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