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