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