1 /*
2 * Copyright (c) 2005, 2019, 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/classLoaderDataGraph.hpp"
28 #include "classfile/javaClasses.inline.hpp"
29 #include "classfile/stringTable.hpp"
30 #include "classfile/symbolTable.hpp"
31 #include "classfile/systemDictionary.hpp"
32 #include "code/codeCache.hpp"
33 #include "gc/parallel/parallelArguments.hpp"
34 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
35 #include "gc/parallel/parMarkBitMap.inline.hpp"
36 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
37 #include "gc/parallel/psCompactionManager.inline.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/psRootType.hpp"
42 #include "gc/parallel/psScavenge.hpp"
43 #include "gc/parallel/psYoungGen.hpp"
44 #include "gc/shared/gcCause.hpp"
45 #include "gc/shared/gcHeapSummary.hpp"
46 #include "gc/shared/gcId.hpp"
47 #include "gc/shared/gcLocker.hpp"
48 #include "gc/shared/gcTimer.hpp"
49 #include "gc/shared/gcTrace.hpp"
50 #include "gc/shared/gcTraceTime.inline.hpp"
51 #include "gc/shared/isGCActiveMark.hpp"
52 #include "gc/shared/referencePolicy.hpp"
53 #include "gc/shared/referenceProcessor.hpp"
54 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
55 #include "gc/shared/spaceDecorator.inline.hpp"
56 #include "gc/shared/weakProcessor.hpp"
57 #include "gc/shared/workerPolicy.hpp"
58 #include "gc/shared/workgroup.hpp"
59 #include "logging/log.hpp"
60 #include "memory/iterator.inline.hpp"
61 #include "memory/resourceArea.hpp"
62 #include "memory/universe.hpp"
63 #include "oops/access.inline.hpp"
64 #include "oops/instanceClassLoaderKlass.inline.hpp"
65 #include "oops/instanceKlass.inline.hpp"
66 #include "oops/instanceMirrorKlass.inline.hpp"
67 #include "oops/methodData.hpp"
68 #include "oops/objArrayKlass.inline.hpp"
69 #include "oops/oop.inline.hpp"
70 #include "runtime/atomic.hpp"
71 #include "runtime/handles.inline.hpp"
72 #include "runtime/safepoint.hpp"
73 #include "runtime/vmThread.hpp"
74 #include "services/management.hpp"
75 #include "services/memTracker.hpp"
76 #include "services/memoryService.hpp"
77 #include "utilities/align.hpp"
78 #include "utilities/debug.hpp"
79 #include "utilities/events.hpp"
80 #include "utilities/formatBuffer.hpp"
81 #include "utilities/macros.hpp"
82 #include "utilities/stack.inline.hpp"
83 #if INCLUDE_JVMCI
84 #include "jvmci/jvmci.hpp"
85 #endif
86
87 #include <math.h>
88
89 // All sizes are in HeapWords.
90 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
91 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
92 const size_t ParallelCompactData::RegionSizeBytes =
93 RegionSize << LogHeapWordSize;
94 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
95 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
96 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
97
98 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
99 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
100 const size_t ParallelCompactData::BlockSizeBytes =
101 BlockSize << LogHeapWordSize;
102 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
103 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
104 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
105
106 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
107 const size_t ParallelCompactData::Log2BlocksPerRegion =
108 Log2RegionSize - Log2BlockSize;
109
110 const ParallelCompactData::RegionData::region_sz_t
111 ParallelCompactData::RegionData::dc_shift = 27;
112
113 const ParallelCompactData::RegionData::region_sz_t
114 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
115
116 const ParallelCompactData::RegionData::region_sz_t
117 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
118
119 const ParallelCompactData::RegionData::region_sz_t
120 ParallelCompactData::RegionData::los_mask = ~dc_mask;
121
122 const ParallelCompactData::RegionData::region_sz_t
123 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
124
125 const ParallelCompactData::RegionData::region_sz_t
126 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
127
128 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
129
130 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
131 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
132
133 double PSParallelCompact::_dwl_mean;
134 double PSParallelCompact::_dwl_std_dev;
135 double PSParallelCompact::_dwl_first_term;
136 double PSParallelCompact::_dwl_adjustment;
137 #ifdef ASSERT
138 bool PSParallelCompact::_dwl_initialized = false;
139 #endif // #ifdef ASSERT
140
record(size_t src_region_idx,size_t partial_obj_size,HeapWord * destination)141 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
142 HeapWord* destination)
143 {
144 assert(src_region_idx != 0, "invalid src_region_idx");
145 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
146 assert(destination != NULL, "invalid destination argument");
147
148 _src_region_idx = src_region_idx;
149 _partial_obj_size = partial_obj_size;
150 _destination = destination;
151
152 // These fields may not be updated below, so make sure they're clear.
153 assert(_dest_region_addr == NULL, "should have been cleared");
154 assert(_first_src_addr == NULL, "should have been cleared");
155
156 // Determine the number of destination regions for the partial object.
157 HeapWord* const last_word = destination + partial_obj_size - 1;
158 const ParallelCompactData& sd = PSParallelCompact::summary_data();
159 HeapWord* const beg_region_addr = sd.region_align_down(destination);
160 HeapWord* const end_region_addr = sd.region_align_down(last_word);
161
162 if (beg_region_addr == end_region_addr) {
163 // One destination region.
164 _destination_count = 1;
165 if (end_region_addr == destination) {
166 // The destination falls on a region boundary, thus the first word of the
167 // partial object will be the first word copied to the destination region.
168 _dest_region_addr = end_region_addr;
169 _first_src_addr = sd.region_to_addr(src_region_idx);
170 }
171 } else {
172 // Two destination regions. When copied, the partial object will cross a
173 // destination region boundary, so a word somewhere within the partial
174 // object will be the first word copied to the second destination region.
175 _destination_count = 2;
176 _dest_region_addr = end_region_addr;
177 const size_t ofs = pointer_delta(end_region_addr, destination);
178 assert(ofs < _partial_obj_size, "sanity");
179 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
180 }
181 }
182
clear()183 void SplitInfo::clear()
184 {
185 _src_region_idx = 0;
186 _partial_obj_size = 0;
187 _destination = NULL;
188 _destination_count = 0;
189 _dest_region_addr = NULL;
190 _first_src_addr = NULL;
191 assert(!is_valid(), "sanity");
192 }
193
194 #ifdef ASSERT
verify_clear()195 void SplitInfo::verify_clear()
196 {
197 assert(_src_region_idx == 0, "not clear");
198 assert(_partial_obj_size == 0, "not clear");
199 assert(_destination == NULL, "not clear");
200 assert(_destination_count == 0, "not clear");
201 assert(_dest_region_addr == NULL, "not clear");
202 assert(_first_src_addr == NULL, "not clear");
203 }
204 #endif // #ifdef ASSERT
205
206
print_on_error(outputStream * st)207 void PSParallelCompact::print_on_error(outputStream* st) {
208 _mark_bitmap.print_on_error(st);
209 }
210
211 #ifndef PRODUCT
212 const char* PSParallelCompact::space_names[] = {
213 "old ", "eden", "from", "to "
214 };
215
print_region_ranges()216 void PSParallelCompact::print_region_ranges() {
217 if (!log_develop_is_enabled(Trace, gc, compaction)) {
218 return;
219 }
220 Log(gc, compaction) log;
221 ResourceMark rm;
222 LogStream ls(log.trace());
223 Universe::print_on(&ls);
224 log.trace("space bottom top end new_top");
225 log.trace("------ ---------- ---------- ---------- ----------");
226
227 for (unsigned int id = 0; id < last_space_id; ++id) {
228 const MutableSpace* space = _space_info[id].space();
229 log.trace("%u %s "
230 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
231 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
232 id, space_names[id],
233 summary_data().addr_to_region_idx(space->bottom()),
234 summary_data().addr_to_region_idx(space->top()),
235 summary_data().addr_to_region_idx(space->end()),
236 summary_data().addr_to_region_idx(_space_info[id].new_top()));
237 }
238 }
239
240 void
print_generic_summary_region(size_t i,const ParallelCompactData::RegionData * c)241 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
242 {
243 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
244 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
245
246 ParallelCompactData& sd = PSParallelCompact::summary_data();
247 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
248 log_develop_trace(gc, compaction)(
249 REGION_IDX_FORMAT " " PTR_FORMAT " "
250 REGION_IDX_FORMAT " " PTR_FORMAT " "
251 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
252 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
253 i, p2i(c->data_location()), dci, p2i(c->destination()),
254 c->partial_obj_size(), c->live_obj_size(),
255 c->data_size(), c->source_region(), c->destination_count());
256
257 #undef REGION_IDX_FORMAT
258 #undef REGION_DATA_FORMAT
259 }
260
261 void
print_generic_summary_data(ParallelCompactData & summary_data,HeapWord * const beg_addr,HeapWord * const end_addr)262 print_generic_summary_data(ParallelCompactData& summary_data,
263 HeapWord* const beg_addr,
264 HeapWord* const end_addr)
265 {
266 size_t total_words = 0;
267 size_t i = summary_data.addr_to_region_idx(beg_addr);
268 const size_t last = summary_data.addr_to_region_idx(end_addr);
269 HeapWord* pdest = 0;
270
271 while (i < last) {
272 ParallelCompactData::RegionData* c = summary_data.region(i);
273 if (c->data_size() != 0 || c->destination() != pdest) {
274 print_generic_summary_region(i, c);
275 total_words += c->data_size();
276 pdest = c->destination();
277 }
278 ++i;
279 }
280
281 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
282 }
283
284 void
print_generic_summary_data(ParallelCompactData & summary_data,HeapWord * const beg_addr,HeapWord * const end_addr)285 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
286 HeapWord* const beg_addr,
287 HeapWord* const end_addr) {
288 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
289 }
290
291 void
print_generic_summary_data(ParallelCompactData & summary_data,SpaceInfo * space_info)292 print_generic_summary_data(ParallelCompactData& summary_data,
293 SpaceInfo* space_info)
294 {
295 if (!log_develop_is_enabled(Trace, gc, compaction)) {
296 return;
297 }
298
299 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
300 const MutableSpace* space = space_info[id].space();
301 print_generic_summary_data(summary_data, space->bottom(),
302 MAX2(space->top(), space_info[id].new_top()));
303 }
304 }
305
306 void
print_initial_summary_data(ParallelCompactData & summary_data,const MutableSpace * space)307 print_initial_summary_data(ParallelCompactData& summary_data,
308 const MutableSpace* space) {
309 if (space->top() == space->bottom()) {
310 return;
311 }
312
313 const size_t region_size = ParallelCompactData::RegionSize;
314 typedef ParallelCompactData::RegionData RegionData;
315 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
316 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
317 const RegionData* c = summary_data.region(end_region - 1);
318 HeapWord* end_addr = c->destination() + c->data_size();
319 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
320
321 // Print (and count) the full regions at the beginning of the space.
322 size_t full_region_count = 0;
323 size_t i = summary_data.addr_to_region_idx(space->bottom());
324 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
325 ParallelCompactData::RegionData* c = summary_data.region(i);
326 log_develop_trace(gc, compaction)(
327 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
328 i, p2i(c->destination()),
329 c->partial_obj_size(), c->live_obj_size(),
330 c->data_size(), c->source_region(), c->destination_count());
331 ++full_region_count;
332 ++i;
333 }
334
335 size_t live_to_right = live_in_space - full_region_count * region_size;
336
337 double max_reclaimed_ratio = 0.0;
338 size_t max_reclaimed_ratio_region = 0;
339 size_t max_dead_to_right = 0;
340 size_t max_live_to_right = 0;
341
342 // Print the 'reclaimed ratio' for regions while there is something live in
343 // the region or to the right of it. The remaining regions are empty (and
344 // uninteresting), and computing the ratio will result in division by 0.
345 while (i < end_region && live_to_right > 0) {
346 c = summary_data.region(i);
347 HeapWord* const region_addr = summary_data.region_to_addr(i);
348 const size_t used_to_right = pointer_delta(space->top(), region_addr);
349 const size_t dead_to_right = used_to_right - live_to_right;
350 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
351
352 if (reclaimed_ratio > max_reclaimed_ratio) {
353 max_reclaimed_ratio = reclaimed_ratio;
354 max_reclaimed_ratio_region = i;
355 max_dead_to_right = dead_to_right;
356 max_live_to_right = live_to_right;
357 }
358
359 ParallelCompactData::RegionData* c = summary_data.region(i);
360 log_develop_trace(gc, compaction)(
361 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
362 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
363 i, p2i(c->destination()),
364 c->partial_obj_size(), c->live_obj_size(),
365 c->data_size(), c->source_region(), c->destination_count(),
366 reclaimed_ratio, dead_to_right, live_to_right);
367
368
369 live_to_right -= c->data_size();
370 ++i;
371 }
372
373 // Any remaining regions are empty. Print one more if there is one.
374 if (i < end_region) {
375 ParallelCompactData::RegionData* c = summary_data.region(i);
376 log_develop_trace(gc, compaction)(
377 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
378 i, p2i(c->destination()),
379 c->partial_obj_size(), c->live_obj_size(),
380 c->data_size(), c->source_region(), c->destination_count());
381 }
382
383 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
384 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
385 }
386
387 void
print_initial_summary_data(ParallelCompactData & summary_data,SpaceInfo * space_info)388 print_initial_summary_data(ParallelCompactData& summary_data,
389 SpaceInfo* space_info) {
390 if (!log_develop_is_enabled(Trace, gc, compaction)) {
391 return;
392 }
393
394 unsigned int id = PSParallelCompact::old_space_id;
395 const MutableSpace* space;
396 do {
397 space = space_info[id].space();
398 print_initial_summary_data(summary_data, space);
399 } while (++id < PSParallelCompact::eden_space_id);
400
401 do {
402 space = space_info[id].space();
403 print_generic_summary_data(summary_data, space->bottom(), space->top());
404 } while (++id < PSParallelCompact::last_space_id);
405 }
406 #endif // #ifndef PRODUCT
407
408 #ifdef ASSERT
409 size_t add_obj_count;
410 size_t add_obj_size;
411 size_t mark_bitmap_count;
412 size_t mark_bitmap_size;
413 #endif // #ifdef ASSERT
414
ParallelCompactData()415 ParallelCompactData::ParallelCompactData() :
416 _region_start(NULL),
417 DEBUG_ONLY(_region_end(NULL) COMMA)
418 _region_vspace(NULL),
419 _reserved_byte_size(0),
420 _region_data(NULL),
421 _region_count(0),
422 _block_vspace(NULL),
423 _block_data(NULL),
424 _block_count(0) {}
425
initialize(MemRegion covered_region)426 bool ParallelCompactData::initialize(MemRegion covered_region)
427 {
428 _region_start = covered_region.start();
429 const size_t region_size = covered_region.word_size();
430 DEBUG_ONLY(_region_end = _region_start + region_size;)
431
432 assert(region_align_down(_region_start) == _region_start,
433 "region start not aligned");
434 assert((region_size & RegionSizeOffsetMask) == 0,
435 "region size not a multiple of RegionSize");
436
437 bool result = initialize_region_data(region_size) && initialize_block_data();
438 return result;
439 }
440
441 PSVirtualSpace*
create_vspace(size_t count,size_t element_size)442 ParallelCompactData::create_vspace(size_t count, size_t element_size)
443 {
444 const size_t raw_bytes = count * element_size;
445 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
446 const size_t granularity = os::vm_allocation_granularity();
447 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
448
449 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
450 MAX2(page_sz, granularity);
451 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
452 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
453 rs.size());
454
455 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
456
457 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
458 if (vspace != 0) {
459 if (vspace->expand_by(_reserved_byte_size)) {
460 return vspace;
461 }
462 delete vspace;
463 // Release memory reserved in the space.
464 rs.release();
465 }
466
467 return 0;
468 }
469
initialize_region_data(size_t region_size)470 bool ParallelCompactData::initialize_region_data(size_t region_size)
471 {
472 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
473 _region_vspace = create_vspace(count, sizeof(RegionData));
474 if (_region_vspace != 0) {
475 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
476 _region_count = count;
477 return true;
478 }
479 return false;
480 }
481
initialize_block_data()482 bool ParallelCompactData::initialize_block_data()
483 {
484 assert(_region_count != 0, "region data must be initialized first");
485 const size_t count = _region_count << Log2BlocksPerRegion;
486 _block_vspace = create_vspace(count, sizeof(BlockData));
487 if (_block_vspace != 0) {
488 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
489 _block_count = count;
490 return true;
491 }
492 return false;
493 }
494
clear()495 void ParallelCompactData::clear()
496 {
497 memset(_region_data, 0, _region_vspace->committed_size());
498 memset(_block_data, 0, _block_vspace->committed_size());
499 }
500
clear_range(size_t beg_region,size_t end_region)501 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
502 assert(beg_region <= _region_count, "beg_region out of range");
503 assert(end_region <= _region_count, "end_region out of range");
504 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
505
506 const size_t region_cnt = end_region - beg_region;
507 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
508
509 const size_t beg_block = beg_region * BlocksPerRegion;
510 const size_t block_cnt = region_cnt * BlocksPerRegion;
511 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
512 }
513
partial_obj_end(size_t region_idx) const514 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
515 {
516 const RegionData* cur_cp = region(region_idx);
517 const RegionData* const end_cp = region(region_count() - 1);
518
519 HeapWord* result = region_to_addr(region_idx);
520 if (cur_cp < end_cp) {
521 do {
522 result += cur_cp->partial_obj_size();
523 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
524 }
525 return result;
526 }
527
add_obj(HeapWord * addr,size_t len)528 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
529 {
530 const size_t obj_ofs = pointer_delta(addr, _region_start);
531 const size_t beg_region = obj_ofs >> Log2RegionSize;
532 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
533
534 DEBUG_ONLY(Atomic::inc(&add_obj_count);)
535 DEBUG_ONLY(Atomic::add(&add_obj_size, len);)
536
537 if (beg_region == end_region) {
538 // All in one region.
539 _region_data[beg_region].add_live_obj(len);
540 return;
541 }
542
543 // First region.
544 const size_t beg_ofs = region_offset(addr);
545 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
546
547 Klass* klass = ((oop)addr)->klass();
548 // Middle regions--completely spanned by this object.
549 for (size_t region = beg_region + 1; region < end_region; ++region) {
550 _region_data[region].set_partial_obj_size(RegionSize);
551 _region_data[region].set_partial_obj_addr(addr);
552 }
553
554 // Last region.
555 const size_t end_ofs = region_offset(addr + len - 1);
556 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
557 _region_data[end_region].set_partial_obj_addr(addr);
558 }
559
560 void
summarize_dense_prefix(HeapWord * beg,HeapWord * end)561 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
562 {
563 assert(region_offset(beg) == 0, "not RegionSize aligned");
564 assert(region_offset(end) == 0, "not RegionSize aligned");
565
566 size_t cur_region = addr_to_region_idx(beg);
567 const size_t end_region = addr_to_region_idx(end);
568 HeapWord* addr = beg;
569 while (cur_region < end_region) {
570 _region_data[cur_region].set_destination(addr);
571 _region_data[cur_region].set_destination_count(0);
572 _region_data[cur_region].set_source_region(cur_region);
573 _region_data[cur_region].set_data_location(addr);
574
575 // Update live_obj_size so the region appears completely full.
576 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
577 _region_data[cur_region].set_live_obj_size(live_size);
578
579 ++cur_region;
580 addr += RegionSize;
581 }
582 }
583
584 // Find the point at which a space can be split and, if necessary, record the
585 // split point.
586 //
587 // If the current src region (which overflowed the destination space) doesn't
588 // have a partial object, the split point is at the beginning of the current src
589 // region (an "easy" split, no extra bookkeeping required).
590 //
591 // If the current src region has a partial object, the split point is in the
592 // region where that partial object starts (call it the split_region). If
593 // split_region has a partial object, then the split point is just after that
594 // partial object (a "hard" split where we have to record the split data and
595 // zero the partial_obj_size field). With a "hard" split, we know that the
596 // partial_obj ends within split_region because the partial object that caused
597 // the overflow starts in split_region. If split_region doesn't have a partial
598 // obj, then the split is at the beginning of split_region (another "easy"
599 // split).
600 HeapWord*
summarize_split_space(size_t src_region,SplitInfo & split_info,HeapWord * destination,HeapWord * target_end,HeapWord ** target_next)601 ParallelCompactData::summarize_split_space(size_t src_region,
602 SplitInfo& split_info,
603 HeapWord* destination,
604 HeapWord* target_end,
605 HeapWord** target_next)
606 {
607 assert(destination <= target_end, "sanity");
608 assert(destination + _region_data[src_region].data_size() > target_end,
609 "region should not fit into target space");
610 assert(is_region_aligned(target_end), "sanity");
611
612 size_t split_region = src_region;
613 HeapWord* split_destination = destination;
614 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
615
616 if (destination + partial_obj_size > target_end) {
617 // The split point is just after the partial object (if any) in the
618 // src_region that contains the start of the object that overflowed the
619 // destination space.
620 //
621 // Find the start of the "overflow" object and set split_region to the
622 // region containing it.
623 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
624 split_region = addr_to_region_idx(overflow_obj);
625
626 // Clear the source_region field of all destination regions whose first word
627 // came from data after the split point (a non-null source_region field
628 // implies a region must be filled).
629 //
630 // An alternative to the simple loop below: clear during post_compact(),
631 // which uses memcpy instead of individual stores, and is easy to
632 // parallelize. (The downside is that it clears the entire RegionData
633 // object as opposed to just one field.)
634 //
635 // post_compact() would have to clear the summary data up to the highest
636 // address that was written during the summary phase, which would be
637 //
638 // max(top, max(new_top, clear_top))
639 //
640 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
641 // to target_end.
642 const RegionData* const sr = region(split_region);
643 const size_t beg_idx =
644 addr_to_region_idx(region_align_up(sr->destination() +
645 sr->partial_obj_size()));
646 const size_t end_idx = addr_to_region_idx(target_end);
647
648 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
649 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
650 _region_data[idx].set_source_region(0);
651 }
652
653 // Set split_destination and partial_obj_size to reflect the split region.
654 split_destination = sr->destination();
655 partial_obj_size = sr->partial_obj_size();
656 }
657
658 // The split is recorded only if a partial object extends onto the region.
659 if (partial_obj_size != 0) {
660 _region_data[split_region].set_partial_obj_size(0);
661 split_info.record(split_region, partial_obj_size, split_destination);
662 }
663
664 // Setup the continuation addresses.
665 *target_next = split_destination + partial_obj_size;
666 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
667
668 if (log_develop_is_enabled(Trace, gc, compaction)) {
669 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
670 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
671 split_type, p2i(source_next), split_region, partial_obj_size);
672 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
673 split_type, p2i(split_destination),
674 addr_to_region_idx(split_destination),
675 p2i(*target_next));
676
677 if (partial_obj_size != 0) {
678 HeapWord* const po_beg = split_info.destination();
679 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
680 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
681 split_type,
682 p2i(po_beg), addr_to_region_idx(po_beg),
683 p2i(po_end), addr_to_region_idx(po_end));
684 }
685 }
686
687 return source_next;
688 }
689
summarize(SplitInfo & split_info,HeapWord * source_beg,HeapWord * source_end,HeapWord ** source_next,HeapWord * target_beg,HeapWord * target_end,HeapWord ** target_next)690 bool ParallelCompactData::summarize(SplitInfo& split_info,
691 HeapWord* source_beg, HeapWord* source_end,
692 HeapWord** source_next,
693 HeapWord* target_beg, HeapWord* target_end,
694 HeapWord** target_next)
695 {
696 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
697 log_develop_trace(gc, compaction)(
698 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
699 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
700 p2i(source_beg), p2i(source_end), p2i(source_next_val),
701 p2i(target_beg), p2i(target_end), p2i(*target_next));
702
703 size_t cur_region = addr_to_region_idx(source_beg);
704 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
705
706 HeapWord *dest_addr = target_beg;
707 while (cur_region < end_region) {
708 // The destination must be set even if the region has no data.
709 _region_data[cur_region].set_destination(dest_addr);
710
711 size_t words = _region_data[cur_region].data_size();
712 if (words > 0) {
713 // If cur_region does not fit entirely into the target space, find a point
714 // at which the source space can be 'split' so that part is copied to the
715 // target space and the rest is copied elsewhere.
716 if (dest_addr + words > target_end) {
717 assert(source_next != NULL, "source_next is NULL when splitting");
718 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
719 target_end, target_next);
720 return false;
721 }
722
723 // Compute the destination_count for cur_region, and if necessary, update
724 // source_region for a destination region. The source_region field is
725 // updated if cur_region is the first (left-most) region to be copied to a
726 // destination region.
727 //
728 // The destination_count calculation is a bit subtle. A region that has
729 // data that compacts into itself does not count itself as a destination.
730 // This maintains the invariant that a zero count means the region is
731 // available and can be claimed and then filled.
732 uint destination_count = 0;
733 if (split_info.is_split(cur_region)) {
734 // The current region has been split: the partial object will be copied
735 // to one destination space and the remaining data will be copied to
736 // another destination space. Adjust the initial destination_count and,
737 // if necessary, set the source_region field if the partial object will
738 // cross a destination region boundary.
739 destination_count = split_info.destination_count();
740 if (destination_count == 2) {
741 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
742 _region_data[dest_idx].set_source_region(cur_region);
743 }
744 }
745
746 HeapWord* const last_addr = dest_addr + words - 1;
747 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
748 const size_t dest_region_2 = addr_to_region_idx(last_addr);
749
750 // Initially assume that the destination regions will be the same and
751 // adjust the value below if necessary. Under this assumption, if
752 // cur_region == dest_region_2, then cur_region will be compacted
753 // completely into itself.
754 destination_count += cur_region == dest_region_2 ? 0 : 1;
755 if (dest_region_1 != dest_region_2) {
756 // Destination regions differ; adjust destination_count.
757 destination_count += 1;
758 // Data from cur_region will be copied to the start of dest_region_2.
759 _region_data[dest_region_2].set_source_region(cur_region);
760 } else if (region_offset(dest_addr) == 0) {
761 // Data from cur_region will be copied to the start of the destination
762 // region.
763 _region_data[dest_region_1].set_source_region(cur_region);
764 }
765
766 _region_data[cur_region].set_destination_count(destination_count);
767 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
768 dest_addr += words;
769 }
770
771 ++cur_region;
772 }
773
774 *target_next = dest_addr;
775 return true;
776 }
777
calc_new_pointer(HeapWord * addr,ParCompactionManager * cm)778 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
779 assert(addr != NULL, "Should detect NULL oop earlier");
780 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
781 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
782
783 // Region covering the object.
784 RegionData* const region_ptr = addr_to_region_ptr(addr);
785 HeapWord* result = region_ptr->destination();
786
787 // If the entire Region is live, the new location is region->destination + the
788 // offset of the object within in the Region.
789
790 // Run some performance tests to determine if this special case pays off. It
791 // is worth it for pointers into the dense prefix. If the optimization to
792 // avoid pointer updates in regions that only point to the dense prefix is
793 // ever implemented, this should be revisited.
794 if (region_ptr->data_size() == RegionSize) {
795 result += region_offset(addr);
796 return result;
797 }
798
799 // Otherwise, the new location is region->destination + block offset + the
800 // number of live words in the Block that are (a) to the left of addr and (b)
801 // due to objects that start in the Block.
802
803 // Fill in the block table if necessary. This is unsynchronized, so multiple
804 // threads may fill the block table for a region (harmless, since it is
805 // idempotent).
806 if (!region_ptr->blocks_filled()) {
807 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
808 region_ptr->set_blocks_filled();
809 }
810
811 HeapWord* const search_start = block_align_down(addr);
812 const size_t block_offset = addr_to_block_ptr(addr)->offset();
813
814 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
815 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr));
816 result += block_offset + live;
817 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
818 return result;
819 }
820
821 #ifdef ASSERT
verify_clear(const PSVirtualSpace * vspace)822 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
823 {
824 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
825 const size_t* const end = (const size_t*)vspace->committed_high_addr();
826 for (const size_t* p = beg; p < end; ++p) {
827 assert(*p == 0, "not zero");
828 }
829 }
830
verify_clear()831 void ParallelCompactData::verify_clear()
832 {
833 verify_clear(_region_vspace);
834 verify_clear(_block_vspace);
835 }
836 #endif // #ifdef ASSERT
837
838 STWGCTimer PSParallelCompact::_gc_timer;
839 ParallelOldTracer PSParallelCompact::_gc_tracer;
840 elapsedTimer PSParallelCompact::_accumulated_time;
841 unsigned int PSParallelCompact::_total_invocations = 0;
842 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
843 jlong PSParallelCompact::_time_of_last_gc = 0;
844 CollectorCounters* PSParallelCompact::_counters = NULL;
845 ParMarkBitMap PSParallelCompact::_mark_bitmap;
846 ParallelCompactData PSParallelCompact::_summary_data;
847
848 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
849
do_object_b(oop p)850 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
851
852 class PCReferenceProcessor: public ReferenceProcessor {
853 public:
PCReferenceProcessor(BoolObjectClosure * is_subject_to_discovery,BoolObjectClosure * is_alive_non_header)854 PCReferenceProcessor(
855 BoolObjectClosure* is_subject_to_discovery,
856 BoolObjectClosure* is_alive_non_header) :
857 ReferenceProcessor(is_subject_to_discovery,
858 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
859 ParallelGCThreads, // mt processing degree
860 true, // mt discovery
861 ParallelGCThreads, // mt discovery degree
862 true, // atomic_discovery
863 is_alive_non_header) {
864 }
865
discover(oop obj,ReferenceType type)866 template<typename T> bool discover(oop obj, ReferenceType type) {
867 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
868 T heap_oop = RawAccess<>::oop_load(referent_addr);
869 oop referent = CompressedOops::decode_not_null(heap_oop);
870 return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
871 && ReferenceProcessor::discover_reference(obj, type);
872 }
discover_reference(oop obj,ReferenceType type)873 virtual bool discover_reference(oop obj, ReferenceType type) {
874 if (UseCompressedOops) {
875 return discover<narrowOop>(obj, type);
876 } else {
877 return discover<oop>(obj, type);
878 }
879 }
880 };
881
post_initialize()882 void PSParallelCompact::post_initialize() {
883 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
884 _span_based_discoverer.set_span(heap->reserved_region());
885 _ref_processor =
886 new PCReferenceProcessor(&_span_based_discoverer,
887 &_is_alive_closure); // non-header is alive closure
888
889 _counters = new CollectorCounters("Parallel full collection pauses", 1);
890
891 // Initialize static fields in ParCompactionManager.
892 ParCompactionManager::initialize(mark_bitmap());
893 }
894
initialize()895 bool PSParallelCompact::initialize() {
896 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
897 MemRegion mr = heap->reserved_region();
898
899 // Was the old gen get allocated successfully?
900 if (!heap->old_gen()->is_allocated()) {
901 return false;
902 }
903
904 initialize_space_info();
905 initialize_dead_wood_limiter();
906
907 if (!_mark_bitmap.initialize(mr)) {
908 vm_shutdown_during_initialization(
909 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
910 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
911 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
912 return false;
913 }
914
915 if (!_summary_data.initialize(mr)) {
916 vm_shutdown_during_initialization(
917 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
918 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
919 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
920 return false;
921 }
922
923 return true;
924 }
925
initialize_space_info()926 void PSParallelCompact::initialize_space_info()
927 {
928 memset(&_space_info, 0, sizeof(_space_info));
929
930 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
931 PSYoungGen* young_gen = heap->young_gen();
932
933 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
934 _space_info[eden_space_id].set_space(young_gen->eden_space());
935 _space_info[from_space_id].set_space(young_gen->from_space());
936 _space_info[to_space_id].set_space(young_gen->to_space());
937
938 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
939 }
940
initialize_dead_wood_limiter()941 void PSParallelCompact::initialize_dead_wood_limiter()
942 {
943 const size_t max = 100;
944 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
945 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
946 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
947 DEBUG_ONLY(_dwl_initialized = true;)
948 _dwl_adjustment = normal_distribution(1.0);
949 }
950
951 void
clear_data_covering_space(SpaceId id)952 PSParallelCompact::clear_data_covering_space(SpaceId id)
953 {
954 // At this point, top is the value before GC, new_top() is the value that will
955 // be set at the end of GC. The marking bitmap is cleared to top; nothing
956 // should be marked above top. The summary data is cleared to the larger of
957 // top & new_top.
958 MutableSpace* const space = _space_info[id].space();
959 HeapWord* const bot = space->bottom();
960 HeapWord* const top = space->top();
961 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
962
963 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
964 const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
965 _mark_bitmap.clear_range(beg_bit, end_bit);
966
967 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
968 const size_t end_region =
969 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
970 _summary_data.clear_range(beg_region, end_region);
971
972 // Clear the data used to 'split' regions.
973 SplitInfo& split_info = _space_info[id].split_info();
974 if (split_info.is_valid()) {
975 split_info.clear();
976 }
977 DEBUG_ONLY(split_info.verify_clear();)
978 }
979
pre_compact()980 void PSParallelCompact::pre_compact()
981 {
982 // Update the from & to space pointers in space_info, since they are swapped
983 // at each young gen gc. Do the update unconditionally (even though a
984 // promotion failure does not swap spaces) because an unknown number of young
985 // collections will have swapped the spaces an unknown number of times.
986 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
987 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
988 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
989 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
990
991 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
992 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
993
994 // Increment the invocation count
995 heap->increment_total_collections(true);
996
997 // We need to track unique mark sweep invocations as well.
998 _total_invocations++;
999
1000 heap->print_heap_before_gc();
1001 heap->trace_heap_before_gc(&_gc_tracer);
1002
1003 // Fill in TLABs
1004 heap->ensure_parsability(true); // retire TLABs
1005
1006 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1007 HandleMark hm; // Discard invalid handles created during verification
1008 Universe::verify("Before GC");
1009 }
1010
1011 // Verify object start arrays
1012 if (VerifyObjectStartArray &&
1013 VerifyBeforeGC) {
1014 heap->old_gen()->verify_object_start_array();
1015 }
1016
1017 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1018 DEBUG_ONLY(summary_data().verify_clear();)
1019
1020 ParCompactionManager::reset_all_bitmap_query_caches();
1021 }
1022
post_compact()1023 void PSParallelCompact::post_compact()
1024 {
1025 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1026 ParCompactionManager::remove_all_shadow_regions();
1027
1028 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1029 // Clear the marking bitmap, summary data and split info.
1030 clear_data_covering_space(SpaceId(id));
1031 // Update top(). Must be done after clearing the bitmap and summary data.
1032 _space_info[id].publish_new_top();
1033 }
1034
1035 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1036 MutableSpace* const from_space = _space_info[from_space_id].space();
1037 MutableSpace* const to_space = _space_info[to_space_id].space();
1038
1039 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1040 bool eden_empty = eden_space->is_empty();
1041 if (!eden_empty) {
1042 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1043 heap->young_gen(), heap->old_gen());
1044 }
1045
1046 // Update heap occupancy information which is used as input to the soft ref
1047 // clearing policy at the next gc.
1048 Universe::update_heap_info_at_gc();
1049
1050 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1051 to_space->is_empty();
1052
1053 PSCardTable* ct = heap->card_table();
1054 MemRegion old_mr = heap->old_gen()->reserved();
1055 if (young_gen_empty) {
1056 ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1057 } else {
1058 ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1059 }
1060
1061 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1062 ClassLoaderDataGraph::purge();
1063 MetaspaceUtils::verify_metrics();
1064
1065 heap->prune_scavengable_nmethods();
1066
1067 #if COMPILER2_OR_JVMCI
1068 DerivedPointerTable::update_pointers();
1069 #endif
1070
1071 if (ZapUnusedHeapArea) {
1072 heap->gen_mangle_unused_area();
1073 }
1074
1075 // Update time of last GC
1076 reset_millis_since_last_gc();
1077 }
1078
1079 HeapWord*
compute_dense_prefix_via_density(const SpaceId id,bool maximum_compaction)1080 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1081 bool maximum_compaction)
1082 {
1083 const size_t region_size = ParallelCompactData::RegionSize;
1084 const ParallelCompactData& sd = summary_data();
1085
1086 const MutableSpace* const space = _space_info[id].space();
1087 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1088 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1089 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1090
1091 // Skip full regions at the beginning of the space--they are necessarily part
1092 // of the dense prefix.
1093 size_t full_count = 0;
1094 const RegionData* cp;
1095 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1096 ++full_count;
1097 }
1098
1099 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1100 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1101 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1102 if (maximum_compaction || cp == end_cp || interval_ended) {
1103 _maximum_compaction_gc_num = total_invocations();
1104 return sd.region_to_addr(cp);
1105 }
1106
1107 HeapWord* const new_top = _space_info[id].new_top();
1108 const size_t space_live = pointer_delta(new_top, space->bottom());
1109 const size_t space_used = space->used_in_words();
1110 const size_t space_capacity = space->capacity_in_words();
1111
1112 const double cur_density = double(space_live) / space_capacity;
1113 const double deadwood_density =
1114 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1115 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1116
1117 if (TraceParallelOldGCDensePrefix) {
1118 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1119 cur_density, deadwood_density, deadwood_goal);
1120 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1121 "space_cap=" SIZE_FORMAT,
1122 space_live, space_used,
1123 space_capacity);
1124 }
1125
1126 // XXX - Use binary search?
1127 HeapWord* dense_prefix = sd.region_to_addr(cp);
1128 const RegionData* full_cp = cp;
1129 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1130 while (cp < end_cp) {
1131 HeapWord* region_destination = cp->destination();
1132 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1133 if (TraceParallelOldGCDensePrefix && Verbose) {
1134 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1135 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1136 sd.region(cp), p2i(region_destination),
1137 p2i(dense_prefix), cur_deadwood);
1138 }
1139
1140 if (cur_deadwood >= deadwood_goal) {
1141 // Found the region that has the correct amount of deadwood to the left.
1142 // This typically occurs after crossing a fairly sparse set of regions, so
1143 // iterate backwards over those sparse regions, looking for the region
1144 // that has the lowest density of live objects 'to the right.'
1145 size_t space_to_left = sd.region(cp) * region_size;
1146 size_t live_to_left = space_to_left - cur_deadwood;
1147 size_t space_to_right = space_capacity - space_to_left;
1148 size_t live_to_right = space_live - live_to_left;
1149 double density_to_right = double(live_to_right) / space_to_right;
1150 while (cp > full_cp) {
1151 --cp;
1152 const size_t prev_region_live_to_right = live_to_right -
1153 cp->data_size();
1154 const size_t prev_region_space_to_right = space_to_right + region_size;
1155 double prev_region_density_to_right =
1156 double(prev_region_live_to_right) / prev_region_space_to_right;
1157 if (density_to_right <= prev_region_density_to_right) {
1158 return dense_prefix;
1159 }
1160 if (TraceParallelOldGCDensePrefix && Verbose) {
1161 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1162 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1163 prev_region_density_to_right);
1164 }
1165 dense_prefix -= region_size;
1166 live_to_right = prev_region_live_to_right;
1167 space_to_right = prev_region_space_to_right;
1168 density_to_right = prev_region_density_to_right;
1169 }
1170 return dense_prefix;
1171 }
1172
1173 dense_prefix += region_size;
1174 ++cp;
1175 }
1176
1177 return dense_prefix;
1178 }
1179
1180 #ifndef PRODUCT
print_dense_prefix_stats(const char * const algorithm,const SpaceId id,const bool maximum_compaction,HeapWord * const addr)1181 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1182 const SpaceId id,
1183 const bool maximum_compaction,
1184 HeapWord* const addr)
1185 {
1186 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1187 RegionData* const cp = summary_data().region(region_idx);
1188 const MutableSpace* const space = _space_info[id].space();
1189 HeapWord* const new_top = _space_info[id].new_top();
1190
1191 const size_t space_live = pointer_delta(new_top, space->bottom());
1192 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1193 const size_t space_cap = space->capacity_in_words();
1194 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1195 const size_t live_to_right = new_top - cp->destination();
1196 const size_t dead_to_right = space->top() - addr - live_to_right;
1197
1198 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1199 "spl=" SIZE_FORMAT " "
1200 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1201 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1202 " ratio=%10.8f",
1203 algorithm, p2i(addr), region_idx,
1204 space_live,
1205 dead_to_left, dead_to_left_pct,
1206 dead_to_right, live_to_right,
1207 double(dead_to_right) / live_to_right);
1208 }
1209 #endif // #ifndef PRODUCT
1210
1211 // Return a fraction indicating how much of the generation can be treated as
1212 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1213 // based on the density of live objects in the generation to determine a limit,
1214 // which is then adjusted so the return value is min_percent when the density is
1215 // 1.
1216 //
1217 // The following table shows some return values for a different values of the
1218 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1219 // min_percent is 1.
1220 //
1221 // fraction allowed as dead wood
1222 // -----------------------------------------------------------------
1223 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1224 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1225 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1226 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1227 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1228 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1229 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1230 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1231 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1232 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1233 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1234 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1235 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1236 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1237 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1238 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1239 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1240 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1241 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1242 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1243 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1244 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1245 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1246
dead_wood_limiter(double density,size_t min_percent)1247 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1248 {
1249 assert(_dwl_initialized, "uninitialized");
1250
1251 // The raw limit is the value of the normal distribution at x = density.
1252 const double raw_limit = normal_distribution(density);
1253
1254 // Adjust the raw limit so it becomes the minimum when the density is 1.
1255 //
1256 // First subtract the adjustment value (which is simply the precomputed value
1257 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1258 // Then add the minimum value, so the minimum is returned when the density is
1259 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1260 const double min = double(min_percent) / 100.0;
1261 const double limit = raw_limit - _dwl_adjustment + min;
1262 return MAX2(limit, 0.0);
1263 }
1264
1265 ParallelCompactData::RegionData*
first_dead_space_region(const RegionData * beg,const RegionData * end)1266 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1267 const RegionData* end)
1268 {
1269 const size_t region_size = ParallelCompactData::RegionSize;
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 if (middle > left && dest < addr) {
1285 right = middle - 1;
1286 } else if (middle < right && middle_ptr->data_size() == region_size) {
1287 left = middle + 1;
1288 } else {
1289 return middle_ptr;
1290 }
1291 }
1292 return sd.region(left);
1293 }
1294
1295 ParallelCompactData::RegionData*
dead_wood_limit_region(const RegionData * beg,const RegionData * end,size_t dead_words)1296 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1297 const RegionData* end,
1298 size_t dead_words)
1299 {
1300 ParallelCompactData& sd = summary_data();
1301 size_t left = sd.region(beg);
1302 size_t right = end > beg ? sd.region(end) - 1 : left;
1303
1304 // Binary search.
1305 while (left < right) {
1306 // Equivalent to (left + right) / 2, but does not overflow.
1307 const size_t middle = left + (right - left) / 2;
1308 RegionData* const middle_ptr = sd.region(middle);
1309 HeapWord* const dest = middle_ptr->destination();
1310 HeapWord* const addr = sd.region_to_addr(middle);
1311 assert(dest != NULL, "sanity");
1312 assert(dest <= addr, "must move left");
1313
1314 const size_t dead_to_left = pointer_delta(addr, dest);
1315 if (middle > left && dead_to_left > dead_words) {
1316 right = middle - 1;
1317 } else if (middle < right && dead_to_left < dead_words) {
1318 left = middle + 1;
1319 } else {
1320 return middle_ptr;
1321 }
1322 }
1323 return sd.region(left);
1324 }
1325
1326 // The result is valid during the summary phase, after the initial summarization
1327 // of each space into itself, and before final summarization.
1328 inline double
reclaimed_ratio(const RegionData * const cp,HeapWord * const bottom,HeapWord * const top,HeapWord * const new_top)1329 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1330 HeapWord* const bottom,
1331 HeapWord* const top,
1332 HeapWord* const new_top)
1333 {
1334 ParallelCompactData& sd = summary_data();
1335
1336 assert(cp != NULL, "sanity");
1337 assert(bottom != NULL, "sanity");
1338 assert(top != NULL, "sanity");
1339 assert(new_top != NULL, "sanity");
1340 assert(top >= new_top, "summary data problem?");
1341 assert(new_top > bottom, "space is empty; should not be here");
1342 assert(new_top >= cp->destination(), "sanity");
1343 assert(top >= sd.region_to_addr(cp), "sanity");
1344
1345 HeapWord* const destination = cp->destination();
1346 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1347 const size_t compacted_region_live = pointer_delta(new_top, destination);
1348 const size_t compacted_region_used = pointer_delta(top,
1349 sd.region_to_addr(cp));
1350 const size_t reclaimable = compacted_region_used - compacted_region_live;
1351
1352 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1353 return double(reclaimable) / divisor;
1354 }
1355
1356 // Return the address of the end of the dense prefix, a.k.a. the start of the
1357 // compacted region. The address is always on a region boundary.
1358 //
1359 // Completely full regions at the left are skipped, since no compaction can
1360 // occur in those regions. Then the maximum amount of dead wood to allow is
1361 // computed, based on the density (amount live / capacity) of the generation;
1362 // the region with approximately that amount of dead space to the left is
1363 // identified as the limit region. Regions between the last completely full
1364 // region and the limit region are scanned and the one that has the best
1365 // (maximum) reclaimed_ratio() is selected.
1366 HeapWord*
compute_dense_prefix(const SpaceId id,bool maximum_compaction)1367 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1368 bool maximum_compaction)
1369 {
1370 const size_t region_size = ParallelCompactData::RegionSize;
1371 const ParallelCompactData& sd = summary_data();
1372
1373 const MutableSpace* const space = _space_info[id].space();
1374 HeapWord* const top = space->top();
1375 HeapWord* const top_aligned_up = sd.region_align_up(top);
1376 HeapWord* const new_top = _space_info[id].new_top();
1377 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1378 HeapWord* const bottom = space->bottom();
1379 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1380 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1381 const RegionData* const new_top_cp =
1382 sd.addr_to_region_ptr(new_top_aligned_up);
1383
1384 // Skip full regions at the beginning of the space--they are necessarily part
1385 // of the dense prefix.
1386 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1387 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1388 space->is_empty(), "no dead space allowed to the left");
1389 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1390 "region must have dead space");
1391
1392 // The gc number is saved whenever a maximum compaction is done, and used to
1393 // determine when the maximum compaction interval has expired. This avoids
1394 // successive max compactions for different reasons.
1395 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1396 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1397 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1398 total_invocations() == HeapFirstMaximumCompactionCount;
1399 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1400 _maximum_compaction_gc_num = total_invocations();
1401 return sd.region_to_addr(full_cp);
1402 }
1403
1404 const size_t space_live = pointer_delta(new_top, bottom);
1405 const size_t space_used = space->used_in_words();
1406 const size_t space_capacity = space->capacity_in_words();
1407
1408 const double density = double(space_live) / double(space_capacity);
1409 const size_t min_percent_free = MarkSweepDeadRatio;
1410 const double limiter = dead_wood_limiter(density, min_percent_free);
1411 const size_t dead_wood_max = space_used - space_live;
1412 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1413 dead_wood_max);
1414
1415 if (TraceParallelOldGCDensePrefix) {
1416 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1417 "space_cap=" SIZE_FORMAT,
1418 space_live, space_used,
1419 space_capacity);
1420 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1421 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1422 density, min_percent_free, limiter,
1423 dead_wood_max, dead_wood_limit);
1424 }
1425
1426 // Locate the region with the desired amount of dead space to the left.
1427 const RegionData* const limit_cp =
1428 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1429
1430 // Scan from the first region with dead space to the limit region and find the
1431 // one with the best (largest) reclaimed ratio.
1432 double best_ratio = 0.0;
1433 const RegionData* best_cp = full_cp;
1434 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1435 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1436 if (tmp_ratio > best_ratio) {
1437 best_cp = cp;
1438 best_ratio = tmp_ratio;
1439 }
1440 }
1441
1442 return sd.region_to_addr(best_cp);
1443 }
1444
summarize_spaces_quick()1445 void PSParallelCompact::summarize_spaces_quick()
1446 {
1447 for (unsigned int i = 0; i < last_space_id; ++i) {
1448 const MutableSpace* space = _space_info[i].space();
1449 HeapWord** nta = _space_info[i].new_top_addr();
1450 bool result = _summary_data.summarize(_space_info[i].split_info(),
1451 space->bottom(), space->top(), NULL,
1452 space->bottom(), space->end(), nta);
1453 assert(result, "space must fit into itself");
1454 _space_info[i].set_dense_prefix(space->bottom());
1455 }
1456 }
1457
fill_dense_prefix_end(SpaceId id)1458 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1459 {
1460 HeapWord* const dense_prefix_end = dense_prefix(id);
1461 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1462 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1463 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1464 // Only enough dead space is filled so that any remaining dead space to the
1465 // left is larger than the minimum filler object. (The remainder is filled
1466 // during the copy/update phase.)
1467 //
1468 // The size of the dead space to the right of the boundary is not a
1469 // concern, since compaction will be able to use whatever space is
1470 // available.
1471 //
1472 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1473 // surrounds the space to be filled with an object.
1474 //
1475 // In the 32-bit VM, each bit represents two 32-bit words:
1476 // +---+
1477 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1478 // end_bits: ... x x x | 0 | || 0 x x ...
1479 // +---+
1480 //
1481 // In the 64-bit VM, each bit represents one 64-bit word:
1482 // +------------+
1483 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1484 // end_bits: ... x x 1 | 0 || 0 | x x ...
1485 // +------------+
1486 // +-------+
1487 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1488 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1489 // +-------+
1490 // +-----------+
1491 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1492 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1493 // +-----------+
1494 // +-------+
1495 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1496 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1497 // +-------+
1498
1499 // Initially assume case a, c or e will apply.
1500 size_t obj_len = CollectedHeap::min_fill_size();
1501 HeapWord* obj_beg = dense_prefix_end - obj_len;
1502
1503 #ifdef _LP64
1504 if (MinObjAlignment > 1) { // object alignment > heap word size
1505 // Cases a, c or e.
1506 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1507 // Case b above.
1508 obj_beg = dense_prefix_end - 1;
1509 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1510 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1511 // Case d above.
1512 obj_beg = dense_prefix_end - 3;
1513 obj_len = 3;
1514 }
1515 #endif // #ifdef _LP64
1516
1517 CollectedHeap::fill_with_object(obj_beg, obj_len);
1518 _mark_bitmap.mark_obj(obj_beg, obj_len);
1519 _summary_data.add_obj(obj_beg, obj_len);
1520 assert(start_array(id) != NULL, "sanity");
1521 start_array(id)->allocate_block(obj_beg);
1522 }
1523 }
1524
1525 void
summarize_space(SpaceId id,bool maximum_compaction)1526 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1527 {
1528 assert(id < last_space_id, "id out of range");
1529 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1530 "should have been reset in summarize_spaces_quick()");
1531
1532 const MutableSpace* space = _space_info[id].space();
1533 if (_space_info[id].new_top() != space->bottom()) {
1534 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1535 _space_info[id].set_dense_prefix(dense_prefix_end);
1536
1537 #ifndef PRODUCT
1538 if (TraceParallelOldGCDensePrefix) {
1539 print_dense_prefix_stats("ratio", id, maximum_compaction,
1540 dense_prefix_end);
1541 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1542 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1543 }
1544 #endif // #ifndef PRODUCT
1545
1546 // Recompute the summary data, taking into account the dense prefix. If
1547 // every last byte will be reclaimed, then the existing summary data which
1548 // compacts everything can be left in place.
1549 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1550 // If dead space crosses the dense prefix boundary, it is (at least
1551 // partially) filled with a dummy object, marked live and added to the
1552 // summary data. This simplifies the copy/update phase and must be done
1553 // before the final locations of objects are determined, to prevent
1554 // leaving a fragment of dead space that is too small to fill.
1555 fill_dense_prefix_end(id);
1556
1557 // Compute the destination of each Region, and thus each object.
1558 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1559 _summary_data.summarize(_space_info[id].split_info(),
1560 dense_prefix_end, space->top(), NULL,
1561 dense_prefix_end, space->end(),
1562 _space_info[id].new_top_addr());
1563 }
1564 }
1565
1566 if (log_develop_is_enabled(Trace, gc, compaction)) {
1567 const size_t region_size = ParallelCompactData::RegionSize;
1568 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1569 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1570 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1571 HeapWord* const new_top = _space_info[id].new_top();
1572 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1573 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1574 log_develop_trace(gc, compaction)(
1575 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1576 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1577 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1578 id, space->capacity_in_words(), p2i(dense_prefix_end),
1579 dp_region, dp_words / region_size,
1580 cr_words / region_size, p2i(new_top));
1581 }
1582 }
1583
1584 #ifndef PRODUCT
summary_phase_msg(SpaceId dst_space_id,HeapWord * dst_beg,HeapWord * dst_end,SpaceId src_space_id,HeapWord * src_beg,HeapWord * src_end)1585 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1586 HeapWord* dst_beg, HeapWord* dst_end,
1587 SpaceId src_space_id,
1588 HeapWord* src_beg, HeapWord* src_end)
1589 {
1590 log_develop_trace(gc, compaction)(
1591 "Summarizing %d [%s] into %d [%s]: "
1592 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1593 SIZE_FORMAT "-" SIZE_FORMAT " "
1594 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1595 SIZE_FORMAT "-" SIZE_FORMAT,
1596 src_space_id, space_names[src_space_id],
1597 dst_space_id, space_names[dst_space_id],
1598 p2i(src_beg), p2i(src_end),
1599 _summary_data.addr_to_region_idx(src_beg),
1600 _summary_data.addr_to_region_idx(src_end),
1601 p2i(dst_beg), p2i(dst_end),
1602 _summary_data.addr_to_region_idx(dst_beg),
1603 _summary_data.addr_to_region_idx(dst_end));
1604 }
1605 #endif // #ifndef PRODUCT
1606
summary_phase(ParCompactionManager * cm,bool maximum_compaction)1607 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1608 bool maximum_compaction)
1609 {
1610 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1611
1612 #ifdef ASSERT
1613 if (TraceParallelOldGCMarkingPhase) {
1614 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1615 "add_obj_bytes=" SIZE_FORMAT,
1616 add_obj_count, add_obj_size * HeapWordSize);
1617 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1618 "mark_bitmap_bytes=" SIZE_FORMAT,
1619 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1620 }
1621 #endif // #ifdef ASSERT
1622
1623 // Quick summarization of each space into itself, to see how much is live.
1624 summarize_spaces_quick();
1625
1626 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self");
1627 NOT_PRODUCT(print_region_ranges());
1628 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1629
1630 // The amount of live data that will end up in old space (assuming it fits).
1631 size_t old_space_total_live = 0;
1632 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1633 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1634 _space_info[id].space()->bottom());
1635 }
1636
1637 MutableSpace* const old_space = _space_info[old_space_id].space();
1638 const size_t old_capacity = old_space->capacity_in_words();
1639 if (old_space_total_live > old_capacity) {
1640 // XXX - should also try to expand
1641 maximum_compaction = true;
1642 }
1643
1644 // Old generations.
1645 summarize_space(old_space_id, maximum_compaction);
1646
1647 // Summarize the remaining spaces in the young gen. The initial target space
1648 // is the old gen. If a space does not fit entirely into the target, then the
1649 // remainder is compacted into the space itself and that space becomes the new
1650 // target.
1651 SpaceId dst_space_id = old_space_id;
1652 HeapWord* dst_space_end = old_space->end();
1653 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1654 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1655 const MutableSpace* space = _space_info[id].space();
1656 const size_t live = pointer_delta(_space_info[id].new_top(),
1657 space->bottom());
1658 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1659
1660 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1661 SpaceId(id), space->bottom(), space->top());)
1662 if (live > 0 && live <= available) {
1663 // All the live data will fit.
1664 bool done = _summary_data.summarize(_space_info[id].split_info(),
1665 space->bottom(), space->top(),
1666 NULL,
1667 *new_top_addr, dst_space_end,
1668 new_top_addr);
1669 assert(done, "space must fit into old gen");
1670
1671 // Reset the new_top value for the space.
1672 _space_info[id].set_new_top(space->bottom());
1673 } else if (live > 0) {
1674 // Attempt to fit part of the source space into the target space.
1675 HeapWord* next_src_addr = NULL;
1676 bool done = _summary_data.summarize(_space_info[id].split_info(),
1677 space->bottom(), space->top(),
1678 &next_src_addr,
1679 *new_top_addr, dst_space_end,
1680 new_top_addr);
1681 assert(!done, "space should not fit into old gen");
1682 assert(next_src_addr != NULL, "sanity");
1683
1684 // The source space becomes the new target, so the remainder is compacted
1685 // within the space itself.
1686 dst_space_id = SpaceId(id);
1687 dst_space_end = space->end();
1688 new_top_addr = _space_info[id].new_top_addr();
1689 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1690 space->bottom(), dst_space_end,
1691 SpaceId(id), next_src_addr, space->top());)
1692 done = _summary_data.summarize(_space_info[id].split_info(),
1693 next_src_addr, space->top(),
1694 NULL,
1695 space->bottom(), dst_space_end,
1696 new_top_addr);
1697 assert(done, "space must fit when compacted into itself");
1698 assert(*new_top_addr <= space->top(), "usage should not grow");
1699 }
1700 }
1701
1702 log_develop_trace(gc, compaction)("Summary_phase: after final summarization");
1703 NOT_PRODUCT(print_region_ranges());
1704 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1705 }
1706
1707 // This method should contain all heap-specific policy for invoking a full
1708 // collection. invoke_no_policy() will only attempt to compact the heap; it
1709 // will do nothing further. If we need to bail out for policy reasons, scavenge
1710 // before full gc, or any other specialized behavior, it needs to be added here.
1711 //
1712 // Note that this method should only be called from the vm_thread while at a
1713 // safepoint.
1714 //
1715 // Note that the all_soft_refs_clear flag in the soft ref policy
1716 // may be true because this method can be called without intervening
1717 // activity. For example when the heap space is tight and full measure
1718 // are being taken to free space.
invoke(bool maximum_heap_compaction)1719 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1720 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1721 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1722 "should be in vm thread");
1723
1724 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1725 GCCause::Cause gc_cause = heap->gc_cause();
1726 assert(!heap->is_gc_active(), "not reentrant");
1727
1728 PSAdaptiveSizePolicy* policy = heap->size_policy();
1729 IsGCActiveMark mark;
1730
1731 if (ScavengeBeforeFullGC) {
1732 PSScavenge::invoke_no_policy();
1733 }
1734
1735 const bool clear_all_soft_refs =
1736 heap->soft_ref_policy()->should_clear_all_soft_refs();
1737
1738 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1739 maximum_heap_compaction);
1740 }
1741
1742 // This method contains no policy. You should probably
1743 // be calling invoke() instead.
invoke_no_policy(bool maximum_heap_compaction)1744 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1745 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1746 assert(ref_processor() != NULL, "Sanity");
1747
1748 if (GCLocker::check_active_before_gc()) {
1749 return false;
1750 }
1751
1752 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1753
1754 GCIdMark gc_id_mark;
1755 _gc_timer.register_gc_start();
1756 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1757
1758 TimeStamp marking_start;
1759 TimeStamp compaction_start;
1760 TimeStamp collection_exit;
1761
1762 GCCause::Cause gc_cause = heap->gc_cause();
1763 PSYoungGen* young_gen = heap->young_gen();
1764 PSOldGen* old_gen = heap->old_gen();
1765 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1766
1767 // The scope of casr should end after code that can change
1768 // SoftRefPolicy::_should_clear_all_soft_refs.
1769 ClearedAllSoftRefs casr(maximum_heap_compaction,
1770 heap->soft_ref_policy());
1771
1772 if (ZapUnusedHeapArea) {
1773 // Save information needed to minimize mangling
1774 heap->record_gen_tops_before_GC();
1775 }
1776
1777 // Make sure data structures are sane, make the heap parsable, and do other
1778 // miscellaneous bookkeeping.
1779 pre_compact();
1780
1781 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1782
1783 // Get the compaction manager reserved for the VM thread.
1784 ParCompactionManager* const vmthread_cm =
1785 ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers());
1786
1787 {
1788 ResourceMark rm;
1789 HandleMark hm;
1790
1791 const uint active_workers =
1792 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(),
1793 ParallelScavengeHeap::heap()->workers().active_workers(),
1794 Threads::number_of_non_daemon_threads());
1795 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers);
1796
1797 GCTraceCPUTime tcpu;
1798 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1799
1800 heap->pre_full_gc_dump(&_gc_timer);
1801
1802 TraceCollectorStats tcs(counters());
1803 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1804
1805 if (log_is_enabled(Debug, gc, heap, exit)) {
1806 accumulated_time()->start();
1807 }
1808
1809 // Let the size policy know we're starting
1810 size_policy->major_collection_begin();
1811
1812 #if COMPILER2_OR_JVMCI
1813 DerivedPointerTable::clear();
1814 #endif
1815
1816 ref_processor()->enable_discovery();
1817 ref_processor()->setup_policy(maximum_heap_compaction);
1818
1819 bool marked_for_unloading = false;
1820
1821 marking_start.update();
1822 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
1823
1824 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1825 && GCCause::is_user_requested_gc(gc_cause);
1826 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1827
1828 #if COMPILER2_OR_JVMCI
1829 assert(DerivedPointerTable::is_active(), "Sanity");
1830 DerivedPointerTable::set_active(false);
1831 #endif
1832
1833 // adjust_roots() updates Universe::_intArrayKlassObj which is
1834 // needed by the compaction for filling holes in the dense prefix.
1835 adjust_roots(vmthread_cm);
1836
1837 compaction_start.update();
1838 compact();
1839
1840 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
1841 // done before resizing.
1842 post_compact();
1843
1844 // Let the size policy know we're done
1845 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1846
1847 if (UseAdaptiveSizePolicy) {
1848 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1849 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1850 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1851
1852 // Don't check if the size_policy is ready here. Let
1853 // the size_policy check that internally.
1854 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1855 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1856 // Swap the survivor spaces if from_space is empty. The
1857 // resize_young_gen() called below is normally used after
1858 // a successful young GC and swapping of survivor spaces;
1859 // otherwise, it will fail to resize the young gen with
1860 // the current implementation.
1861 if (young_gen->from_space()->is_empty()) {
1862 young_gen->from_space()->clear(SpaceDecorator::Mangle);
1863 young_gen->swap_spaces();
1864 }
1865
1866 // Calculate optimal free space amounts
1867 assert(young_gen->max_size() >
1868 young_gen->from_space()->capacity_in_bytes() +
1869 young_gen->to_space()->capacity_in_bytes(),
1870 "Sizes of space in young gen are out-of-bounds");
1871
1872 size_t young_live = young_gen->used_in_bytes();
1873 size_t eden_live = young_gen->eden_space()->used_in_bytes();
1874 size_t old_live = old_gen->used_in_bytes();
1875 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1876 size_t max_old_gen_size = old_gen->max_gen_size();
1877 size_t max_eden_size = young_gen->max_size() -
1878 young_gen->from_space()->capacity_in_bytes() -
1879 young_gen->to_space()->capacity_in_bytes();
1880
1881 // Used for diagnostics
1882 size_policy->clear_generation_free_space_flags();
1883
1884 size_policy->compute_generations_free_space(young_live,
1885 eden_live,
1886 old_live,
1887 cur_eden,
1888 max_old_gen_size,
1889 max_eden_size,
1890 true /* full gc*/);
1891
1892 size_policy->check_gc_overhead_limit(eden_live,
1893 max_old_gen_size,
1894 max_eden_size,
1895 true /* full gc*/,
1896 gc_cause,
1897 heap->soft_ref_policy());
1898
1899 size_policy->decay_supplemental_growth(true /* full gc*/);
1900
1901 heap->resize_old_gen(
1902 size_policy->calculated_old_free_size_in_bytes());
1903
1904 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1905 size_policy->calculated_survivor_size_in_bytes());
1906 }
1907
1908 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1909 }
1910
1911 if (UsePerfData) {
1912 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1913 counters->update_counters();
1914 counters->update_old_capacity(old_gen->capacity_in_bytes());
1915 counters->update_young_capacity(young_gen->capacity_in_bytes());
1916 }
1917
1918 heap->resize_all_tlabs();
1919
1920 // Resize the metaspace capacity after a collection
1921 MetaspaceGC::compute_new_size();
1922
1923 if (log_is_enabled(Debug, gc, heap, exit)) {
1924 accumulated_time()->stop();
1925 }
1926
1927 heap->print_heap_change(pre_gc_values);
1928
1929 // Track memory usage and detect low memory
1930 MemoryService::track_memory_usage();
1931 heap->update_counters();
1932
1933 heap->post_full_gc_dump(&_gc_timer);
1934 }
1935
1936 #ifdef ASSERT
1937 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1938 ParCompactionManager* const cm =
1939 ParCompactionManager::manager_array(int(i));
1940 assert(cm->marking_stack()->is_empty(), "should be empty");
1941 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i);
1942 }
1943 #endif // ASSERT
1944
1945 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1946 HandleMark hm; // Discard invalid handles created during verification
1947 Universe::verify("After GC");
1948 }
1949
1950 // Re-verify object start arrays
1951 if (VerifyObjectStartArray &&
1952 VerifyAfterGC) {
1953 old_gen->verify_object_start_array();
1954 }
1955
1956 if (ZapUnusedHeapArea) {
1957 old_gen->object_space()->check_mangled_unused_area_complete();
1958 }
1959
1960 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1961
1962 collection_exit.update();
1963
1964 heap->print_heap_after_gc();
1965 heap->trace_heap_after_gc(&_gc_tracer);
1966
1967 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1968 marking_start.ticks(), compaction_start.ticks(),
1969 collection_exit.ticks());
1970
1971 #ifdef TRACESPINNING
1972 ParallelTaskTerminator::print_termination_counts();
1973 #endif
1974
1975 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1976
1977 _gc_timer.register_gc_end();
1978
1979 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1980 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1981
1982 return true;
1983 }
1984
absorb_live_data_from_eden(PSAdaptiveSizePolicy * size_policy,PSYoungGen * young_gen,PSOldGen * old_gen)1985 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1986 PSYoungGen* young_gen,
1987 PSOldGen* old_gen) {
1988 MutableSpace* const eden_space = young_gen->eden_space();
1989 assert(!eden_space->is_empty(), "eden must be non-empty");
1990 assert(young_gen->virtual_space()->alignment() ==
1991 old_gen->virtual_space()->alignment(), "alignments do not match");
1992
1993 // We also return false when it's a heterogeneous heap because old generation cannot absorb data from eden
1994 // when it is allocated on different memory (example, nv-dimm) than young.
1995 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) ||
1996 ParallelArguments::is_heterogeneous_heap()) {
1997 return false;
1998 }
1999
2000 // Both generations must be completely committed.
2001 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2002 return false;
2003 }
2004 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2005 return false;
2006 }
2007
2008 // Figure out how much to take from eden. Include the average amount promoted
2009 // in the total; otherwise the next young gen GC will simply bail out to a
2010 // full GC.
2011 const size_t alignment = old_gen->virtual_space()->alignment();
2012 const size_t eden_used = eden_space->used_in_bytes();
2013 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2014 const size_t absorb_size = align_up(eden_used + promoted, alignment);
2015 const size_t eden_capacity = eden_space->capacity_in_bytes();
2016
2017 if (absorb_size >= eden_capacity) {
2018 return false; // Must leave some space in eden.
2019 }
2020
2021 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2022 if (new_young_size < young_gen->min_gen_size()) {
2023 return false; // Respect young gen minimum size.
2024 }
2025
2026 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: "
2027 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2028 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2029 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2030 absorb_size / K,
2031 eden_capacity / K, (eden_capacity - absorb_size) / K,
2032 young_gen->from_space()->used_in_bytes() / K,
2033 young_gen->to_space()->used_in_bytes() / K,
2034 young_gen->capacity_in_bytes() / K, new_young_size / K);
2035
2036 // Fill the unused part of the old gen.
2037 MutableSpace* const old_space = old_gen->object_space();
2038 HeapWord* const unused_start = old_space->top();
2039 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2040
2041 if (unused_words > 0) {
2042 if (unused_words < CollectedHeap::min_fill_size()) {
2043 return false; // If the old gen cannot be filled, must give up.
2044 }
2045 CollectedHeap::fill_with_objects(unused_start, unused_words);
2046 }
2047
2048 // Take the live data from eden and set both top and end in the old gen to
2049 // eden top. (Need to set end because reset_after_change() mangles the region
2050 // from end to virtual_space->high() in debug builds).
2051 HeapWord* const new_top = eden_space->top();
2052 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2053 absorb_size);
2054 young_gen->reset_after_change();
2055 old_space->set_top(new_top);
2056 old_space->set_end(new_top);
2057 old_gen->reset_after_change();
2058
2059 // Update the object start array for the filler object and the data from eden.
2060 ObjectStartArray* const start_array = old_gen->start_array();
2061 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2062 start_array->allocate_block(p);
2063 }
2064
2065 // Could update the promoted average here, but it is not typically updated at
2066 // full GCs and the value to use is unclear. Something like
2067 //
2068 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2069
2070 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2071 return true;
2072 }
2073
2074 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
2075 private:
2076 uint _worker_id;
2077
2078 public:
PCAddThreadRootsMarkingTaskClosure(uint worker_id)2079 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
do_thread(Thread * thread)2080 void do_thread(Thread* thread) {
2081 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2082
2083 ResourceMark rm;
2084
2085 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
2086
2087 PCMarkAndPushClosure mark_and_push_closure(cm);
2088 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
2089
2090 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
2091
2092 // Do the real work
2093 cm->follow_marking_stacks();
2094 }
2095 };
2096
mark_from_roots_work(ParallelRootType::Value root_type,uint worker_id)2097 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
2098 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2099
2100 ParCompactionManager* cm =
2101 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2102 PCMarkAndPushClosure mark_and_push_closure(cm);
2103
2104 switch (root_type) {
2105 case ParallelRootType::universe:
2106 Universe::oops_do(&mark_and_push_closure);
2107 break;
2108
2109 case ParallelRootType::jni_handles:
2110 JNIHandles::oops_do(&mark_and_push_closure);
2111 break;
2112
2113 case ParallelRootType::object_synchronizer:
2114 ObjectSynchronizer::oops_do(&mark_and_push_closure);
2115 break;
2116
2117 case ParallelRootType::management:
2118 Management::oops_do(&mark_and_push_closure);
2119 break;
2120
2121 case ParallelRootType::jvmti:
2122 JvmtiExport::oops_do(&mark_and_push_closure);
2123 break;
2124
2125 case ParallelRootType::system_dictionary:
2126 SystemDictionary::oops_do(&mark_and_push_closure);
2127 break;
2128
2129 case ParallelRootType::class_loader_data:
2130 {
2131 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
2132 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
2133 }
2134 break;
2135
2136 case ParallelRootType::code_cache:
2137 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
2138 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
2139 AOTLoader::oops_do(&mark_and_push_closure);
2140 break;
2141
2142 case ParallelRootType::sentinel:
2143 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
2144 fatal("Bad enumeration value: %u", root_type);
2145 break;
2146 }
2147
2148 // Do the real work
2149 cm->follow_marking_stacks();
2150 }
2151
steal_marking_work(ParallelTaskTerminator & terminator,uint worker_id)2152 static void steal_marking_work(ParallelTaskTerminator& terminator, uint worker_id) {
2153 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2154
2155 ParCompactionManager* cm =
2156 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2157
2158 oop obj = NULL;
2159 ObjArrayTask task;
2160 do {
2161 while (ParCompactionManager::steal_objarray(worker_id, task)) {
2162 cm->follow_array((objArrayOop)task.obj(), task.index());
2163 cm->follow_marking_stacks();
2164 }
2165 while (ParCompactionManager::steal(worker_id, obj)) {
2166 cm->follow_contents(obj);
2167 cm->follow_marking_stacks();
2168 }
2169 } while (!terminator.offer_termination());
2170 }
2171
2172 class MarkFromRootsTask : public AbstractGangTask {
2173 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2174 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2175 SequentialSubTasksDone _subtasks;
2176 TaskTerminator _terminator;
2177 uint _active_workers;
2178
2179 public:
MarkFromRootsTask(uint active_workers)2180 MarkFromRootsTask(uint active_workers) :
2181 AbstractGangTask("MarkFromRootsTask"),
2182 _strong_roots_scope(active_workers),
2183 _subtasks(),
2184 _terminator(active_workers, ParCompactionManager::stack_array()),
2185 _active_workers(active_workers) {
2186 _subtasks.set_n_threads(active_workers);
2187 _subtasks.set_n_tasks(ParallelRootType::sentinel);
2188 }
2189
work(uint worker_id)2190 virtual void work(uint worker_id) {
2191 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2192 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2193 }
2194 _subtasks.all_tasks_completed();
2195
2196 PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2197 Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2198
2199 if (_active_workers > 1) {
2200 steal_marking_work(*_terminator.terminator(), worker_id);
2201 }
2202 }
2203 };
2204
2205 class PCRefProcTask : public AbstractGangTask {
2206 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2207 ProcessTask& _task;
2208 uint _ergo_workers;
2209 TaskTerminator _terminator;
2210
2211 public:
PCRefProcTask(ProcessTask & task,uint ergo_workers)2212 PCRefProcTask(ProcessTask& task, uint ergo_workers) :
2213 AbstractGangTask("PCRefProcTask"),
2214 _task(task),
2215 _ergo_workers(ergo_workers),
2216 _terminator(_ergo_workers, ParCompactionManager::stack_array()) {
2217 }
2218
work(uint worker_id)2219 virtual void work(uint worker_id) {
2220 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2221 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2222
2223 ParCompactionManager* cm =
2224 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2225 PCMarkAndPushClosure mark_and_push_closure(cm);
2226 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2227 _task.work(worker_id, *PSParallelCompact::is_alive_closure(),
2228 mark_and_push_closure, follow_stack_closure);
2229
2230 steal_marking_work(*_terminator.terminator(), worker_id);
2231 }
2232 };
2233
2234 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor {
execute(ProcessTask & process_task,uint ergo_workers)2235 void execute(ProcessTask& process_task, uint ergo_workers) {
2236 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers,
2237 "Ergonomically chosen workers (%u) must be equal to active workers (%u)",
2238 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers());
2239
2240 PCRefProcTask task(process_task, ergo_workers);
2241 ParallelScavengeHeap::heap()->workers().run_task(&task);
2242 }
2243 };
2244
marking_phase(ParCompactionManager * cm,bool maximum_heap_compaction,ParallelOldTracer * gc_tracer)2245 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2246 bool maximum_heap_compaction,
2247 ParallelOldTracer *gc_tracer) {
2248 // Recursively traverse all live objects and mark them
2249 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2250
2251 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2252 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2253
2254 PCMarkAndPushClosure mark_and_push_closure(cm);
2255 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2256
2257 // Need new claim bits before marking starts.
2258 ClassLoaderDataGraph::clear_claimed_marks();
2259
2260 {
2261 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2262
2263 MarkFromRootsTask task(active_gc_threads);
2264 ParallelScavengeHeap::heap()->workers().run_task(&task);
2265 }
2266
2267 // Process reference objects found during marking
2268 {
2269 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2270
2271 ReferenceProcessorStats stats;
2272 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2273
2274 if (ref_processor()->processing_is_mt()) {
2275 ref_processor()->set_active_mt_degree(active_gc_threads);
2276
2277 RefProcTaskExecutor task_executor;
2278 stats = ref_processor()->process_discovered_references(
2279 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2280 &task_executor, &pt);
2281 } else {
2282 stats = ref_processor()->process_discovered_references(
2283 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2284 &pt);
2285 }
2286
2287 gc_tracer->report_gc_reference_stats(stats);
2288 pt.print_all_references();
2289 }
2290
2291 // This is the point where the entire marking should have completed.
2292 assert(cm->marking_stacks_empty(), "Marking should have completed");
2293
2294 {
2295 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2296 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl);
2297 }
2298
2299 {
2300 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2301
2302 // Follow system dictionary roots and unload classes.
2303 bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2304
2305 // Unload nmethods.
2306 CodeCache::do_unloading(is_alive_closure(), purged_class);
2307
2308 // Prune dead klasses from subklass/sibling/implementor lists.
2309 Klass::clean_weak_klass_links(purged_class);
2310
2311 // Clean JVMCI metadata handles.
2312 JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2313 }
2314
2315 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2316 }
2317
adjust_roots(ParCompactionManager * cm)2318 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2319 // Adjust the pointers to reflect the new locations
2320 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2321
2322 // Need new claim bits when tracing through and adjusting pointers.
2323 ClassLoaderDataGraph::clear_claimed_marks();
2324
2325 PCAdjustPointerClosure oop_closure(cm);
2326
2327 // General strong roots.
2328 Universe::oops_do(&oop_closure);
2329 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2330 Threads::oops_do(&oop_closure, NULL);
2331 ObjectSynchronizer::oops_do(&oop_closure);
2332 Management::oops_do(&oop_closure);
2333 JvmtiExport::oops_do(&oop_closure);
2334 SystemDictionary::oops_do(&oop_closure);
2335 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong);
2336 ClassLoaderDataGraph::cld_do(&cld_closure);
2337
2338 // Now adjust pointers in remaining weak roots. (All of which should
2339 // have been cleared if they pointed to non-surviving objects.)
2340 WeakProcessor::oops_do(&oop_closure);
2341
2342 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2343 CodeCache::blobs_do(&adjust_from_blobs);
2344 AOT_ONLY(AOTLoader::oops_do(&oop_closure);)
2345
2346 ref_processor()->weak_oops_do(&oop_closure);
2347 // Roots were visited so references into the young gen in roots
2348 // may have been scanned. Process them also.
2349 // Should the reference processor have a span that excludes
2350 // young gen objects?
2351 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2352 }
2353
2354 // Helper class to print 8 region numbers per line and then print the total at the end.
2355 class FillableRegionLogger : public StackObj {
2356 private:
2357 Log(gc, compaction) log;
2358 static const int LineLength = 8;
2359 size_t _regions[LineLength];
2360 int _next_index;
2361 bool _enabled;
2362 size_t _total_regions;
2363 public:
FillableRegionLogger()2364 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
~FillableRegionLogger()2365 ~FillableRegionLogger() {
2366 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2367 }
2368
print_line()2369 void print_line() {
2370 if (!_enabled || _next_index == 0) {
2371 return;
2372 }
2373 FormatBuffer<> line("Fillable: ");
2374 for (int i = 0; i < _next_index; i++) {
2375 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2376 }
2377 log.trace("%s", line.buffer());
2378 _next_index = 0;
2379 }
2380
handle(size_t region)2381 void handle(size_t region) {
2382 if (!_enabled) {
2383 return;
2384 }
2385 _regions[_next_index++] = region;
2386 if (_next_index == LineLength) {
2387 print_line();
2388 }
2389 _total_regions++;
2390 }
2391 };
2392
prepare_region_draining_tasks(uint parallel_gc_threads)2393 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2394 {
2395 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2396
2397 // Find the threads that are active
2398 uint worker_id = 0;
2399
2400 // Find all regions that are available (can be filled immediately) and
2401 // distribute them to the thread stacks. The iteration is done in reverse
2402 // order (high to low) so the regions will be removed in ascending order.
2403
2404 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2405
2406 // id + 1 is used to test termination so unsigned can
2407 // be used with an old_space_id == 0.
2408 FillableRegionLogger region_logger;
2409 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2410 SpaceInfo* const space_info = _space_info + id;
2411 MutableSpace* const space = space_info->space();
2412 HeapWord* const new_top = space_info->new_top();
2413
2414 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2415 const size_t end_region =
2416 sd.addr_to_region_idx(sd.region_align_up(new_top));
2417
2418 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2419 if (sd.region(cur)->claim_unsafe()) {
2420 ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id);
2421 bool result = sd.region(cur)->mark_normal();
2422 assert(result, "Must succeed at this point.");
2423 cm->region_stack()->push(cur);
2424 region_logger.handle(cur);
2425 // Assign regions to tasks in round-robin fashion.
2426 if (++worker_id == parallel_gc_threads) {
2427 worker_id = 0;
2428 }
2429 }
2430 }
2431 region_logger.print_line();
2432 }
2433 }
2434
2435 class TaskQueue : StackObj {
2436 volatile uint _counter;
2437 uint _size;
2438 uint _insert_index;
2439 PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2440 public:
TaskQueue(uint size)2441 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2442 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2443 }
~TaskQueue()2444 ~TaskQueue() {
2445 assert(_counter >= _insert_index, "not all queue elements were claimed");
2446 FREE_C_HEAP_ARRAY(T, _backing_array);
2447 }
2448
push(const PSParallelCompact::UpdateDensePrefixTask & value)2449 void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2450 assert(_insert_index < _size, "too small backing array");
2451 _backing_array[_insert_index++] = value;
2452 }
2453
try_claim(PSParallelCompact::UpdateDensePrefixTask & reference)2454 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2455 uint claimed = Atomic::add(&_counter, 1u) - 1; // -1 is so that we start with zero
2456 if (claimed < _insert_index) {
2457 reference = _backing_array[claimed];
2458 return true;
2459 } else {
2460 return false;
2461 }
2462 }
2463 };
2464
2465 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2466
enqueue_dense_prefix_tasks(TaskQueue & task_queue,uint parallel_gc_threads)2467 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2468 uint parallel_gc_threads) {
2469 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2470
2471 ParallelCompactData& sd = PSParallelCompact::summary_data();
2472
2473 // Iterate over all the spaces adding tasks for updating
2474 // regions in the dense prefix. Assume that 1 gc thread
2475 // will work on opening the gaps and the remaining gc threads
2476 // will work on the dense prefix.
2477 unsigned int space_id;
2478 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2479 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2480 const MutableSpace* const space = _space_info[space_id].space();
2481
2482 if (dense_prefix_end == space->bottom()) {
2483 // There is no dense prefix for this space.
2484 continue;
2485 }
2486
2487 // The dense prefix is before this region.
2488 size_t region_index_end_dense_prefix =
2489 sd.addr_to_region_idx(dense_prefix_end);
2490 RegionData* const dense_prefix_cp =
2491 sd.region(region_index_end_dense_prefix);
2492 assert(dense_prefix_end == space->end() ||
2493 dense_prefix_cp->available() ||
2494 dense_prefix_cp->claimed(),
2495 "The region after the dense prefix should always be ready to fill");
2496
2497 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2498
2499 // Is there dense prefix work?
2500 size_t total_dense_prefix_regions =
2501 region_index_end_dense_prefix - region_index_start;
2502 // How many regions of the dense prefix should be given to
2503 // each thread?
2504 if (total_dense_prefix_regions > 0) {
2505 uint tasks_for_dense_prefix = 1;
2506 if (total_dense_prefix_regions <=
2507 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2508 // Don't over partition. This assumes that
2509 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2510 // so there are not many regions to process.
2511 tasks_for_dense_prefix = parallel_gc_threads;
2512 } else {
2513 // Over partition
2514 tasks_for_dense_prefix = parallel_gc_threads *
2515 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2516 }
2517 size_t regions_per_thread = total_dense_prefix_regions /
2518 tasks_for_dense_prefix;
2519 // Give each thread at least 1 region.
2520 if (regions_per_thread == 0) {
2521 regions_per_thread = 1;
2522 }
2523
2524 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2525 if (region_index_start >= region_index_end_dense_prefix) {
2526 break;
2527 }
2528 // region_index_end is not processed
2529 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2530 region_index_end_dense_prefix);
2531 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2532 region_index_start,
2533 region_index_end));
2534 region_index_start = region_index_end;
2535 }
2536 }
2537 // This gets any part of the dense prefix that did not
2538 // fit evenly.
2539 if (region_index_start < region_index_end_dense_prefix) {
2540 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2541 region_index_start,
2542 region_index_end_dense_prefix));
2543 }
2544 }
2545 }
2546
2547 #ifdef ASSERT
2548 // Write a histogram of the number of times the block table was filled for a
2549 // region.
write_block_fill_histogram()2550 void PSParallelCompact::write_block_fill_histogram()
2551 {
2552 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2553 return;
2554 }
2555
2556 Log(gc, compaction) log;
2557 ResourceMark rm;
2558 LogStream ls(log.trace());
2559 outputStream* out = &ls;
2560
2561 typedef ParallelCompactData::RegionData rd_t;
2562 ParallelCompactData& sd = summary_data();
2563
2564 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2565 MutableSpace* const spc = _space_info[id].space();
2566 if (spc->bottom() != spc->top()) {
2567 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2568 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2569 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2570
2571 size_t histo[5] = { 0, 0, 0, 0, 0 };
2572 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2573 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2574
2575 for (const rd_t* cur = beg; cur < end; ++cur) {
2576 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2577 }
2578 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2579 for (size_t i = 0; i < histo_len; ++i) {
2580 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2581 histo[i], 100.0 * histo[i] / region_cnt);
2582 }
2583 out->cr();
2584 }
2585 }
2586 }
2587 #endif // #ifdef ASSERT
2588
compaction_with_stealing_work(ParallelTaskTerminator * terminator,uint worker_id)2589 static void compaction_with_stealing_work(ParallelTaskTerminator* terminator, uint worker_id) {
2590 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2591
2592 ParCompactionManager* cm =
2593 ParCompactionManager::gc_thread_compaction_manager(worker_id);
2594
2595 // Drain the stacks that have been preloaded with regions
2596 // that are ready to fill.
2597
2598 cm->drain_region_stacks();
2599
2600 guarantee(cm->region_stack()->is_empty(), "Not empty");
2601
2602 size_t region_index = 0;
2603
2604 while (true) {
2605 if (ParCompactionManager::steal(worker_id, region_index)) {
2606 PSParallelCompact::fill_and_update_region(cm, region_index);
2607 cm->drain_region_stacks();
2608 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2609 // Fill and update an unavailable region with the help of a shadow region
2610 PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2611 cm->drain_region_stacks();
2612 } else {
2613 if (terminator->offer_termination()) {
2614 break;
2615 }
2616 // Go around again.
2617 }
2618 }
2619 return;
2620 }
2621
2622 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
2623 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2624 TaskQueue& _tq;
2625 TaskTerminator _terminator;
2626 uint _active_workers;
2627
2628 public:
UpdateDensePrefixAndCompactionTask(TaskQueue & tq,uint active_workers)2629 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2630 AbstractGangTask("UpdateDensePrefixAndCompactionTask"),
2631 _tq(tq),
2632 _terminator(active_workers, ParCompactionManager::region_array()),
2633 _active_workers(active_workers) {
2634 }
work(uint worker_id)2635 virtual void work(uint worker_id) {
2636 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2637
2638 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2639 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2640 task._space_id,
2641 task._region_index_start,
2642 task._region_index_end);
2643 }
2644
2645 // Once a thread has drained it's stack, it should try to steal regions from
2646 // other threads.
2647 compaction_with_stealing_work(_terminator.terminator(), worker_id);
2648 }
2649 };
2650
compact()2651 void PSParallelCompact::compact() {
2652 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2653
2654 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2655 PSOldGen* old_gen = heap->old_gen();
2656 old_gen->start_array()->reset();
2657 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2658
2659 // for [0..last_space_id)
2660 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2661 // push
2662 // push
2663 //
2664 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2665 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2666 initialize_shadow_regions(active_gc_threads);
2667 prepare_region_draining_tasks(active_gc_threads);
2668 enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2669
2670 {
2671 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2672
2673 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2674 ParallelScavengeHeap::heap()->workers().run_task(&task);
2675
2676 #ifdef ASSERT
2677 // Verify that all regions have been processed before the deferred updates.
2678 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2679 verify_complete(SpaceId(id));
2680 }
2681 #endif
2682 }
2683
2684 {
2685 // Update the deferred objects, if any. Any compaction manager can be used.
2686 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2687 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2688 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2689 update_deferred_objects(cm, SpaceId(id));
2690 }
2691 }
2692
2693 DEBUG_ONLY(write_block_fill_histogram());
2694 }
2695
2696 #ifdef ASSERT
verify_complete(SpaceId space_id)2697 void PSParallelCompact::verify_complete(SpaceId space_id) {
2698 // All Regions between space bottom() to new_top() should be marked as filled
2699 // and all Regions between new_top() and top() should be available (i.e.,
2700 // should have been emptied).
2701 ParallelCompactData& sd = summary_data();
2702 SpaceInfo si = _space_info[space_id];
2703 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2704 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2705 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2706 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2707 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2708
2709 bool issued_a_warning = false;
2710
2711 size_t cur_region;
2712 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2713 const RegionData* const c = sd.region(cur_region);
2714 if (!c->completed()) {
2715 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2716 cur_region, c->destination_count());
2717 issued_a_warning = true;
2718 }
2719 }
2720
2721 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2722 const RegionData* const c = sd.region(cur_region);
2723 if (!c->available()) {
2724 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2725 cur_region, c->destination_count());
2726 issued_a_warning = true;
2727 }
2728 }
2729
2730 if (issued_a_warning) {
2731 print_region_ranges();
2732 }
2733 }
2734 #endif // #ifdef ASSERT
2735
do_addr(HeapWord * addr)2736 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2737 _start_array->allocate_block(addr);
2738 compaction_manager()->update_contents(oop(addr));
2739 }
2740
2741 // Update interior oops in the ranges of regions [beg_region, end_region).
2742 void
update_and_deadwood_in_dense_prefix(ParCompactionManager * cm,SpaceId space_id,size_t beg_region,size_t end_region)2743 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2744 SpaceId space_id,
2745 size_t beg_region,
2746 size_t end_region) {
2747 ParallelCompactData& sd = summary_data();
2748 ParMarkBitMap* const mbm = mark_bitmap();
2749
2750 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2751 HeapWord* const end_addr = sd.region_to_addr(end_region);
2752 assert(beg_region <= end_region, "bad region range");
2753 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2754
2755 #ifdef ASSERT
2756 // Claim the regions to avoid triggering an assert when they are marked as
2757 // filled.
2758 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2759 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2760 }
2761 #endif // #ifdef ASSERT
2762
2763 if (beg_addr != space(space_id)->bottom()) {
2764 // Find the first live object or block of dead space that *starts* in this
2765 // range of regions. If a partial object crosses onto the region, skip it;
2766 // it will be marked for 'deferred update' when the object head is
2767 // processed. If dead space crosses onto the region, it is also skipped; it
2768 // will be filled when the prior region is processed. If neither of those
2769 // apply, the first word in the region is the start of a live object or dead
2770 // space.
2771 assert(beg_addr > space(space_id)->bottom(), "sanity");
2772 const RegionData* const cp = sd.region(beg_region);
2773 if (cp->partial_obj_size() != 0) {
2774 beg_addr = sd.partial_obj_end(beg_region);
2775 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2776 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2777 }
2778 }
2779
2780 if (beg_addr < end_addr) {
2781 // A live object or block of dead space starts in this range of Regions.
2782 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2783
2784 // Create closures and iterate.
2785 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2786 FillClosure fill_closure(cm, space_id);
2787 ParMarkBitMap::IterationStatus status;
2788 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2789 dense_prefix_end);
2790 if (status == ParMarkBitMap::incomplete) {
2791 update_closure.do_addr(update_closure.source());
2792 }
2793 }
2794
2795 // Mark the regions as filled.
2796 RegionData* const beg_cp = sd.region(beg_region);
2797 RegionData* const end_cp = sd.region(end_region);
2798 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2799 cp->set_completed();
2800 }
2801 }
2802
2803 // Return the SpaceId for the space containing addr. If addr is not in the
2804 // heap, last_space_id is returned. In debug mode it expects the address to be
2805 // in the heap and asserts such.
space_id(HeapWord * addr)2806 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2807 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2808
2809 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2810 if (_space_info[id].space()->contains(addr)) {
2811 return SpaceId(id);
2812 }
2813 }
2814
2815 assert(false, "no space contains the addr");
2816 return last_space_id;
2817 }
2818
update_deferred_objects(ParCompactionManager * cm,SpaceId id)2819 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2820 SpaceId id) {
2821 assert(id < last_space_id, "bad space id");
2822
2823 ParallelCompactData& sd = summary_data();
2824 const SpaceInfo* const space_info = _space_info + id;
2825 ObjectStartArray* const start_array = space_info->start_array();
2826
2827 const MutableSpace* const space = space_info->space();
2828 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2829 HeapWord* const beg_addr = space_info->dense_prefix();
2830 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2831
2832 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2833 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2834 const RegionData* cur_region;
2835 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2836 HeapWord* const addr = cur_region->deferred_obj_addr();
2837 if (addr != NULL) {
2838 if (start_array != NULL) {
2839 start_array->allocate_block(addr);
2840 }
2841 cm->update_contents(oop(addr));
2842 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2843 }
2844 }
2845 }
2846
2847 // Skip over count live words starting from beg, and return the address of the
2848 // next live word. Unless marked, the word corresponding to beg is assumed to
2849 // be dead. Callers must either ensure beg does not correspond to the middle of
2850 // an object, or account for those live words in some other way. Callers must
2851 // also ensure that there are enough live words in the range [beg, end) to skip.
2852 HeapWord*
skip_live_words(HeapWord * beg,HeapWord * end,size_t count)2853 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2854 {
2855 assert(count > 0, "sanity");
2856
2857 ParMarkBitMap* m = mark_bitmap();
2858 idx_t bits_to_skip = m->words_to_bits(count);
2859 idx_t cur_beg = m->addr_to_bit(beg);
2860 const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2861
2862 do {
2863 cur_beg = m->find_obj_beg(cur_beg, search_end);
2864 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2865 const size_t obj_bits = cur_end - cur_beg + 1;
2866 if (obj_bits > bits_to_skip) {
2867 return m->bit_to_addr(cur_beg + bits_to_skip);
2868 }
2869 bits_to_skip -= obj_bits;
2870 cur_beg = cur_end + 1;
2871 } while (bits_to_skip > 0);
2872
2873 // Skipping the desired number of words landed just past the end of an object.
2874 // Find the start of the next object.
2875 cur_beg = m->find_obj_beg(cur_beg, search_end);
2876 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2877 return m->bit_to_addr(cur_beg);
2878 }
2879
first_src_addr(HeapWord * const dest_addr,SpaceId src_space_id,size_t src_region_idx)2880 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2881 SpaceId src_space_id,
2882 size_t src_region_idx)
2883 {
2884 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2885
2886 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2887 if (split_info.dest_region_addr() == dest_addr) {
2888 // The partial object ending at the split point contains the first word to
2889 // be copied to dest_addr.
2890 return split_info.first_src_addr();
2891 }
2892
2893 const ParallelCompactData& sd = summary_data();
2894 ParMarkBitMap* const bitmap = mark_bitmap();
2895 const size_t RegionSize = ParallelCompactData::RegionSize;
2896
2897 assert(sd.is_region_aligned(dest_addr), "not aligned");
2898 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2899 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2900 HeapWord* const src_region_destination = src_region_ptr->destination();
2901
2902 assert(dest_addr >= src_region_destination, "wrong src region");
2903 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2904
2905 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2906 HeapWord* const src_region_end = src_region_beg + RegionSize;
2907
2908 HeapWord* addr = src_region_beg;
2909 if (dest_addr == src_region_destination) {
2910 // Return the first live word in the source region.
2911 if (partial_obj_size == 0) {
2912 addr = bitmap->find_obj_beg(addr, src_region_end);
2913 assert(addr < src_region_end, "no objects start in src region");
2914 }
2915 return addr;
2916 }
2917
2918 // Must skip some live data.
2919 size_t words_to_skip = dest_addr - src_region_destination;
2920 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2921
2922 if (partial_obj_size >= words_to_skip) {
2923 // All the live words to skip are part of the partial object.
2924 addr += words_to_skip;
2925 if (partial_obj_size == words_to_skip) {
2926 // Find the first live word past the partial object.
2927 addr = bitmap->find_obj_beg(addr, src_region_end);
2928 assert(addr < src_region_end, "wrong src region");
2929 }
2930 return addr;
2931 }
2932
2933 // Skip over the partial object (if any).
2934 if (partial_obj_size != 0) {
2935 words_to_skip -= partial_obj_size;
2936 addr += partial_obj_size;
2937 }
2938
2939 // Skip over live words due to objects that start in the region.
2940 addr = skip_live_words(addr, src_region_end, words_to_skip);
2941 assert(addr < src_region_end, "wrong src region");
2942 return addr;
2943 }
2944
decrement_destination_counts(ParCompactionManager * cm,SpaceId src_space_id,size_t beg_region,HeapWord * end_addr)2945 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2946 SpaceId src_space_id,
2947 size_t beg_region,
2948 HeapWord* end_addr)
2949 {
2950 ParallelCompactData& sd = summary_data();
2951
2952 #ifdef ASSERT
2953 MutableSpace* const src_space = _space_info[src_space_id].space();
2954 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2955 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2956 "src_space_id does not match beg_addr");
2957 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2958 "src_space_id does not match end_addr");
2959 #endif // #ifdef ASSERT
2960
2961 RegionData* const beg = sd.region(beg_region);
2962 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2963
2964 // Regions up to new_top() are enqueued if they become available.
2965 HeapWord* const new_top = _space_info[src_space_id].new_top();
2966 RegionData* const enqueue_end =
2967 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2968
2969 for (RegionData* cur = beg; cur < end; ++cur) {
2970 assert(cur->data_size() > 0, "region must have live data");
2971 cur->decrement_destination_count();
2972 if (cur < enqueue_end && cur->available() && cur->claim()) {
2973 if (cur->mark_normal()) {
2974 cm->push_region(sd.region(cur));
2975 } else if (cur->mark_copied()) {
2976 // Try to copy the content of the shadow region back to its corresponding
2977 // heap region if the shadow region is filled. Otherwise, the GC thread
2978 // fills the shadow region will copy the data back (see
2979 // MoveAndUpdateShadowClosure::complete_region).
2980 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2981 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2982 cur->set_completed();
2983 }
2984 }
2985 }
2986 }
2987
next_src_region(MoveAndUpdateClosure & closure,SpaceId & src_space_id,HeapWord * & src_space_top,HeapWord * end_addr)2988 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2989 SpaceId& src_space_id,
2990 HeapWord*& src_space_top,
2991 HeapWord* end_addr)
2992 {
2993 typedef ParallelCompactData::RegionData RegionData;
2994
2995 ParallelCompactData& sd = PSParallelCompact::summary_data();
2996 const size_t region_size = ParallelCompactData::RegionSize;
2997
2998 size_t src_region_idx = 0;
2999
3000 // Skip empty regions (if any) up to the top of the space.
3001 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3002 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3003 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3004 const RegionData* const top_region_ptr =
3005 sd.addr_to_region_ptr(top_aligned_up);
3006 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3007 ++src_region_ptr;
3008 }
3009
3010 if (src_region_ptr < top_region_ptr) {
3011 // The next source region is in the current space. Update src_region_idx
3012 // and the source address to match src_region_ptr.
3013 src_region_idx = sd.region(src_region_ptr);
3014 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3015 if (src_region_addr > closure.source()) {
3016 closure.set_source(src_region_addr);
3017 }
3018 return src_region_idx;
3019 }
3020
3021 // Switch to a new source space and find the first non-empty region.
3022 unsigned int space_id = src_space_id + 1;
3023 assert(space_id < last_space_id, "not enough spaces");
3024
3025 HeapWord* const destination = closure.destination();
3026
3027 do {
3028 MutableSpace* space = _space_info[space_id].space();
3029 HeapWord* const bottom = space->bottom();
3030 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3031
3032 // Iterate over the spaces that do not compact into themselves.
3033 if (bottom_cp->destination() != bottom) {
3034 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3035 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3036
3037 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3038 if (src_cp->live_obj_size() > 0) {
3039 // Found it.
3040 assert(src_cp->destination() == destination,
3041 "first live obj in the space must match the destination");
3042 assert(src_cp->partial_obj_size() == 0,
3043 "a space cannot begin with a partial obj");
3044
3045 src_space_id = SpaceId(space_id);
3046 src_space_top = space->top();
3047 const size_t src_region_idx = sd.region(src_cp);
3048 closure.set_source(sd.region_to_addr(src_region_idx));
3049 return src_region_idx;
3050 } else {
3051 assert(src_cp->data_size() == 0, "sanity");
3052 }
3053 }
3054 }
3055 } while (++space_id < last_space_id);
3056
3057 assert(false, "no source region was found");
3058 return 0;
3059 }
3060
fill_region(ParCompactionManager * cm,MoveAndUpdateClosure & closure,size_t region_idx)3061 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
3062 {
3063 typedef ParMarkBitMap::IterationStatus IterationStatus;
3064 ParMarkBitMap* const bitmap = mark_bitmap();
3065 ParallelCompactData& sd = summary_data();
3066 RegionData* const region_ptr = sd.region(region_idx);
3067
3068 // Get the source region and related info.
3069 size_t src_region_idx = region_ptr->source_region();
3070 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3071 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3072 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3073
3074 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3075
3076 // Adjust src_region_idx to prepare for decrementing destination counts (the
3077 // destination count is not decremented when a region is copied to itself).
3078 if (src_region_idx == region_idx) {
3079 src_region_idx += 1;
3080 }
3081
3082 if (bitmap->is_unmarked(closure.source())) {
3083 // The first source word is in the middle of an object; copy the remainder
3084 // of the object or as much as will fit. The fact that pointer updates were
3085 // deferred will be noted when the object header is processed.
3086 HeapWord* const old_src_addr = closure.source();
3087 closure.copy_partial_obj();
3088 if (closure.is_full()) {
3089 decrement_destination_counts(cm, src_space_id, src_region_idx,
3090 closure.source());
3091 region_ptr->set_deferred_obj_addr(NULL);
3092 closure.complete_region(cm, dest_addr, region_ptr);
3093 return;
3094 }
3095
3096 HeapWord* const end_addr = sd.region_align_down(closure.source());
3097 if (sd.region_align_down(old_src_addr) != end_addr) {
3098 // The partial object was copied from more than one source region.
3099 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3100
3101 // Move to the next source region, possibly switching spaces as well. All
3102 // args except end_addr may be modified.
3103 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3104 end_addr);
3105 }
3106 }
3107
3108 do {
3109 HeapWord* const cur_addr = closure.source();
3110 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3111 src_space_top);
3112 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3113
3114 if (status == ParMarkBitMap::incomplete) {
3115 // The last obj that starts in the source region does not end in the
3116 // region.
3117 assert(closure.source() < end_addr, "sanity");
3118 HeapWord* const obj_beg = closure.source();
3119 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3120 src_space_top);
3121 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3122 if (obj_end < range_end) {
3123 // The end was found; the entire object will fit.
3124 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3125 assert(status != ParMarkBitMap::would_overflow, "sanity");
3126 } else {
3127 // The end was not found; the object will not fit.
3128 assert(range_end < src_space_top, "obj cannot cross space boundary");
3129 status = ParMarkBitMap::would_overflow;
3130 }
3131 }
3132
3133 if (status == ParMarkBitMap::would_overflow) {
3134 // The last object did not fit. Note that interior oop updates were
3135 // deferred, then copy enough of the object to fill the region.
3136 region_ptr->set_deferred_obj_addr(closure.destination());
3137 status = closure.copy_until_full(); // copies from closure.source()
3138
3139 decrement_destination_counts(cm, src_space_id, src_region_idx,
3140 closure.source());
3141 closure.complete_region(cm, dest_addr, region_ptr);
3142 return;
3143 }
3144
3145 if (status == ParMarkBitMap::full) {
3146 decrement_destination_counts(cm, src_space_id, src_region_idx,
3147 closure.source());
3148 region_ptr->set_deferred_obj_addr(NULL);
3149 closure.complete_region(cm, dest_addr, region_ptr);
3150 return;
3151 }
3152
3153 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3154
3155 // Move to the next source region, possibly switching spaces as well. All
3156 // args except end_addr may be modified.
3157 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3158 end_addr);
3159 } while (true);
3160 }
3161
fill_and_update_region(ParCompactionManager * cm,size_t region_idx)3162 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3163 {
3164 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3165 fill_region(cm, cl, region_idx);
3166 }
3167
fill_and_update_shadow_region(ParCompactionManager * cm,size_t region_idx)3168 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3169 {
3170 // Get a shadow region first
3171 ParallelCompactData& sd = summary_data();
3172 RegionData* const region_ptr = sd.region(region_idx);
3173 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3174 // The InvalidShadow return value indicates the corresponding heap region is available,
3175 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3176 // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3177 if (shadow_region == ParCompactionManager::InvalidShadow) {
3178 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3179 region_ptr->shadow_to_normal();
3180 return fill_region(cm, cl, region_idx);
3181 } else {
3182 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3183 return fill_region(cm, cl, region_idx);
3184 }
3185 }
3186
copy_back(HeapWord * shadow_addr,HeapWord * region_addr)3187 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3188 {
3189 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3190 }
3191
steal_unavailable_region(ParCompactionManager * cm,size_t & region_idx)3192 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx)
3193 {
3194 size_t next = cm->next_shadow_region();
3195 ParallelCompactData& sd = summary_data();
3196 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3197 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3198
3199 while (next < old_new_top) {
3200 if (sd.region(next)->mark_shadow()) {
3201 region_idx = next;
3202 return true;
3203 }
3204 next = cm->move_next_shadow_region_by(active_gc_threads);
3205 }
3206
3207 return false;
3208 }
3209
3210 // The shadow region is an optimization to address region dependencies in full GC. The basic
3211 // idea is making more regions available by temporally storing their live objects in empty
3212 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3213 // GC threads need not wait destination regions to be available before processing sources.
3214 //
3215 // A typical workflow would be:
3216 // After draining its own stack and failing to steal from others, a GC worker would pick an
3217 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3218 // the shadow region by copying live objects from source regions of the unavailable one. Once
3219 // the unavailable region becomes available, the data in the shadow region will be copied back.
3220 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3221 //
3222 // For more details, please refer to §4.2 of the VEE'19 paper:
3223 // Haoyu Li, Mingyu Wu, Binyu Zang, and Haibo Chen. 2019. ScissorGC: scalable and efficient
3224 // compaction for Java full garbage collection. In Proceedings of the 15th ACM SIGPLAN/SIGOPS
3225 // International Conference on Virtual Execution Environments (VEE 2019). ACM, New York, NY, USA,
3226 // 108-121. DOI: https://doi.org/10.1145/3313808.3313820
initialize_shadow_regions(uint parallel_gc_threads)3227 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3228 {
3229 const ParallelCompactData& sd = PSParallelCompact::summary_data();
3230
3231 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3232 SpaceInfo* const space_info = _space_info + id;
3233 MutableSpace* const space = space_info->space();
3234
3235 const size_t beg_region =
3236 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3237 const size_t end_region =
3238 sd.addr_to_region_idx(sd.region_align_down(space->end()));
3239
3240 for (size_t cur = beg_region; cur < end_region; ++cur) {
3241 ParCompactionManager::push_shadow_region(cur);
3242 }
3243 }
3244
3245 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3246 for (uint i = 0; i < parallel_gc_threads; i++) {
3247 ParCompactionManager *cm = ParCompactionManager::manager_array(i);
3248 cm->set_next_shadow_region(beg_region + i);
3249 }
3250 }
3251
fill_blocks(size_t region_idx)3252 void PSParallelCompact::fill_blocks(size_t region_idx)
3253 {
3254 // Fill in the block table elements for the specified region. Each block
3255 // table element holds the number of live words in the region that are to the
3256 // left of the first object that starts in the block. Thus only blocks in
3257 // which an object starts need to be filled.
3258 //
3259 // The algorithm scans the section of the bitmap that corresponds to the
3260 // region, keeping a running total of the live words. When an object start is
3261 // found, if it's the first to start in the block that contains it, the
3262 // current total is written to the block table element.
3263 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3264 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3265 const size_t RegionSize = ParallelCompactData::RegionSize;
3266
3267 ParallelCompactData& sd = summary_data();
3268 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3269 if (partial_obj_size >= RegionSize) {
3270 return; // No objects start in this region.
3271 }
3272
3273 // Ensure the first loop iteration decides that the block has changed.
3274 size_t cur_block = sd.block_count();
3275
3276 const ParMarkBitMap* const bitmap = mark_bitmap();
3277
3278 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3279 assert((size_t)1 << Log2BitsPerBlock ==
3280 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3281
3282 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3283 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3284 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3285 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3286 while (beg_bit < range_end) {
3287 const size_t new_block = beg_bit >> Log2BitsPerBlock;
3288 if (new_block != cur_block) {
3289 cur_block = new_block;
3290 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3291 }
3292
3293 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3294 if (end_bit < range_end - 1) {
3295 live_bits += end_bit - beg_bit + 1;
3296 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3297 } else {
3298 return;
3299 }
3300 }
3301 }
3302
millis_since_last_gc()3303 jlong PSParallelCompact::millis_since_last_gc() {
3304 // We need a monotonically non-decreasing time in ms but
3305 // os::javaTimeMillis() does not guarantee monotonicity.
3306 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3307 jlong ret_val = now - _time_of_last_gc;
3308 // XXX See note in genCollectedHeap::millis_since_last_gc().
3309 if (ret_val < 0) {
3310 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);)
3311 return 0;
3312 }
3313 return ret_val;
3314 }
3315
reset_millis_since_last_gc()3316 void PSParallelCompact::reset_millis_since_last_gc() {
3317 // We need a monotonically non-decreasing time in ms but
3318 // os::javaTimeMillis() does not guarantee monotonicity.
3319 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3320 }
3321
copy_until_full()3322 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3323 {
3324 if (source() != copy_destination()) {
3325 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3326 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3327 }
3328 update_state(words_remaining());
3329 assert(is_full(), "sanity");
3330 return ParMarkBitMap::full;
3331 }
3332
copy_partial_obj()3333 void MoveAndUpdateClosure::copy_partial_obj()
3334 {
3335 size_t words = words_remaining();
3336
3337 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3338 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3339 if (end_addr < range_end) {
3340 words = bitmap()->obj_size(source(), end_addr);
3341 }
3342
3343 // This test is necessary; if omitted, the pointer updates to a partial object
3344 // that crosses the dense prefix boundary could be overwritten.
3345 if (source() != copy_destination()) {
3346 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3347 Copy::aligned_conjoint_words(source(), copy_destination(), words);
3348 }
3349 update_state(words);
3350 }
3351
complete_region(ParCompactionManager * cm,HeapWord * dest_addr,PSParallelCompact::RegionData * region_ptr)3352 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3353 PSParallelCompact::RegionData *region_ptr) {
3354 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3355 region_ptr->set_completed();
3356 }
3357
3358 ParMarkBitMapClosure::IterationStatus
do_addr(HeapWord * addr,size_t words)3359 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3360 assert(destination() != NULL, "sanity");
3361 assert(bitmap()->obj_size(addr) == words, "bad size");
3362
3363 _source = addr;
3364 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3365 destination(), "wrong destination");
3366
3367 if (words > words_remaining()) {
3368 return ParMarkBitMap::would_overflow;
3369 }
3370
3371 // The start_array must be updated even if the object is not moving.
3372 if (_start_array != NULL) {
3373 _start_array->allocate_block(destination());
3374 }
3375
3376 if (copy_destination() != source()) {
3377 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3378 Copy::aligned_conjoint_words(source(), copy_destination(), words);
3379 }
3380
3381 oop moved_oop = (oop) copy_destination();
3382 compaction_manager()->update_contents(moved_oop);
3383 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3384
3385 update_state(words);
3386 assert(copy_destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3387 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3388 }
3389
complete_region(ParCompactionManager * cm,HeapWord * dest_addr,PSParallelCompact::RegionData * region_ptr)3390 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3391 PSParallelCompact::RegionData *region_ptr) {
3392 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3393 // Record the shadow region index
3394 region_ptr->set_shadow_region(_shadow);
3395 // Mark the shadow region as filled to indicate the data is ready to be
3396 // copied back
3397 region_ptr->mark_filled();
3398 // Try to copy the content of the shadow region back to its corresponding
3399 // heap region if available; the GC thread that decreases the destination
3400 // count to zero will do the copying otherwise (see
3401 // PSParallelCompact::decrement_destination_counts).
3402 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3403 region_ptr->set_completed();
3404 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3405 ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3406 }
3407 }
3408
UpdateOnlyClosure(ParMarkBitMap * mbm,ParCompactionManager * cm,PSParallelCompact::SpaceId space_id)3409 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3410 ParCompactionManager* cm,
3411 PSParallelCompact::SpaceId space_id) :
3412 ParMarkBitMapClosure(mbm, cm),
3413 _space_id(space_id),
3414 _start_array(PSParallelCompact::start_array(space_id))
3415 {
3416 }
3417
3418 // Updates the references in the object to their new values.
3419 ParMarkBitMapClosure::IterationStatus
do_addr(HeapWord * addr,size_t words)3420 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3421 do_addr(addr);
3422 return ParMarkBitMap::incomplete;
3423 }
3424
FillClosure(ParCompactionManager * cm,PSParallelCompact::SpaceId space_id)3425 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3426 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3427 _start_array(PSParallelCompact::start_array(space_id))
3428 {
3429 assert(space_id == PSParallelCompact::old_space_id,
3430 "cannot use FillClosure in the young gen");
3431 }
3432
3433 ParMarkBitMapClosure::IterationStatus
do_addr(HeapWord * addr,size_t size)3434 FillClosure::do_addr(HeapWord* addr, size_t size) {
3435 CollectedHeap::fill_with_objects(addr, size);
3436 HeapWord* const end = addr + size;
3437 do {
3438 _start_array->allocate_block(addr);
3439 addr += oop(addr)->size();
3440 } while (addr < end);
3441 return ParMarkBitMap::incomplete;
3442 }
3443