1 /*
2 * Copyright (c) 2005, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/stringTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
30 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
31 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
32 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
33 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
34 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
35 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
36 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
37 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
39 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
41 #include "gc_implementation/shared/gcHeapSummary.hpp"
42 #include "gc_implementation/shared/gcTimer.hpp"
43 #include "gc_implementation/shared/gcTrace.hpp"
44 #include "gc_implementation/shared/gcTraceTime.hpp"
45 #include "gc_implementation/shared/isGCActiveMark.hpp"
46 #include "gc_implementation/shared/spaceDecorator.hpp"
47 #include "gc_interface/gcCause.hpp"
48 #include "memory/gcLocker.inline.hpp"
49 #include "memory/referencePolicy.hpp"
50 #include "memory/referenceProcessor.hpp"
51 #include "oops/methodData.hpp"
52 #include "oops/oop.inline.hpp"
53 #include "oops/oop.pcgc.inline.hpp"
54 #include "runtime/fprofiler.hpp"
55 #include "runtime/safepoint.hpp"
56 #include "runtime/vmThread.hpp"
57 #include "services/management.hpp"
58 #include "services/memoryService.hpp"
59 #include "services/memTracker.hpp"
60 #include "utilities/events.hpp"
61 #include "utilities/stack.inline.hpp"
62
63 #include <math.h>
64
65 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
66
67 // All sizes are in HeapWords.
68 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
69 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
70 const size_t ParallelCompactData::RegionSizeBytes =
71 RegionSize << LogHeapWordSize;
72 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
73 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
74 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
75
76 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
77 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
78 const size_t ParallelCompactData::BlockSizeBytes =
79 BlockSize << LogHeapWordSize;
80 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
81 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
82 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
83
84 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
85 const size_t ParallelCompactData::Log2BlocksPerRegion =
86 Log2RegionSize - Log2BlockSize;
87
88 const ParallelCompactData::RegionData::region_sz_t
89 ParallelCompactData::RegionData::dc_shift = 27;
90
91 const ParallelCompactData::RegionData::region_sz_t
92 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
93
94 const ParallelCompactData::RegionData::region_sz_t
95 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
96
97 const ParallelCompactData::RegionData::region_sz_t
98 ParallelCompactData::RegionData::los_mask = ~dc_mask;
99
100 const ParallelCompactData::RegionData::region_sz_t
101 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
102
103 const ParallelCompactData::RegionData::region_sz_t
104 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
105
106 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
107 bool PSParallelCompact::_print_phases = false;
108
109 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
110 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL;
111
112 double PSParallelCompact::_dwl_mean;
113 double PSParallelCompact::_dwl_std_dev;
114 double PSParallelCompact::_dwl_first_term;
115 double PSParallelCompact::_dwl_adjustment;
116 #ifdef ASSERT
117 bool PSParallelCompact::_dwl_initialized = false;
118 #endif // #ifdef ASSERT
119
120 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
121 HeapWord* destination)
122 {
123 assert(src_region_idx != 0, "invalid src_region_idx");
124 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
125 assert(destination != NULL, "invalid destination argument");
126
127 _src_region_idx = src_region_idx;
128 _partial_obj_size = partial_obj_size;
129 _destination = destination;
130
131 // These fields may not be updated below, so make sure they're clear.
132 assert(_dest_region_addr == NULL, "should have been cleared");
133 assert(_first_src_addr == NULL, "should have been cleared");
134
135 // Determine the number of destination regions for the partial object.
136 HeapWord* const last_word = destination + partial_obj_size - 1;
137 const ParallelCompactData& sd = PSParallelCompact::summary_data();
138 HeapWord* const beg_region_addr = sd.region_align_down(destination);
139 HeapWord* const end_region_addr = sd.region_align_down(last_word);
140
141 if (beg_region_addr == end_region_addr) {
142 // One destination region.
143 _destination_count = 1;
144 if (end_region_addr == destination) {
145 // The destination falls on a region boundary, thus the first word of the
146 // partial object will be the first word copied to the destination region.
147 _dest_region_addr = end_region_addr;
148 _first_src_addr = sd.region_to_addr(src_region_idx);
149 }
150 } else {
151 // Two destination regions. When copied, the partial object will cross a
152 // destination region boundary, so a word somewhere within the partial
153 // object will be the first word copied to the second destination region.
154 _destination_count = 2;
155 _dest_region_addr = end_region_addr;
156 const size_t ofs = pointer_delta(end_region_addr, destination);
157 assert(ofs < _partial_obj_size, "sanity");
158 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
159 }
160 }
161
162 void SplitInfo::clear()
163 {
164 _src_region_idx = 0;
165 _partial_obj_size = 0;
166 _destination = NULL;
167 _destination_count = 0;
168 _dest_region_addr = NULL;
169 _first_src_addr = NULL;
170 assert(!is_valid(), "sanity");
171 }
172
173 #ifdef ASSERT
174 void SplitInfo::verify_clear()
175 {
176 assert(_src_region_idx == 0, "not clear");
177 assert(_partial_obj_size == 0, "not clear");
178 assert(_destination == NULL, "not clear");
179 assert(_destination_count == 0, "not clear");
180 assert(_dest_region_addr == NULL, "not clear");
181 assert(_first_src_addr == NULL, "not clear");
182 }
183 #endif // #ifdef ASSERT
184
185
186 void PSParallelCompact::print_on_error(outputStream* st) {
187 _mark_bitmap.print_on_error(st);
188 }
189
190 #ifndef PRODUCT
191 const char* PSParallelCompact::space_names[] = {
192 "old ", "eden", "from", "to "
193 };
194
195 void PSParallelCompact::print_region_ranges()
196 {
197 tty->print_cr("space bottom top end new_top");
198 tty->print_cr("------ ---------- ---------- ---------- ----------");
199
200 for (unsigned int id = 0; id < last_space_id; ++id) {
201 const MutableSpace* space = _space_info[id].space();
202 tty->print_cr("%u %s "
203 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
205 id, space_names[id],
206 summary_data().addr_to_region_idx(space->bottom()),
207 summary_data().addr_to_region_idx(space->top()),
208 summary_data().addr_to_region_idx(space->end()),
209 summary_data().addr_to_region_idx(_space_info[id].new_top()));
210 }
211 }
212
213 void
214 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
215 {
216 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
217 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
218
219 ParallelCompactData& sd = PSParallelCompact::summary_data();
220 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
221 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
222 REGION_IDX_FORMAT " " PTR_FORMAT " "
223 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
224 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
225 i, c->data_location(), dci, c->destination(),
226 c->partial_obj_size(), c->live_obj_size(),
227 c->data_size(), c->source_region(), c->destination_count());
228
229 #undef REGION_IDX_FORMAT
230 #undef REGION_DATA_FORMAT
231 }
232
233 void
234 print_generic_summary_data(ParallelCompactData& summary_data,
235 HeapWord* const beg_addr,
236 HeapWord* const end_addr)
237 {
238 size_t total_words = 0;
239 size_t i = summary_data.addr_to_region_idx(beg_addr);
240 const size_t last = summary_data.addr_to_region_idx(end_addr);
241 HeapWord* pdest = 0;
242
243 while (i <= last) {
244 ParallelCompactData::RegionData* c = summary_data.region(i);
245 if (c->data_size() != 0 || c->destination() != pdest) {
246 print_generic_summary_region(i, c);
247 total_words += c->data_size();
248 pdest = c->destination();
249 }
250 ++i;
251 }
252
253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
254 }
255
256 void
257 print_generic_summary_data(ParallelCompactData& summary_data,
258 SpaceInfo* space_info)
259 {
260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
261 const MutableSpace* space = space_info[id].space();
262 print_generic_summary_data(summary_data, space->bottom(),
263 MAX2(space->top(), space_info[id].new_top()));
264 }
265 }
266
267 void
268 print_initial_summary_region(size_t i,
269 const ParallelCompactData::RegionData* c,
270 bool newline = true)
271 {
272 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
273 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
275 i, c->destination(),
276 c->partial_obj_size(), c->live_obj_size(),
277 c->data_size(), c->source_region(), c->destination_count());
278 if (newline) tty->cr();
279 }
280
281 void
282 print_initial_summary_data(ParallelCompactData& summary_data,
283 const MutableSpace* space) {
284 if (space->top() == space->bottom()) {
285 return;
286 }
287
288 const size_t region_size = ParallelCompactData::RegionSize;
289 typedef ParallelCompactData::RegionData RegionData;
290 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
291 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
292 const RegionData* c = summary_data.region(end_region - 1);
293 HeapWord* end_addr = c->destination() + c->data_size();
294 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
295
296 // Print (and count) the full regions at the beginning of the space.
297 size_t full_region_count = 0;
298 size_t i = summary_data.addr_to_region_idx(space->bottom());
299 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
300 print_initial_summary_region(i, summary_data.region(i));
301 ++full_region_count;
302 ++i;
303 }
304
305 size_t live_to_right = live_in_space - full_region_count * region_size;
306
307 double max_reclaimed_ratio = 0.0;
308 size_t max_reclaimed_ratio_region = 0;
309 size_t max_dead_to_right = 0;
310 size_t max_live_to_right = 0;
311
312 // Print the 'reclaimed ratio' for regions while there is something live in
313 // the region or to the right of it. The remaining regions are empty (and
314 // uninteresting), and computing the ratio will result in division by 0.
315 while (i < end_region && live_to_right > 0) {
316 c = summary_data.region(i);
317 HeapWord* const region_addr = summary_data.region_to_addr(i);
318 const size_t used_to_right = pointer_delta(space->top(), region_addr);
319 const size_t dead_to_right = used_to_right - live_to_right;
320 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
321
322 if (reclaimed_ratio > max_reclaimed_ratio) {
323 max_reclaimed_ratio = reclaimed_ratio;
324 max_reclaimed_ratio_region = i;
325 max_dead_to_right = dead_to_right;
326 max_live_to_right = live_to_right;
327 }
328
329 print_initial_summary_region(i, c, false);
330 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
331 reclaimed_ratio, dead_to_right, live_to_right);
332
333 live_to_right -= c->data_size();
334 ++i;
335 }
336
337 // Any remaining regions are empty. Print one more if there is one.
338 if (i < end_region) {
339 print_initial_summary_region(i, summary_data.region(i));
340 }
341
342 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
343 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
344 max_reclaimed_ratio_region, max_dead_to_right,
345 max_live_to_right, max_reclaimed_ratio);
346 }
347
348 void
349 print_initial_summary_data(ParallelCompactData& summary_data,
350 SpaceInfo* space_info) {
351 unsigned int id = PSParallelCompact::old_space_id;
352 const MutableSpace* space;
353 do {
354 space = space_info[id].space();
355 print_initial_summary_data(summary_data, space);
356 } while (++id < PSParallelCompact::eden_space_id);
357
358 do {
359 space = space_info[id].space();
360 print_generic_summary_data(summary_data, space->bottom(), space->top());
361 } while (++id < PSParallelCompact::last_space_id);
362 }
363 #endif // #ifndef PRODUCT
364
365 #ifdef ASSERT
366 size_t add_obj_count;
367 size_t add_obj_size;
368 size_t mark_bitmap_count;
369 size_t mark_bitmap_size;
370 #endif // #ifdef ASSERT
371
372 ParallelCompactData::ParallelCompactData()
373 {
374 _region_start = 0;
375
376 _region_vspace = 0;
377 _reserved_byte_size = 0;
378 _region_data = 0;
379 _region_count = 0;
380
381 _block_vspace = 0;
382 _block_data = 0;
383 _block_count = 0;
384 }
385
386 bool ParallelCompactData::initialize(MemRegion covered_region)
387 {
388 _region_start = covered_region.start();
389 const size_t region_size = covered_region.word_size();
390 DEBUG_ONLY(_region_end = _region_start + region_size;)
391
392 assert(region_align_down(_region_start) == _region_start,
393 "region start not aligned");
394 assert((region_size & RegionSizeOffsetMask) == 0,
395 "region size not a multiple of RegionSize");
396
397 bool result = initialize_region_data(region_size) && initialize_block_data();
398 return result;
399 }
400
401 PSVirtualSpace*
402 ParallelCompactData::create_vspace(size_t count, size_t element_size)
403 {
404 const size_t raw_bytes = count * element_size;
405 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
406 const size_t granularity = os::vm_allocation_granularity();
407 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
408
409 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
410 MAX2(page_sz, granularity);
411 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
412 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
413 rs.size());
414
415 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
416
417 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
418 if (vspace != 0) {
419 if (vspace->expand_by(_reserved_byte_size)) {
420 return vspace;
421 }
422 delete vspace;
423 // Release memory reserved in the space.
424 rs.release();
425 }
426
427 return 0;
428 }
429
430 bool ParallelCompactData::initialize_region_data(size_t region_size)
431 {
432 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
433 _region_vspace = create_vspace(count, sizeof(RegionData));
434 if (_region_vspace != 0) {
435 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
436 _region_count = count;
437 return true;
438 }
439 return false;
440 }
441
442 bool ParallelCompactData::initialize_block_data()
443 {
444 assert(_region_count != 0, "region data must be initialized first");
445 const size_t count = _region_count << Log2BlocksPerRegion;
446 _block_vspace = create_vspace(count, sizeof(BlockData));
447 if (_block_vspace != 0) {
448 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
449 _block_count = count;
450 return true;
451 }
452 return false;
453 }
454
455 void ParallelCompactData::clear()
456 {
457 memset(_region_data, 0, _region_vspace->committed_size());
458 memset(_block_data, 0, _block_vspace->committed_size());
459 }
460
461 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
462 assert(beg_region <= _region_count, "beg_region out of range");
463 assert(end_region <= _region_count, "end_region out of range");
464 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
465
466 const size_t region_cnt = end_region - beg_region;
467 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
468
469 const size_t beg_block = beg_region * BlocksPerRegion;
470 const size_t block_cnt = region_cnt * BlocksPerRegion;
471 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
472 }
473
474 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
475 {
476 const RegionData* cur_cp = region(region_idx);
477 const RegionData* const end_cp = region(region_count() - 1);
478
479 HeapWord* result = region_to_addr(region_idx);
480 if (cur_cp < end_cp) {
481 do {
482 result += cur_cp->partial_obj_size();
483 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
484 }
485 return result;
486 }
487
488 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
489 {
490 const size_t obj_ofs = pointer_delta(addr, _region_start);
491 const size_t beg_region = obj_ofs >> Log2RegionSize;
492 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
493
494 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
495 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
496
497 if (beg_region == end_region) {
498 // All in one region.
499 _region_data[beg_region].add_live_obj(len);
500 return;
501 }
502
503 // First region.
504 const size_t beg_ofs = region_offset(addr);
505 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
506
507 Klass* klass = ((oop)addr)->klass();
508 // Middle regions--completely spanned by this object.
509 for (size_t region = beg_region + 1; region < end_region; ++region) {
510 _region_data[region].set_partial_obj_size(RegionSize);
511 _region_data[region].set_partial_obj_addr(addr);
512 }
513
514 // Last region.
515 const size_t end_ofs = region_offset(addr + len - 1);
516 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
517 _region_data[end_region].set_partial_obj_addr(addr);
518 }
519
520 void
521 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
522 {
523 assert(region_offset(beg) == 0, "not RegionSize aligned");
524 assert(region_offset(end) == 0, "not RegionSize aligned");
525
526 size_t cur_region = addr_to_region_idx(beg);
527 const size_t end_region = addr_to_region_idx(end);
528 HeapWord* addr = beg;
529 while (cur_region < end_region) {
530 _region_data[cur_region].set_destination(addr);
531 _region_data[cur_region].set_destination_count(0);
532 _region_data[cur_region].set_source_region(cur_region);
533 _region_data[cur_region].set_data_location(addr);
534
535 // Update live_obj_size so the region appears completely full.
536 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
537 _region_data[cur_region].set_live_obj_size(live_size);
538
539 ++cur_region;
540 addr += RegionSize;
541 }
542 }
543
544 // Find the point at which a space can be split and, if necessary, record the
545 // split point.
546 //
547 // If the current src region (which overflowed the destination space) doesn't
548 // have a partial object, the split point is at the beginning of the current src
549 // region (an "easy" split, no extra bookkeeping required).
550 //
551 // If the current src region has a partial object, the split point is in the
552 // region where that partial object starts (call it the split_region). If
553 // split_region has a partial object, then the split point is just after that
554 // partial object (a "hard" split where we have to record the split data and
555 // zero the partial_obj_size field). With a "hard" split, we know that the
556 // partial_obj ends within split_region because the partial object that caused
557 // the overflow starts in split_region. If split_region doesn't have a partial
558 // obj, then the split is at the beginning of split_region (another "easy"
559 // split).
560 HeapWord*
561 ParallelCompactData::summarize_split_space(size_t src_region,
562 SplitInfo& split_info,
563 HeapWord* destination,
564 HeapWord* target_end,
565 HeapWord** target_next)
566 {
567 assert(destination <= target_end, "sanity");
568 assert(destination + _region_data[src_region].data_size() > target_end,
569 "region should not fit into target space");
570 assert(is_region_aligned(target_end), "sanity");
571
572 size_t split_region = src_region;
573 HeapWord* split_destination = destination;
574 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
575
576 if (destination + partial_obj_size > target_end) {
577 // The split point is just after the partial object (if any) in the
578 // src_region that contains the start of the object that overflowed the
579 // destination space.
580 //
581 // Find the start of the "overflow" object and set split_region to the
582 // region containing it.
583 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
584 split_region = addr_to_region_idx(overflow_obj);
585
586 // Clear the source_region field of all destination regions whose first word
587 // came from data after the split point (a non-null source_region field
588 // implies a region must be filled).
589 //
590 // An alternative to the simple loop below: clear during post_compact(),
591 // which uses memcpy instead of individual stores, and is easy to
592 // parallelize. (The downside is that it clears the entire RegionData
593 // object as opposed to just one field.)
594 //
595 // post_compact() would have to clear the summary data up to the highest
596 // address that was written during the summary phase, which would be
597 //
598 // max(top, max(new_top, clear_top))
599 //
600 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
601 // to target_end.
602 const RegionData* const sr = region(split_region);
603 const size_t beg_idx =
604 addr_to_region_idx(region_align_up(sr->destination() +
605 sr->partial_obj_size()));
606 const size_t end_idx = addr_to_region_idx(target_end);
607
608 if (TraceParallelOldGCSummaryPhase) {
609 gclog_or_tty->print_cr("split: clearing source_region field in ["
610 SIZE_FORMAT ", " SIZE_FORMAT ")",
611 beg_idx, end_idx);
612 }
613 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
614 _region_data[idx].set_source_region(0);
615 }
616
617 // Set split_destination and partial_obj_size to reflect the split region.
618 split_destination = sr->destination();
619 partial_obj_size = sr->partial_obj_size();
620 }
621
622 // The split is recorded only if a partial object extends onto the region.
623 if (partial_obj_size != 0) {
624 _region_data[split_region].set_partial_obj_size(0);
625 split_info.record(split_region, partial_obj_size, split_destination);
626 }
627
628 // Setup the continuation addresses.
629 *target_next = split_destination + partial_obj_size;
630 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
631
632 if (TraceParallelOldGCSummaryPhase) {
633 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
634 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
635 " pos=" SIZE_FORMAT,
636 split_type, source_next, split_region,
637 partial_obj_size);
638 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
639 " tn=" PTR_FORMAT,
640 split_type, split_destination,
641 addr_to_region_idx(split_destination),
642 *target_next);
643
644 if (partial_obj_size != 0) {
645 HeapWord* const po_beg = split_info.destination();
646 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
647 gclog_or_tty->print_cr("%s split: "
648 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
649 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
650 split_type,
651 po_beg, addr_to_region_idx(po_beg),
652 po_end, addr_to_region_idx(po_end));
653 }
654 }
655
656 return source_next;
657 }
658
659 bool ParallelCompactData::summarize(SplitInfo& split_info,
660 HeapWord* source_beg, HeapWord* source_end,
661 HeapWord** source_next,
662 HeapWord* target_beg, HeapWord* target_end,
663 HeapWord** target_next)
664 {
665 if (TraceParallelOldGCSummaryPhase) {
666 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
667 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
668 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
669 source_beg, source_end, source_next_val,
670 target_beg, target_end, *target_next);
671 }
672
673 size_t cur_region = addr_to_region_idx(source_beg);
674 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
675
676 HeapWord *dest_addr = target_beg;
677 while (cur_region < end_region) {
678 // The destination must be set even if the region has no data.
679 _region_data[cur_region].set_destination(dest_addr);
680
681 size_t words = _region_data[cur_region].data_size();
682 if (words > 0) {
683 // If cur_region does not fit entirely into the target space, find a point
684 // at which the source space can be 'split' so that part is copied to the
685 // target space and the rest is copied elsewhere.
686 if (dest_addr + words > target_end) {
687 assert(source_next != NULL, "source_next is NULL when splitting");
688 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
689 target_end, target_next);
690 return false;
691 }
692
693 // Compute the destination_count for cur_region, and if necessary, update
694 // source_region for a destination region. The source_region field is
695 // updated if cur_region is the first (left-most) region to be copied to a
696 // destination region.
697 //
698 // The destination_count calculation is a bit subtle. A region that has
699 // data that compacts into itself does not count itself as a destination.
700 // This maintains the invariant that a zero count means the region is
701 // available and can be claimed and then filled.
702 uint destination_count = 0;
703 if (split_info.is_split(cur_region)) {
704 // The current region has been split: the partial object will be copied
705 // to one destination space and the remaining data will be copied to
706 // another destination space. Adjust the initial destination_count and,
707 // if necessary, set the source_region field if the partial object will
708 // cross a destination region boundary.
709 destination_count = split_info.destination_count();
710 if (destination_count == 2) {
711 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
712 _region_data[dest_idx].set_source_region(cur_region);
713 }
714 }
715
716 HeapWord* const last_addr = dest_addr + words - 1;
717 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
718 const size_t dest_region_2 = addr_to_region_idx(last_addr);
719
720 // Initially assume that the destination regions will be the same and
721 // adjust the value below if necessary. Under this assumption, if
722 // cur_region == dest_region_2, then cur_region will be compacted
723 // completely into itself.
724 destination_count += cur_region == dest_region_2 ? 0 : 1;
725 if (dest_region_1 != dest_region_2) {
726 // Destination regions differ; adjust destination_count.
727 destination_count += 1;
728 // Data from cur_region will be copied to the start of dest_region_2.
729 _region_data[dest_region_2].set_source_region(cur_region);
730 } else if (region_offset(dest_addr) == 0) {
731 // Data from cur_region will be copied to the start of the destination
732 // region.
733 _region_data[dest_region_1].set_source_region(cur_region);
734 }
735
736 _region_data[cur_region].set_destination_count(destination_count);
737 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
738 dest_addr += words;
739 }
740
741 ++cur_region;
742 }
743
744 *target_next = dest_addr;
745 return true;
746 }
747
748 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
749 assert(addr != NULL, "Should detect NULL oop earlier");
750 assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap");
751 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
752
753 // Region covering the object.
754 RegionData* const region_ptr = addr_to_region_ptr(addr);
755 HeapWord* result = region_ptr->destination();
756
757 // If the entire Region is live, the new location is region->destination + the
758 // offset of the object within in the Region.
759
760 // Run some performance tests to determine if this special case pays off. It
761 // is worth it for pointers into the dense prefix. If the optimization to
762 // avoid pointer updates in regions that only point to the dense prefix is
763 // ever implemented, this should be revisited.
764 if (region_ptr->data_size() == RegionSize) {
765 result += region_offset(addr);
766 return result;
767 }
768
769 // Otherwise, the new location is region->destination + block offset + the
770 // number of live words in the Block that are (a) to the left of addr and (b)
771 // due to objects that start in the Block.
772
773 // Fill in the block table if necessary. This is unsynchronized, so multiple
774 // threads may fill the block table for a region (harmless, since it is
775 // idempotent).
776 if (!region_ptr->blocks_filled()) {
777 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
778 region_ptr->set_blocks_filled();
779 }
780
781 HeapWord* const search_start = block_align_down(addr);
782 const size_t block_offset = addr_to_block_ptr(addr)->offset();
783
784 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
785 const size_t live = bitmap->live_words_in_range(search_start, oop(addr));
786 result += block_offset + live;
787 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
788 return result;
789 }
790
791 #ifdef ASSERT
792 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
793 {
794 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
795 const size_t* const end = (const size_t*)vspace->committed_high_addr();
796 for (const size_t* p = beg; p < end; ++p) {
797 assert(*p == 0, "not zero");
798 }
799 }
800
801 void ParallelCompactData::verify_clear()
802 {
803 verify_clear(_region_vspace);
804 verify_clear(_block_vspace);
805 }
806 #endif // #ifdef ASSERT
807
808 STWGCTimer PSParallelCompact::_gc_timer;
809 ParallelOldTracer PSParallelCompact::_gc_tracer;
810 elapsedTimer PSParallelCompact::_accumulated_time;
811 unsigned int PSParallelCompact::_total_invocations = 0;
812 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
813 jlong PSParallelCompact::_time_of_last_gc = 0;
814 CollectorCounters* PSParallelCompact::_counters = NULL;
815 ParMarkBitMap PSParallelCompact::_mark_bitmap;
816 ParallelCompactData PSParallelCompact::_summary_data;
817
818 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
819
820 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
821
822 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
823 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
824
825 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
826 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
827
828 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); }
829 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
830
831 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
832
833 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) {
834 mark_and_push(_compaction_manager, p);
835 }
836 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
837
838 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
839 klass->oops_do(_mark_and_push_closure);
840 }
841 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
842 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
843 }
844
845 void PSParallelCompact::post_initialize() {
846 ParallelScavengeHeap* heap = gc_heap();
847 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
848
849 MemRegion mr = heap->reserved_region();
850 _ref_processor =
851 new ReferenceProcessor(mr, // span
852 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
853 (int) ParallelGCThreads, // mt processing degree
854 true, // mt discovery
855 (int) ParallelGCThreads, // mt discovery degree
856 true, // atomic_discovery
857 &_is_alive_closure, // non-header is alive closure
858 false); // write barrier for next field updates
859 _counters = new CollectorCounters("PSParallelCompact", 1);
860
861 // Initialize static fields in ParCompactionManager.
862 ParCompactionManager::initialize(mark_bitmap());
863 }
864
865 bool PSParallelCompact::initialize() {
866 ParallelScavengeHeap* heap = gc_heap();
867 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
868 MemRegion mr = heap->reserved_region();
869
870 // Was the old gen get allocated successfully?
871 if (!heap->old_gen()->is_allocated()) {
872 return false;
873 }
874
875 initialize_space_info();
876 initialize_dead_wood_limiter();
877
878 if (!_mark_bitmap.initialize(mr)) {
879 vm_shutdown_during_initialization(
880 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
881 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
882 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
883 return false;
884 }
885
886 if (!_summary_data.initialize(mr)) {
887 vm_shutdown_during_initialization(
888 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
889 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
890 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
891 return false;
892 }
893
894 return true;
895 }
896
897 void PSParallelCompact::initialize_space_info()
898 {
899 memset(&_space_info, 0, sizeof(_space_info));
900
901 ParallelScavengeHeap* heap = gc_heap();
902 PSYoungGen* young_gen = heap->young_gen();
903
904 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
905 _space_info[eden_space_id].set_space(young_gen->eden_space());
906 _space_info[from_space_id].set_space(young_gen->from_space());
907 _space_info[to_space_id].set_space(young_gen->to_space());
908
909 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
910 }
911
912 void PSParallelCompact::initialize_dead_wood_limiter()
913 {
914 const size_t max = 100;
915 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
916 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
917 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
918 DEBUG_ONLY(_dwl_initialized = true;)
919 _dwl_adjustment = normal_distribution(1.0);
920 }
921
922 // Simple class for storing info about the heap at the start of GC, to be used
923 // after GC for comparison/printing.
924 class PreGCValues {
925 public:
926 PreGCValues() { }
927 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
928
929 void fill(ParallelScavengeHeap* heap) {
930 _heap_used = heap->used();
931 _young_gen_used = heap->young_gen()->used_in_bytes();
932 _old_gen_used = heap->old_gen()->used_in_bytes();
933 _metadata_used = MetaspaceAux::used_bytes();
934 };
935
936 size_t heap_used() const { return _heap_used; }
937 size_t young_gen_used() const { return _young_gen_used; }
938 size_t old_gen_used() const { return _old_gen_used; }
939 size_t metadata_used() const { return _metadata_used; }
940
941 private:
942 size_t _heap_used;
943 size_t _young_gen_used;
944 size_t _old_gen_used;
945 size_t _metadata_used;
946 };
947
948 void
949 PSParallelCompact::clear_data_covering_space(SpaceId id)
950 {
951 // At this point, top is the value before GC, new_top() is the value that will
952 // be set at the end of GC. The marking bitmap is cleared to top; nothing
953 // should be marked above top. The summary data is cleared to the larger of
954 // top & new_top.
955 MutableSpace* const space = _space_info[id].space();
956 HeapWord* const bot = space->bottom();
957 HeapWord* const top = space->top();
958 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
959
960 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
961 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
962 _mark_bitmap.clear_range(beg_bit, end_bit);
963
964 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
965 const size_t end_region =
966 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
967 _summary_data.clear_range(beg_region, end_region);
968
969 // Clear the data used to 'split' regions.
970 SplitInfo& split_info = _space_info[id].split_info();
971 if (split_info.is_valid()) {
972 split_info.clear();
973 }
974 DEBUG_ONLY(split_info.verify_clear();)
975 }
976
977 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
978 {
979 // Update the from & to space pointers in space_info, since they are swapped
980 // at each young gen gc. Do the update unconditionally (even though a
981 // promotion failure does not swap spaces) because an unknown number of minor
982 // collections will have swapped the spaces an unknown number of times.
983 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer);
984 ParallelScavengeHeap* heap = gc_heap();
985 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
986 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
987
988 pre_gc_values->fill(heap);
989
990 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
991 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
992
993 // Increment the invocation count
994 heap->increment_total_collections(true);
995
996 // We need to track unique mark sweep invocations as well.
997 _total_invocations++;
998
999 heap->print_heap_before_gc();
1000 heap->trace_heap_before_gc(&_gc_tracer);
1001
1002 // Fill in TLABs
1003 heap->accumulate_statistics_all_tlabs();
1004 heap->ensure_parsability(true); // retire TLABs
1005
1006 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1007 HandleMark hm; // Discard invalid handles created during verification
1008 Universe::verify(" VerifyBeforeGC:");
1009 }
1010
1011 // Verify object start arrays
1012 if (VerifyObjectStartArray &&
1013 VerifyBeforeGC) {
1014 heap->old_gen()->verify_object_start_array();
1015 }
1016
1017 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1018 DEBUG_ONLY(summary_data().verify_clear();)
1019
1020 // Have worker threads release resources the next time they run a task.
1021 gc_task_manager()->release_all_resources();
1022 }
1023
1024 void PSParallelCompact::post_compact()
1025 {
1026 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer);
1027
1028 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1029 // Clear the marking bitmap, summary data and split info.
1030 clear_data_covering_space(SpaceId(id));
1031 // Update top(). Must be done after clearing the bitmap and summary data.
1032 _space_info[id].publish_new_top();
1033 }
1034
1035 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1036 MutableSpace* const from_space = _space_info[from_space_id].space();
1037 MutableSpace* const to_space = _space_info[to_space_id].space();
1038
1039 ParallelScavengeHeap* heap = gc_heap();
1040 bool eden_empty = eden_space->is_empty();
1041 if (!eden_empty) {
1042 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1043 heap->young_gen(), heap->old_gen());
1044 }
1045
1046 // Update heap occupancy information which is used as input to the soft ref
1047 // clearing policy at the next gc.
1048 Universe::update_heap_info_at_gc();
1049
1050 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1051 to_space->is_empty();
1052
1053 BarrierSet* bs = heap->barrier_set();
1054 if (bs->is_a(BarrierSet::ModRef)) {
1055 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1056 MemRegion old_mr = heap->old_gen()->reserved();
1057
1058 if (young_gen_empty) {
1059 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1060 } else {
1061 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1062 }
1063 }
1064
1065 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1066 ClassLoaderDataGraph::purge();
1067 MetaspaceAux::verify_metrics();
1068
1069 Threads::gc_epilogue();
1070 CodeCache::gc_epilogue();
1071 JvmtiExport::gc_epilogue();
1072
1073 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1074
1075 ref_processor()->enqueue_discovered_references(NULL);
1076
1077 if (ZapUnusedHeapArea) {
1078 heap->gen_mangle_unused_area();
1079 }
1080
1081 // Update time of last GC
1082 reset_millis_since_last_gc();
1083 }
1084
1085 HeapWord*
1086 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1087 bool maximum_compaction)
1088 {
1089 const size_t region_size = ParallelCompactData::RegionSize;
1090 const ParallelCompactData& sd = summary_data();
1091
1092 const MutableSpace* const space = _space_info[id].space();
1093 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1094 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1095 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1096
1097 // Skip full regions at the beginning of the space--they are necessarily part
1098 // of the dense prefix.
1099 size_t full_count = 0;
1100 const RegionData* cp;
1101 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1102 ++full_count;
1103 }
1104
1105 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1106 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1107 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1108 if (maximum_compaction || cp == end_cp || interval_ended) {
1109 _maximum_compaction_gc_num = total_invocations();
1110 return sd.region_to_addr(cp);
1111 }
1112
1113 HeapWord* const new_top = _space_info[id].new_top();
1114 const size_t space_live = pointer_delta(new_top, space->bottom());
1115 const size_t space_used = space->used_in_words();
1116 const size_t space_capacity = space->capacity_in_words();
1117
1118 const double cur_density = double(space_live) / space_capacity;
1119 const double deadwood_density =
1120 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1121 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1122
1123 if (TraceParallelOldGCDensePrefix) {
1124 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1125 cur_density, deadwood_density, deadwood_goal);
1126 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1127 "space_cap=" SIZE_FORMAT,
1128 space_live, space_used,
1129 space_capacity);
1130 }
1131
1132 // XXX - Use binary search?
1133 HeapWord* dense_prefix = sd.region_to_addr(cp);
1134 const RegionData* full_cp = cp;
1135 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1136 while (cp < end_cp) {
1137 HeapWord* region_destination = cp->destination();
1138 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1139 if (TraceParallelOldGCDensePrefix && Verbose) {
1140 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1141 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1142 sd.region(cp), region_destination,
1143 dense_prefix, cur_deadwood);
1144 }
1145
1146 if (cur_deadwood >= deadwood_goal) {
1147 // Found the region that has the correct amount of deadwood to the left.
1148 // This typically occurs after crossing a fairly sparse set of regions, so
1149 // iterate backwards over those sparse regions, looking for the region
1150 // that has the lowest density of live objects 'to the right.'
1151 size_t space_to_left = sd.region(cp) * region_size;
1152 size_t live_to_left = space_to_left - cur_deadwood;
1153 size_t space_to_right = space_capacity - space_to_left;
1154 size_t live_to_right = space_live - live_to_left;
1155 double density_to_right = double(live_to_right) / space_to_right;
1156 while (cp > full_cp) {
1157 --cp;
1158 const size_t prev_region_live_to_right = live_to_right -
1159 cp->data_size();
1160 const size_t prev_region_space_to_right = space_to_right + region_size;
1161 double prev_region_density_to_right =
1162 double(prev_region_live_to_right) / prev_region_space_to_right;
1163 if (density_to_right <= prev_region_density_to_right) {
1164 return dense_prefix;
1165 }
1166 if (TraceParallelOldGCDensePrefix && Verbose) {
1167 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1168 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1169 prev_region_density_to_right);
1170 }
1171 dense_prefix -= region_size;
1172 live_to_right = prev_region_live_to_right;
1173 space_to_right = prev_region_space_to_right;
1174 density_to_right = prev_region_density_to_right;
1175 }
1176 return dense_prefix;
1177 }
1178
1179 dense_prefix += region_size;
1180 ++cp;
1181 }
1182
1183 return dense_prefix;
1184 }
1185
1186 #ifndef PRODUCT
1187 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1188 const SpaceId id,
1189 const bool maximum_compaction,
1190 HeapWord* const addr)
1191 {
1192 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1193 RegionData* const cp = summary_data().region(region_idx);
1194 const MutableSpace* const space = _space_info[id].space();
1195 HeapWord* const new_top = _space_info[id].new_top();
1196
1197 const size_t space_live = pointer_delta(new_top, space->bottom());
1198 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1199 const size_t space_cap = space->capacity_in_words();
1200 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1201 const size_t live_to_right = new_top - cp->destination();
1202 const size_t dead_to_right = space->top() - addr - live_to_right;
1203
1204 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1205 "spl=" SIZE_FORMAT " "
1206 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1207 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1208 " ratio=%10.8f",
1209 algorithm, addr, region_idx,
1210 space_live,
1211 dead_to_left, dead_to_left_pct,
1212 dead_to_right, live_to_right,
1213 double(dead_to_right) / live_to_right);
1214 }
1215 #endif // #ifndef PRODUCT
1216
1217 // Return a fraction indicating how much of the generation can be treated as
1218 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1219 // based on the density of live objects in the generation to determine a limit,
1220 // which is then adjusted so the return value is min_percent when the density is
1221 // 1.
1222 //
1223 // The following table shows some return values for a different values of the
1224 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1225 // min_percent is 1.
1226 //
1227 // fraction allowed as dead wood
1228 // -----------------------------------------------------------------
1229 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1230 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1231 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1232 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1233 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1234 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1235 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1236 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1237 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1238 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1239 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1240 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1241 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1242 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1243 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1244 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1245 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1246 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1247 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1248 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1249 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1250 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1251 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1252
1253 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1254 {
1255 assert(_dwl_initialized, "uninitialized");
1256
1257 // The raw limit is the value of the normal distribution at x = density.
1258 const double raw_limit = normal_distribution(density);
1259
1260 // Adjust the raw limit so it becomes the minimum when the density is 1.
1261 //
1262 // First subtract the adjustment value (which is simply the precomputed value
1263 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1264 // Then add the minimum value, so the minimum is returned when the density is
1265 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1266 const double min = double(min_percent) / 100.0;
1267 const double limit = raw_limit - _dwl_adjustment + min;
1268 return MAX2(limit, 0.0);
1269 }
1270
1271 ParallelCompactData::RegionData*
1272 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1273 const RegionData* end)
1274 {
1275 const size_t region_size = ParallelCompactData::RegionSize;
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 if (middle > left && dest < addr) {
1291 right = middle - 1;
1292 } else if (middle < right && middle_ptr->data_size() == region_size) {
1293 left = middle + 1;
1294 } else {
1295 return middle_ptr;
1296 }
1297 }
1298 return sd.region(left);
1299 }
1300
1301 ParallelCompactData::RegionData*
1302 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1303 const RegionData* end,
1304 size_t dead_words)
1305 {
1306 ParallelCompactData& sd = summary_data();
1307 size_t left = sd.region(beg);
1308 size_t right = end > beg ? sd.region(end) - 1 : left;
1309
1310 // Binary search.
1311 while (left < right) {
1312 // Equivalent to (left + right) / 2, but does not overflow.
1313 const size_t middle = left + (right - left) / 2;
1314 RegionData* const middle_ptr = sd.region(middle);
1315 HeapWord* const dest = middle_ptr->destination();
1316 HeapWord* const addr = sd.region_to_addr(middle);
1317 assert(dest != NULL, "sanity");
1318 assert(dest <= addr, "must move left");
1319
1320 const size_t dead_to_left = pointer_delta(addr, dest);
1321 if (middle > left && dead_to_left > dead_words) {
1322 right = middle - 1;
1323 } else if (middle < right && dead_to_left < dead_words) {
1324 left = middle + 1;
1325 } else {
1326 return middle_ptr;
1327 }
1328 }
1329 return sd.region(left);
1330 }
1331
1332 // The result is valid during the summary phase, after the initial summarization
1333 // of each space into itself, and before final summarization.
1334 inline double
1335 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1336 HeapWord* const bottom,
1337 HeapWord* const top,
1338 HeapWord* const new_top)
1339 {
1340 ParallelCompactData& sd = summary_data();
1341
1342 assert(cp != NULL, "sanity");
1343 assert(bottom != NULL, "sanity");
1344 assert(top != NULL, "sanity");
1345 assert(new_top != NULL, "sanity");
1346 assert(top >= new_top, "summary data problem?");
1347 assert(new_top > bottom, "space is empty; should not be here");
1348 assert(new_top >= cp->destination(), "sanity");
1349 assert(top >= sd.region_to_addr(cp), "sanity");
1350
1351 HeapWord* const destination = cp->destination();
1352 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1353 const size_t compacted_region_live = pointer_delta(new_top, destination);
1354 const size_t compacted_region_used = pointer_delta(top,
1355 sd.region_to_addr(cp));
1356 const size_t reclaimable = compacted_region_used - compacted_region_live;
1357
1358 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1359 return double(reclaimable) / divisor;
1360 }
1361
1362 // Return the address of the end of the dense prefix, a.k.a. the start of the
1363 // compacted region. The address is always on a region boundary.
1364 //
1365 // Completely full regions at the left are skipped, since no compaction can
1366 // occur in those regions. Then the maximum amount of dead wood to allow is
1367 // computed, based on the density (amount live / capacity) of the generation;
1368 // the region with approximately that amount of dead space to the left is
1369 // identified as the limit region. Regions between the last completely full
1370 // region and the limit region are scanned and the one that has the best
1371 // (maximum) reclaimed_ratio() is selected.
1372 HeapWord*
1373 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1374 bool maximum_compaction)
1375 {
1376 if (ParallelOldGCSplitALot) {
1377 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1378 // The value was chosen to provoke splitting a young gen space; use it.
1379 return _space_info[id].dense_prefix();
1380 }
1381 }
1382
1383 const size_t region_size = ParallelCompactData::RegionSize;
1384 const ParallelCompactData& sd = summary_data();
1385
1386 const MutableSpace* const space = _space_info[id].space();
1387 HeapWord* const top = space->top();
1388 HeapWord* const top_aligned_up = sd.region_align_up(top);
1389 HeapWord* const new_top = _space_info[id].new_top();
1390 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1391 HeapWord* const bottom = space->bottom();
1392 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1393 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1394 const RegionData* const new_top_cp =
1395 sd.addr_to_region_ptr(new_top_aligned_up);
1396
1397 // Skip full regions at the beginning of the space--they are necessarily part
1398 // of the dense prefix.
1399 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1400 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1401 space->is_empty(), "no dead space allowed to the left");
1402 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1403 "region must have dead space");
1404
1405 // The gc number is saved whenever a maximum compaction is done, and used to
1406 // determine when the maximum compaction interval has expired. This avoids
1407 // successive max compactions for different reasons.
1408 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1409 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1410 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1411 total_invocations() == HeapFirstMaximumCompactionCount;
1412 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1413 _maximum_compaction_gc_num = total_invocations();
1414 return sd.region_to_addr(full_cp);
1415 }
1416
1417 const size_t space_live = pointer_delta(new_top, bottom);
1418 const size_t space_used = space->used_in_words();
1419 const size_t space_capacity = space->capacity_in_words();
1420
1421 const double density = double(space_live) / double(space_capacity);
1422 const size_t min_percent_free = MarkSweepDeadRatio;
1423 const double limiter = dead_wood_limiter(density, min_percent_free);
1424 const size_t dead_wood_max = space_used - space_live;
1425 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1426 dead_wood_max);
1427
1428 if (TraceParallelOldGCDensePrefix) {
1429 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1430 "space_cap=" SIZE_FORMAT,
1431 space_live, space_used,
1432 space_capacity);
1433 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1434 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1435 density, min_percent_free, limiter,
1436 dead_wood_max, dead_wood_limit);
1437 }
1438
1439 // Locate the region with the desired amount of dead space to the left.
1440 const RegionData* const limit_cp =
1441 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1442
1443 // Scan from the first region with dead space to the limit region and find the
1444 // one with the best (largest) reclaimed ratio.
1445 double best_ratio = 0.0;
1446 const RegionData* best_cp = full_cp;
1447 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1448 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1449 if (tmp_ratio > best_ratio) {
1450 best_cp = cp;
1451 best_ratio = tmp_ratio;
1452 }
1453 }
1454
1455 #if 0
1456 // Something to consider: if the region with the best ratio is 'close to' the
1457 // first region w/free space, choose the first region with free space
1458 // ("first-free"). The first-free region is usually near the start of the
1459 // heap, which means we are copying most of the heap already, so copy a bit
1460 // more to get complete compaction.
1461 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1462 _maximum_compaction_gc_num = total_invocations();
1463 best_cp = full_cp;
1464 }
1465 #endif // #if 0
1466
1467 return sd.region_to_addr(best_cp);
1468 }
1469
1470 #ifndef PRODUCT
1471 void
1472 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1473 size_t words)
1474 {
1475 if (TraceParallelOldGCSummaryPhase) {
1476 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1477 SIZE_FORMAT, start, start + words, words);
1478 }
1479
1480 ObjectStartArray* const start_array = _space_info[id].start_array();
1481 CollectedHeap::fill_with_objects(start, words);
1482 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1483 _mark_bitmap.mark_obj(p, words);
1484 _summary_data.add_obj(p, words);
1485 start_array->allocate_block(p);
1486 }
1487 }
1488
1489 void
1490 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1491 {
1492 ParallelCompactData& sd = summary_data();
1493 MutableSpace* space = _space_info[id].space();
1494
1495 // Find the source and destination start addresses.
1496 HeapWord* const src_addr = sd.region_align_down(start);
1497 HeapWord* dst_addr;
1498 if (src_addr < start) {
1499 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1500 } else if (src_addr > space->bottom()) {
1501 // The start (the original top() value) is aligned to a region boundary so
1502 // the associated region does not have a destination. Compute the
1503 // destination from the previous region.
1504 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1505 dst_addr = cp->destination() + cp->data_size();
1506 } else {
1507 // Filling the entire space.
1508 dst_addr = space->bottom();
1509 }
1510 assert(dst_addr != NULL, "sanity");
1511
1512 // Update the summary data.
1513 bool result = _summary_data.summarize(_space_info[id].split_info(),
1514 src_addr, space->top(), NULL,
1515 dst_addr, space->end(),
1516 _space_info[id].new_top_addr());
1517 assert(result, "should not fail: bad filler object size");
1518 }
1519
1520 void
1521 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1522 {
1523 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1524 return;
1525 }
1526
1527 MutableSpace* const space = _space_info[id].space();
1528 if (space->is_empty()) {
1529 HeapWord* b = space->bottom();
1530 HeapWord* t = b + space->capacity_in_words() / 2;
1531 space->set_top(t);
1532 if (ZapUnusedHeapArea) {
1533 space->set_top_for_allocations();
1534 }
1535
1536 size_t min_size = CollectedHeap::min_fill_size();
1537 size_t obj_len = min_size;
1538 while (b + obj_len <= t) {
1539 CollectedHeap::fill_with_object(b, obj_len);
1540 mark_bitmap()->mark_obj(b, obj_len);
1541 summary_data().add_obj(b, obj_len);
1542 b += obj_len;
1543 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1544 }
1545 if (b < t) {
1546 // The loop didn't completely fill to t (top); adjust top downward.
1547 space->set_top(b);
1548 if (ZapUnusedHeapArea) {
1549 space->set_top_for_allocations();
1550 }
1551 }
1552
1553 HeapWord** nta = _space_info[id].new_top_addr();
1554 bool result = summary_data().summarize(_space_info[id].split_info(),
1555 space->bottom(), space->top(), NULL,
1556 space->bottom(), space->end(), nta);
1557 assert(result, "space must fit into itself");
1558 }
1559 }
1560
1561 void
1562 PSParallelCompact::provoke_split(bool & max_compaction)
1563 {
1564 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1565 return;
1566 }
1567
1568 const size_t region_size = ParallelCompactData::RegionSize;
1569 ParallelCompactData& sd = summary_data();
1570
1571 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1572 MutableSpace* const from_space = _space_info[from_space_id].space();
1573 const size_t eden_live = pointer_delta(eden_space->top(),
1574 _space_info[eden_space_id].new_top());
1575 const size_t from_live = pointer_delta(from_space->top(),
1576 _space_info[from_space_id].new_top());
1577
1578 const size_t min_fill_size = CollectedHeap::min_fill_size();
1579 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1580 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1581 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1582 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1583
1584 // Choose the space to split; need at least 2 regions live (or fillable).
1585 SpaceId id;
1586 MutableSpace* space;
1587 size_t live_words;
1588 size_t fill_words;
1589 if (eden_live + eden_fillable >= region_size * 2) {
1590 id = eden_space_id;
1591 space = eden_space;
1592 live_words = eden_live;
1593 fill_words = eden_fillable;
1594 } else if (from_live + from_fillable >= region_size * 2) {
1595 id = from_space_id;
1596 space = from_space;
1597 live_words = from_live;
1598 fill_words = from_fillable;
1599 } else {
1600 return; // Give up.
1601 }
1602 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1603
1604 if (live_words < region_size * 2) {
1605 // Fill from top() to end() w/live objects of mixed sizes.
1606 HeapWord* const fill_start = space->top();
1607 live_words += fill_words;
1608
1609 space->set_top(fill_start + fill_words);
1610 if (ZapUnusedHeapArea) {
1611 space->set_top_for_allocations();
1612 }
1613
1614 HeapWord* cur_addr = fill_start;
1615 while (fill_words > 0) {
1616 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1617 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1618 if (fill_words - cur_size < min_fill_size) {
1619 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1620 }
1621
1622 CollectedHeap::fill_with_object(cur_addr, cur_size);
1623 mark_bitmap()->mark_obj(cur_addr, cur_size);
1624 sd.add_obj(cur_addr, cur_size);
1625
1626 cur_addr += cur_size;
1627 fill_words -= cur_size;
1628 }
1629
1630 summarize_new_objects(id, fill_start);
1631 }
1632
1633 max_compaction = false;
1634
1635 // Manipulate the old gen so that it has room for about half of the live data
1636 // in the target young gen space (live_words / 2).
1637 id = old_space_id;
1638 space = _space_info[id].space();
1639 const size_t free_at_end = space->free_in_words();
1640 const size_t free_target = align_object_size(live_words / 2);
1641 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1642
1643 if (free_at_end >= free_target + min_fill_size) {
1644 // Fill space above top() and set the dense prefix so everything survives.
1645 HeapWord* const fill_start = space->top();
1646 const size_t fill_size = free_at_end - free_target;
1647 space->set_top(space->top() + fill_size);
1648 if (ZapUnusedHeapArea) {
1649 space->set_top_for_allocations();
1650 }
1651 fill_with_live_objects(id, fill_start, fill_size);
1652 summarize_new_objects(id, fill_start);
1653 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1654 } else if (dead + free_at_end > free_target) {
1655 // Find a dense prefix that makes the right amount of space available.
1656 HeapWord* cur = sd.region_align_down(space->top());
1657 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1658 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1659 while (dead_to_right < free_target) {
1660 cur -= region_size;
1661 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1662 dead_to_right = pointer_delta(space->end(), cur_destination);
1663 }
1664 _space_info[id].set_dense_prefix(cur);
1665 }
1666 }
1667 #endif // #ifndef PRODUCT
1668
1669 void PSParallelCompact::summarize_spaces_quick()
1670 {
1671 for (unsigned int i = 0; i < last_space_id; ++i) {
1672 const MutableSpace* space = _space_info[i].space();
1673 HeapWord** nta = _space_info[i].new_top_addr();
1674 bool result = _summary_data.summarize(_space_info[i].split_info(),
1675 space->bottom(), space->top(), NULL,
1676 space->bottom(), space->end(), nta);
1677 assert(result, "space must fit into itself");
1678 _space_info[i].set_dense_prefix(space->bottom());
1679 }
1680
1681 #ifndef PRODUCT
1682 if (ParallelOldGCSplitALot) {
1683 provoke_split_fill_survivor(to_space_id);
1684 }
1685 #endif // #ifndef PRODUCT
1686 }
1687
1688 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1689 {
1690 HeapWord* const dense_prefix_end = dense_prefix(id);
1691 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1692 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1693 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1694 // Only enough dead space is filled so that any remaining dead space to the
1695 // left is larger than the minimum filler object. (The remainder is filled
1696 // during the copy/update phase.)
1697 //
1698 // The size of the dead space to the right of the boundary is not a
1699 // concern, since compaction will be able to use whatever space is
1700 // available.
1701 //
1702 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1703 // surrounds the space to be filled with an object.
1704 //
1705 // In the 32-bit VM, each bit represents two 32-bit words:
1706 // +---+
1707 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1708 // end_bits: ... x x x | 0 | || 0 x x ...
1709 // +---+
1710 //
1711 // In the 64-bit VM, each bit represents one 64-bit word:
1712 // +------------+
1713 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1714 // end_bits: ... x x 1 | 0 || 0 | x x ...
1715 // +------------+
1716 // +-------+
1717 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1718 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1719 // +-------+
1720 // +-----------+
1721 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1722 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1723 // +-----------+
1724 // +-------+
1725 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1726 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1727 // +-------+
1728
1729 // Initially assume case a, c or e will apply.
1730 size_t obj_len = CollectedHeap::min_fill_size();
1731 HeapWord* obj_beg = dense_prefix_end - obj_len;
1732
1733 #ifdef _LP64
1734 if (MinObjAlignment > 1) { // object alignment > heap word size
1735 // Cases a, c or e.
1736 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1737 // Case b above.
1738 obj_beg = dense_prefix_end - 1;
1739 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1740 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1741 // Case d above.
1742 obj_beg = dense_prefix_end - 3;
1743 obj_len = 3;
1744 }
1745 #endif // #ifdef _LP64
1746
1747 CollectedHeap::fill_with_object(obj_beg, obj_len);
1748 _mark_bitmap.mark_obj(obj_beg, obj_len);
1749 _summary_data.add_obj(obj_beg, obj_len);
1750 assert(start_array(id) != NULL, "sanity");
1751 start_array(id)->allocate_block(obj_beg);
1752 }
1753 }
1754
1755 void
1756 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1757 {
1758 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1759 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1760 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1761 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1762 cur->set_source_region(0);
1763 }
1764 }
1765
1766 void
1767 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1768 {
1769 assert(id < last_space_id, "id out of range");
1770 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1771 ParallelOldGCSplitALot && id == old_space_id,
1772 "should have been reset in summarize_spaces_quick()");
1773
1774 const MutableSpace* space = _space_info[id].space();
1775 if (_space_info[id].new_top() != space->bottom()) {
1776 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1777 _space_info[id].set_dense_prefix(dense_prefix_end);
1778
1779 #ifndef PRODUCT
1780 if (TraceParallelOldGCDensePrefix) {
1781 print_dense_prefix_stats("ratio", id, maximum_compaction,
1782 dense_prefix_end);
1783 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1784 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1785 }
1786 #endif // #ifndef PRODUCT
1787
1788 // Recompute the summary data, taking into account the dense prefix. If
1789 // every last byte will be reclaimed, then the existing summary data which
1790 // compacts everything can be left in place.
1791 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1792 // If dead space crosses the dense prefix boundary, it is (at least
1793 // partially) filled with a dummy object, marked live and added to the
1794 // summary data. This simplifies the copy/update phase and must be done
1795 // before the final locations of objects are determined, to prevent
1796 // leaving a fragment of dead space that is too small to fill.
1797 fill_dense_prefix_end(id);
1798
1799 // Compute the destination of each Region, and thus each object.
1800 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1801 _summary_data.summarize(_space_info[id].split_info(),
1802 dense_prefix_end, space->top(), NULL,
1803 dense_prefix_end, space->end(),
1804 _space_info[id].new_top_addr());
1805 }
1806 }
1807
1808 if (TraceParallelOldGCSummaryPhase) {
1809 const size_t region_size = ParallelCompactData::RegionSize;
1810 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1811 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1812 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1813 HeapWord* const new_top = _space_info[id].new_top();
1814 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1815 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1816 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1817 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1818 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1819 id, space->capacity_in_words(), dense_prefix_end,
1820 dp_region, dp_words / region_size,
1821 cr_words / region_size, new_top);
1822 }
1823 }
1824
1825 #ifndef PRODUCT
1826 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1827 HeapWord* dst_beg, HeapWord* dst_end,
1828 SpaceId src_space_id,
1829 HeapWord* src_beg, HeapWord* src_end)
1830 {
1831 if (TraceParallelOldGCSummaryPhase) {
1832 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1833 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1834 SIZE_FORMAT "-" SIZE_FORMAT " "
1835 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1836 SIZE_FORMAT "-" SIZE_FORMAT,
1837 src_space_id, space_names[src_space_id],
1838 dst_space_id, space_names[dst_space_id],
1839 src_beg, src_end,
1840 _summary_data.addr_to_region_idx(src_beg),
1841 _summary_data.addr_to_region_idx(src_end),
1842 dst_beg, dst_end,
1843 _summary_data.addr_to_region_idx(dst_beg),
1844 _summary_data.addr_to_region_idx(dst_end));
1845 }
1846 }
1847 #endif // #ifndef PRODUCT
1848
1849 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1850 bool maximum_compaction)
1851 {
1852 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer);
1853 // trace("2");
1854
1855 #ifdef ASSERT
1856 if (TraceParallelOldGCMarkingPhase) {
1857 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1858 "add_obj_bytes=" SIZE_FORMAT,
1859 add_obj_count, add_obj_size * HeapWordSize);
1860 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1861 "mark_bitmap_bytes=" SIZE_FORMAT,
1862 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1863 }
1864 #endif // #ifdef ASSERT
1865
1866 // Quick summarization of each space into itself, to see how much is live.
1867 summarize_spaces_quick();
1868
1869 if (TraceParallelOldGCSummaryPhase) {
1870 tty->print_cr("summary_phase: after summarizing each space to self");
1871 Universe::print();
1872 NOT_PRODUCT(print_region_ranges());
1873 if (Verbose) {
1874 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1875 }
1876 }
1877
1878 // The amount of live data that will end up in old space (assuming it fits).
1879 size_t old_space_total_live = 0;
1880 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1881 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1882 _space_info[id].space()->bottom());
1883 }
1884
1885 MutableSpace* const old_space = _space_info[old_space_id].space();
1886 const size_t old_capacity = old_space->capacity_in_words();
1887 if (old_space_total_live > old_capacity) {
1888 // XXX - should also try to expand
1889 maximum_compaction = true;
1890 }
1891 #ifndef PRODUCT
1892 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1893 provoke_split(maximum_compaction);
1894 }
1895 #endif // #ifndef PRODUCT
1896
1897 // Old generations.
1898 summarize_space(old_space_id, maximum_compaction);
1899
1900 // Summarize the remaining spaces in the young gen. The initial target space
1901 // is the old gen. If a space does not fit entirely into the target, then the
1902 // remainder is compacted into the space itself and that space becomes the new
1903 // target.
1904 SpaceId dst_space_id = old_space_id;
1905 HeapWord* dst_space_end = old_space->end();
1906 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1907 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1908 const MutableSpace* space = _space_info[id].space();
1909 const size_t live = pointer_delta(_space_info[id].new_top(),
1910 space->bottom());
1911 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1912
1913 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1914 SpaceId(id), space->bottom(), space->top());)
1915 if (live > 0 && live <= available) {
1916 // All the live data will fit.
1917 bool done = _summary_data.summarize(_space_info[id].split_info(),
1918 space->bottom(), space->top(),
1919 NULL,
1920 *new_top_addr, dst_space_end,
1921 new_top_addr);
1922 assert(done, "space must fit into old gen");
1923
1924 // Reset the new_top value for the space.
1925 _space_info[id].set_new_top(space->bottom());
1926 } else if (live > 0) {
1927 // Attempt to fit part of the source space into the target space.
1928 HeapWord* next_src_addr = NULL;
1929 bool done = _summary_data.summarize(_space_info[id].split_info(),
1930 space->bottom(), space->top(),
1931 &next_src_addr,
1932 *new_top_addr, dst_space_end,
1933 new_top_addr);
1934 assert(!done, "space should not fit into old gen");
1935 assert(next_src_addr != NULL, "sanity");
1936
1937 // The source space becomes the new target, so the remainder is compacted
1938 // within the space itself.
1939 dst_space_id = SpaceId(id);
1940 dst_space_end = space->end();
1941 new_top_addr = _space_info[id].new_top_addr();
1942 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1943 space->bottom(), dst_space_end,
1944 SpaceId(id), next_src_addr, space->top());)
1945 done = _summary_data.summarize(_space_info[id].split_info(),
1946 next_src_addr, space->top(),
1947 NULL,
1948 space->bottom(), dst_space_end,
1949 new_top_addr);
1950 assert(done, "space must fit when compacted into itself");
1951 assert(*new_top_addr <= space->top(), "usage should not grow");
1952 }
1953 }
1954
1955 if (TraceParallelOldGCSummaryPhase) {
1956 tty->print_cr("summary_phase: after final summarization");
1957 Universe::print();
1958 NOT_PRODUCT(print_region_ranges());
1959 if (Verbose) {
1960 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1961 }
1962 }
1963 }
1964
1965 // This method should contain all heap-specific policy for invoking a full
1966 // collection. invoke_no_policy() will only attempt to compact the heap; it
1967 // will do nothing further. If we need to bail out for policy reasons, scavenge
1968 // before full gc, or any other specialized behavior, it needs to be added here.
1969 //
1970 // Note that this method should only be called from the vm_thread while at a
1971 // safepoint.
1972 //
1973 // Note that the all_soft_refs_clear flag in the collector policy
1974 // may be true because this method can be called without intervening
1975 // activity. For example when the heap space is tight and full measure
1976 // are being taken to free space.
1977 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1978 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1979 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1980 "should be in vm thread");
1981
1982 ParallelScavengeHeap* heap = gc_heap();
1983 GCCause::Cause gc_cause = heap->gc_cause();
1984 assert(!heap->is_gc_active(), "not reentrant");
1985
1986 PSAdaptiveSizePolicy* policy = heap->size_policy();
1987 IsGCActiveMark mark;
1988
1989 if (ScavengeBeforeFullGC) {
1990 PSScavenge::invoke_no_policy();
1991 }
1992
1993 const bool clear_all_soft_refs =
1994 heap->collector_policy()->should_clear_all_soft_refs();
1995
1996 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1997 maximum_heap_compaction);
1998 }
1999
2000 // This method contains no policy. You should probably
2001 // be calling invoke() instead.
2002 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
2003 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
2004 assert(ref_processor() != NULL, "Sanity");
2005
2006 if (GC_locker::check_active_before_gc()) {
2007 return false;
2008 }
2009
2010 ParallelScavengeHeap* heap = gc_heap();
2011
2012 _gc_timer.register_gc_start();
2013 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
2014
2015 TimeStamp marking_start;
2016 TimeStamp compaction_start;
2017 TimeStamp collection_exit;
2018
2019 GCCause::Cause gc_cause = heap->gc_cause();
2020 PSYoungGen* young_gen = heap->young_gen();
2021 PSOldGen* old_gen = heap->old_gen();
2022 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2023
2024 // The scope of casr should end after code that can change
2025 // CollectorPolicy::_should_clear_all_soft_refs.
2026 ClearedAllSoftRefs casr(maximum_heap_compaction,
2027 heap->collector_policy());
2028
2029 if (ZapUnusedHeapArea) {
2030 // Save information needed to minimize mangling
2031 heap->record_gen_tops_before_GC();
2032 }
2033
2034 heap->pre_full_gc_dump(&_gc_timer);
2035
2036 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2037
2038 // Make sure data structures are sane, make the heap parsable, and do other
2039 // miscellaneous bookkeeping.
2040 PreGCValues pre_gc_values;
2041 pre_compact(&pre_gc_values);
2042
2043 // Get the compaction manager reserved for the VM thread.
2044 ParCompactionManager* const vmthread_cm =
2045 ParCompactionManager::manager_array(gc_task_manager()->workers());
2046
2047 // Place after pre_compact() where the number of invocations is incremented.
2048 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2049
2050 {
2051 ResourceMark rm;
2052 HandleMark hm;
2053
2054 // Set the number of GC threads to be used in this collection
2055 gc_task_manager()->set_active_gang();
2056 gc_task_manager()->task_idle_workers();
2057 heap->set_par_threads(gc_task_manager()->active_workers());
2058
2059 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2060 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2061 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
2062 TraceCollectorStats tcs(counters());
2063 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2064
2065 if (TraceGen1Time) accumulated_time()->start();
2066
2067 // Let the size policy know we're starting
2068 size_policy->major_collection_begin();
2069
2070 CodeCache::gc_prologue();
2071 Threads::gc_prologue();
2072
2073 COMPILER2_PRESENT(DerivedPointerTable::clear());
2074
2075 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2076 ref_processor()->setup_policy(maximum_heap_compaction);
2077
2078 bool marked_for_unloading = false;
2079
2080 marking_start.update();
2081 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
2082
2083 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2084 && gc_cause == GCCause::_java_lang_system_gc;
2085 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2086
2087 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2088 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2089
2090 // adjust_roots() updates Universe::_intArrayKlassObj which is
2091 // needed by the compaction for filling holes in the dense prefix.
2092 adjust_roots();
2093
2094 compaction_start.update();
2095 compact();
2096
2097 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2098 // done before resizing.
2099 post_compact();
2100
2101 // Let the size policy know we're done
2102 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2103
2104 if (UseAdaptiveSizePolicy) {
2105 if (PrintAdaptiveSizePolicy) {
2106 gclog_or_tty->print("AdaptiveSizeStart: ");
2107 gclog_or_tty->stamp();
2108 gclog_or_tty->print_cr(" collection: %d ",
2109 heap->total_collections());
2110 if (Verbose) {
2111 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT
2112 " young_gen_capacity: " SIZE_FORMAT,
2113 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2114 }
2115 }
2116
2117 // Don't check if the size_policy is ready here. Let
2118 // the size_policy check that internally.
2119 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2120 ((gc_cause != GCCause::_java_lang_system_gc) ||
2121 UseAdaptiveSizePolicyWithSystemGC)) {
2122 // Swap the survivor spaces if from_space is empty. The
2123 // resize_young_gen() called below is normally used after
2124 // a successful young GC and swapping of survivor spaces;
2125 // otherwise, it will fail to resize the young gen with
2126 // the current implementation.
2127 if (young_gen->from_space()->is_empty()) {
2128 young_gen->from_space()->clear(SpaceDecorator::Mangle);
2129 young_gen->swap_spaces();
2130 }
2131
2132 // Calculate optimal free space amounts
2133 assert(young_gen->max_size() >
2134 young_gen->from_space()->capacity_in_bytes() +
2135 young_gen->to_space()->capacity_in_bytes(),
2136 "Sizes of space in young gen are out-of-bounds");
2137
2138 size_t young_live = young_gen->used_in_bytes();
2139 size_t eden_live = young_gen->eden_space()->used_in_bytes();
2140 size_t old_live = old_gen->used_in_bytes();
2141 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2142 size_t max_old_gen_size = old_gen->max_gen_size();
2143 size_t max_eden_size = young_gen->max_size() -
2144 young_gen->from_space()->capacity_in_bytes() -
2145 young_gen->to_space()->capacity_in_bytes();
2146
2147 // Used for diagnostics
2148 size_policy->clear_generation_free_space_flags();
2149
2150 size_policy->compute_generations_free_space(young_live,
2151 eden_live,
2152 old_live,
2153 cur_eden,
2154 max_old_gen_size,
2155 max_eden_size,
2156 true /* full gc*/);
2157
2158 size_policy->check_gc_overhead_limit(young_live,
2159 eden_live,
2160 max_old_gen_size,
2161 max_eden_size,
2162 true /* full gc*/,
2163 gc_cause,
2164 heap->collector_policy());
2165
2166 size_policy->decay_supplemental_growth(true /* full gc*/);
2167
2168 heap->resize_old_gen(
2169 size_policy->calculated_old_free_size_in_bytes());
2170
2171 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
2172 size_policy->calculated_survivor_size_in_bytes());
2173 }
2174 if (PrintAdaptiveSizePolicy) {
2175 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2176 heap->total_collections());
2177 }
2178 }
2179
2180 if (UsePerfData) {
2181 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2182 counters->update_counters();
2183 counters->update_old_capacity(old_gen->capacity_in_bytes());
2184 counters->update_young_capacity(young_gen->capacity_in_bytes());
2185 }
2186
2187 heap->resize_all_tlabs();
2188
2189 // Resize the metaspace capacity after a collection
2190 MetaspaceGC::compute_new_size();
2191
2192 if (TraceGen1Time) accumulated_time()->stop();
2193
2194 if (PrintGC) {
2195 if (PrintGCDetails) {
2196 // No GC timestamp here. This is after GC so it would be confusing.
2197 young_gen->print_used_change(pre_gc_values.young_gen_used());
2198 old_gen->print_used_change(pre_gc_values.old_gen_used());
2199 heap->print_heap_change(pre_gc_values.heap_used());
2200 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2201 } else {
2202 heap->print_heap_change(pre_gc_values.heap_used());
2203 }
2204 }
2205
2206 // Track memory usage and detect low memory
2207 MemoryService::track_memory_usage();
2208 heap->update_counters();
2209 gc_task_manager()->release_idle_workers();
2210 }
2211
2212 #ifdef ASSERT
2213 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2214 ParCompactionManager* const cm =
2215 ParCompactionManager::manager_array(int(i));
2216 assert(cm->marking_stack()->is_empty(), "should be empty");
2217 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2218 }
2219 #endif // ASSERT
2220
2221 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2222 HandleMark hm; // Discard invalid handles created during verification
2223 Universe::verify(" VerifyAfterGC:");
2224 }
2225
2226 // Re-verify object start arrays
2227 if (VerifyObjectStartArray &&
2228 VerifyAfterGC) {
2229 old_gen->verify_object_start_array();
2230 }
2231
2232 if (ZapUnusedHeapArea) {
2233 old_gen->object_space()->check_mangled_unused_area_complete();
2234 }
2235
2236 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2237
2238 collection_exit.update();
2239
2240 heap->print_heap_after_gc();
2241 heap->trace_heap_after_gc(&_gc_tracer);
2242
2243 if (PrintGCTaskTimeStamps) {
2244 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2245 INT64_FORMAT,
2246 marking_start.ticks(), compaction_start.ticks(),
2247 collection_exit.ticks());
2248 gc_task_manager()->print_task_time_stamps();
2249 }
2250
2251 heap->post_full_gc_dump(&_gc_timer);
2252
2253 #ifdef TRACESPINNING
2254 ParallelTaskTerminator::print_termination_counts();
2255 #endif
2256
2257 _gc_timer.register_gc_end();
2258
2259 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2260 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2261
2262 return true;
2263 }
2264
2265 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2266 PSYoungGen* young_gen,
2267 PSOldGen* old_gen) {
2268 MutableSpace* const eden_space = young_gen->eden_space();
2269 assert(!eden_space->is_empty(), "eden must be non-empty");
2270 assert(young_gen->virtual_space()->alignment() ==
2271 old_gen->virtual_space()->alignment(), "alignments do not match");
2272
2273 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2274 return false;
2275 }
2276
2277 // Both generations must be completely committed.
2278 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2279 return false;
2280 }
2281 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2282 return false;
2283 }
2284
2285 // Figure out how much to take from eden. Include the average amount promoted
2286 // in the total; otherwise the next young gen GC will simply bail out to a
2287 // full GC.
2288 const size_t alignment = old_gen->virtual_space()->alignment();
2289 const size_t eden_used = eden_space->used_in_bytes();
2290 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2291 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2292 const size_t eden_capacity = eden_space->capacity_in_bytes();
2293
2294 if (absorb_size >= eden_capacity) {
2295 return false; // Must leave some space in eden.
2296 }
2297
2298 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2299 if (new_young_size < young_gen->min_gen_size()) {
2300 return false; // Respect young gen minimum size.
2301 }
2302
2303 if (TraceAdaptiveGCBoundary && Verbose) {
2304 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2305 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2306 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2307 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2308 absorb_size / K,
2309 eden_capacity / K, (eden_capacity - absorb_size) / K,
2310 young_gen->from_space()->used_in_bytes() / K,
2311 young_gen->to_space()->used_in_bytes() / K,
2312 young_gen->capacity_in_bytes() / K, new_young_size / K);
2313 }
2314
2315 // Fill the unused part of the old gen.
2316 MutableSpace* const old_space = old_gen->object_space();
2317 HeapWord* const unused_start = old_space->top();
2318 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2319
2320 if (unused_words > 0) {
2321 if (unused_words < CollectedHeap::min_fill_size()) {
2322 return false; // If the old gen cannot be filled, must give up.
2323 }
2324 CollectedHeap::fill_with_objects(unused_start, unused_words);
2325 }
2326
2327 // Take the live data from eden and set both top and end in the old gen to
2328 // eden top. (Need to set end because reset_after_change() mangles the region
2329 // from end to virtual_space->high() in debug builds).
2330 HeapWord* const new_top = eden_space->top();
2331 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2332 absorb_size);
2333 young_gen->reset_after_change();
2334 old_space->set_top(new_top);
2335 old_space->set_end(new_top);
2336 old_gen->reset_after_change();
2337
2338 // Update the object start array for the filler object and the data from eden.
2339 ObjectStartArray* const start_array = old_gen->start_array();
2340 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2341 start_array->allocate_block(p);
2342 }
2343
2344 // Could update the promoted average here, but it is not typically updated at
2345 // full GCs and the value to use is unclear. Something like
2346 //
2347 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2348
2349 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2350 return true;
2351 }
2352
2353 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2354 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2355 "shouldn't return NULL");
2356 return ParallelScavengeHeap::gc_task_manager();
2357 }
2358
2359 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2360 bool maximum_heap_compaction,
2361 ParallelOldTracer *gc_tracer) {
2362 // Recursively traverse all live objects and mark them
2363 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer);
2364
2365 ParallelScavengeHeap* heap = gc_heap();
2366 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2367 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2368 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2369 ParallelTaskTerminator terminator(active_gc_threads, qset);
2370
2371 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2372 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2373
2374 // Need new claim bits before marking starts.
2375 ClassLoaderDataGraph::clear_claimed_marks();
2376
2377 {
2378 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer);
2379
2380 ParallelScavengeHeap::ParStrongRootsScope psrs;
2381
2382 GCTaskQueue* q = GCTaskQueue::create();
2383
2384 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2385 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2386 // We scan the thread roots in parallel
2387 Threads::create_thread_roots_marking_tasks(q);
2388 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2389 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2390 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2391 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2392 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2393 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2394 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2395
2396 if (active_gc_threads > 1) {
2397 for (uint j = 0; j < active_gc_threads; j++) {
2398 q->enqueue(new StealMarkingTask(&terminator));
2399 }
2400 }
2401
2402 gc_task_manager()->execute_and_wait(q);
2403 }
2404
2405 // Process reference objects found during marking
2406 {
2407 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer);
2408
2409 ReferenceProcessorStats stats;
2410 if (ref_processor()->processing_is_mt()) {
2411 RefProcTaskExecutor task_executor;
2412 stats = ref_processor()->process_discovered_references(
2413 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2414 &task_executor, &_gc_timer);
2415 } else {
2416 stats = ref_processor()->process_discovered_references(
2417 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2418 &_gc_timer);
2419 }
2420
2421 gc_tracer->report_gc_reference_stats(stats);
2422 }
2423
2424 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer);
2425
2426 // This is the point where the entire marking should have completed.
2427 assert(cm->marking_stacks_empty(), "Marking should have completed");
2428
2429 // Follow system dictionary roots and unload classes.
2430 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2431
2432 // Unload nmethods.
2433 CodeCache::do_unloading(is_alive_closure(), purged_class);
2434
2435 // Prune dead klasses from subklass/sibling/implementor lists.
2436 Klass::clean_weak_klass_links(is_alive_closure());
2437
2438 // Delete entries for dead interned strings.
2439 StringTable::unlink(is_alive_closure());
2440
2441 // Clean up unreferenced symbols in symbol table.
2442 SymbolTable::unlink();
2443 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2444 }
2445
2446 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2447 ClassLoaderData* cld) {
2448 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2449 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2450
2451 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2452 }
2453
2454 // This should be moved to the shared markSweep code!
2455 class PSAlwaysTrueClosure: public BoolObjectClosure {
2456 public:
2457 bool do_object_b(oop p) { return true; }
2458 };
2459 static PSAlwaysTrueClosure always_true;
2460
2461 void PSParallelCompact::adjust_roots() {
2462 // Adjust the pointers to reflect the new locations
2463 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer);
2464
2465 // Need new claim bits when tracing through and adjusting pointers.
2466 ClassLoaderDataGraph::clear_claimed_marks();
2467
2468 // General strong roots.
2469 Universe::oops_do(adjust_pointer_closure());
2470 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles
2471 CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2472 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2473 ObjectSynchronizer::oops_do(adjust_pointer_closure());
2474 FlatProfiler::oops_do(adjust_pointer_closure());
2475 Management::oops_do(adjust_pointer_closure());
2476 JvmtiExport::oops_do(adjust_pointer_closure());
2477 // SO_AllClasses
2478 SystemDictionary::oops_do(adjust_pointer_closure());
2479 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2480
2481 // Now adjust pointers in remaining weak roots. (All of which should
2482 // have been cleared if they pointed to non-surviving objects.)
2483 // Global (weak) JNI handles
2484 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2485
2486 CodeCache::oops_do(adjust_pointer_closure());
2487 StringTable::oops_do(adjust_pointer_closure());
2488 ref_processor()->weak_oops_do(adjust_pointer_closure());
2489 // Roots were visited so references into the young gen in roots
2490 // may have been scanned. Process them also.
2491 // Should the reference processor have a span that excludes
2492 // young gen objects?
2493 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2494 }
2495
2496 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2497 uint parallel_gc_threads)
2498 {
2499 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer);
2500
2501 // Find the threads that are active
2502 unsigned int which = 0;
2503
2504 const uint task_count = MAX2(parallel_gc_threads, 1U);
2505 for (uint j = 0; j < task_count; j++) {
2506 q->enqueue(new DrainStacksCompactionTask(j));
2507 ParCompactionManager::verify_region_list_empty(j);
2508 // Set the region stacks variables to "no" region stack values
2509 // so that they will be recognized and needing a region stack
2510 // in the stealing tasks if they do not get one by executing
2511 // a draining stack.
2512 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2513 cm->set_region_stack(NULL);
2514 cm->set_region_stack_index((uint)max_uintx);
2515 }
2516 ParCompactionManager::reset_recycled_stack_index();
2517
2518 // Find all regions that are available (can be filled immediately) and
2519 // distribute them to the thread stacks. The iteration is done in reverse
2520 // order (high to low) so the regions will be removed in ascending order.
2521
2522 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2523
2524 size_t fillable_regions = 0; // A count for diagnostic purposes.
2525 // A region index which corresponds to the tasks created above.
2526 // "which" must be 0 <= which < task_count
2527
2528 which = 0;
2529 // id + 1 is used to test termination so unsigned can
2530 // be used with an old_space_id == 0.
2531 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2532 SpaceInfo* const space_info = _space_info + id;
2533 MutableSpace* const space = space_info->space();
2534 HeapWord* const new_top = space_info->new_top();
2535
2536 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2537 const size_t end_region =
2538 sd.addr_to_region_idx(sd.region_align_up(new_top));
2539
2540 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2541 if (sd.region(cur)->claim_unsafe()) {
2542 ParCompactionManager::region_list_push(which, cur);
2543
2544 if (TraceParallelOldGCCompactionPhase && Verbose) {
2545 const size_t count_mod_8 = fillable_regions & 7;
2546 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2547 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2548 if (count_mod_8 == 7) gclog_or_tty->cr();
2549 }
2550
2551 NOT_PRODUCT(++fillable_regions;)
2552
2553 // Assign regions to tasks in round-robin fashion.
2554 if (++which == task_count) {
2555 assert(which <= parallel_gc_threads,
2556 "Inconsistent number of workers");
2557 which = 0;
2558 }
2559 }
2560 }
2561 }
2562
2563 if (TraceParallelOldGCCompactionPhase) {
2564 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2565 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions);
2566 }
2567 }
2568
2569 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2570
2571 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2572 uint parallel_gc_threads) {
2573 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer);
2574
2575 ParallelCompactData& sd = PSParallelCompact::summary_data();
2576
2577 // Iterate over all the spaces adding tasks for updating
2578 // regions in the dense prefix. Assume that 1 gc thread
2579 // will work on opening the gaps and the remaining gc threads
2580 // will work on the dense prefix.
2581 unsigned int space_id;
2582 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2583 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2584 const MutableSpace* const space = _space_info[space_id].space();
2585
2586 if (dense_prefix_end == space->bottom()) {
2587 // There is no dense prefix for this space.
2588 continue;
2589 }
2590
2591 // The dense prefix is before this region.
2592 size_t region_index_end_dense_prefix =
2593 sd.addr_to_region_idx(dense_prefix_end);
2594 RegionData* const dense_prefix_cp =
2595 sd.region(region_index_end_dense_prefix);
2596 assert(dense_prefix_end == space->end() ||
2597 dense_prefix_cp->available() ||
2598 dense_prefix_cp->claimed(),
2599 "The region after the dense prefix should always be ready to fill");
2600
2601 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2602
2603 // Is there dense prefix work?
2604 size_t total_dense_prefix_regions =
2605 region_index_end_dense_prefix - region_index_start;
2606 // How many regions of the dense prefix should be given to
2607 // each thread?
2608 if (total_dense_prefix_regions > 0) {
2609 uint tasks_for_dense_prefix = 1;
2610 if (total_dense_prefix_regions <=
2611 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2612 // Don't over partition. This assumes that
2613 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2614 // so there are not many regions to process.
2615 tasks_for_dense_prefix = parallel_gc_threads;
2616 } else {
2617 // Over partition
2618 tasks_for_dense_prefix = parallel_gc_threads *
2619 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2620 }
2621 size_t regions_per_thread = total_dense_prefix_regions /
2622 tasks_for_dense_prefix;
2623 // Give each thread at least 1 region.
2624 if (regions_per_thread == 0) {
2625 regions_per_thread = 1;
2626 }
2627
2628 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2629 if (region_index_start >= region_index_end_dense_prefix) {
2630 break;
2631 }
2632 // region_index_end is not processed
2633 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2634 region_index_end_dense_prefix);
2635 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2636 region_index_start,
2637 region_index_end));
2638 region_index_start = region_index_end;
2639 }
2640 }
2641 // This gets any part of the dense prefix that did not
2642 // fit evenly.
2643 if (region_index_start < region_index_end_dense_prefix) {
2644 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2645 region_index_start,
2646 region_index_end_dense_prefix));
2647 }
2648 }
2649 }
2650
2651 void PSParallelCompact::enqueue_region_stealing_tasks(
2652 GCTaskQueue* q,
2653 ParallelTaskTerminator* terminator_ptr,
2654 uint parallel_gc_threads) {
2655 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer);
2656
2657 // Once a thread has drained it's stack, it should try to steal regions from
2658 // other threads.
2659 if (parallel_gc_threads > 1) {
2660 for (uint j = 0; j < parallel_gc_threads; j++) {
2661 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2662 }
2663 }
2664 }
2665
2666 #ifdef ASSERT
2667 // Write a histogram of the number of times the block table was filled for a
2668 // region.
2669 void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
2670 {
2671 if (!TraceParallelOldGCCompactionPhase) return;
2672
2673 typedef ParallelCompactData::RegionData rd_t;
2674 ParallelCompactData& sd = summary_data();
2675
2676 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2677 MutableSpace* const spc = _space_info[id].space();
2678 if (spc->bottom() != spc->top()) {
2679 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2680 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2681 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2682
2683 size_t histo[5] = { 0, 0, 0, 0, 0 };
2684 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2685 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2686
2687 for (const rd_t* cur = beg; cur < end; ++cur) {
2688 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2689 }
2690 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2691 for (size_t i = 0; i < histo_len; ++i) {
2692 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2693 histo[i], 100.0 * histo[i] / region_cnt);
2694 }
2695 out->cr();
2696 }
2697 }
2698 }
2699 #endif // #ifdef ASSERT
2700
2701 void PSParallelCompact::compact() {
2702 // trace("5");
2703 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer);
2704
2705 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2706 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2707 PSOldGen* old_gen = heap->old_gen();
2708 old_gen->start_array()->reset();
2709 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2710 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2711 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2712 ParallelTaskTerminator terminator(active_gc_threads, qset);
2713
2714 GCTaskQueue* q = GCTaskQueue::create();
2715 enqueue_region_draining_tasks(q, active_gc_threads);
2716 enqueue_dense_prefix_tasks(q, active_gc_threads);
2717 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2718
2719 {
2720 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer);
2721
2722 gc_task_manager()->execute_and_wait(q);
2723
2724 #ifdef ASSERT
2725 // Verify that all regions have been processed before the deferred updates.
2726 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2727 verify_complete(SpaceId(id));
2728 }
2729 #endif
2730 }
2731
2732 {
2733 // Update the deferred objects, if any. Any compaction manager can be used.
2734 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer);
2735 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2736 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2737 update_deferred_objects(cm, SpaceId(id));
2738 }
2739 }
2740
2741 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
2742 }
2743
2744 #ifdef ASSERT
2745 void PSParallelCompact::verify_complete(SpaceId space_id) {
2746 // All Regions between space bottom() to new_top() should be marked as filled
2747 // and all Regions between new_top() and top() should be available (i.e.,
2748 // should have been emptied).
2749 ParallelCompactData& sd = summary_data();
2750 SpaceInfo si = _space_info[space_id];
2751 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2752 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2753 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2754 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2755 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2756
2757 bool issued_a_warning = false;
2758
2759 size_t cur_region;
2760 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2761 const RegionData* const c = sd.region(cur_region);
2762 if (!c->completed()) {
2763 warning("region " SIZE_FORMAT " not filled: "
2764 "destination_count=" SIZE_FORMAT,
2765 cur_region, c->destination_count());
2766 issued_a_warning = true;
2767 }
2768 }
2769
2770 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2771 const RegionData* const c = sd.region(cur_region);
2772 if (!c->available()) {
2773 warning("region " SIZE_FORMAT " not empty: "
2774 "destination_count=" SIZE_FORMAT,
2775 cur_region, c->destination_count());
2776 issued_a_warning = true;
2777 }
2778 }
2779
2780 if (issued_a_warning) {
2781 print_region_ranges();
2782 }
2783 }
2784 #endif // #ifdef ASSERT
2785
2786 // Update interior oops in the ranges of regions [beg_region, end_region).
2787 void
2788 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2789 SpaceId space_id,
2790 size_t beg_region,
2791 size_t end_region) {
2792 ParallelCompactData& sd = summary_data();
2793 ParMarkBitMap* const mbm = mark_bitmap();
2794
2795 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2796 HeapWord* const end_addr = sd.region_to_addr(end_region);
2797 assert(beg_region <= end_region, "bad region range");
2798 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2799
2800 #ifdef ASSERT
2801 // Claim the regions to avoid triggering an assert when they are marked as
2802 // filled.
2803 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2804 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2805 }
2806 #endif // #ifdef ASSERT
2807
2808 if (beg_addr != space(space_id)->bottom()) {
2809 // Find the first live object or block of dead space that *starts* in this
2810 // range of regions. If a partial object crosses onto the region, skip it;
2811 // it will be marked for 'deferred update' when the object head is
2812 // processed. If dead space crosses onto the region, it is also skipped; it
2813 // will be filled when the prior region is processed. If neither of those
2814 // apply, the first word in the region is the start of a live object or dead
2815 // space.
2816 assert(beg_addr > space(space_id)->bottom(), "sanity");
2817 const RegionData* const cp = sd.region(beg_region);
2818 if (cp->partial_obj_size() != 0) {
2819 beg_addr = sd.partial_obj_end(beg_region);
2820 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2821 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2822 }
2823 }
2824
2825 if (beg_addr < end_addr) {
2826 // A live object or block of dead space starts in this range of Regions.
2827 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2828
2829 // Create closures and iterate.
2830 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2831 FillClosure fill_closure(cm, space_id);
2832 ParMarkBitMap::IterationStatus status;
2833 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2834 dense_prefix_end);
2835 if (status == ParMarkBitMap::incomplete) {
2836 update_closure.do_addr(update_closure.source());
2837 }
2838 }
2839
2840 // Mark the regions as filled.
2841 RegionData* const beg_cp = sd.region(beg_region);
2842 RegionData* const end_cp = sd.region(end_region);
2843 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2844 cp->set_completed();
2845 }
2846 }
2847
2848 // Return the SpaceId for the space containing addr. If addr is not in the
2849 // heap, last_space_id is returned. In debug mode it expects the address to be
2850 // in the heap and asserts such.
2851 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2852 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2853
2854 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2855 if (_space_info[id].space()->contains(addr)) {
2856 return SpaceId(id);
2857 }
2858 }
2859
2860 assert(false, "no space contains the addr");
2861 return last_space_id;
2862 }
2863
2864 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2865 SpaceId id) {
2866 assert(id < last_space_id, "bad space id");
2867
2868 ParallelCompactData& sd = summary_data();
2869 const SpaceInfo* const space_info = _space_info + id;
2870 ObjectStartArray* const start_array = space_info->start_array();
2871
2872 const MutableSpace* const space = space_info->space();
2873 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2874 HeapWord* const beg_addr = space_info->dense_prefix();
2875 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2876
2877 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2878 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2879 const RegionData* cur_region;
2880 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2881 HeapWord* const addr = cur_region->deferred_obj_addr();
2882 if (addr != NULL) {
2883 if (start_array != NULL) {
2884 start_array->allocate_block(addr);
2885 }
2886 oop(addr)->update_contents(cm);
2887 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2888 }
2889 }
2890 }
2891
2892 // Skip over count live words starting from beg, and return the address of the
2893 // next live word. Unless marked, the word corresponding to beg is assumed to
2894 // be dead. Callers must either ensure beg does not correspond to the middle of
2895 // an object, or account for those live words in some other way. Callers must
2896 // also ensure that there are enough live words in the range [beg, end) to skip.
2897 HeapWord*
2898 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2899 {
2900 assert(count > 0, "sanity");
2901
2902 ParMarkBitMap* m = mark_bitmap();
2903 idx_t bits_to_skip = m->words_to_bits(count);
2904 idx_t cur_beg = m->addr_to_bit(beg);
2905 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2906
2907 do {
2908 cur_beg = m->find_obj_beg(cur_beg, search_end);
2909 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2910 const size_t obj_bits = cur_end - cur_beg + 1;
2911 if (obj_bits > bits_to_skip) {
2912 return m->bit_to_addr(cur_beg + bits_to_skip);
2913 }
2914 bits_to_skip -= obj_bits;
2915 cur_beg = cur_end + 1;
2916 } while (bits_to_skip > 0);
2917
2918 // Skipping the desired number of words landed just past the end of an object.
2919 // Find the start of the next object.
2920 cur_beg = m->find_obj_beg(cur_beg, search_end);
2921 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2922 return m->bit_to_addr(cur_beg);
2923 }
2924
2925 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2926 SpaceId src_space_id,
2927 size_t src_region_idx)
2928 {
2929 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2930
2931 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2932 if (split_info.dest_region_addr() == dest_addr) {
2933 // The partial object ending at the split point contains the first word to
2934 // be copied to dest_addr.
2935 return split_info.first_src_addr();
2936 }
2937
2938 const ParallelCompactData& sd = summary_data();
2939 ParMarkBitMap* const bitmap = mark_bitmap();
2940 const size_t RegionSize = ParallelCompactData::RegionSize;
2941
2942 assert(sd.is_region_aligned(dest_addr), "not aligned");
2943 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2944 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2945 HeapWord* const src_region_destination = src_region_ptr->destination();
2946
2947 assert(dest_addr >= src_region_destination, "wrong src region");
2948 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2949
2950 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2951 HeapWord* const src_region_end = src_region_beg + RegionSize;
2952
2953 HeapWord* addr = src_region_beg;
2954 if (dest_addr == src_region_destination) {
2955 // Return the first live word in the source region.
2956 if (partial_obj_size == 0) {
2957 addr = bitmap->find_obj_beg(addr, src_region_end);
2958 assert(addr < src_region_end, "no objects start in src region");
2959 }
2960 return addr;
2961 }
2962
2963 // Must skip some live data.
2964 size_t words_to_skip = dest_addr - src_region_destination;
2965 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2966
2967 if (partial_obj_size >= words_to_skip) {
2968 // All the live words to skip are part of the partial object.
2969 addr += words_to_skip;
2970 if (partial_obj_size == words_to_skip) {
2971 // Find the first live word past the partial object.
2972 addr = bitmap->find_obj_beg(addr, src_region_end);
2973 assert(addr < src_region_end, "wrong src region");
2974 }
2975 return addr;
2976 }
2977
2978 // Skip over the partial object (if any).
2979 if (partial_obj_size != 0) {
2980 words_to_skip -= partial_obj_size;
2981 addr += partial_obj_size;
2982 }
2983
2984 // Skip over live words due to objects that start in the region.
2985 addr = skip_live_words(addr, src_region_end, words_to_skip);
2986 assert(addr < src_region_end, "wrong src region");
2987 return addr;
2988 }
2989
2990 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2991 SpaceId src_space_id,
2992 size_t beg_region,
2993 HeapWord* end_addr)
2994 {
2995 ParallelCompactData& sd = summary_data();
2996
2997 #ifdef ASSERT
2998 MutableSpace* const src_space = _space_info[src_space_id].space();
2999 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3000 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3001 "src_space_id does not match beg_addr");
3002 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3003 "src_space_id does not match end_addr");
3004 #endif // #ifdef ASSERT
3005
3006 RegionData* const beg = sd.region(beg_region);
3007 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3008
3009 // Regions up to new_top() are enqueued if they become available.
3010 HeapWord* const new_top = _space_info[src_space_id].new_top();
3011 RegionData* const enqueue_end =
3012 sd.addr_to_region_ptr(sd.region_align_up(new_top));
3013
3014 for (RegionData* cur = beg; cur < end; ++cur) {
3015 assert(cur->data_size() > 0, "region must have live data");
3016 cur->decrement_destination_count();
3017 if (cur < enqueue_end && cur->available() && cur->claim()) {
3018 cm->push_region(sd.region(cur));
3019 }
3020 }
3021 }
3022
3023 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3024 SpaceId& src_space_id,
3025 HeapWord*& src_space_top,
3026 HeapWord* end_addr)
3027 {
3028 typedef ParallelCompactData::RegionData RegionData;
3029
3030 ParallelCompactData& sd = PSParallelCompact::summary_data();
3031 const size_t region_size = ParallelCompactData::RegionSize;
3032
3033 size_t src_region_idx = 0;
3034
3035 // Skip empty regions (if any) up to the top of the space.
3036 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3037 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3038 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3039 const RegionData* const top_region_ptr =
3040 sd.addr_to_region_ptr(top_aligned_up);
3041 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3042 ++src_region_ptr;
3043 }
3044
3045 if (src_region_ptr < top_region_ptr) {
3046 // The next source region is in the current space. Update src_region_idx
3047 // and the source address to match src_region_ptr.
3048 src_region_idx = sd.region(src_region_ptr);
3049 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3050 if (src_region_addr > closure.source()) {
3051 closure.set_source(src_region_addr);
3052 }
3053 return src_region_idx;
3054 }
3055
3056 // Switch to a new source space and find the first non-empty region.
3057 unsigned int space_id = src_space_id + 1;
3058 assert(space_id < last_space_id, "not enough spaces");
3059
3060 HeapWord* const destination = closure.destination();
3061
3062 do {
3063 MutableSpace* space = _space_info[space_id].space();
3064 HeapWord* const bottom = space->bottom();
3065 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3066
3067 // Iterate over the spaces that do not compact into themselves.
3068 if (bottom_cp->destination() != bottom) {
3069 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3070 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3071
3072 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3073 if (src_cp->live_obj_size() > 0) {
3074 // Found it.
3075 assert(src_cp->destination() == destination,
3076 "first live obj in the space must match the destination");
3077 assert(src_cp->partial_obj_size() == 0,
3078 "a space cannot begin with a partial obj");
3079
3080 src_space_id = SpaceId(space_id);
3081 src_space_top = space->top();
3082 const size_t src_region_idx = sd.region(src_cp);
3083 closure.set_source(sd.region_to_addr(src_region_idx));
3084 return src_region_idx;
3085 } else {
3086 assert(src_cp->data_size() == 0, "sanity");
3087 }
3088 }
3089 }
3090 } while (++space_id < last_space_id);
3091
3092 assert(false, "no source region was found");
3093 return 0;
3094 }
3095
3096 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3097 {
3098 typedef ParMarkBitMap::IterationStatus IterationStatus;
3099 const size_t RegionSize = ParallelCompactData::RegionSize;
3100 ParMarkBitMap* const bitmap = mark_bitmap();
3101 ParallelCompactData& sd = summary_data();
3102 RegionData* const region_ptr = sd.region(region_idx);
3103
3104 // Get the items needed to construct the closure.
3105 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3106 SpaceId dest_space_id = space_id(dest_addr);
3107 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3108 HeapWord* new_top = _space_info[dest_space_id].new_top();
3109 assert(dest_addr < new_top, "sanity");
3110 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3111
3112 // Get the source region and related info.
3113 size_t src_region_idx = region_ptr->source_region();
3114 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3115 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3116
3117 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3118 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3119
3120 // Adjust src_region_idx to prepare for decrementing destination counts (the
3121 // destination count is not decremented when a region is copied to itself).
3122 if (src_region_idx == region_idx) {
3123 src_region_idx += 1;
3124 }
3125
3126 if (bitmap->is_unmarked(closure.source())) {
3127 // The first source word is in the middle of an object; copy the remainder
3128 // of the object or as much as will fit. The fact that pointer updates were
3129 // deferred will be noted when the object header is processed.
3130 HeapWord* const old_src_addr = closure.source();
3131 closure.copy_partial_obj();
3132 if (closure.is_full()) {
3133 decrement_destination_counts(cm, src_space_id, src_region_idx,
3134 closure.source());
3135 region_ptr->set_deferred_obj_addr(NULL);
3136 region_ptr->set_completed();
3137 return;
3138 }
3139
3140 HeapWord* const end_addr = sd.region_align_down(closure.source());
3141 if (sd.region_align_down(old_src_addr) != end_addr) {
3142 // The partial object was copied from more than one source region.
3143 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3144
3145 // Move to the next source region, possibly switching spaces as well. All
3146 // args except end_addr may be modified.
3147 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3148 end_addr);
3149 }
3150 }
3151
3152 do {
3153 HeapWord* const cur_addr = closure.source();
3154 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3155 src_space_top);
3156 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3157
3158 if (status == ParMarkBitMap::incomplete) {
3159 // The last obj that starts in the source region does not end in the
3160 // region.
3161 assert(closure.source() < end_addr, "sanity");
3162 HeapWord* const obj_beg = closure.source();
3163 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3164 src_space_top);
3165 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3166 if (obj_end < range_end) {
3167 // The end was found; the entire object will fit.
3168 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3169 assert(status != ParMarkBitMap::would_overflow, "sanity");
3170 } else {
3171 // The end was not found; the object will not fit.
3172 assert(range_end < src_space_top, "obj cannot cross space boundary");
3173 status = ParMarkBitMap::would_overflow;
3174 }
3175 }
3176
3177 if (status == ParMarkBitMap::would_overflow) {
3178 // The last object did not fit. Note that interior oop updates were
3179 // deferred, then copy enough of the object to fill the region.
3180 region_ptr->set_deferred_obj_addr(closure.destination());
3181 status = closure.copy_until_full(); // copies from closure.source()
3182
3183 decrement_destination_counts(cm, src_space_id, src_region_idx,
3184 closure.source());
3185 region_ptr->set_completed();
3186 return;
3187 }
3188
3189 if (status == ParMarkBitMap::full) {
3190 decrement_destination_counts(cm, src_space_id, src_region_idx,
3191 closure.source());
3192 region_ptr->set_deferred_obj_addr(NULL);
3193 region_ptr->set_completed();
3194 return;
3195 }
3196
3197 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3198
3199 // Move to the next source region, possibly switching spaces as well. All
3200 // args except end_addr may be modified.
3201 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3202 end_addr);
3203 } while (true);
3204 }
3205
3206 void PSParallelCompact::fill_blocks(size_t region_idx)
3207 {
3208 // Fill in the block table elements for the specified region. Each block
3209 // table element holds the number of live words in the region that are to the
3210 // left of the first object that starts in the block. Thus only blocks in
3211 // which an object starts need to be filled.
3212 //
3213 // The algorithm scans the section of the bitmap that corresponds to the
3214 // region, keeping a running total of the live words. When an object start is
3215 // found, if it's the first to start in the block that contains it, the
3216 // current total is written to the block table element.
3217 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3218 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3219 const size_t RegionSize = ParallelCompactData::RegionSize;
3220
3221 ParallelCompactData& sd = summary_data();
3222 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3223 if (partial_obj_size >= RegionSize) {
3224 return; // No objects start in this region.
3225 }
3226
3227 // Ensure the first loop iteration decides that the block has changed.
3228 size_t cur_block = sd.block_count();
3229
3230 const ParMarkBitMap* const bitmap = mark_bitmap();
3231
3232 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3233 assert((size_t)1 << Log2BitsPerBlock ==
3234 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3235
3236 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3237 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3238 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3239 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3240 while (beg_bit < range_end) {
3241 const size_t new_block = beg_bit >> Log2BitsPerBlock;
3242 if (new_block != cur_block) {
3243 cur_block = new_block;
3244 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3245 }
3246
3247 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3248 if (end_bit < range_end - 1) {
3249 live_bits += end_bit - beg_bit + 1;
3250 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3251 } else {
3252 return;
3253 }
3254 }
3255 }
3256
3257 void
3258 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3259 const MutableSpace* sp = space(space_id);
3260 if (sp->is_empty()) {
3261 return;
3262 }
3263
3264 ParallelCompactData& sd = PSParallelCompact::summary_data();
3265 ParMarkBitMap* const bitmap = mark_bitmap();
3266 HeapWord* const dp_addr = dense_prefix(space_id);
3267 HeapWord* beg_addr = sp->bottom();
3268 HeapWord* end_addr = sp->top();
3269
3270 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3271
3272 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3273 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3274 if (beg_region < dp_region) {
3275 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3276 }
3277
3278 // The destination of the first live object that starts in the region is one
3279 // past the end of the partial object entering the region (if any).
3280 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3281 HeapWord* const new_top = _space_info[space_id].new_top();
3282 assert(new_top >= dest_addr, "bad new_top value");
3283 const size_t words = pointer_delta(new_top, dest_addr);
3284
3285 if (words > 0) {
3286 ObjectStartArray* start_array = _space_info[space_id].start_array();
3287 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3288
3289 ParMarkBitMap::IterationStatus status;
3290 status = bitmap->iterate(&closure, dest_addr, end_addr);
3291 assert(status == ParMarkBitMap::full, "iteration not complete");
3292 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3293 "live objects skipped because closure is full");
3294 }
3295 }
3296
3297 jlong PSParallelCompact::millis_since_last_gc() {
3298 // We need a monotonically non-decreasing time in ms but
3299 // os::javaTimeMillis() does not guarantee monotonicity.
3300 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3301 jlong ret_val = now - _time_of_last_gc;
3302 // XXX See note in genCollectedHeap::millis_since_last_gc().
3303 if (ret_val < 0) {
3304 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3305 return 0;
3306 }
3307 return ret_val;
3308 }
3309
3310 void PSParallelCompact::reset_millis_since_last_gc() {
3311 // We need a monotonically non-decreasing time in ms but
3312 // os::javaTimeMillis() does not guarantee monotonicity.
3313 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3314 }
3315
3316 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3317 {
3318 if (source() != destination()) {
3319 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3320 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3321 }
3322 update_state(words_remaining());
3323 assert(is_full(), "sanity");
3324 return ParMarkBitMap::full;
3325 }
3326
3327 void MoveAndUpdateClosure::copy_partial_obj()
3328 {
3329 size_t words = words_remaining();
3330
3331 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3332 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3333 if (end_addr < range_end) {
3334 words = bitmap()->obj_size(source(), end_addr);
3335 }
3336
3337 // This test is necessary; if omitted, the pointer updates to a partial object
3338 // that crosses the dense prefix boundary could be overwritten.
3339 if (source() != destination()) {
3340 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3341 Copy::aligned_conjoint_words(source(), destination(), words);
3342 }
3343 update_state(words);
3344 }
3345
3346 ParMarkBitMapClosure::IterationStatus
3347 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3348 assert(destination() != NULL, "sanity");
3349 assert(bitmap()->obj_size(addr) == words, "bad size");
3350
3351 _source = addr;
3352 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3353 destination(), "wrong destination");
3354
3355 if (words > words_remaining()) {
3356 return ParMarkBitMap::would_overflow;
3357 }
3358
3359 // The start_array must be updated even if the object is not moving.
3360 if (_start_array != NULL) {
3361 _start_array->allocate_block(destination());
3362 }
3363
3364 if (destination() != source()) {
3365 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3366 Copy::aligned_conjoint_words(source(), destination(), words);
3367 }
3368
3369 oop moved_oop = (oop) destination();
3370 moved_oop->update_contents(compaction_manager());
3371 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3372
3373 update_state(words);
3374 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3375 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3376 }
3377
3378 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3379 ParCompactionManager* cm,
3380 PSParallelCompact::SpaceId space_id) :
3381 ParMarkBitMapClosure(mbm, cm),
3382 _space_id(space_id),
3383 _start_array(PSParallelCompact::start_array(space_id))
3384 {
3385 }
3386
3387 // Updates the references in the object to their new values.
3388 ParMarkBitMapClosure::IterationStatus
3389 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3390 do_addr(addr);
3391 return ParMarkBitMap::incomplete;
3392 }