_source(sd.region_to_addr(src_region_idx));
2797 return src_region_idx;
2798 } else {
2799 assert(src_cp->data_size() == 0, "sanity");
2800 }
2801 }
2802 }
2803 } while (++space_id < last_space_id);
2804
2805 assert(false, "no source region was found");
2806 return 0;
2807 }
2808
2809 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
2810 {
2811 typedef ParMarkBitMap::IterationStatus IterationStatus;
2812 const size_t RegionSize = ParallelCompactData::RegionSize;
2813 ParMarkBitMap* const bitmap = mark_bitmap();
2814 ParallelCompactData& sd = summary_data();
2815 RegionData* const region_ptr = sd.region(region_idx);
2816
2817 // Get the items needed to construct the closure.
2818 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2819 SpaceId dest_space_id = space_id(dest_addr);
2820 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
2821 HeapWord* new_top = _space_info[dest_space_id].new_top();
2822 assert(dest_addr < new_top, "sanity");
2823 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
2824
2825 // Get the source region and related info.
2826 size_t src_region_idx = region_ptr->source_region();
2827 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2828 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2829
2830 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2831 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2832
2833 // Adjust src_region_idx to prepare for decrementing destination counts (the
2834 // destination count is not decremented when a region is copied to itself).
2835 if (src_region_idx == region_idx) {
2836 src_region_idx += 1;
2837 }
2838
2839 if (bitmap->is_unmarked(closure.source())) {
2840 // The first source word is in the middle of an object; copy the remainder
2841 // of the object or as much as will fit. The fact that pointer updates were
2842 // deferred will be noted when the object header is processed.
2843 HeapWord* const old_src_addr = closure.source();
2844 closure.copy_partial_obj();
2845 if (closure.is_full()) {
2846 decrement_destination_counts(cm, src_space_id, src_region_idx,
2847 closure.source());
2848 region_ptr->set_deferred_obj_addr(NULL);
2849 region_ptr->set_completed();
2850 return;
2851 }
2852
2853 HeapWord* const end_addr = sd.region_align_down(closure.source());
2854 if (sd.region_align_down(old_src_addr) != end_addr) {
2855 // The partial object was copied from more than one source region.
2856 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2857
2858 // Move to the next source region, possibly switching spaces as well. All
2859 // args except end_addr may be modified.
2860 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2861 end_addr);
2862 }
2863 }
2864
2865 do {
2866 HeapWord* const cur_addr = closure.source();
2867 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2868 src_space_top);
2869 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2870
2871 if (status == ParMarkBitMap::incomplete) {
2872 // The last obj that starts in the source region does not end in the
2873 // region.
2874 assert(closure.source() < end_addr, "sanity");
2875 HeapWord* const obj_beg = closure.source();
2876 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2877 src_space_top);
2878 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2879 if (obj_end < range_end) {
2880 // The end was found; the entire object will fit.
2881 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2882 assert(status != ParMarkBitMap::would_overflow, "sanity");
2883 } else {
2884 // The end was not found; the object will not fit.
2885 assert(range_end < src_space_top, "obj cannot cross space boundary");
2886 status = ParMarkBitMap::would_overflow;
2887 }
2888 }
2889
2890 if (status == ParMarkBitMap::would_overflow) {
2891 // The last object did not fit. Note that interior oop updates were
2892 // deferred, then copy enough of the object to fill the region.
2893 region_ptr->set_deferred_obj_addr(closure.destination());
2894 status = closure.copy_until_full(); // copies from closure.source()
2895
2896 decrement_destination_counts(cm, src_space_id, src_region_idx,
2897 closure.source());
2898 region_ptr->set_completed();
2899 return;
2900 }
2901
2902 if (status == ParMarkBitMap::full) {
2903 decrement_destination_counts(cm, src_space_id, src_region_idx,
2904 closure.source());
2905 region_ptr->set_deferred_obj_addr(NULL);
2906 region_ptr->set_completed();
2907 return;
2908 }
2909
2910 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2911
2912 // Move to the next source region, possibly switching spaces as well. All
2913 // args except end_addr may be modified.
2914 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2915 end_addr);
2916 } while (true);
2917 }
2918
2919 void PSParallelCompact::fill_blocks(size_t region_idx)
2920 {
2921 // Fill in the block table elements for the specified region. Each block
2922 // table element holds the number of live words in the region that are to the
2923 // left of the first object that starts in the block. Thus only blocks in
2924 // which an object starts need to be filled.
2925 //
2926 // The algorithm scans the section of the bitmap that corresponds to the
2927 // region, keeping a running total of the live words. When an object start is
2928 // found, if it's the first to start in the block that contains it, the
2929 // current total is written to the block table element.
2930 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
2931 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
2932 const size_t RegionSize = ParallelCompactData::RegionSize;
2933
2934 ParallelCompactData& sd = summary_data();
2935 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
2936 if (partial_obj_size >= RegionSize) {
2937 return; // No objects start in this region.
2938 }
2939
2940 // Ensure the first loop iteration decides that the block has changed.
2941 size_t cur_block = sd.block_count();
2942
2943 const ParMarkBitMap* const bitmap = mark_bitmap();
2944
2945 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
2946 assert((size_t)1 << Log2BitsPerBlock ==
2947 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
2948
2949 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
2950 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
2951 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
2952 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
2953 while (beg_bit < range_end) {
2954 const size_t new_block = beg_bit >> Log2BitsPerBlock;
2955 if (new_block != cur_block) {
2956 cur_block = new_block;
2957 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
2958 }
2959
2960 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
2961 if (end_bit < range_end - 1) {
2962 live_bits += end_bit - beg_bit + 1;
2963 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
2964 } else {
2965 return;
2966 }
2967 }
2968 }
2969
2970 void
2971 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
2972 const MutableSpace* sp = space(space_id);
2973 if (sp->is_empty()) {
2974 return;
2975 }
2976
2977 ParallelCompactData& sd = PSParallelCompact::summary_data();
2978 ParMarkBitMap* const bitmap = mark_bitmap();
2979 HeapWord* const dp_addr = dense_prefix(space_id);
2980 HeapWord* beg_addr = sp->bottom();
2981 HeapWord* end_addr = sp->top();
2982
2983 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
2984
2985 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
2986 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
2987 if (beg_region < dp_region) {
2988 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
2989 }
2990
2991 // The destination of the first live object that starts in the region is one
2992 // past the end of the partial object entering the region (if any).
2993 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
2994 HeapWord* const new_top = _space_info[space_id].new_top();
2995 assert(new_top >= dest_addr, "bad new_top value");
2996 const size_t words = pointer_delta(new_top, dest_addr);
2997
2998 if (words > 0) {
2999 ObjectStartArray* start_array = _space_info[space_id].start_array();
3000 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3001
3002 ParMarkBitMap::IterationStatus status;
3003 status = bitmap->iterate(&closure, dest_addr, end_addr);
3004 assert(status == ParMarkBitMap::full, "iteration not complete");
3005 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3006 "live objects skipped because closure is full");
3007 }
3008 }
3009
3010 jlong PSParallelCompact::millis_since_last_gc() {
3011 // We need a monotonically non-decreasing time in ms but
3012 // os::javaTimeMillis() does not guarantee monotonicity.
3013 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3014 jlong ret_val = now - _time_of_last_gc;
3015 // XXX See note in genCollectedHeap::millis_since_last_gc().
3016 if (ret_val < 0) {
3017 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);)
3018 return 0;
3019 }
3020 return ret_val;
3021 }
3022
3023 void PSParallelCompact::reset_millis_since_last_gc() {
3024 // We need a monotonically non-decreasing time in ms but
3025 // os::javaTimeMillis() does not guarantee monotonicity.
3026 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3027 }
3028
3029 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3030 {
3031 if (source() != destination()) {
3032 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3033 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3034 }
3035 update_state(words_remaining());
3036 assert(is_full(), "sanity");
3037 return ParMarkBitMap::full;
3038 }
3039
3040 void MoveAndUpdateClosure::copy_partial_obj()
3041 {
3042 size_t words = words_remaining();
3043
3044 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3045 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3046 if (end_addr < range_end) {
3047 words = bitmap()->obj_size(source(), end_addr);
3048 }
3049
3050 // This test is necessary; if omitted, the pointer updates to a partial object
3051 // that crosses the dense prefix boundary could be overwritten.
3052 if (source() != destination()) {
3053 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3054 Copy::aligned_conjoint_words(source(), destination(), words);
3055 }
3056 update_state(words);
3057 }
3058
3059 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3060 PSParallelCompact::AdjustPointerClosure closure(cm);
3061 oop_oop_iterate_oop_maps<true>(obj, &closure);
3062 }
3063
3064 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3065 InstanceKlass::oop_pc_update_pointers(obj, cm);
3066
3067 PSParallelCompact::AdjustPointerClosure closure(cm);
3068 oop_oop_iterate_statics<true>(obj, &closure);
3069 }
3070
3071 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3072 InstanceKlass::oop_pc_update_pointers(obj, cm);
3073 }
3074
3075 #ifdef ASSERT
3076 template <class T> static void trace_reference_gc(const char *s, oop obj,
3077 T* referent_addr,
3078 T* next_addr,
3079 T* discovered_addr) {
3080 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj));
3081 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3082 p2i(referent_addr), referent_addr ? p2i((oop)RawAccess<>::oop_load(referent_addr)) : NULL);
3083 log_develop_trace(gc, ref)(" next_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3084 p2i(next_addr), next_addr ? p2i((oop)RawAccess<>::oop_load(next_addr)) : NULL);
3085 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3086 p2i(discovered_addr), discovered_addr ? p2i((oop)RawAccess<>::oop_load(discovered_addr)) : NULL);
3087 }
3088 #endif
3089
3090 template <class T>
3091 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) {
3092 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr_raw(obj);
3093 PSParallelCompact::adjust_pointer(referent_addr, cm);
3094 T* next_addr = (T*)java_lang_ref_Reference::next_addr_raw(obj);
3095 PSParallelCompact::adjust_pointer(next_addr, cm);
3096 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr_raw(obj);
3097 PSParallelCompact::adjust_pointer(discovered_addr, cm);
3098 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj,
3099 referent_addr, next_addr, discovered_addr);)
3100 }
3101
3102 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3103 InstanceKlass::oop_pc_update_pointers(obj, cm);
3104
3105 if (UseCompressedOops) {
3106 oop_pc_update_pointers_specialized<narrowOop>(obj, cm);
3107 } else {
3108 oop_pc_update_pointers_specialized<oop>(obj, cm);
3109 }
3110 }
3111
3112 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3113 assert(obj->is_objArray(), "obj must be obj array");
3114 PSParallelCompact::AdjustPointerClosure closure(cm);
3115 oop_oop_iterate_elements<true>(objArrayOop(obj), &closure);
3116 }
3117
3118 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3119 assert(obj->is_typeArray(),"must be a type array");
3120 }
3121
3122 ParMarkBitMapClosure::IterationStatus
3123 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3124 assert(destination() != NULL, "sanity");
3125 assert(bitmap()->obj_size(addr) == words, "bad size");
3126
3127 _source = addr;
3128 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3129 destination(), "wrong destination");
3130
3131 if (words > words_remaining()) {
3132 return ParMarkBitMap::would_overflow;
3133 }
3134
3135 // The start_array must be updated even if the object is not moving.
3136 if (_start_array != NULL) {
3137 _start_array->allocate_block(destination());
3138 }
3139
3140 if (destination() != source()) {
3141 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3142 Copy::aligned_conjoint_words(source(), destination(), words);
3143 }
3144
3145 oop moved_oop = (oop) destination();
3146 compaction_manager()->update_contents(moved_oop);
3147 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3148
3149 update_state(words);
3150 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3151 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3152 }
3153
3154 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3155 ParCompactionManager* cm,
3156 PSParallelCompact::SpaceId space_id) :
3157 ParMarkBitMapClosure(mbm, cm),
3158 _space_id(space_id),
3159 _start_array(PSParallelCompact::start_array(space_id))
3160 {
3161 }
3162
3163 // Updates the references in the object to their new values.
3164 ParMarkBitMapClosure::IterationStatus
3165 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3166 do_addr(addr);
3167 return ParMarkBitMap::incomplete;
3168 }
3169
3170 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3171 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3172 _start_array(PSParallelCompact::start_array(space_id))
3173 {
3174 assert(space_id == PSParallelCompact::old_space_id,
3175 "cannot use FillClosure in the young gen");
3176 }
3177
3178 ParMarkBitMapClosure::IterationStatus
3179 FillClosure::do_addr(HeapWord* addr, size_t size) {
3180 CollectedHeap::fill_with_objects(addr, size);
3181 HeapWord* const end = addr + size;
3182 do {
3183 _start_array->allocate_block(addr);
3184 addr += oop(addr)->size();
3185 } while (addr < end);
3186 return ParMarkBitMap::incomplete;
3187 }
--- EOF ---