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 "classfile/stringTable.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc/parallel/gcTaskManager.hpp"
31 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
32 #include "gc/parallel/pcTasks.hpp"
33 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
34 #include "gc/parallel/psCompactionManager.inline.hpp"
35 #include "gc/parallel/psMarkSweep.hpp"
36 #include "gc/parallel/psMarkSweepDecorator.hpp"
37 #include "gc/parallel/psOldGen.hpp"
38 #include "gc/parallel/psParallelCompact.inline.hpp"
39 #include "gc/parallel/psPromotionManager.inline.hpp"
40 #include "gc/parallel/psScavenge.hpp"
41 #include "gc/parallel/psYoungGen.hpp"
42 #include "gc/shared/gcCause.hpp"
43 #include "gc/shared/gcHeapSummary.hpp"
44 #include "gc/shared/gcId.hpp"
45 #include "gc/shared/gcLocker.inline.hpp"
46 #include "gc/shared/gcTimer.hpp"
47 #include "gc/shared/gcTrace.hpp"
48 #include "gc/shared/gcTraceTime.inline.hpp"
49 #include "gc/shared/isGCActiveMark.hpp"
50 #include "gc/shared/referencePolicy.hpp"
51 #include "gc/shared/referenceProcessor.hpp"
52 #include "gc/shared/spaceDecorator.hpp"
53 #include "logging/log.hpp"
54 #include "oops/instanceKlass.inline.hpp"
55 #include "oops/instanceMirrorKlass.inline.hpp"
56 #include "oops/methodData.hpp"
57 #include "oops/objArrayKlass.inline.hpp"
58 #include "oops/oop.inline.hpp"
59 #include "oops/valueArrayKlass.inline.hpp"
60 #include "runtime/atomic.inline.hpp"
61 #include "runtime/fprofiler.hpp"
62 #include "runtime/safepoint.hpp"
63 #include "runtime/vmThread.hpp"
64 #include "services/management.hpp"
65 #include "services/memTracker.hpp"
66 #include "services/memoryService.hpp"
67 #include "utilities/events.hpp"
68 #include "utilities/stack.inline.hpp"
69
70 #include <math.h>
71
72 // All sizes are in HeapWords.
73 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
74 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
75 const size_t ParallelCompactData::RegionSizeBytes =
76 RegionSize << LogHeapWordSize;
77 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
78 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
79 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
80
81 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
82 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
83 const size_t ParallelCompactData::BlockSizeBytes =
84 BlockSize << LogHeapWordSize;
85 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
86 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
87 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
88
89 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
90 const size_t ParallelCompactData::Log2BlocksPerRegion =
91 Log2RegionSize - Log2BlockSize;
92
93 const ParallelCompactData::RegionData::region_sz_t
94 ParallelCompactData::RegionData::dc_shift = 27;
95
96 const ParallelCompactData::RegionData::region_sz_t
97 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
98
99 const ParallelCompactData::RegionData::region_sz_t
100 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
101
102 const ParallelCompactData::RegionData::region_sz_t
103 ParallelCompactData::RegionData::los_mask = ~dc_mask;
104
105 const ParallelCompactData::RegionData::region_sz_t
106 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
107
108 const ParallelCompactData::RegionData::region_sz_t
109 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
110
111 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
112
113 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
114
115 double PSParallelCompact::_dwl_mean;
116 double PSParallelCompact::_dwl_std_dev;
117 double PSParallelCompact::_dwl_first_term;
118 double PSParallelCompact::_dwl_adjustment;
119 #ifdef ASSERT
120 bool PSParallelCompact::_dwl_initialized = false;
121 #endif // #ifdef ASSERT
122
123 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
124 HeapWord* destination)
125 {
126 assert(src_region_idx != 0, "invalid src_region_idx");
127 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
128 assert(destination != NULL, "invalid destination argument");
129
130 _src_region_idx = src_region_idx;
131 _partial_obj_size = partial_obj_size;
132 _destination = destination;
133
134 // These fields may not be updated below, so make sure they're clear.
135 assert(_dest_region_addr == NULL, "should have been cleared");
136 assert(_first_src_addr == NULL, "should have been cleared");
137
138 // Determine the number of destination regions for the partial object.
139 HeapWord* const last_word = destination + partial_obj_size - 1;
140 const ParallelCompactData& sd = PSParallelCompact::summary_data();
141 HeapWord* const beg_region_addr = sd.region_align_down(destination);
142 HeapWord* const end_region_addr = sd.region_align_down(last_word);
143
144 if (beg_region_addr == end_region_addr) {
145 // One destination region.
146 _destination_count = 1;
147 if (end_region_addr == destination) {
148 // The destination falls on a region boundary, thus the first word of the
149 // partial object will be the first word copied to the destination region.
150 _dest_region_addr = end_region_addr;
151 _first_src_addr = sd.region_to_addr(src_region_idx);
152 }
153 } else {
154 // Two destination regions. When copied, the partial object will cross a
155 // destination region boundary, so a word somewhere within the partial
156 // object will be the first word copied to the second destination region.
157 _destination_count = 2;
158 _dest_region_addr = end_region_addr;
159 const size_t ofs = pointer_delta(end_region_addr, destination);
160 assert(ofs < _partial_obj_size, "sanity");
161 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
162 }
163 }
164
165 void SplitInfo::clear()
166 {
167 _src_region_idx = 0;
168 _partial_obj_size = 0;
169 _destination = NULL;
170 _destination_count = 0;
171 _dest_region_addr = NULL;
172 _first_src_addr = NULL;
173 assert(!is_valid(), "sanity");
174 }
175
176 #ifdef ASSERT
177 void SplitInfo::verify_clear()
178 {
179 assert(_src_region_idx == 0, "not clear");
180 assert(_partial_obj_size == 0, "not clear");
181 assert(_destination == NULL, "not clear");
182 assert(_destination_count == 0, "not clear");
183 assert(_dest_region_addr == NULL, "not clear");
184 assert(_first_src_addr == NULL, "not clear");
185 }
186 #endif // #ifdef ASSERT
187
188
189 void PSParallelCompact::print_on_error(outputStream* st) {
190 _mark_bitmap.print_on_error(st);
191 }
192
193 #ifndef PRODUCT
194 const char* PSParallelCompact::space_names[] = {
195 "old ", "eden", "from", "to "
196 };
197
198 void PSParallelCompact::print_region_ranges() {
199 if (!log_develop_is_enabled(Trace, gc, compaction, phases)) {
200 return;
201 }
202 LogHandle(gc, compaction, phases) log;
203 ResourceMark rm;
204 Universe::print_on(log.trace_stream());
205 log.trace("space bottom top end new_top");
206 log.trace("------ ---------- ---------- ---------- ----------");
207
208 for (unsigned int id = 0; id < last_space_id; ++id) {
209 const MutableSpace* space = _space_info[id].space();
210 log.trace("%u %s "
211 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
212 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
213 id, space_names[id],
214 summary_data().addr_to_region_idx(space->bottom()),
215 summary_data().addr_to_region_idx(space->top()),
216 summary_data().addr_to_region_idx(space->end()),
217 summary_data().addr_to_region_idx(_space_info[id].new_top()));
218 }
219 }
220
221 void
222 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
223 {
224 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
225 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
226
227 ParallelCompactData& sd = PSParallelCompact::summary_data();
228 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
229 log_develop_trace(gc, compaction, phases)(
230 REGION_IDX_FORMAT " " PTR_FORMAT " "
231 REGION_IDX_FORMAT " " PTR_FORMAT " "
232 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
233 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
234 i, p2i(c->data_location()), dci, p2i(c->destination()),
235 c->partial_obj_size(), c->live_obj_size(),
236 c->data_size(), c->source_region(), c->destination_count());
237
238 #undef REGION_IDX_FORMAT
239 #undef REGION_DATA_FORMAT
240 }
241
242 void
243 print_generic_summary_data(ParallelCompactData& summary_data,
244 HeapWord* const beg_addr,
245 HeapWord* const end_addr)
246 {
247 size_t total_words = 0;
248 size_t i = summary_data.addr_to_region_idx(beg_addr);
249 const size_t last = summary_data.addr_to_region_idx(end_addr);
250 HeapWord* pdest = 0;
251
252 while (i <= last) {
253 ParallelCompactData::RegionData* c = summary_data.region(i);
254 if (c->data_size() != 0 || c->destination() != pdest) {
255 print_generic_summary_region(i, c);
256 total_words += c->data_size();
257 pdest = c->destination();
258 }
259 ++i;
260 }
261
262 log_develop_trace(gc, compaction, phases)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
263 }
264
265 void
266 print_generic_summary_data(ParallelCompactData& summary_data,
267 SpaceInfo* space_info)
268 {
269 if (!log_develop_is_enabled(Trace, gc, compaction, phases)) {
270 return;
271 }
272
273 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
274 const MutableSpace* space = space_info[id].space();
275 print_generic_summary_data(summary_data, space->bottom(),
276 MAX2(space->top(), space_info[id].new_top()));
277 }
278 }
279
280 void
281 print_initial_summary_data(ParallelCompactData& summary_data,
282 const MutableSpace* space) {
283 if (space->top() == space->bottom()) {
284 return;
285 }
286
287 const size_t region_size = ParallelCompactData::RegionSize;
288 typedef ParallelCompactData::RegionData RegionData;
289 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
290 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
291 const RegionData* c = summary_data.region(end_region - 1);
292 HeapWord* end_addr = c->destination() + c->data_size();
293 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
294
295 // Print (and count) the full regions at the beginning of the space.
296 size_t full_region_count = 0;
297 size_t i = summary_data.addr_to_region_idx(space->bottom());
298 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
299 ParallelCompactData::RegionData* c = summary_data.region(i);
300 log_develop_trace(gc, compaction, phases)(
301 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
302 i, p2i(c->destination()),
303 c->partial_obj_size(), c->live_obj_size(),
304 c->data_size(), c->source_region(), c->destination_count());
305 ++full_region_count;
306 ++i;
307 }
308
309 size_t live_to_right = live_in_space - full_region_count * region_size;
310
311 double max_reclaimed_ratio = 0.0;
312 size_t max_reclaimed_ratio_region = 0;
313 size_t max_dead_to_right = 0;
314 size_t max_live_to_right = 0;
315
316 // Print the 'reclaimed ratio' for regions while there is something live in
317 // the region or to the right of it. The remaining regions are empty (and
318 // uninteresting), and computing the ratio will result in division by 0.
319 while (i < end_region && live_to_right > 0) {
320 c = summary_data.region(i);
321 HeapWord* const region_addr = summary_data.region_to_addr(i);
322 const size_t used_to_right = pointer_delta(space->top(), region_addr);
323 const size_t dead_to_right = used_to_right - live_to_right;
324 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
325
326 if (reclaimed_ratio > max_reclaimed_ratio) {
327 max_reclaimed_ratio = reclaimed_ratio;
328 max_reclaimed_ratio_region = i;
329 max_dead_to_right = dead_to_right;
330 max_live_to_right = live_to_right;
331 }
332
333 ParallelCompactData::RegionData* c = summary_data.region(i);
334 log_develop_trace(gc, compaction, phases)(
335 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
336 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
337 i, p2i(c->destination()),
338 c->partial_obj_size(), c->live_obj_size(),
339 c->data_size(), c->source_region(), c->destination_count(),
340 reclaimed_ratio, dead_to_right, live_to_right);
341
342
343 live_to_right -= c->data_size();
344 ++i;
345 }
346
347 // Any remaining regions are empty. Print one more if there is one.
348 if (i < end_region) {
349 ParallelCompactData::RegionData* c = summary_data.region(i);
350 log_develop_trace(gc, compaction, phases)(
351 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
352 i, p2i(c->destination()),
353 c->partial_obj_size(), c->live_obj_size(),
354 c->data_size(), c->source_region(), c->destination_count());
355 }
356
357 log_develop_trace(gc, compaction, phases)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
358 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
359 }
360
361 void
362 print_initial_summary_data(ParallelCompactData& summary_data,
363 SpaceInfo* space_info) {
364 if (!log_develop_is_enabled(Trace, gc, compaction, phases)) {
365 return;
366 }
367
368 unsigned int id = PSParallelCompact::old_space_id;
369 const MutableSpace* space;
370 do {
371 space = space_info[id].space();
372 print_initial_summary_data(summary_data, space);
373 } while (++id < PSParallelCompact::eden_space_id);
374
375 do {
376 space = space_info[id].space();
377 print_generic_summary_data(summary_data, space->bottom(), space->top());
378 } while (++id < PSParallelCompact::last_space_id);
379 }
380 #endif // #ifndef PRODUCT
381
382 #ifdef ASSERT
383 size_t add_obj_count;
384 size_t add_obj_size;
385 size_t mark_bitmap_count;
386 size_t mark_bitmap_size;
387 #endif // #ifdef ASSERT
388
389 ParallelCompactData::ParallelCompactData()
390 {
391 _region_start = 0;
392
393 _region_vspace = 0;
394 _reserved_byte_size = 0;
395 _region_data = 0;
396 _region_count = 0;
397
398 _block_vspace = 0;
399 _block_data = 0;
400 _block_count = 0;
401 }
402
403 bool ParallelCompactData::initialize(MemRegion covered_region)
404 {
405 _region_start = covered_region.start();
406 const size_t region_size = covered_region.word_size();
407 DEBUG_ONLY(_region_end = _region_start + region_size;)
408
409 assert(region_align_down(_region_start) == _region_start,
410 "region start not aligned");
411 assert((region_size & RegionSizeOffsetMask) == 0,
412 "region size not a multiple of RegionSize");
413
414 bool result = initialize_region_data(region_size) && initialize_block_data();
415 return result;
416 }
417
418 PSVirtualSpace*
419 ParallelCompactData::create_vspace(size_t count, size_t element_size)
420 {
421 const size_t raw_bytes = count * element_size;
422 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
423 const size_t granularity = os::vm_allocation_granularity();
424 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
425
426 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
427 MAX2(page_sz, granularity);
428 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
429 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
430 rs.size());
431
432 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
433
434 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
435 if (vspace != 0) {
436 if (vspace->expand_by(_reserved_byte_size)) {
437 return vspace;
438 }
439 delete vspace;
440 // Release memory reserved in the space.
441 rs.release();
442 }
443
444 return 0;
445 }
446
447 bool ParallelCompactData::initialize_region_data(size_t region_size)
448 {
449 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
450 _region_vspace = create_vspace(count, sizeof(RegionData));
451 if (_region_vspace != 0) {
452 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
453 _region_count = count;
454 return true;
455 }
456 return false;
457 }
458
459 bool ParallelCompactData::initialize_block_data()
460 {
461 assert(_region_count != 0, "region data must be initialized first");
462 const size_t count = _region_count << Log2BlocksPerRegion;
463 _block_vspace = create_vspace(count, sizeof(BlockData));
464 if (_block_vspace != 0) {
465 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
466 _block_count = count;
467 return true;
468 }
469 return false;
470 }
471
472 void ParallelCompactData::clear()
473 {
474 memset(_region_data, 0, _region_vspace->committed_size());
475 memset(_block_data, 0, _block_vspace->committed_size());
476 }
477
478 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
479 assert(beg_region <= _region_count, "beg_region out of range");
480 assert(end_region <= _region_count, "end_region out of range");
481 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
482
483 const size_t region_cnt = end_region - beg_region;
484 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
485
486 const size_t beg_block = beg_region * BlocksPerRegion;
487 const size_t block_cnt = region_cnt * BlocksPerRegion;
488 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
489 }
490
491 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
492 {
493 const RegionData* cur_cp = region(region_idx);
494 const RegionData* const end_cp = region(region_count() - 1);
495
496 HeapWord* result = region_to_addr(region_idx);
497 if (cur_cp < end_cp) {
498 do {
499 result += cur_cp->partial_obj_size();
500 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
501 }
502 return result;
503 }
504
505 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
506 {
507 const size_t obj_ofs = pointer_delta(addr, _region_start);
508 const size_t beg_region = obj_ofs >> Log2RegionSize;
509 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
510
511 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
512 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
513
514 if (beg_region == end_region) {
515 // All in one region.
516 _region_data[beg_region].add_live_obj(len);
517 return;
518 }
519
520 // First region.
521 const size_t beg_ofs = region_offset(addr);
522 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
523
524 Klass* klass = ((oop)addr)->klass();
525 // Middle regions--completely spanned by this object.
526 for (size_t region = beg_region + 1; region < end_region; ++region) {
527 _region_data[region].set_partial_obj_size(RegionSize);
528 _region_data[region].set_partial_obj_addr(addr);
529 }
530
531 // Last region.
532 const size_t end_ofs = region_offset(addr + len - 1);
533 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
534 _region_data[end_region].set_partial_obj_addr(addr);
535 }
536
537 void
538 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
539 {
540 assert(region_offset(beg) == 0, "not RegionSize aligned");
541 assert(region_offset(end) == 0, "not RegionSize aligned");
542
543 size_t cur_region = addr_to_region_idx(beg);
544 const size_t end_region = addr_to_region_idx(end);
545 HeapWord* addr = beg;
546 while (cur_region < end_region) {
547 _region_data[cur_region].set_destination(addr);
548 _region_data[cur_region].set_destination_count(0);
549 _region_data[cur_region].set_source_region(cur_region);
550 _region_data[cur_region].set_data_location(addr);
551
552 // Update live_obj_size so the region appears completely full.
553 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
554 _region_data[cur_region].set_live_obj_size(live_size);
555
556 ++cur_region;
557 addr += RegionSize;
558 }
559 }
560
561 // Find the point at which a space can be split and, if necessary, record the
562 // split point.
563 //
564 // If the current src region (which overflowed the destination space) doesn't
565 // have a partial object, the split point is at the beginning of the current src
566 // region (an "easy" split, no extra bookkeeping required).
567 //
568 // If the current src region has a partial object, the split point is in the
569 // region where that partial object starts (call it the split_region). If
570 // split_region has a partial object, then the split point is just after that
571 // partial object (a "hard" split where we have to record the split data and
572 // zero the partial_obj_size field). With a "hard" split, we know that the
573 // partial_obj ends within split_region because the partial object that caused
574 // the overflow starts in split_region. If split_region doesn't have a partial
575 // obj, then the split is at the beginning of split_region (another "easy"
576 // split).
577 HeapWord*
578 ParallelCompactData::summarize_split_space(size_t src_region,
579 SplitInfo& split_info,
580 HeapWord* destination,
581 HeapWord* target_end,
582 HeapWord** target_next)
583 {
584 assert(destination <= target_end, "sanity");
585 assert(destination + _region_data[src_region].data_size() > target_end,
586 "region should not fit into target space");
587 assert(is_region_aligned(target_end), "sanity");
588
589 size_t split_region = src_region;
590 HeapWord* split_destination = destination;
591 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
592
593 if (destination + partial_obj_size > target_end) {
594 // The split point is just after the partial object (if any) in the
595 // src_region that contains the start of the object that overflowed the
596 // destination space.
597 //
598 // Find the start of the "overflow" object and set split_region to the
599 // region containing it.
600 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
601 split_region = addr_to_region_idx(overflow_obj);
602
603 // Clear the source_region field of all destination regions whose first word
604 // came from data after the split point (a non-null source_region field
605 // implies a region must be filled).
606 //
607 // An alternative to the simple loop below: clear during post_compact(),
608 // which uses memcpy instead of individual stores, and is easy to
609 // parallelize. (The downside is that it clears the entire RegionData
610 // object as opposed to just one field.)
611 //
612 // post_compact() would have to clear the summary data up to the highest
613 // address that was written during the summary phase, which would be
614 //
615 // max(top, max(new_top, clear_top))
616 //
617 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
618 // to target_end.
619 const RegionData* const sr = region(split_region);
620 const size_t beg_idx =
621 addr_to_region_idx(region_align_up(sr->destination() +
622 sr->partial_obj_size()));
623 const size_t end_idx = addr_to_region_idx(target_end);
624
625 log_develop_trace(gc, compaction, phases)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
626 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
627 _region_data[idx].set_source_region(0);
628 }
629
630 // Set split_destination and partial_obj_size to reflect the split region.
631 split_destination = sr->destination();
632 partial_obj_size = sr->partial_obj_size();
633 }
634
635 // The split is recorded only if a partial object extends onto the region.
636 if (partial_obj_size != 0) {
637 _region_data[split_region].set_partial_obj_size(0);
638 split_info.record(split_region, partial_obj_size, split_destination);
639 }
640
641 // Setup the continuation addresses.
642 *target_next = split_destination + partial_obj_size;
643 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
644
645 if (log_develop_is_enabled(Trace, gc, compaction, phases)) {
646 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
647 log_develop_trace(gc, compaction, phases)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
648 split_type, p2i(source_next), split_region, partial_obj_size);
649 log_develop_trace(gc, compaction, phases)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
650 split_type, p2i(split_destination),
651 addr_to_region_idx(split_destination),
652 p2i(*target_next));
653
654 if (partial_obj_size != 0) {
655 HeapWord* const po_beg = split_info.destination();
656 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
657 log_develop_trace(gc, compaction, phases)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
658 split_type,
659 p2i(po_beg), addr_to_region_idx(po_beg),
660 p2i(po_end), addr_to_region_idx(po_end));
661 }
662 }
663
664 return source_next;
665 }
666
667 bool ParallelCompactData::summarize(SplitInfo& split_info,
668 HeapWord* source_beg, HeapWord* source_end,
669 HeapWord** source_next,
670 HeapWord* target_beg, HeapWord* target_end,
671 HeapWord** target_next)
672 {
673 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
674 log_develop_trace(gc, compaction, phases)(
675 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
676 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
677 p2i(source_beg), p2i(source_end), p2i(source_next_val),
678 p2i(target_beg), p2i(target_end), p2i(*target_next));
679
680 size_t cur_region = addr_to_region_idx(source_beg);
681 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
682
683 HeapWord *dest_addr = target_beg;
684 while (cur_region < end_region) {
685 // The destination must be set even if the region has no data.
686 _region_data[cur_region].set_destination(dest_addr);
687
688 size_t words = _region_data[cur_region].data_size();
689 if (words > 0) {
690 // If cur_region does not fit entirely into the target space, find a point
691 // at which the source space can be 'split' so that part is copied to the
692 // target space and the rest is copied elsewhere.
693 if (dest_addr + words > target_end) {
694 assert(source_next != NULL, "source_next is NULL when splitting");
695 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
696 target_end, target_next);
697 return false;
698 }
699
700 // Compute the destination_count for cur_region, and if necessary, update
701 // source_region for a destination region. The source_region field is
702 // updated if cur_region is the first (left-most) region to be copied to a
703 // destination region.
704 //
705 // The destination_count calculation is a bit subtle. A region that has
706 // data that compacts into itself does not count itself as a destination.
707 // This maintains the invariant that a zero count means the region is
708 // available and can be claimed and then filled.
709 uint destination_count = 0;
710 if (split_info.is_split(cur_region)) {
711 // The current region has been split: the partial object will be copied
712 // to one destination space and the remaining data will be copied to
713 // another destination space. Adjust the initial destination_count and,
714 // if necessary, set the source_region field if the partial object will
715 // cross a destination region boundary.
716 destination_count = split_info.destination_count();
717 if (destination_count == 2) {
718 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
719 _region_data[dest_idx].set_source_region(cur_region);
720 }
721 }
722
723 HeapWord* const last_addr = dest_addr + words - 1;
724 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
725 const size_t dest_region_2 = addr_to_region_idx(last_addr);
726
727 // Initially assume that the destination regions will be the same and
728 // adjust the value below if necessary. Under this assumption, if
729 // cur_region == dest_region_2, then cur_region will be compacted
730 // completely into itself.
731 destination_count += cur_region == dest_region_2 ? 0 : 1;
732 if (dest_region_1 != dest_region_2) {
733 // Destination regions differ; adjust destination_count.
734 destination_count += 1;
735 // Data from cur_region will be copied to the start of dest_region_2.
736 _region_data[dest_region_2].set_source_region(cur_region);
737 } else if (region_offset(dest_addr) == 0) {
738 // Data from cur_region will be copied to the start of the destination
739 // region.
740 _region_data[dest_region_1].set_source_region(cur_region);
741 }
742
743 _region_data[cur_region].set_destination_count(destination_count);
744 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
745 dest_addr += words;
746 }
747
748 ++cur_region;
749 }
750
751 *target_next = dest_addr;
752 return true;
753 }
754
755 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
756 assert(addr != NULL, "Should detect NULL oop earlier");
757 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
758 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
759
760 // Region covering the object.
761 RegionData* const region_ptr = addr_to_region_ptr(addr);
762 HeapWord* result = region_ptr->destination();
763
764 // If the entire Region is live, the new location is region->destination + the
765 // offset of the object within in the Region.
766
767 // Run some performance tests to determine if this special case pays off. It
768 // is worth it for pointers into the dense prefix. If the optimization to
769 // avoid pointer updates in regions that only point to the dense prefix is
770 // ever implemented, this should be revisited.
771 if (region_ptr->data_size() == RegionSize) {
772 result += region_offset(addr);
773 return result;
774 }
775
776 // Otherwise, the new location is region->destination + block offset + the
777 // number of live words in the Block that are (a) to the left of addr and (b)
778 // due to objects that start in the Block.
779
780 // Fill in the block table if necessary. This is unsynchronized, so multiple
781 // threads may fill the block table for a region (harmless, since it is
782 // idempotent).
783 if (!region_ptr->blocks_filled()) {
784 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
785 region_ptr->set_blocks_filled();
786 }
787
788 HeapWord* const search_start = block_align_down(addr);
789 const size_t block_offset = addr_to_block_ptr(addr)->offset();
790
791 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
792 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr));
793 result += block_offset + live;
794 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
795 return result;
796 }
797
798 #ifdef ASSERT
799 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
800 {
801 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
802 const size_t* const end = (const size_t*)vspace->committed_high_addr();
803 for (const size_t* p = beg; p < end; ++p) {
804 assert(*p == 0, "not zero");
805 }
806 }
807
808 void ParallelCompactData::verify_clear()
809 {
810 verify_clear(_region_vspace);
811 verify_clear(_block_vspace);
812 }
813 #endif // #ifdef ASSERT
814
815 STWGCTimer PSParallelCompact::_gc_timer;
816 ParallelOldTracer PSParallelCompact::_gc_tracer;
817 elapsedTimer PSParallelCompact::_accumulated_time;
818 unsigned int PSParallelCompact::_total_invocations = 0;
819 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
820 jlong PSParallelCompact::_time_of_last_gc = 0;
821 CollectorCounters* PSParallelCompact::_counters = NULL;
822 ParMarkBitMap PSParallelCompact::_mark_bitmap;
823 ParallelCompactData PSParallelCompact::_summary_data;
824
825 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
826
827 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
828
829 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
830 PSParallelCompact::AdjustPointerClosure closure(_cm);
831 klass->oops_do(&closure);
832 }
833
834 void PSParallelCompact::post_initialize() {
835 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
836 MemRegion mr = heap->reserved_region();
837 _ref_processor =
838 new ReferenceProcessor(mr, // span
839 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
840 ParallelGCThreads, // mt processing degree
841 true, // mt discovery
842 ParallelGCThreads, // mt discovery degree
843 true, // atomic_discovery
844 &_is_alive_closure); // non-header is alive closure
845 _counters = new CollectorCounters("PSParallelCompact", 1);
846
847 // Initialize static fields in ParCompactionManager.
848 ParCompactionManager::initialize(mark_bitmap());
849 }
850
851 bool PSParallelCompact::initialize() {
852 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
853 MemRegion mr = heap->reserved_region();
854
855 // Was the old gen get allocated successfully?
856 if (!heap->old_gen()->is_allocated()) {
857 return false;
858 }
859
860 initialize_space_info();
861 initialize_dead_wood_limiter();
862
863 if (!_mark_bitmap.initialize(mr)) {
864 vm_shutdown_during_initialization(
865 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
866 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
867 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
868 return false;
869 }
870
871 if (!_summary_data.initialize(mr)) {
872 vm_shutdown_during_initialization(
873 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
874 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
875 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
876 return false;
877 }
878
879 return true;
880 }
881
882 void PSParallelCompact::initialize_space_info()
883 {
884 memset(&_space_info, 0, sizeof(_space_info));
885
886 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
887 PSYoungGen* young_gen = heap->young_gen();
888
889 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
890 _space_info[eden_space_id].set_space(young_gen->eden_space());
891 _space_info[from_space_id].set_space(young_gen->from_space());
892 _space_info[to_space_id].set_space(young_gen->to_space());
893
894 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
895 }
896
897 void PSParallelCompact::initialize_dead_wood_limiter()
898 {
899 const size_t max = 100;
900 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
901 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
902 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
903 DEBUG_ONLY(_dwl_initialized = true;)
904 _dwl_adjustment = normal_distribution(1.0);
905 }
906
907 void
908 PSParallelCompact::clear_data_covering_space(SpaceId id)
909 {
910 // At this point, top is the value before GC, new_top() is the value that will
911 // be set at the end of GC. The marking bitmap is cleared to top; nothing
912 // should be marked above top. The summary data is cleared to the larger of
913 // top & new_top.
914 MutableSpace* const space = _space_info[id].space();
915 HeapWord* const bot = space->bottom();
916 HeapWord* const top = space->top();
917 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
918
919 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
920 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
921 _mark_bitmap.clear_range(beg_bit, end_bit);
922
923 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
924 const size_t end_region =
925 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
926 _summary_data.clear_range(beg_region, end_region);
927
928 // Clear the data used to 'split' regions.
929 SplitInfo& split_info = _space_info[id].split_info();
930 if (split_info.is_valid()) {
931 split_info.clear();
932 }
933 DEBUG_ONLY(split_info.verify_clear();)
934 }
935
936 void PSParallelCompact::pre_compact()
937 {
938 // Update the from & to space pointers in space_info, since they are swapped
939 // at each young gen gc. Do the update unconditionally (even though a
940 // promotion failure does not swap spaces) because an unknown number of young
941 // collections will have swapped the spaces an unknown number of times.
942 GCTraceTime(Trace, gc, phases) tm("Pre Compact", &_gc_timer);
943 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
944 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
945 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
946
947 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
948 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
949
950 // Increment the invocation count
951 heap->increment_total_collections(true);
952
953 // We need to track unique mark sweep invocations as well.
954 _total_invocations++;
955
956 heap->print_heap_before_gc();
957 heap->trace_heap_before_gc(&_gc_tracer);
958
959 // Fill in TLABs
960 heap->accumulate_statistics_all_tlabs();
961 heap->ensure_parsability(true); // retire TLABs
962
963 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
964 HandleMark hm; // Discard invalid handles created during verification
965 Universe::verify("Before GC");
966 }
967
968 // Verify object start arrays
969 if (VerifyObjectStartArray &&
970 VerifyBeforeGC) {
971 heap->old_gen()->verify_object_start_array();
972 }
973
974 DEBUG_ONLY(mark_bitmap()->verify_clear();)
975 DEBUG_ONLY(summary_data().verify_clear();)
976
977 // Have worker threads release resources the next time they run a task.
978 gc_task_manager()->release_all_resources();
979
980 ParCompactionManager::reset_all_bitmap_query_caches();
981 }
982
983 void PSParallelCompact::post_compact()
984 {
985 GCTraceTime(Trace, gc, phases) tm("Post Compact", &_gc_timer);
986
987 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
988 // Clear the marking bitmap, summary data and split info.
989 clear_data_covering_space(SpaceId(id));
990 // Update top(). Must be done after clearing the bitmap and summary data.
991 _space_info[id].publish_new_top();
992 }
993
994 MutableSpace* const eden_space = _space_info[eden_space_id].space();
995 MutableSpace* const from_space = _space_info[from_space_id].space();
996 MutableSpace* const to_space = _space_info[to_space_id].space();
997
998 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
999 bool eden_empty = eden_space->is_empty();
1000 if (!eden_empty) {
1001 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1002 heap->young_gen(), heap->old_gen());
1003 }
1004
1005 // Update heap occupancy information which is used as input to the soft ref
1006 // clearing policy at the next gc.
1007 Universe::update_heap_info_at_gc();
1008
1009 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1010 to_space->is_empty();
1011
1012 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set());
1013 MemRegion old_mr = heap->old_gen()->reserved();
1014 if (young_gen_empty) {
1015 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1016 } else {
1017 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1018 }
1019
1020 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1021 ClassLoaderDataGraph::purge();
1022 MetaspaceAux::verify_metrics();
1023
1024 CodeCache::gc_epilogue();
1025 JvmtiExport::gc_epilogue();
1026
1027 #if defined(COMPILER2) || INCLUDE_JVMCI
1028 DerivedPointerTable::update_pointers();
1029 #endif
1030
1031 ref_processor()->enqueue_discovered_references(NULL);
1032
1033 if (ZapUnusedHeapArea) {
1034 heap->gen_mangle_unused_area();
1035 }
1036
1037 // Update time of last GC
1038 reset_millis_since_last_gc();
1039 }
1040
1041 HeapWord*
1042 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1043 bool maximum_compaction)
1044 {
1045 const size_t region_size = ParallelCompactData::RegionSize;
1046 const ParallelCompactData& sd = summary_data();
1047
1048 const MutableSpace* const space = _space_info[id].space();
1049 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1050 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1051 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1052
1053 // Skip full regions at the beginning of the space--they are necessarily part
1054 // of the dense prefix.
1055 size_t full_count = 0;
1056 const RegionData* cp;
1057 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1058 ++full_count;
1059 }
1060
1061 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1062 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1063 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1064 if (maximum_compaction || cp == end_cp || interval_ended) {
1065 _maximum_compaction_gc_num = total_invocations();
1066 return sd.region_to_addr(cp);
1067 }
1068
1069 HeapWord* const new_top = _space_info[id].new_top();
1070 const size_t space_live = pointer_delta(new_top, space->bottom());
1071 const size_t space_used = space->used_in_words();
1072 const size_t space_capacity = space->capacity_in_words();
1073
1074 const double cur_density = double(space_live) / space_capacity;
1075 const double deadwood_density =
1076 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1077 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1078
1079 if (TraceParallelOldGCDensePrefix) {
1080 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1081 cur_density, deadwood_density, deadwood_goal);
1082 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1083 "space_cap=" SIZE_FORMAT,
1084 space_live, space_used,
1085 space_capacity);
1086 }
1087
1088 // XXX - Use binary search?
1089 HeapWord* dense_prefix = sd.region_to_addr(cp);
1090 const RegionData* full_cp = cp;
1091 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1092 while (cp < end_cp) {
1093 HeapWord* region_destination = cp->destination();
1094 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1095 if (TraceParallelOldGCDensePrefix && Verbose) {
1096 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1097 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1098 sd.region(cp), p2i(region_destination),
1099 p2i(dense_prefix), cur_deadwood);
1100 }
1101
1102 if (cur_deadwood >= deadwood_goal) {
1103 // Found the region that has the correct amount of deadwood to the left.
1104 // This typically occurs after crossing a fairly sparse set of regions, so
1105 // iterate backwards over those sparse regions, looking for the region
1106 // that has the lowest density of live objects 'to the right.'
1107 size_t space_to_left = sd.region(cp) * region_size;
1108 size_t live_to_left = space_to_left - cur_deadwood;
1109 size_t space_to_right = space_capacity - space_to_left;
1110 size_t live_to_right = space_live - live_to_left;
1111 double density_to_right = double(live_to_right) / space_to_right;
1112 while (cp > full_cp) {
1113 --cp;
1114 const size_t prev_region_live_to_right = live_to_right -
1115 cp->data_size();
1116 const size_t prev_region_space_to_right = space_to_right + region_size;
1117 double prev_region_density_to_right =
1118 double(prev_region_live_to_right) / prev_region_space_to_right;
1119 if (density_to_right <= prev_region_density_to_right) {
1120 return dense_prefix;
1121 }
1122 if (TraceParallelOldGCDensePrefix && Verbose) {
1123 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1124 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1125 prev_region_density_to_right);
1126 }
1127 dense_prefix -= region_size;
1128 live_to_right = prev_region_live_to_right;
1129 space_to_right = prev_region_space_to_right;
1130 density_to_right = prev_region_density_to_right;
1131 }
1132 return dense_prefix;
1133 }
1134
1135 dense_prefix += region_size;
1136 ++cp;
1137 }
1138
1139 return dense_prefix;
1140 }
1141
1142 #ifndef PRODUCT
1143 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1144 const SpaceId id,
1145 const bool maximum_compaction,
1146 HeapWord* const addr)
1147 {
1148 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1149 RegionData* const cp = summary_data().region(region_idx);
1150 const MutableSpace* const space = _space_info[id].space();
1151 HeapWord* const new_top = _space_info[id].new_top();
1152
1153 const size_t space_live = pointer_delta(new_top, space->bottom());
1154 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1155 const size_t space_cap = space->capacity_in_words();
1156 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1157 const size_t live_to_right = new_top - cp->destination();
1158 const size_t dead_to_right = space->top() - addr - live_to_right;
1159
1160 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1161 "spl=" SIZE_FORMAT " "
1162 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1163 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1164 " ratio=%10.8f",
1165 algorithm, p2i(addr), region_idx,
1166 space_live,
1167 dead_to_left, dead_to_left_pct,
1168 dead_to_right, live_to_right,
1169 double(dead_to_right) / live_to_right);
1170 }
1171 #endif // #ifndef PRODUCT
1172
1173 // Return a fraction indicating how much of the generation can be treated as
1174 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1175 // based on the density of live objects in the generation to determine a limit,
1176 // which is then adjusted so the return value is min_percent when the density is
1177 // 1.
1178 //
1179 // The following table shows some return values for a different values of the
1180 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1181 // min_percent is 1.
1182 //
1183 // fraction allowed as dead wood
1184 // -----------------------------------------------------------------
1185 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1186 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1187 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1188 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1189 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1190 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1191 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1192 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1193 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1194 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1195 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1196 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1197 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1198 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1199 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1200 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1201 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1202 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1203 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1204 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1205 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1206 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1207 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1208
1209 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1210 {
1211 assert(_dwl_initialized, "uninitialized");
1212
1213 // The raw limit is the value of the normal distribution at x = density.
1214 const double raw_limit = normal_distribution(density);
1215
1216 // Adjust the raw limit so it becomes the minimum when the density is 1.
1217 //
1218 // First subtract the adjustment value (which is simply the precomputed value
1219 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1220 // Then add the minimum value, so the minimum is returned when the density is
1221 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1222 const double min = double(min_percent) / 100.0;
1223 const double limit = raw_limit - _dwl_adjustment + min;
1224 return MAX2(limit, 0.0);
1225 }
1226
1227 ParallelCompactData::RegionData*
1228 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1229 const RegionData* end)
1230 {
1231 const size_t region_size = ParallelCompactData::RegionSize;
1232 ParallelCompactData& sd = summary_data();
1233 size_t left = sd.region(beg);
1234 size_t right = end > beg ? sd.region(end) - 1 : left;
1235
1236 // Binary search.
1237 while (left < right) {
1238 // Equivalent to (left + right) / 2, but does not overflow.
1239 const size_t middle = left + (right - left) / 2;
1240 RegionData* const middle_ptr = sd.region(middle);
1241 HeapWord* const dest = middle_ptr->destination();
1242 HeapWord* const addr = sd.region_to_addr(middle);
1243 assert(dest != NULL, "sanity");
1244 assert(dest <= addr, "must move left");
1245
1246 if (middle > left && dest < addr) {
1247 right = middle - 1;
1248 } else if (middle < right && middle_ptr->data_size() == region_size) {
1249 left = middle + 1;
1250 } else {
1251 return middle_ptr;
1252 }
1253 }
1254 return sd.region(left);
1255 }
1256
1257 ParallelCompactData::RegionData*
1258 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1259 const RegionData* end,
1260 size_t dead_words)
1261 {
1262 ParallelCompactData& sd = summary_data();
1263 size_t left = sd.region(beg);
1264 size_t right = end > beg ? sd.region(end) - 1 : left;
1265
1266 // Binary search.
1267 while (left < right) {
1268 // Equivalent to (left + right) / 2, but does not overflow.
1269 const size_t middle = left + (right - left) / 2;
1270 RegionData* const middle_ptr = sd.region(middle);
1271 HeapWord* const dest = middle_ptr->destination();
1272 HeapWord* const addr = sd.region_to_addr(middle);
1273 assert(dest != NULL, "sanity");
1274 assert(dest <= addr, "must move left");
1275
1276 const size_t dead_to_left = pointer_delta(addr, dest);
1277 if (middle > left && dead_to_left > dead_words) {
1278 right = middle - 1;
1279 } else if (middle < right && dead_to_left < dead_words) {
1280 left = middle + 1;
1281 } else {
1282 return middle_ptr;
1283 }
1284 }
1285 return sd.region(left);
1286 }
1287
1288 // The result is valid during the summary phase, after the initial summarization
1289 // of each space into itself, and before final summarization.
1290 inline double
1291 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1292 HeapWord* const bottom,
1293 HeapWord* const top,
1294 HeapWord* const new_top)
1295 {
1296 ParallelCompactData& sd = summary_data();
1297
1298 assert(cp != NULL, "sanity");
1299 assert(bottom != NULL, "sanity");
1300 assert(top != NULL, "sanity");
1301 assert(new_top != NULL, "sanity");
1302 assert(top >= new_top, "summary data problem?");
1303 assert(new_top > bottom, "space is empty; should not be here");
1304 assert(new_top >= cp->destination(), "sanity");
1305 assert(top >= sd.region_to_addr(cp), "sanity");
1306
1307 HeapWord* const destination = cp->destination();
1308 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1309 const size_t compacted_region_live = pointer_delta(new_top, destination);
1310 const size_t compacted_region_used = pointer_delta(top,
1311 sd.region_to_addr(cp));
1312 const size_t reclaimable = compacted_region_used - compacted_region_live;
1313
1314 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1315 return double(reclaimable) / divisor;
1316 }
1317
1318 // Return the address of the end of the dense prefix, a.k.a. the start of the
1319 // compacted region. The address is always on a region boundary.
1320 //
1321 // Completely full regions at the left are skipped, since no compaction can
1322 // occur in those regions. Then the maximum amount of dead wood to allow is
1323 // computed, based on the density (amount live / capacity) of the generation;
1324 // the region with approximately that amount of dead space to the left is
1325 // identified as the limit region. Regions between the last completely full
1326 // region and the limit region are scanned and the one that has the best
1327 // (maximum) reclaimed_ratio() is selected.
1328 HeapWord*
1329 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1330 bool maximum_compaction)
1331 {
1332 const size_t region_size = ParallelCompactData::RegionSize;
1333 const ParallelCompactData& sd = summary_data();
1334
1335 const MutableSpace* const space = _space_info[id].space();
1336 HeapWord* const top = space->top();
1337 HeapWord* const top_aligned_up = sd.region_align_up(top);
1338 HeapWord* const new_top = _space_info[id].new_top();
1339 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1340 HeapWord* const bottom = space->bottom();
1341 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1342 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1343 const RegionData* const new_top_cp =
1344 sd.addr_to_region_ptr(new_top_aligned_up);
1345
1346 // Skip full regions at the beginning of the space--they are necessarily part
1347 // of the dense prefix.
1348 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1349 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1350 space->is_empty(), "no dead space allowed to the left");
1351 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1352 "region must have dead space");
1353
1354 // The gc number is saved whenever a maximum compaction is done, and used to
1355 // determine when the maximum compaction interval has expired. This avoids
1356 // successive max compactions for different reasons.
1357 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1358 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1359 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1360 total_invocations() == HeapFirstMaximumCompactionCount;
1361 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1362 _maximum_compaction_gc_num = total_invocations();
1363 return sd.region_to_addr(full_cp);
1364 }
1365
1366 const size_t space_live = pointer_delta(new_top, bottom);
1367 const size_t space_used = space->used_in_words();
1368 const size_t space_capacity = space->capacity_in_words();
1369
1370 const double density = double(space_live) / double(space_capacity);
1371 const size_t min_percent_free = MarkSweepDeadRatio;
1372 const double limiter = dead_wood_limiter(density, min_percent_free);
1373 const size_t dead_wood_max = space_used - space_live;
1374 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1375 dead_wood_max);
1376
1377 if (TraceParallelOldGCDensePrefix) {
1378 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1379 "space_cap=" SIZE_FORMAT,
1380 space_live, space_used,
1381 space_capacity);
1382 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1383 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1384 density, min_percent_free, limiter,
1385 dead_wood_max, dead_wood_limit);
1386 }
1387
1388 // Locate the region with the desired amount of dead space to the left.
1389 const RegionData* const limit_cp =
1390 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1391
1392 // Scan from the first region with dead space to the limit region and find the
1393 // one with the best (largest) reclaimed ratio.
1394 double best_ratio = 0.0;
1395 const RegionData* best_cp = full_cp;
1396 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1397 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1398 if (tmp_ratio > best_ratio) {
1399 best_cp = cp;
1400 best_ratio = tmp_ratio;
1401 }
1402 }
1403
1404 return sd.region_to_addr(best_cp);
1405 }
1406
1407 void PSParallelCompact::summarize_spaces_quick()
1408 {
1409 for (unsigned int i = 0; i < last_space_id; ++i) {
1410 const MutableSpace* space = _space_info[i].space();
1411 HeapWord** nta = _space_info[i].new_top_addr();
1412 bool result = _summary_data.summarize(_space_info[i].split_info(),
1413 space->bottom(), space->top(), NULL,
1414 space->bottom(), space->end(), nta);
1415 assert(result, "space must fit into itself");
1416 _space_info[i].set_dense_prefix(space->bottom());
1417 }
1418 }
1419
1420 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1421 {
1422 HeapWord* const dense_prefix_end = dense_prefix(id);
1423 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1424 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1425 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1426 // Only enough dead space is filled so that any remaining dead space to the
1427 // left is larger than the minimum filler object. (The remainder is filled
1428 // during the copy/update phase.)
1429 //
1430 // The size of the dead space to the right of the boundary is not a
1431 // concern, since compaction will be able to use whatever space is
1432 // available.
1433 //
1434 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1435 // surrounds the space to be filled with an object.
1436 //
1437 // In the 32-bit VM, each bit represents two 32-bit words:
1438 // +---+
1439 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1440 // end_bits: ... x x x | 0 | || 0 x x ...
1441 // +---+
1442 //
1443 // In the 64-bit VM, each bit represents one 64-bit word:
1444 // +------------+
1445 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1446 // end_bits: ... x x 1 | 0 || 0 | x x ...
1447 // +------------+
1448 // +-------+
1449 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1450 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1451 // +-------+
1452 // +-----------+
1453 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1454 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1455 // +-----------+
1456 // +-------+
1457 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1458 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1459 // +-------+
1460
1461 // Initially assume case a, c or e will apply.
1462 size_t obj_len = CollectedHeap::min_fill_size();
1463 HeapWord* obj_beg = dense_prefix_end - obj_len;
1464
1465 #ifdef _LP64
1466 if (MinObjAlignment > 1) { // object alignment > heap word size
1467 // Cases a, c or e.
1468 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1469 // Case b above.
1470 obj_beg = dense_prefix_end - 1;
1471 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1472 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1473 // Case d above.
1474 obj_beg = dense_prefix_end - 3;
1475 obj_len = 3;
1476 }
1477 #endif // #ifdef _LP64
1478
1479 CollectedHeap::fill_with_object(obj_beg, obj_len);
1480 _mark_bitmap.mark_obj(obj_beg, obj_len);
1481 _summary_data.add_obj(obj_beg, obj_len);
1482 assert(start_array(id) != NULL, "sanity");
1483 start_array(id)->allocate_block(obj_beg);
1484 }
1485 }
1486
1487 void
1488 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1489 {
1490 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1491 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1492 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1493 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1494 cur->set_source_region(0);
1495 }
1496 }
1497
1498 void
1499 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1500 {
1501 assert(id < last_space_id, "id out of range");
1502 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1503 "should have been reset in summarize_spaces_quick()");
1504
1505 const MutableSpace* space = _space_info[id].space();
1506 if (_space_info[id].new_top() != space->bottom()) {
1507 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1508 _space_info[id].set_dense_prefix(dense_prefix_end);
1509
1510 #ifndef PRODUCT
1511 if (TraceParallelOldGCDensePrefix) {
1512 print_dense_prefix_stats("ratio", id, maximum_compaction,
1513 dense_prefix_end);
1514 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1515 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1516 }
1517 #endif // #ifndef PRODUCT
1518
1519 // Recompute the summary data, taking into account the dense prefix. If
1520 // every last byte will be reclaimed, then the existing summary data which
1521 // compacts everything can be left in place.
1522 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1523 // If dead space crosses the dense prefix boundary, it is (at least
1524 // partially) filled with a dummy object, marked live and added to the
1525 // summary data. This simplifies the copy/update phase and must be done
1526 // before the final locations of objects are determined, to prevent
1527 // leaving a fragment of dead space that is too small to fill.
1528 fill_dense_prefix_end(id);
1529
1530 // Compute the destination of each Region, and thus each object.
1531 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1532 _summary_data.summarize(_space_info[id].split_info(),
1533 dense_prefix_end, space->top(), NULL,
1534 dense_prefix_end, space->end(),
1535 _space_info[id].new_top_addr());
1536 }
1537 }
1538
1539 if (log_develop_is_enabled(Trace, gc, compaction, phases)) {
1540 const size_t region_size = ParallelCompactData::RegionSize;
1541 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1542 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1543 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1544 HeapWord* const new_top = _space_info[id].new_top();
1545 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1546 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1547 log_develop_trace(gc, compaction, phases)(
1548 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1549 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1550 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1551 id, space->capacity_in_words(), p2i(dense_prefix_end),
1552 dp_region, dp_words / region_size,
1553 cr_words / region_size, p2i(new_top));
1554 }
1555 }
1556
1557 #ifndef PRODUCT
1558 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1559 HeapWord* dst_beg, HeapWord* dst_end,
1560 SpaceId src_space_id,
1561 HeapWord* src_beg, HeapWord* src_end)
1562 {
1563 log_develop_trace(gc, compaction, phases)(
1564 "Summarizing %d [%s] into %d [%s]: "
1565 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1566 SIZE_FORMAT "-" SIZE_FORMAT " "
1567 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1568 SIZE_FORMAT "-" SIZE_FORMAT,
1569 src_space_id, space_names[src_space_id],
1570 dst_space_id, space_names[dst_space_id],
1571 p2i(src_beg), p2i(src_end),
1572 _summary_data.addr_to_region_idx(src_beg),
1573 _summary_data.addr_to_region_idx(src_end),
1574 p2i(dst_beg), p2i(dst_end),
1575 _summary_data.addr_to_region_idx(dst_beg),
1576 _summary_data.addr_to_region_idx(dst_end));
1577 }
1578 #endif // #ifndef PRODUCT
1579
1580 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1581 bool maximum_compaction)
1582 {
1583 GCTraceTime(Trace, gc, phases) tm("Summary Phase", &_gc_timer);
1584
1585 #ifdef ASSERT
1586 if (TraceParallelOldGCMarkingPhase) {
1587 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1588 "add_obj_bytes=" SIZE_FORMAT,
1589 add_obj_count, add_obj_size * HeapWordSize);
1590 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1591 "mark_bitmap_bytes=" SIZE_FORMAT,
1592 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1593 }
1594 #endif // #ifdef ASSERT
1595
1596 // Quick summarization of each space into itself, to see how much is live.
1597 summarize_spaces_quick();
1598
1599 log_develop_trace(gc, compaction, phases)("summary phase: after summarizing each space to self");
1600 NOT_PRODUCT(print_region_ranges());
1601 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1602
1603 // The amount of live data that will end up in old space (assuming it fits).
1604 size_t old_space_total_live = 0;
1605 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1606 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1607 _space_info[id].space()->bottom());
1608 }
1609
1610 MutableSpace* const old_space = _space_info[old_space_id].space();
1611 const size_t old_capacity = old_space->capacity_in_words();
1612 if (old_space_total_live > old_capacity) {
1613 // XXX - should also try to expand
1614 maximum_compaction = true;
1615 }
1616
1617 // Old generations.
1618 summarize_space(old_space_id, maximum_compaction);
1619
1620 // Summarize the remaining spaces in the young gen. The initial target space
1621 // is the old gen. If a space does not fit entirely into the target, then the
1622 // remainder is compacted into the space itself and that space becomes the new
1623 // target.
1624 SpaceId dst_space_id = old_space_id;
1625 HeapWord* dst_space_end = old_space->end();
1626 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1627 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1628 const MutableSpace* space = _space_info[id].space();
1629 const size_t live = pointer_delta(_space_info[id].new_top(),
1630 space->bottom());
1631 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1632
1633 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1634 SpaceId(id), space->bottom(), space->top());)
1635 if (live > 0 && live <= available) {
1636 // All the live data will fit.
1637 bool done = _summary_data.summarize(_space_info[id].split_info(),
1638 space->bottom(), space->top(),
1639 NULL,
1640 *new_top_addr, dst_space_end,
1641 new_top_addr);
1642 assert(done, "space must fit into old gen");
1643
1644 // Reset the new_top value for the space.
1645 _space_info[id].set_new_top(space->bottom());
1646 } else if (live > 0) {
1647 // Attempt to fit part of the source space into the target space.
1648 HeapWord* next_src_addr = NULL;
1649 bool done = _summary_data.summarize(_space_info[id].split_info(),
1650 space->bottom(), space->top(),
1651 &next_src_addr,
1652 *new_top_addr, dst_space_end,
1653 new_top_addr);
1654 assert(!done, "space should not fit into old gen");
1655 assert(next_src_addr != NULL, "sanity");
1656
1657 // The source space becomes the new target, so the remainder is compacted
1658 // within the space itself.
1659 dst_space_id = SpaceId(id);
1660 dst_space_end = space->end();
1661 new_top_addr = _space_info[id].new_top_addr();
1662 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1663 space->bottom(), dst_space_end,
1664 SpaceId(id), next_src_addr, space->top());)
1665 done = _summary_data.summarize(_space_info[id].split_info(),
1666 next_src_addr, space->top(),
1667 NULL,
1668 space->bottom(), dst_space_end,
1669 new_top_addr);
1670 assert(done, "space must fit when compacted into itself");
1671 assert(*new_top_addr <= space->top(), "usage should not grow");
1672 }
1673 }
1674
1675 log_develop_trace(gc, compaction, phases)("Summary_phase: after final summarization");
1676 NOT_PRODUCT(print_region_ranges());
1677 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1678 }
1679
1680 // This method should contain all heap-specific policy for invoking a full
1681 // collection. invoke_no_policy() will only attempt to compact the heap; it
1682 // will do nothing further. If we need to bail out for policy reasons, scavenge
1683 // before full gc, or any other specialized behavior, it needs to be added here.
1684 //
1685 // Note that this method should only be called from the vm_thread while at a
1686 // safepoint.
1687 //
1688 // Note that the all_soft_refs_clear flag in the collector policy
1689 // may be true because this method can be called without intervening
1690 // activity. For example when the heap space is tight and full measure
1691 // are being taken to free space.
1692 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1693 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1694 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1695 "should be in vm thread");
1696
1697 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1698 GCCause::Cause gc_cause = heap->gc_cause();
1699 assert(!heap->is_gc_active(), "not reentrant");
1700
1701 PSAdaptiveSizePolicy* policy = heap->size_policy();
1702 IsGCActiveMark mark;
1703
1704 if (ScavengeBeforeFullGC) {
1705 PSScavenge::invoke_no_policy();
1706 }
1707
1708 const bool clear_all_soft_refs =
1709 heap->collector_policy()->should_clear_all_soft_refs();
1710
1711 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1712 maximum_heap_compaction);
1713 }
1714
1715 // This method contains no policy. You should probably
1716 // be calling invoke() instead.
1717 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1718 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1719 assert(ref_processor() != NULL, "Sanity");
1720
1721 if (GCLocker::check_active_before_gc()) {
1722 return false;
1723 }
1724
1725 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1726
1727 GCIdMark gc_id_mark;
1728 _gc_timer.register_gc_start();
1729 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1730
1731 TimeStamp marking_start;
1732 TimeStamp compaction_start;
1733 TimeStamp collection_exit;
1734
1735 GCCause::Cause gc_cause = heap->gc_cause();
1736 PSYoungGen* young_gen = heap->young_gen();
1737 PSOldGen* old_gen = heap->old_gen();
1738 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1739
1740 // The scope of casr should end after code that can change
1741 // CollectorPolicy::_should_clear_all_soft_refs.
1742 ClearedAllSoftRefs casr(maximum_heap_compaction,
1743 heap->collector_policy());
1744
1745 if (ZapUnusedHeapArea) {
1746 // Save information needed to minimize mangling
1747 heap->record_gen_tops_before_GC();
1748 }
1749
1750 // Make sure data structures are sane, make the heap parsable, and do other
1751 // miscellaneous bookkeeping.
1752 pre_compact();
1753
1754 PreGCValues pre_gc_values(heap);
1755
1756 // Get the compaction manager reserved for the VM thread.
1757 ParCompactionManager* const vmthread_cm =
1758 ParCompactionManager::manager_array(gc_task_manager()->workers());
1759
1760 {
1761 ResourceMark rm;
1762 HandleMark hm;
1763
1764 // Set the number of GC threads to be used in this collection
1765 gc_task_manager()->set_active_gang();
1766 gc_task_manager()->task_idle_workers();
1767
1768 GCTraceCPUTime tcpu;
1769 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1770
1771 heap->pre_full_gc_dump(&_gc_timer);
1772
1773 TraceCollectorStats tcs(counters());
1774 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
1775
1776 if (TraceOldGenTime) accumulated_time()->start();
1777
1778 // Let the size policy know we're starting
1779 size_policy->major_collection_begin();
1780
1781 CodeCache::gc_prologue();
1782
1783 #if defined(COMPILER2) || INCLUDE_JVMCI
1784 DerivedPointerTable::clear();
1785 #endif
1786
1787 ref_processor()->enable_discovery();
1788 ref_processor()->setup_policy(maximum_heap_compaction);
1789
1790 bool marked_for_unloading = false;
1791
1792 marking_start.update();
1793 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
1794
1795 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1796 && GCCause::is_user_requested_gc(gc_cause);
1797 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1798
1799 #if defined(COMPILER2) || INCLUDE_JVMCI
1800 assert(DerivedPointerTable::is_active(), "Sanity");
1801 DerivedPointerTable::set_active(false);
1802 #endif
1803
1804 // adjust_roots() updates Universe::_intArrayKlassObj which is
1805 // needed by the compaction for filling holes in the dense prefix.
1806 adjust_roots(vmthread_cm);
1807
1808 compaction_start.update();
1809 compact();
1810
1811 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1812 // done before resizing.
1813 post_compact();
1814
1815 // Let the size policy know we're done
1816 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1817
1818 if (UseAdaptiveSizePolicy) {
1819 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1820 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1821 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1822
1823 // Don't check if the size_policy is ready here. Let
1824 // the size_policy check that internally.
1825 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1826 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1827 // Swap the survivor spaces if from_space is empty. The
1828 // resize_young_gen() called below is normally used after
1829 // a successful young GC and swapping of survivor spaces;
1830 // otherwise, it will fail to resize the young gen with
1831 // the current implementation.
1832 if (young_gen->from_space()->is_empty()) {
1833 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1834 young_gen->swap_spaces();
1835 }
1836
1837 // Calculate optimal free space amounts
1838 assert(young_gen->max_size() >
1839 young_gen->from_space()->capacity_in_bytes() +
1840 young_gen->to_space()->capacity_in_bytes(),
1841 "Sizes of space in young gen are out-of-bounds");
1842
1843 size_t young_live = young_gen->used_in_bytes();
1844 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1845 size_t old_live = old_gen->used_in_bytes();
1846 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1847 size_t max_old_gen_size = old_gen->max_gen_size();
1848 size_t max_eden_size = young_gen->max_size() -
1849 young_gen->from_space()->capacity_in_bytes() -
1850 young_gen->to_space()->capacity_in_bytes();
1851
1852 // Used for diagnostics
1853 size_policy->clear_generation_free_space_flags();
1854
1855 size_policy->compute_generations_free_space(young_live,
1856 eden_live,
1857 old_live,
1858 cur_eden,
1859 max_old_gen_size,
1860 max_eden_size,
1861 true /* full gc*/);
1862
1863 size_policy->check_gc_overhead_limit(young_live,
1864 eden_live,
1865 max_old_gen_size,
1866 max_eden_size,
1867 true /* full gc*/,
1868 gc_cause,
1869 heap->collector_policy());
1870
1871 size_policy->decay_supplemental_growth(true /* full gc*/);
1872
1873 heap->resize_old_gen(
1874 size_policy->calculated_old_free_size_in_bytes());
1875
1876 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1877 size_policy->calculated_survivor_size_in_bytes());
1878 }
1879
1880 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1881 }
1882
1883 if (UsePerfData) {
1884 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1885 counters->update_counters();
1886 counters->update_old_capacity(old_gen->capacity_in_bytes());
1887 counters->update_young_capacity(young_gen->capacity_in_bytes());
1888 }
1889
1890 heap->resize_all_tlabs();
1891
1892 // Resize the metaspace capacity after a collection
1893 MetaspaceGC::compute_new_size();
1894
1895 if (TraceOldGenTime) {
1896 accumulated_time()->stop();
1897 }
1898
1899 young_gen->print_used_change(pre_gc_values.young_gen_used());
1900 old_gen->print_used_change(pre_gc_values.old_gen_used());
1901 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
1902
1903 // Track memory usage and detect low memory
1904 MemoryService::track_memory_usage();
1905 heap->update_counters();
1906 gc_task_manager()->release_idle_workers();
1907
1908 heap->post_full_gc_dump(&_gc_timer);
1909 }
1910
1911 #ifdef ASSERT
1912 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1913 ParCompactionManager* const cm =
1914 ParCompactionManager::manager_array(int(i));
1915 assert(cm->marking_stack()->is_empty(), "should be empty");
1916 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
1917 }
1918 #endif // ASSERT
1919
1920 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1921 HandleMark hm; // Discard invalid handles created during verification
1922 Universe::verify("After GC");
1923 }
1924
1925 // Re-verify object start arrays
1926 if (VerifyObjectStartArray &&
1927 VerifyAfterGC) {
1928 old_gen->verify_object_start_array();
1929 }
1930
1931 if (ZapUnusedHeapArea) {
1932 old_gen->object_space()->check_mangled_unused_area_complete();
1933 }
1934
1935 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1936
1937 collection_exit.update();
1938
1939 heap->print_heap_after_gc();
1940 heap->trace_heap_after_gc(&_gc_tracer);
1941
1942 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1943 marking_start.ticks(), compaction_start.ticks(),
1944 collection_exit.ticks());
1945 gc_task_manager()->print_task_time_stamps();
1946
1947 #ifdef TRACESPINNING
1948 ParallelTaskTerminator::print_termination_counts();
1949 #endif
1950
1951 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1952
1953 _gc_timer.register_gc_end();
1954
1955 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1956 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1957
1958 return true;
1959 }
1960
1961 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1962 PSYoungGen* young_gen,
1963 PSOldGen* old_gen) {
1964 MutableSpace* const eden_space = young_gen->eden_space();
1965 assert(!eden_space->is_empty(), "eden must be non-empty");
1966 assert(young_gen->virtual_space()->alignment() ==
1967 old_gen->virtual_space()->alignment(), "alignments do not match");
1968
1969 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
1970 return false;
1971 }
1972
1973 // Both generations must be completely committed.
1974 if (young_gen->virtual_space()->uncommitted_size() != 0) {
1975 return false;
1976 }
1977 if (old_gen->virtual_space()->uncommitted_size() != 0) {
1978 return false;
1979 }
1980
1981 // Figure out how much to take from eden. Include the average amount promoted
1982 // in the total; otherwise the next young gen GC will simply bail out to a
1983 // full GC.
1984 const size_t alignment = old_gen->virtual_space()->alignment();
1985 const size_t eden_used = eden_space->used_in_bytes();
1986 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
1987 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
1988 const size_t eden_capacity = eden_space->capacity_in_bytes();
1989
1990 if (absorb_size >= eden_capacity) {
1991 return false; // Must leave some space in eden.
1992 }
1993
1994 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
1995 if (new_young_size < young_gen->min_gen_size()) {
1996 return false; // Respect young gen minimum size.
1997 }
1998
1999 log_trace(heap, ergo)(" absorbing " SIZE_FORMAT "K: "
2000 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2001 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2002 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2003 absorb_size / K,
2004 eden_capacity / K, (eden_capacity - absorb_size) / K,
2005 young_gen->from_space()->used_in_bytes() / K,
2006 young_gen->to_space()->used_in_bytes() / K,
2007 young_gen->capacity_in_bytes() / K, new_young_size / K);
2008
2009 // Fill the unused part of the old gen.
2010 MutableSpace* const old_space = old_gen->object_space();
2011 HeapWord* const unused_start = old_space->top();
2012 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2013
2014 if (unused_words > 0) {
2015 if (unused_words < CollectedHeap::min_fill_size()) {
2016 return false; // If the old gen cannot be filled, must give up.
2017 }
2018 CollectedHeap::fill_with_objects(unused_start, unused_words);
2019 }
2020
2021 // Take the live data from eden and set both top and end in the old gen to
2022 // eden top. (Need to set end because reset_after_change() mangles the region
2023 // from end to virtual_space->high() in debug builds).
2024 HeapWord* const new_top = eden_space->top();
2025 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2026 absorb_size);
2027 young_gen->reset_after_change();
2028 old_space->set_top(new_top);
2029 old_space->set_end(new_top);
2030 old_gen->reset_after_change();
2031
2032 // Update the object start array for the filler object and the data from eden.
2033 ObjectStartArray* const start_array = old_gen->start_array();
2034 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2035 start_array->allocate_block(p);
2036 }
2037
2038 // Could update the promoted average here, but it is not typically updated at
2039 // full GCs and the value to use is unclear. Something like
2040 //
2041 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2042
2043 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2044 return true;
2045 }
2046
2047 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2048 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2049 "shouldn't return NULL");
2050 return ParallelScavengeHeap::gc_task_manager();
2051 }
2052
2053 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2054 bool maximum_heap_compaction,
2055 ParallelOldTracer *gc_tracer) {
2056 // Recursively traverse all live objects and mark them
2057 GCTraceTime(Trace, gc, phases) tm("Marking Phase", &_gc_timer);
2058
2059 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2060 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2061 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2062 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2063 ParallelTaskTerminator terminator(active_gc_threads, qset);
2064
2065 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm);
2066 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2067
2068 // Need new claim bits before marking starts.
2069 ClassLoaderDataGraph::clear_claimed_marks();
2070
2071 {
2072 GCTraceTime(Trace, gc, phases) tm("Par Mark", &_gc_timer);
2073
2074 ParallelScavengeHeap::ParStrongRootsScope psrs;
2075
2076 GCTaskQueue* q = GCTaskQueue::create();
2077
2078 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2079 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2080 // We scan the thread roots in parallel
2081 Threads::create_thread_roots_marking_tasks(q);
2082 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2083 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2084 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2085 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2086 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2087 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2088 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2089
2090 if (active_gc_threads > 1) {
2091 for (uint j = 0; j < active_gc_threads; j++) {
2092 q->enqueue(new StealMarkingTask(&terminator));
2093 }
2094 }
2095
2096 gc_task_manager()->execute_and_wait(q);
2097 }
2098
2099 // Process reference objects found during marking
2100 {
2101 GCTraceTime(Trace, gc, phases) tm("Reference Processing", &_gc_timer);
2102
2103 ReferenceProcessorStats stats;
2104 if (ref_processor()->processing_is_mt()) {
2105 RefProcTaskExecutor task_executor;
2106 stats = ref_processor()->process_discovered_references(
2107 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2108 &task_executor, &_gc_timer);
2109 } else {
2110 stats = ref_processor()->process_discovered_references(
2111 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2112 &_gc_timer);
2113 }
2114
2115 gc_tracer->report_gc_reference_stats(stats);
2116 }
2117
2118 GCTraceTime(Trace, gc) tm_m("Class Unloading", &_gc_timer);
2119
2120 // This is the point where the entire marking should have completed.
2121 assert(cm->marking_stacks_empty(), "Marking should have completed");
2122
2123 // Follow system dictionary roots and unload classes.
2124 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
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 // Delete entries for dead interned strings.
2133 StringTable::unlink(is_alive_closure());
2134
2135 // Clean up unreferenced symbols in symbol table.
2136 SymbolTable::unlink();
2137 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2138 }
2139
2140 // This should be moved to the shared markSweep code!
2141 class PSAlwaysTrueClosure: public BoolObjectClosure {
2142 public:
2143 bool do_object_b(oop p) { return true; }
2144 };
2145 static PSAlwaysTrueClosure always_true;
2146
2147 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2148 // Adjust the pointers to reflect the new locations
2149 GCTraceTime(Trace, gc, phases) tm("Adjust Roots", &_gc_timer);
2150
2151 // Need new claim bits when tracing through and adjusting pointers.
2152 ClassLoaderDataGraph::clear_claimed_marks();
2153
2154 PSParallelCompact::AdjustPointerClosure oop_closure(cm);
2155 PSParallelCompact::AdjustKlassClosure klass_closure(cm);
2156
2157 // General strong roots.
2158 Universe::oops_do(&oop_closure);
2159 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2160 CLDToOopClosure adjust_from_cld(&oop_closure);
2161 Threads::oops_do(&oop_closure, &adjust_from_cld, 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(&always_true, &oop_closure);
2173
2174 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2175 CodeCache::blobs_do(&adjust_from_blobs);
2176 StringTable::oops_do(&oop_closure);
2177 ref_processor()->weak_oops_do(&oop_closure);
2178 // Roots were visited so references into the young gen in roots
2179 // may have been scanned. Process them also.
2180 // Should the reference processor have a span that excludes
2181 // young gen objects?
2182 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2183 }
2184
2185 // Helper class to print 8 region numbers per line and then print the total at the end.
2186 class FillableRegionLogger : public StackObj {
2187 private:
2188 LogHandle(gc, compaction) log;
2189 static const int LineLength = 8;
2190 size_t _regions[LineLength];
2191 int _next_index;
2192 bool _enabled;
2193 size_t _total_regions;
2194 public:
2195 FillableRegionLogger() : _next_index(0), _total_regions(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)) { }
2196 ~FillableRegionLogger() {
2197 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2198 }
2199
2200 void print_line() {
2201 if (!_enabled || _next_index == 0) {
2202 return;
2203 }
2204 FormatBuffer<> line("Fillable: ");
2205 for (int i = 0; i < _next_index; i++) {
2206 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2207 }
2208 log.trace("%s", line.buffer());
2209 _next_index = 0;
2210 }
2211
2212 void handle(size_t region) {
2213 if (!_enabled) {
2214 return;
2215 }
2216 _regions[_next_index++] = region;
2217 if (_next_index == LineLength) {
2218 print_line();
2219 }
2220 _total_regions++;
2221 }
2222 };
2223
2224 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2225 uint parallel_gc_threads)
2226 {
2227 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2228
2229 // Find the threads that are active
2230 unsigned int which = 0;
2231
2232 const uint task_count = MAX2(parallel_gc_threads, 1U);
2233 for (uint j = 0; j < task_count; j++) {
2234 q->enqueue(new DrainStacksCompactionTask(j));
2235 ParCompactionManager::verify_region_list_empty(j);
2236 // Set the region stacks variables to "no" region stack values
2237 // so that they will be recognized and needing a region stack
2238 // in the stealing tasks if they do not get one by executing
2239 // a draining stack.
2240 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2241 cm->set_region_stack(NULL);
2242 cm->set_region_stack_index((uint)max_uintx);
2243 }
2244 ParCompactionManager::reset_recycled_stack_index();
2245
2246 // Find all regions that are available (can be filled immediately) and
2247 // distribute them to the thread stacks. The iteration is done in reverse
2248 // order (high to low) so the regions will be removed in ascending order.
2249
2250 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2251
2252 // A region index which corresponds to the tasks created above.
2253 // "which" must be 0 <= which < task_count
2254
2255 which = 0;
2256 // id + 1 is used to test termination so unsigned can
2257 // be used with an old_space_id == 0.
2258 FillableRegionLogger region_logger;
2259 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2260 SpaceInfo* const space_info = _space_info + id;
2261 MutableSpace* const space = space_info->space();
2262 HeapWord* const new_top = space_info->new_top();
2263
2264 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2265 const size_t end_region =
2266 sd.addr_to_region_idx(sd.region_align_up(new_top));
2267
2268 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2269 if (sd.region(cur)->claim_unsafe()) {
2270 ParCompactionManager::region_list_push(which, cur);
2271 region_logger.handle(cur);
2272 // Assign regions to tasks in round-robin fashion.
2273 if (++which == task_count) {
2274 assert(which <= parallel_gc_threads,
2275 "Inconsistent number of workers");
2276 which = 0;
2277 }
2278 }
2279 }
2280 region_logger.print_line();
2281 }
2282 }
2283
2284 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2285
2286 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2287 uint parallel_gc_threads) {
2288 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2289
2290 ParallelCompactData& sd = PSParallelCompact::summary_data();
2291
2292 // Iterate over all the spaces adding tasks for updating
2293 // regions in the dense prefix. Assume that 1 gc thread
2294 // will work on opening the gaps and the remaining gc threads
2295 // will work on the dense prefix.
2296 unsigned int space_id;
2297 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2298 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2299 const MutableSpace* const space = _space_info[space_id].space();
2300
2301 if (dense_prefix_end == space->bottom()) {
2302 // There is no dense prefix for this space.
2303 continue;
2304 }
2305
2306 // The dense prefix is before this region.
2307 size_t region_index_end_dense_prefix =
2308 sd.addr_to_region_idx(dense_prefix_end);
2309 RegionData* const dense_prefix_cp =
2310 sd.region(region_index_end_dense_prefix);
2311 assert(dense_prefix_end == space->end() ||
2312 dense_prefix_cp->available() ||
2313 dense_prefix_cp->claimed(),
2314 "The region after the dense prefix should always be ready to fill");
2315
2316 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2317
2318 // Is there dense prefix work?
2319 size_t total_dense_prefix_regions =
2320 region_index_end_dense_prefix - region_index_start;
2321 // How many regions of the dense prefix should be given to
2322 // each thread?
2323 if (total_dense_prefix_regions > 0) {
2324 uint tasks_for_dense_prefix = 1;
2325 if (total_dense_prefix_regions <=
2326 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2327 // Don't over partition. This assumes that
2328 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2329 // so there are not many regions to process.
2330 tasks_for_dense_prefix = parallel_gc_threads;
2331 } else {
2332 // Over partition
2333 tasks_for_dense_prefix = parallel_gc_threads *
2334 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2335 }
2336 size_t regions_per_thread = total_dense_prefix_regions /
2337 tasks_for_dense_prefix;
2338 // Give each thread at least 1 region.
2339 if (regions_per_thread == 0) {
2340 regions_per_thread = 1;
2341 }
2342
2343 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2344 if (region_index_start >= region_index_end_dense_prefix) {
2345 break;
2346 }
2347 // region_index_end is not processed
2348 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2349 region_index_end_dense_prefix);
2350 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2351 region_index_start,
2352 region_index_end));
2353 region_index_start = region_index_end;
2354 }
2355 }
2356 // This gets any part of the dense prefix that did not
2357 // fit evenly.
2358 if (region_index_start < region_index_end_dense_prefix) {
2359 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2360 region_index_start,
2361 region_index_end_dense_prefix));
2362 }
2363 }
2364 }
2365
2366 void PSParallelCompact::enqueue_region_stealing_tasks(
2367 GCTaskQueue* q,
2368 ParallelTaskTerminator* terminator_ptr,
2369 uint parallel_gc_threads) {
2370 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer);
2371
2372 // Once a thread has drained it's stack, it should try to steal regions from
2373 // other threads.
2374 if (parallel_gc_threads > 1) {
2375 for (uint j = 0; j < parallel_gc_threads; j++) {
2376 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2377 }
2378 }
2379 }
2380
2381 #ifdef ASSERT
2382 // Write a histogram of the number of times the block table was filled for a
2383 // region.
2384 void PSParallelCompact::write_block_fill_histogram()
2385 {
2386 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2387 return;
2388 }
2389
2390 LogHandle(gc, compaction) log;
2391 ResourceMark rm;
2392 outputStream* out = log.trace_stream();
2393
2394 typedef ParallelCompactData::RegionData rd_t;
2395 ParallelCompactData& sd = summary_data();
2396
2397 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2398 MutableSpace* const spc = _space_info[id].space();
2399 if (spc->bottom() != spc->top()) {
2400 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2401 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2402 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2403
2404 size_t histo[5] = { 0, 0, 0, 0, 0 };
2405 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2406 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2407
2408 for (const rd_t* cur = beg; cur < end; ++cur) {
2409 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2410 }
2411 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2412 for (size_t i = 0; i < histo_len; ++i) {
2413 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2414 histo[i], 100.0 * histo[i] / region_cnt);
2415 }
2416 out->cr();
2417 }
2418 }
2419 }
2420 #endif // #ifdef ASSERT
2421
2422 void PSParallelCompact::compact() {
2423 GCTraceTime(Trace, gc, phases) tm("Compaction Phase", &_gc_timer);
2424
2425 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2426 PSOldGen* old_gen = heap->old_gen();
2427 old_gen->start_array()->reset();
2428 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2429 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2430 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2431 ParallelTaskTerminator terminator(active_gc_threads, qset);
2432
2433 GCTaskQueue* q = GCTaskQueue::create();
2434 enqueue_region_draining_tasks(q, active_gc_threads);
2435 enqueue_dense_prefix_tasks(q, active_gc_threads);
2436 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2437
2438 {
2439 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2440
2441 gc_task_manager()->execute_and_wait(q);
2442
2443 #ifdef ASSERT
2444 // Verify that all regions have been processed before the deferred updates.
2445 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2446 verify_complete(SpaceId(id));
2447 }
2448 #endif
2449 }
2450
2451 {
2452 // Update the deferred objects, if any. Any compaction manager can be used.
2453 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2454 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2455 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2456 update_deferred_objects(cm, SpaceId(id));
2457 }
2458 }
2459
2460 DEBUG_ONLY(write_block_fill_histogram());
2461 }
2462
2463 #ifdef ASSERT
2464 void PSParallelCompact::verify_complete(SpaceId space_id) {
2465 // All Regions between space bottom() to new_top() should be marked as filled
2466 // and all Regions between new_top() and top() should be available (i.e.,
2467 // should have been emptied).
2468 ParallelCompactData& sd = summary_data();
2469 SpaceInfo si = _space_info[space_id];
2470 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2471 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2472 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2473 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2474 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2475
2476 bool issued_a_warning = false;
2477
2478 size_t cur_region;
2479 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2480 const RegionData* const c = sd.region(cur_region);
2481 if (!c->completed()) {
2482 warning("region " SIZE_FORMAT " not filled: "
2483 "destination_count=%u",
2484 cur_region, c->destination_count());
2485 issued_a_warning = true;
2486 }
2487 }
2488
2489 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2490 const RegionData* const c = sd.region(cur_region);
2491 if (!c->available()) {
2492 warning("region " SIZE_FORMAT " not empty: "
2493 "destination_count=%u",
2494 cur_region, c->destination_count());
2495 issued_a_warning = true;
2496 }
2497 }
2498
2499 if (issued_a_warning) {
2500 print_region_ranges();
2501 }
2502 }
2503 #endif // #ifdef ASSERT
2504
2505 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2506 _start_array->allocate_block(addr);
2507 compaction_manager()->update_contents(oop(addr));
2508 }
2509
2510 // Update interior oops in the ranges of regions [beg_region, end_region).
2511 void
2512 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2513 SpaceId space_id,
2514 size_t beg_region,
2515 size_t end_region) {
2516 ParallelCompactData& sd = summary_data();
2517 ParMarkBitMap* const mbm = mark_bitmap();
2518
2519 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2520 HeapWord* const end_addr = sd.region_to_addr(end_region);
2521 assert(beg_region <= end_region, "bad region range");
2522 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2523
2524 #ifdef ASSERT
2525 // Claim the regions to avoid triggering an assert when they are marked as
2526 // filled.
2527 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2528 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2529 }
2530 #endif // #ifdef ASSERT
2531
2532 if (beg_addr != space(space_id)->bottom()) {
2533 // Find the first live object or block of dead space that *starts* in this
2534 // range of regions. If a partial object crosses onto the region, skip it;
2535 // it will be marked for 'deferred update' when the object head is
2536 // processed. If dead space crosses onto the region, it is also skipped; it
2537 // will be filled when the prior region is processed. If neither of those
2538 // apply, the first word in the region is the start of a live object or dead
2539 // space.
2540 assert(beg_addr > space(space_id)->bottom(), "sanity");
2541 const RegionData* const cp = sd.region(beg_region);
2542 if (cp->partial_obj_size() != 0) {
2543 beg_addr = sd.partial_obj_end(beg_region);
2544 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2545 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2546 }
2547 }
2548
2549 if (beg_addr < end_addr) {
2550 // A live object or block of dead space starts in this range of Regions.
2551 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2552
2553 // Create closures and iterate.
2554 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2555 FillClosure fill_closure(cm, space_id);
2556 ParMarkBitMap::IterationStatus status;
2557 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2558 dense_prefix_end);
2559 if (status == ParMarkBitMap::incomplete) {
2560 update_closure.do_addr(update_closure.source());
2561 }
2562 }
2563
2564 // Mark the regions as filled.
2565 RegionData* const beg_cp = sd.region(beg_region);
2566 RegionData* const end_cp = sd.region(end_region);
2567 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2568 cp->set_completed();
2569 }
2570 }
2571
2572 // Return the SpaceId for the space containing addr. If addr is not in the
2573 // heap, last_space_id is returned. In debug mode it expects the address to be
2574 // in the heap and asserts such.
2575 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2576 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2577
2578 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2579 if (_space_info[id].space()->contains(addr)) {
2580 return SpaceId(id);
2581 }
2582 }
2583
2584 assert(false, "no space contains the addr");
2585 return last_space_id;
2586 }
2587
2588 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2589 SpaceId id) {
2590 assert(id < last_space_id, "bad space id");
2591
2592 ParallelCompactData& sd = summary_data();
2593 const SpaceInfo* const space_info = _space_info + id;
2594 ObjectStartArray* const start_array = space_info->start_array();
2595
2596 const MutableSpace* const space = space_info->space();
2597 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2598 HeapWord* const beg_addr = space_info->dense_prefix();
2599 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2600
2601 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2602 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2603 const RegionData* cur_region;
2604 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2605 HeapWord* const addr = cur_region->deferred_obj_addr();
2606 if (addr != NULL) {
2607 if (start_array != NULL) {
2608 start_array->allocate_block(addr);
2609 }
2610 cm->update_contents(oop(addr));
2611 assert(oop(addr)->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2612 }
2613 }
2614 }
2615
2616 // Skip over count live words starting from beg, and return the address of the
2617 // next live word. Unless marked, the word corresponding to beg is assumed to
2618 // be dead. Callers must either ensure beg does not correspond to the middle of
2619 // an object, or account for those live words in some other way. Callers must
2620 // also ensure that there are enough live words in the range [beg, end) to skip.
2621 HeapWord*
2622 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2623 {
2624 assert(count > 0, "sanity");
2625
2626 ParMarkBitMap* m = mark_bitmap();
2627 idx_t bits_to_skip = m->words_to_bits(count);
2628 idx_t cur_beg = m->addr_to_bit(beg);
2629 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2630
2631 do {
2632 cur_beg = m->find_obj_beg(cur_beg, search_end);
2633 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2634 const size_t obj_bits = cur_end - cur_beg + 1;
2635 if (obj_bits > bits_to_skip) {
2636 return m->bit_to_addr(cur_beg + bits_to_skip);
2637 }
2638 bits_to_skip -= obj_bits;
2639 cur_beg = cur_end + 1;
2640 } while (bits_to_skip > 0);
2641
2642 // Skipping the desired number of words landed just past the end of an object.
2643 // Find the start of the next object.
2644 cur_beg = m->find_obj_beg(cur_beg, search_end);
2645 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2646 return m->bit_to_addr(cur_beg);
2647 }
2648
2649 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2650 SpaceId src_space_id,
2651 size_t src_region_idx)
2652 {
2653 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2654
2655 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2656 if (split_info.dest_region_addr() == dest_addr) {
2657 // The partial object ending at the split point contains the first word to
2658 // be copied to dest_addr.
2659 return split_info.first_src_addr();
2660 }
2661
2662 const ParallelCompactData& sd = summary_data();
2663 ParMarkBitMap* const bitmap = mark_bitmap();
2664 const size_t RegionSize = ParallelCompactData::RegionSize;
2665
2666 assert(sd.is_region_aligned(dest_addr), "not aligned");
2667 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2668 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2669 HeapWord* const src_region_destination = src_region_ptr->destination();
2670
2671 assert(dest_addr >= src_region_destination, "wrong src region");
2672 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2673
2674 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2675 HeapWord* const src_region_end = src_region_beg + RegionSize;
2676
2677 HeapWord* addr = src_region_beg;
2678 if (dest_addr == src_region_destination) {
2679 // Return the first live word in the source region.
2680 if (partial_obj_size == 0) {
2681 addr = bitmap->find_obj_beg(addr, src_region_end);
2682 assert(addr < src_region_end, "no objects start in src region");
2683 }
2684 return addr;
2685 }
2686
2687 // Must skip some live data.
2688 size_t words_to_skip = dest_addr - src_region_destination;
2689 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2690
2691 if (partial_obj_size >= words_to_skip) {
2692 // All the live words to skip are part of the partial object.
2693 addr += words_to_skip;
2694 if (partial_obj_size == words_to_skip) {
2695 // Find the first live word past the partial object.
2696 addr = bitmap->find_obj_beg(addr, src_region_end);
2697 assert(addr < src_region_end, "wrong src region");
2698 }
2699 return addr;
2700 }
2701
2702 // Skip over the partial object (if any).
2703 if (partial_obj_size != 0) {
2704 words_to_skip -= partial_obj_size;
2705 addr += partial_obj_size;
2706 }
2707
2708 // Skip over live words due to objects that start in the region.
2709 addr = skip_live_words(addr, src_region_end, words_to_skip);
2710 assert(addr < src_region_end, "wrong src region");
2711 return addr;
2712 }
2713
2714 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2715 SpaceId src_space_id,
2716 size_t beg_region,
2717 HeapWord* end_addr)
2718 {
2719 ParallelCompactData& sd = summary_data();
2720
2721 #ifdef ASSERT
2722 MutableSpace* const src_space = _space_info[src_space_id].space();
2723 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2724 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2725 "src_space_id does not match beg_addr");
2726 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2727 "src_space_id does not match end_addr");
2728 #endif // #ifdef ASSERT
2729
2730 RegionData* const beg = sd.region(beg_region);
2731 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2732
2733 // Regions up to new_top() are enqueued if they become available.
2734 HeapWord* const new_top = _space_info[src_space_id].new_top();
2735 RegionData* const enqueue_end =
2736 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2737
2738 for (RegionData* cur = beg; cur < end; ++cur) {
2739 assert(cur->data_size() > 0, "region must have live data");
2740 cur->decrement_destination_count();
2741 if (cur < enqueue_end && cur->available() && cur->claim()) {
2742 cm->push_region(sd.region(cur));
2743 }
2744 }
2745 }
2746
2747 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2748 SpaceId& src_space_id,
2749 HeapWord*& src_space_top,
2750 HeapWord* end_addr)
2751 {
2752 typedef ParallelCompactData::RegionData RegionData;
2753
2754 ParallelCompactData& sd = PSParallelCompact::summary_data();
2755 const size_t region_size = ParallelCompactData::RegionSize;
2756
2757 size_t src_region_idx = 0;
2758
2759 // Skip empty regions (if any) up to the top of the space.
2760 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2761 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2762 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2763 const RegionData* const top_region_ptr =
2764 sd.addr_to_region_ptr(top_aligned_up);
2765 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2766 ++src_region_ptr;
2767 }
2768
2769 if (src_region_ptr < top_region_ptr) {
2770 // The next source region is in the current space. Update src_region_idx
2771 // and the source address to match src_region_ptr.
2772 src_region_idx = sd.region(src_region_ptr);
2773 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2774 if (src_region_addr > closure.source()) {
2775 closure.set_source(src_region_addr);
2776 }
2777 return src_region_idx;
2778 }
2779
2780 // Switch to a new source space and find the first non-empty region.
2781 unsigned int space_id = src_space_id + 1;
2782 assert(space_id < last_space_id, "not enough spaces");
2783
2784 HeapWord* const destination = closure.destination();
2785
2786 do {
2787 MutableSpace* space = _space_info[space_id].space();
2788 HeapWord* const bottom = space->bottom();
2789 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2790
2791 // Iterate over the spaces that do not compact into themselves.
2792 if (bottom_cp->destination() != bottom) {
2793 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2794 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2795
2796 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2797 if (src_cp->live_obj_size() > 0) {
2798 // Found it.
2799 assert(src_cp->destination() == destination,
2800 "first live obj in the space must match the destination");
2801 assert(src_cp->partial_obj_size() == 0,
2802 "a space cannot begin with a partial obj");
2803
2804 src_space_id = SpaceId(space_id);
2805 src_space_top = space->top();
2806 const size_t src_region_idx = sd.region(src_cp);
2807 closure.set_source(sd.region_to_addr(src_region_idx));
2808 return src_region_idx;
2809 } else {
2810 assert(src_cp->data_size() == 0, "sanity");
2811 }
2812 }
2813 }
2814 } while (++space_id < last_space_id);
2815
2816 assert(false, "no source region was found");
2817 return 0;
2818 }
2819
2820 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
2821 {
2822 typedef ParMarkBitMap::IterationStatus IterationStatus;
2823 const size_t RegionSize = ParallelCompactData::RegionSize;
2824 ParMarkBitMap* const bitmap = mark_bitmap();
2825 ParallelCompactData& sd = summary_data();
2826 RegionData* const region_ptr = sd.region(region_idx);
2827
2828 // Get the items needed to construct the closure.
2829 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2830 SpaceId dest_space_id = space_id(dest_addr);
2831 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
2832 HeapWord* new_top = _space_info[dest_space_id].new_top();
2833 assert(dest_addr < new_top, "sanity");
2834 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
2835
2836 // Get the source region and related info.
2837 size_t src_region_idx = region_ptr->source_region();
2838 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2839 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2840
2841 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2842 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2843
2844 // Adjust src_region_idx to prepare for decrementing destination counts (the
2845 // destination count is not decremented when a region is copied to itself).
2846 if (src_region_idx == region_idx) {
2847 src_region_idx += 1;
2848 }
2849
2850 if (bitmap->is_unmarked(closure.source())) {
2851 // The first source word is in the middle of an object; copy the remainder
2852 // of the object or as much as will fit. The fact that pointer updates were
2853 // deferred will be noted when the object header is processed.
2854 HeapWord* const old_src_addr = closure.source();
2855 closure.copy_partial_obj();
2856 if (closure.is_full()) {
2857 decrement_destination_counts(cm, src_space_id, src_region_idx,
2858 closure.source());
2859 region_ptr->set_deferred_obj_addr(NULL);
2860 region_ptr->set_completed();
2861 return;
2862 }
2863
2864 HeapWord* const end_addr = sd.region_align_down(closure.source());
2865 if (sd.region_align_down(old_src_addr) != end_addr) {
2866 // The partial object was copied from more than one source region.
2867 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2868
2869 // Move to the next source region, possibly switching spaces as well. All
2870 // args except end_addr may be modified.
2871 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2872 end_addr);
2873 }
2874 }
2875
2876 do {
2877 HeapWord* const cur_addr = closure.source();
2878 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2879 src_space_top);
2880 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2881
2882 if (status == ParMarkBitMap::incomplete) {
2883 // The last obj that starts in the source region does not end in the
2884 // region.
2885 assert(closure.source() < end_addr, "sanity");
2886 HeapWord* const obj_beg = closure.source();
2887 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2888 src_space_top);
2889 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2890 if (obj_end < range_end) {
2891 // The end was found; the entire object will fit.
2892 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2893 assert(status != ParMarkBitMap::would_overflow, "sanity");
2894 } else {
2895 // The end was not found; the object will not fit.
2896 assert(range_end < src_space_top, "obj cannot cross space boundary");
2897 status = ParMarkBitMap::would_overflow;
2898 }
2899 }
2900
2901 if (status == ParMarkBitMap::would_overflow) {
2902 // The last object did not fit. Note that interior oop updates were
2903 // deferred, then copy enough of the object to fill the region.
2904 region_ptr->set_deferred_obj_addr(closure.destination());
2905 status = closure.copy_until_full(); // copies from closure.source()
2906
2907 decrement_destination_counts(cm, src_space_id, src_region_idx,
2908 closure.source());
2909 region_ptr->set_completed();
2910 return;
2911 }
2912
2913 if (status == ParMarkBitMap::full) {
2914 decrement_destination_counts(cm, src_space_id, src_region_idx,
2915 closure.source());
2916 region_ptr->set_deferred_obj_addr(NULL);
2917 region_ptr->set_completed();
2918 return;
2919 }
2920
2921 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2922
2923 // Move to the next source region, possibly switching spaces as well. All
2924 // args except end_addr may be modified.
2925 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2926 end_addr);
2927 } while (true);
2928 }
2929
2930 void PSParallelCompact::fill_blocks(size_t region_idx)
2931 {
2932 // Fill in the block table elements for the specified region. Each block
2933 // table element holds the number of live words in the region that are to the
2934 // left of the first object that starts in the block. Thus only blocks in
2935 // which an object starts need to be filled.
2936 //
2937 // The algorithm scans the section of the bitmap that corresponds to the
2938 // region, keeping a running total of the live words. When an object start is
2939 // found, if it's the first to start in the block that contains it, the
2940 // current total is written to the block table element.
2941 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
2942 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
2943 const size_t RegionSize = ParallelCompactData::RegionSize;
2944
2945 ParallelCompactData& sd = summary_data();
2946 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
2947 if (partial_obj_size >= RegionSize) {
2948 return; // No objects start in this region.
2949 }
2950
2951 // Ensure the first loop iteration decides that the block has changed.
2952 size_t cur_block = sd.block_count();
2953
2954 const ParMarkBitMap* const bitmap = mark_bitmap();
2955
2956 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
2957 assert((size_t)1 << Log2BitsPerBlock ==
2958 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
2959
2960 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
2961 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
2962 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
2963 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
2964 while (beg_bit < range_end) {
2965 const size_t new_block = beg_bit >> Log2BitsPerBlock;
2966 if (new_block != cur_block) {
2967 cur_block = new_block;
2968 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
2969 }
2970
2971 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
2972 if (end_bit < range_end - 1) {
2973 live_bits += end_bit - beg_bit + 1;
2974 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
2975 } else {
2976 return;
2977 }
2978 }
2979 }
2980
2981 void
2982 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
2983 const MutableSpace* sp = space(space_id);
2984 if (sp->is_empty()) {
2985 return;
2986 }
2987
2988 ParallelCompactData& sd = PSParallelCompact::summary_data();
2989 ParMarkBitMap* const bitmap = mark_bitmap();
2990 HeapWord* const dp_addr = dense_prefix(space_id);
2991 HeapWord* beg_addr = sp->bottom();
2992 HeapWord* end_addr = sp->top();
2993
2994 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
2995
2996 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
2997 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
2998 if (beg_region < dp_region) {
2999 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3000 }
3001
3002 // The destination of the first live object that starts in the region is one
3003 // past the end of the partial object entering the region (if any).
3004 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3005 HeapWord* const new_top = _space_info[space_id].new_top();
3006 assert(new_top >= dest_addr, "bad new_top value");
3007 const size_t words = pointer_delta(new_top, dest_addr);
3008
3009 if (words > 0) {
3010 ObjectStartArray* start_array = _space_info[space_id].start_array();
3011 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3012
3013 ParMarkBitMap::IterationStatus status;
3014 status = bitmap->iterate(&closure, dest_addr, end_addr);
3015 assert(status == ParMarkBitMap::full, "iteration not complete");
3016 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3017 "live objects skipped because closure is full");
3018 }
3019 }
3020
3021 jlong PSParallelCompact::millis_since_last_gc() {
3022 // We need a monotonically non-decreasing time in ms but
3023 // os::javaTimeMillis() does not guarantee monotonicity.
3024 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3025 jlong ret_val = now - _time_of_last_gc;
3026 // XXX See note in genCollectedHeap::millis_since_last_gc().
3027 if (ret_val < 0) {
3028 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);)
3029 return 0;
3030 }
3031 return ret_val;
3032 }
3033
3034 void PSParallelCompact::reset_millis_since_last_gc() {
3035 // We need a monotonically non-decreasing time in ms but
3036 // os::javaTimeMillis() does not guarantee monotonicity.
3037 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3038 }
3039
3040 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3041 {
3042 if (source() != destination()) {
3043 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3044 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3045 }
3046 update_state(words_remaining());
3047 assert(is_full(), "sanity");
3048 return ParMarkBitMap::full;
3049 }
3050
3051 void MoveAndUpdateClosure::copy_partial_obj()
3052 {
3053 size_t words = words_remaining();
3054
3055 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3056 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3057 if (end_addr < range_end) {
3058 words = bitmap()->obj_size(source(), end_addr);
3059 }
3060
3061 // This test is necessary; if omitted, the pointer updates to a partial object
3062 // that crosses the dense prefix boundary could be overwritten.
3063 if (source() != destination()) {
3064 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3065 Copy::aligned_conjoint_words(source(), destination(), words);
3066 }
3067 update_state(words);
3068 }
3069
3070 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3071 PSParallelCompact::AdjustPointerClosure closure(cm);
3072 oop_oop_iterate_oop_maps<true>(obj, &closure);
3073 }
3074
3075 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3076 InstanceKlass::oop_pc_update_pointers(obj, cm);
3077
3078 PSParallelCompact::AdjustPointerClosure closure(cm);
3079 oop_oop_iterate_statics<true>(obj, &closure);
3080 }
3081
3082 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3083 InstanceKlass::oop_pc_update_pointers(obj, cm);
3084 }
3085
3086 #ifdef ASSERT
3087 template <class T> static void trace_reference_gc(const char *s, oop obj,
3088 T* referent_addr,
3089 T* next_addr,
3090 T* discovered_addr) {
3091 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj));
3092 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3093 p2i(referent_addr), referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL);
3094 log_develop_trace(gc, ref)(" next_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3095 p2i(next_addr), next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL);
3096 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT,
3097 p2i(discovered_addr), discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL);
3098 }
3099 #endif
3100
3101 template <class T>
3102 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) {
3103 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj);
3104 PSParallelCompact::adjust_pointer(referent_addr, cm);
3105 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj);
3106 PSParallelCompact::adjust_pointer(next_addr, cm);
3107 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj);
3108 PSParallelCompact::adjust_pointer(discovered_addr, cm);
3109 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj,
3110 referent_addr, next_addr, discovered_addr);)
3111 }
3112
3113 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3114 InstanceKlass::oop_pc_update_pointers(obj, cm);
3115
3116 if (UseCompressedOops) {
3117 oop_pc_update_pointers_specialized<narrowOop>(obj, cm);
3118 } else {
3119 oop_pc_update_pointers_specialized<oop>(obj, cm);
3120 }
3121 }
3122
3123 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3124 assert(obj->is_objArray(), "obj must be obj array");
3125 PSParallelCompact::AdjustPointerClosure closure(cm);
3126 oop_oop_iterate_elements<true>(objArrayOop(obj), &closure);
3127 }
3128
3129 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3130 assert(obj->is_typeArray(),"must be a type array");
3131 }
3132
3133 void ValueArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) {
3134 assert(obj->is_valueArray(),"must be a value array");
3135 if (contains_oops()) {
3136 Unimplemented();
3137 #if 0
3138 PSParallelCompact::AdjustPointerClosure closure(cm);
3139 oop_oop_iterate_elements<true>(valueArrayOop(obj), &closure);
3140 #endif
3141 }
3142 }
3143
3144 ParMarkBitMapClosure::IterationStatus
3145 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3146 assert(destination() != NULL, "sanity");
3147 assert(bitmap()->obj_size(addr) == words, "bad size");
3148
3149 _source = addr;
3150 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3151 destination(), "wrong destination");
3152
3153 if (words > words_remaining()) {
3154 return ParMarkBitMap::would_overflow;
3155 }
3156
3157 // The start_array must be updated even if the object is not moving.
3158 if (_start_array != NULL) {
3159 _start_array->allocate_block(destination());
3160 }
3161
3162 if (destination() != source()) {
3163 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3164 Copy::aligned_conjoint_words(source(), destination(), words);
3165 }
3166
3167 oop moved_oop = (oop) destination();
3168 compaction_manager()->update_contents(moved_oop);
3169 assert(moved_oop->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3170
3171 update_state(words);
3172 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3173 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3174 }
3175
3176 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3177 ParCompactionManager* cm,
3178 PSParallelCompact::SpaceId space_id) :
3179 ParMarkBitMapClosure(mbm, cm),
3180 _space_id(space_id),
3181 _start_array(PSParallelCompact::start_array(space_id))
3182 {
3183 }
3184
3185 // Updates the references in the object to their new values.
3186 ParMarkBitMapClosure::IterationStatus
3187 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3188 do_addr(addr);
3189 return ParMarkBitMap::incomplete;
3190 }
3191
3192 ParMarkBitMapClosure::IterationStatus
3193 FillClosure::do_addr(HeapWord* addr, size_t size) {
3194 CollectedHeap::fill_with_objects(addr, size);
3195 HeapWord* const end = addr + size;
3196 do {
3197 _start_array->allocate_block(addr);
3198 addr += oop(addr)->size();
3199 } while (addr < end);
3200 return ParMarkBitMap::incomplete;
3201 }