1 /*
2 * Copyright (c) 2005, 2018, 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/gcTaskManager.hpp"
34 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
35 #include "gc/parallel/parMarkBitMap.inline.hpp"
36 #include "gc/parallel/pcTasks.hpp"
37 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
38 #include "gc/parallel/psCompactionManager.inline.hpp"
39 #include "gc/parallel/psOldGen.hpp"
40 #include "gc/parallel/psParallelCompact.inline.hpp"
41 #include "gc/parallel/psPromotionManager.inline.hpp"
42 #include "gc/parallel/psScavenge.hpp"
43 #include "gc/parallel/psYoungGen.hpp"
44 #include "gc/shared/gcCause.hpp"
45 #include "gc/shared/gcHeapSummary.hpp"
46 #include "gc/shared/gcId.hpp"
47 #include "gc/shared/gcLocker.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.hpp"
56 #include "gc/shared/weakProcessor.hpp"
57 #include "logging/log.hpp"
58 #include "memory/iterator.inline.hpp"
59 #include "memory/resourceArea.hpp"
60 #include "oops/access.inline.hpp"
61 #include "oops/instanceClassLoaderKlass.inline.hpp"
62 #include "oops/instanceKlass.inline.hpp"
63 #include "oops/instanceMirrorKlass.inline.hpp"
64 #include "oops/methodData.hpp"
65 #include "oops/objArrayKlass.inline.hpp"
66 #include "oops/oop.inline.hpp"
67 #include "runtime/atomic.hpp"
68 #include "runtime/handles.inline.hpp"
69 #include "runtime/safepoint.hpp"
70 #include "runtime/vmThread.hpp"
71 #include "services/management.hpp"
72 #include "services/memTracker.hpp"
73 #include "services/memoryService.hpp"
74 #include "utilities/align.hpp"
75 #include "utilities/debug.hpp"
76 #include "utilities/events.hpp"
77 #include "utilities/formatBuffer.hpp"
78 #include "utilities/macros.hpp"
79 #include "utilities/stack.inline.hpp"
80
81 #include <math.h>
82
83 // All sizes are in HeapWords.
84 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words
85 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
86 const size_t ParallelCompactData::RegionSizeBytes =
87 RegionSize << LogHeapWordSize;
88 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
89 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
90 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
91
92 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
93 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
94 const size_t ParallelCompactData::BlockSizeBytes =
95 BlockSize << LogHeapWordSize;
96 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
97 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
98 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask;
99
100 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
101 const size_t ParallelCompactData::Log2BlocksPerRegion =
102 Log2RegionSize - Log2BlockSize;
103
104 const ParallelCompactData::RegionData::region_sz_t
105 ParallelCompactData::RegionData::dc_shift = 27;
106
107 const ParallelCompactData::RegionData::region_sz_t
108 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
109
110 const ParallelCompactData::RegionData::region_sz_t
111 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
112
113 const ParallelCompactData::RegionData::region_sz_t
114 ParallelCompactData::RegionData::los_mask = ~dc_mask;
115
116 const ParallelCompactData::RegionData::region_sz_t
117 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
118
119 const ParallelCompactData::RegionData::region_sz_t
120 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
121
122 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
123
124 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
125 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
126
127 double PSParallelCompact::_dwl_mean;
128 double PSParallelCompact::_dwl_std_dev;
129 double PSParallelCompact::_dwl_first_term;
130 double PSParallelCompact::_dwl_adjustment;
131 #ifdef ASSERT
132 bool PSParallelCompact::_dwl_initialized = false;
133 #endif // #ifdef ASSERT
134
record(size_t src_region_idx,size_t partial_obj_size,HeapWord * destination)135 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
136 HeapWord* destination)
137 {
138 assert(src_region_idx != 0, "invalid src_region_idx");
139 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
140 assert(destination != NULL, "invalid destination argument");
141
142 _src_region_idx = src_region_idx;
143 _partial_obj_size = partial_obj_size;
144 _destination = destination;
145
146 // These fields may not be updated below, so make sure they're clear.
147 assert(_dest_region_addr == NULL, "should have been cleared");
148 assert(_first_src_addr == NULL, "should have been cleared");
149
150 // Determine the number of destination regions for the partial object.
151 HeapWord* const last_word = destination + partial_obj_size - 1;
152 const ParallelCompactData& sd = PSParallelCompact::summary_data();
153 HeapWord* const beg_region_addr = sd.region_align_down(destination);
154 HeapWord* const end_region_addr = sd.region_align_down(last_word);
155
156 if (beg_region_addr == end_region_addr) {
157 // One destination region.
158 _destination_count = 1;
159 if (end_region_addr == destination) {
160 // The destination falls on a region boundary, thus the first word of the
161 // partial object will be the first word copied to the destination region.
162 _dest_region_addr = end_region_addr;
163 _first_src_addr = sd.region_to_addr(src_region_idx);
164 }
165 } else {
166 // Two destination regions. When copied, the partial object will cross a
167 // destination region boundary, so a word somewhere within the partial
168 // object will be the first word copied to the second destination region.
169 _destination_count = 2;
170 _dest_region_addr = end_region_addr;
171 const size_t ofs = pointer_delta(end_region_addr, destination);
172 assert(ofs < _partial_obj_size, "sanity");
173 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
174 }
175 }
176
clear()177 void SplitInfo::clear()
178 {
179 _src_region_idx = 0;
180 _partial_obj_size = 0;
181 _destination = NULL;
182 _destination_count = 0;
183 _dest_region_addr = NULL;
184 _first_src_addr = NULL;
185 assert(!is_valid(), "sanity");
186 }
187
188 #ifdef ASSERT
verify_clear()189 void SplitInfo::verify_clear()
190 {
191 assert(_src_region_idx == 0, "not clear");
192 assert(_partial_obj_size == 0, "not clear");
193 assert(_destination == NULL, "not clear");
194 assert(_destination_count == 0, "not clear");
195 assert(_dest_region_addr == NULL, "not clear");
196 assert(_first_src_addr == NULL, "not clear");
197 }
198 #endif // #ifdef ASSERT
199
200
print_on_error(outputStream * st)201 void PSParallelCompact::print_on_error(outputStream* st) {
202 _mark_bitmap.print_on_error(st);
203 }
204
205 #ifndef PRODUCT
206 const char* PSParallelCompact::space_names[] = {
207 "old ", "eden", "from", "to "
208 };
209
print_region_ranges()210 void PSParallelCompact::print_region_ranges() {
211 if (!log_develop_is_enabled(Trace, gc, compaction)) {
212 return;
213 }
214 Log(gc, compaction) log;
215 ResourceMark rm;
216 LogStream ls(log.trace());
217 Universe::print_on(&ls);
218 log.trace("space bottom top end new_top");
219 log.trace("------ ---------- ---------- ---------- ----------");
220
221 for (unsigned int id = 0; id < last_space_id; ++id) {
222 const MutableSpace* space = _space_info[id].space();
223 log.trace("%u %s "
224 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
225 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
226 id, space_names[id],
227 summary_data().addr_to_region_idx(space->bottom()),
228 summary_data().addr_to_region_idx(space->top()),
229 summary_data().addr_to_region_idx(space->end()),
230 summary_data().addr_to_region_idx(_space_info[id].new_top()));
231 }
232 }
233
234 void
print_generic_summary_region(size_t i,const ParallelCompactData::RegionData * c)235 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
236 {
237 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
238 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
239
240 ParallelCompactData& sd = PSParallelCompact::summary_data();
241 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
242 log_develop_trace(gc, compaction)(
243 REGION_IDX_FORMAT " " PTR_FORMAT " "
244 REGION_IDX_FORMAT " " PTR_FORMAT " "
245 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
246 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
247 i, p2i(c->data_location()), dci, p2i(c->destination()),
248 c->partial_obj_size(), c->live_obj_size(),
249 c->data_size(), c->source_region(), c->destination_count());
250
251 #undef REGION_IDX_FORMAT
252 #undef REGION_DATA_FORMAT
253 }
254
255 void
print_generic_summary_data(ParallelCompactData & summary_data,HeapWord * const beg_addr,HeapWord * const end_addr)256 print_generic_summary_data(ParallelCompactData& summary_data,
257 HeapWord* const beg_addr,
258 HeapWord* const end_addr)
259 {
260 size_t total_words = 0;
261 size_t i = summary_data.addr_to_region_idx(beg_addr);
262 const size_t last = summary_data.addr_to_region_idx(end_addr);
263 HeapWord* pdest = 0;
264
265 while (i < last) {
266 ParallelCompactData::RegionData* c = summary_data.region(i);
267 if (c->data_size() != 0 || c->destination() != pdest) {
268 print_generic_summary_region(i, c);
269 total_words += c->data_size();
270 pdest = c->destination();
271 }
272 ++i;
273 }
274
275 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
276 }
277
278 void
print_generic_summary_data(ParallelCompactData & summary_data,HeapWord * const beg_addr,HeapWord * const end_addr)279 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
280 HeapWord* const beg_addr,
281 HeapWord* const end_addr) {
282 ::print_generic_summary_data(summary_data,beg_addr, end_addr);
283 }
284
285 void
print_generic_summary_data(ParallelCompactData & summary_data,SpaceInfo * space_info)286 print_generic_summary_data(ParallelCompactData& summary_data,
287 SpaceInfo* space_info)
288 {
289 if (!log_develop_is_enabled(Trace, gc, compaction)) {
290 return;
291 }
292
293 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
294 const MutableSpace* space = space_info[id].space();
295 print_generic_summary_data(summary_data, space->bottom(),
296 MAX2(space->top(), space_info[id].new_top()));
297 }
298 }
299
300 void
print_initial_summary_data(ParallelCompactData & summary_data,const MutableSpace * space)301 print_initial_summary_data(ParallelCompactData& summary_data,
302 const MutableSpace* space) {
303 if (space->top() == space->bottom()) {
304 return;
305 }
306
307 const size_t region_size = ParallelCompactData::RegionSize;
308 typedef ParallelCompactData::RegionData RegionData;
309 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
310 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
311 const RegionData* c = summary_data.region(end_region - 1);
312 HeapWord* end_addr = c->destination() + c->data_size();
313 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
314
315 // Print (and count) the full regions at the beginning of the space.
316 size_t full_region_count = 0;
317 size_t i = summary_data.addr_to_region_idx(space->bottom());
318 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
319 ParallelCompactData::RegionData* c = summary_data.region(i);
320 log_develop_trace(gc, compaction)(
321 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
322 i, p2i(c->destination()),
323 c->partial_obj_size(), c->live_obj_size(),
324 c->data_size(), c->source_region(), c->destination_count());
325 ++full_region_count;
326 ++i;
327 }
328
329 size_t live_to_right = live_in_space - full_region_count * region_size;
330
331 double max_reclaimed_ratio = 0.0;
332 size_t max_reclaimed_ratio_region = 0;
333 size_t max_dead_to_right = 0;
334 size_t max_live_to_right = 0;
335
336 // Print the 'reclaimed ratio' for regions while there is something live in
337 // the region or to the right of it. The remaining regions are empty (and
338 // uninteresting), and computing the ratio will result in division by 0.
339 while (i < end_region && live_to_right > 0) {
340 c = summary_data.region(i);
341 HeapWord* const region_addr = summary_data.region_to_addr(i);
342 const size_t used_to_right = pointer_delta(space->top(), region_addr);
343 const size_t dead_to_right = used_to_right - live_to_right;
344 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
345
346 if (reclaimed_ratio > max_reclaimed_ratio) {
347 max_reclaimed_ratio = reclaimed_ratio;
348 max_reclaimed_ratio_region = i;
349 max_dead_to_right = dead_to_right;
350 max_live_to_right = live_to_right;
351 }
352
353 ParallelCompactData::RegionData* c = summary_data.region(i);
354 log_develop_trace(gc, compaction)(
355 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
356 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
357 i, p2i(c->destination()),
358 c->partial_obj_size(), c->live_obj_size(),
359 c->data_size(), c->source_region(), c->destination_count(),
360 reclaimed_ratio, dead_to_right, live_to_right);
361
362
363 live_to_right -= c->data_size();
364 ++i;
365 }
366
367 // Any remaining regions are empty. Print one more if there is one.
368 if (i < end_region) {
369 ParallelCompactData::RegionData* c = summary_data.region(i);
370 log_develop_trace(gc, compaction)(
371 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
372 i, p2i(c->destination()),
373 c->partial_obj_size(), c->live_obj_size(),
374 c->data_size(), c->source_region(), c->destination_count());
375 }
376
377 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
378 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
379 }
380
381 void
print_initial_summary_data(ParallelCompactData & summary_data,SpaceInfo * space_info)382 print_initial_summary_data(ParallelCompactData& summary_data,
383 SpaceInfo* space_info) {
384 if (!log_develop_is_enabled(Trace, gc, compaction)) {
385 return;
386 }
387
388 unsigned int id = PSParallelCompact::old_space_id;
389 const MutableSpace* space;
390 do {
391 space = space_info[id].space();
392 print_initial_summary_data(summary_data, space);
393 } while (++id < PSParallelCompact::eden_space_id);
394
395 do {
396 space = space_info[id].space();
397 print_generic_summary_data(summary_data, space->bottom(), space->top());
398 } while (++id < PSParallelCompact::last_space_id);
399 }
400 #endif // #ifndef PRODUCT
401
402 #ifdef ASSERT
403 size_t add_obj_count;
404 size_t add_obj_size;
405 size_t mark_bitmap_count;
406 size_t mark_bitmap_size;
407 #endif // #ifdef ASSERT
408
ParallelCompactData()409 ParallelCompactData::ParallelCompactData()
410 {
411 _region_start = 0;
412
413 _region_vspace = 0;
414 _reserved_byte_size = 0;
415 _region_data = 0;
416 _region_count = 0;
417
418 _block_vspace = 0;
419 _block_data = 0;
420 _block_count = 0;
421 }
422
initialize(MemRegion covered_region)423 bool ParallelCompactData::initialize(MemRegion covered_region)
424 {
425 _region_start = covered_region.start();
426 const size_t region_size = covered_region.word_size();
427 DEBUG_ONLY(_region_end = _region_start + region_size;)
428
429 assert(region_align_down(_region_start) == _region_start,
430 "region start not aligned");
431 assert((region_size & RegionSizeOffsetMask) == 0,
432 "region size not a multiple of RegionSize");
433
434 bool result = initialize_region_data(region_size) && initialize_block_data();
435 return result;
436 }
437
438 PSVirtualSpace*
create_vspace(size_t count,size_t element_size)439 ParallelCompactData::create_vspace(size_t count, size_t element_size)
440 {
441 const size_t raw_bytes = count * element_size;
442 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
443 const size_t granularity = os::vm_allocation_granularity();
444 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
445
446 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
447 MAX2(page_sz, granularity);
448 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
449 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
450 rs.size());
451
452 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
453
454 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
455 if (vspace != 0) {
456 if (vspace->expand_by(_reserved_byte_size)) {
457 return vspace;
458 }
459 delete vspace;
460 // Release memory reserved in the space.
461 rs.release();
462 }
463
464 return 0;
465 }
466
initialize_region_data(size_t region_size)467 bool ParallelCompactData::initialize_region_data(size_t region_size)
468 {
469 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
470 _region_vspace = create_vspace(count, sizeof(RegionData));
471 if (_region_vspace != 0) {
472 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
473 _region_count = count;
474 return true;
475 }
476 return false;
477 }
478
initialize_block_data()479 bool ParallelCompactData::initialize_block_data()
480 {
481 assert(_region_count != 0, "region data must be initialized first");
482 const size_t count = _region_count << Log2BlocksPerRegion;
483 _block_vspace = create_vspace(count, sizeof(BlockData));
484 if (_block_vspace != 0) {
485 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
486 _block_count = count;
487 return true;
488 }
489 return false;
490 }
491
clear()492 void ParallelCompactData::clear()
493 {
494 memset(_region_data, 0, _region_vspace->committed_size());
495 memset(_block_data, 0, _block_vspace->committed_size());
496 }
497
clear_range(size_t beg_region,size_t end_region)498 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
499 assert(beg_region <= _region_count, "beg_region out of range");
500 assert(end_region <= _region_count, "end_region out of range");
501 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
502
503 const size_t region_cnt = end_region - beg_region;
504 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
505
506 const size_t beg_block = beg_region * BlocksPerRegion;
507 const size_t block_cnt = region_cnt * BlocksPerRegion;
508 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
509 }
510
partial_obj_end(size_t region_idx) const511 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
512 {
513 const RegionData* cur_cp = region(region_idx);
514 const RegionData* const end_cp = region(region_count() - 1);
515
516 HeapWord* result = region_to_addr(region_idx);
517 if (cur_cp < end_cp) {
518 do {
519 result += cur_cp->partial_obj_size();
520 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
521 }
522 return result;
523 }
524
add_obj(HeapWord * addr,size_t len)525 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
526 {
527 const size_t obj_ofs = pointer_delta(addr, _region_start);
528 const size_t beg_region = obj_ofs >> Log2RegionSize;
529 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
530
531 DEBUG_ONLY(Atomic::inc(&add_obj_count);)
532 DEBUG_ONLY(Atomic::add(len, &add_obj_size);)
533
534 if (beg_region == end_region) {
535 // All in one region.
536 _region_data[beg_region].add_live_obj(len);
537 return;
538 }
539
540 // First region.
541 const size_t beg_ofs = region_offset(addr);
542 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
543
544 Klass* klass = ((oop)addr)->klass();
545 // Middle regions--completely spanned by this object.
546 for (size_t region = beg_region + 1; region < end_region; ++region) {
547 _region_data[region].set_partial_obj_size(RegionSize);
548 _region_data[region].set_partial_obj_addr(addr);
549 }
550
551 // Last region.
552 const size_t end_ofs = region_offset(addr + len - 1);
553 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
554 _region_data[end_region].set_partial_obj_addr(addr);
555 }
556
557 void
summarize_dense_prefix(HeapWord * beg,HeapWord * end)558 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
559 {
560 assert(region_offset(beg) == 0, "not RegionSize aligned");
561 assert(region_offset(end) == 0, "not RegionSize aligned");
562
563 size_t cur_region = addr_to_region_idx(beg);
564 const size_t end_region = addr_to_region_idx(end);
565 HeapWord* addr = beg;
566 while (cur_region < end_region) {
567 _region_data[cur_region].set_destination(addr);
568 _region_data[cur_region].set_destination_count(0);
569 _region_data[cur_region].set_source_region(cur_region);
570 _region_data[cur_region].set_data_location(addr);
571
572 // Update live_obj_size so the region appears completely full.
573 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
574 _region_data[cur_region].set_live_obj_size(live_size);
575
576 ++cur_region;
577 addr += RegionSize;
578 }
579 }
580
581 // Find the point at which a space can be split and, if necessary, record the
582 // split point.
583 //
584 // If the current src region (which overflowed the destination space) doesn't
585 // have a partial object, the split point is at the beginning of the current src
586 // region (an "easy" split, no extra bookkeeping required).
587 //
588 // If the current src region has a partial object, the split point is in the
589 // region where that partial object starts (call it the split_region). If
590 // split_region has a partial object, then the split point is just after that
591 // partial object (a "hard" split where we have to record the split data and
592 // zero the partial_obj_size field). With a "hard" split, we know that the
593 // partial_obj ends within split_region because the partial object that caused
594 // the overflow starts in split_region. If split_region doesn't have a partial
595 // obj, then the split is at the beginning of split_region (another "easy"
596 // split).
597 HeapWord*
summarize_split_space(size_t src_region,SplitInfo & split_info,HeapWord * destination,HeapWord * target_end,HeapWord ** target_next)598 ParallelCompactData::summarize_split_space(size_t src_region,
599 SplitInfo& split_info,
600 HeapWord* destination,
601 HeapWord* target_end,
602 HeapWord** target_next)
603 {
604 assert(destination <= target_end, "sanity");
605 assert(destination + _region_data[src_region].data_size() > target_end,
606 "region should not fit into target space");
607 assert(is_region_aligned(target_end), "sanity");
608
609 size_t split_region = src_region;
610 HeapWord* split_destination = destination;
611 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
612
613 if (destination + partial_obj_size > target_end) {
614 // The split point is just after the partial object (if any) in the
615 // src_region that contains the start of the object that overflowed the
616 // destination space.
617 //
618 // Find the start of the "overflow" object and set split_region to the
619 // region containing it.
620 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
621 split_region = addr_to_region_idx(overflow_obj);
622
623 // Clear the source_region field of all destination regions whose first word
624 // came from data after the split point (a non-null source_region field
625 // implies a region must be filled).
626 //
627 // An alternative to the simple loop below: clear during post_compact(),
628 // which uses memcpy instead of individual stores, and is easy to
629 // parallelize. (The downside is that it clears the entire RegionData
630 // object as opposed to just one field.)
631 //
632 // post_compact() would have to clear the summary data up to the highest
633 // address that was written during the summary phase, which would be
634 //
635 // max(top, max(new_top, clear_top))
636 //
637 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
638 // to target_end.
639 const RegionData* const sr = region(split_region);
640 const size_t beg_idx =
641 addr_to_region_idx(region_align_up(sr->destination() +
642 sr->partial_obj_size()));
643 const size_t end_idx = addr_to_region_idx(target_end);
644
645 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
646 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
647 _region_data[idx].set_source_region(0);
648 }
649
650 // Set split_destination and partial_obj_size to reflect the split region.
651 split_destination = sr->destination();
652 partial_obj_size = sr->partial_obj_size();
653 }
654
655 // The split is recorded only if a partial object extends onto the region.
656 if (partial_obj_size != 0) {
657 _region_data[split_region].set_partial_obj_size(0);
658 split_info.record(split_region, partial_obj_size, split_destination);
659 }
660
661 // Setup the continuation addresses.
662 *target_next = split_destination + partial_obj_size;
663 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
664
665 if (log_develop_is_enabled(Trace, gc, compaction)) {
666 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
667 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
668 split_type, p2i(source_next), split_region, partial_obj_size);
669 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
670 split_type, p2i(split_destination),
671 addr_to_region_idx(split_destination),
672 p2i(*target_next));
673
674 if (partial_obj_size != 0) {
675 HeapWord* const po_beg = split_info.destination();
676 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
677 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
678 split_type,
679 p2i(po_beg), addr_to_region_idx(po_beg),
680 p2i(po_end), addr_to_region_idx(po_end));
681 }
682 }
683
684 return source_next;
685 }
686
summarize(SplitInfo & split_info,HeapWord * source_beg,HeapWord * source_end,HeapWord ** source_next,HeapWord * target_beg,HeapWord * target_end,HeapWord ** target_next)687 bool ParallelCompactData::summarize(SplitInfo& split_info,
688 HeapWord* source_beg, HeapWord* source_end,
689 HeapWord** source_next,
690 HeapWord* target_beg, HeapWord* target_end,
691 HeapWord** target_next)
692 {
693 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
694 log_develop_trace(gc, compaction)(
695 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
696 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
697 p2i(source_beg), p2i(source_end), p2i(source_next_val),
698 p2i(target_beg), p2i(target_end), p2i(*target_next));
699
700 size_t cur_region = addr_to_region_idx(source_beg);
701 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
702
703 HeapWord *dest_addr = target_beg;
704 while (cur_region < end_region) {
705 // The destination must be set even if the region has no data.
706 _region_data[cur_region].set_destination(dest_addr);
707
708 size_t words = _region_data[cur_region].data_size();
709 if (words > 0) {
710 // If cur_region does not fit entirely into the target space, find a point
711 // at which the source space can be 'split' so that part is copied to the
712 // target space and the rest is copied elsewhere.
713 if (dest_addr + words > target_end) {
714 assert(source_next != NULL, "source_next is NULL when splitting");
715 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
716 target_end, target_next);
717 return false;
718 }
719
720 // Compute the destination_count for cur_region, and if necessary, update
721 // source_region for a destination region. The source_region field is
722 // updated if cur_region is the first (left-most) region to be copied to a
723 // destination region.
724 //
725 // The destination_count calculation is a bit subtle. A region that has
726 // data that compacts into itself does not count itself as a destination.
727 // This maintains the invariant that a zero count means the region is
728 // available and can be claimed and then filled.
729 uint destination_count = 0;
730 if (split_info.is_split(cur_region)) {
731 // The current region has been split: the partial object will be copied
732 // to one destination space and the remaining data will be copied to
733 // another destination space. Adjust the initial destination_count and,
734 // if necessary, set the source_region field if the partial object will
735 // cross a destination region boundary.
736 destination_count = split_info.destination_count();
737 if (destination_count == 2) {
738 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
739 _region_data[dest_idx].set_source_region(cur_region);
740 }
741 }
742
743 HeapWord* const last_addr = dest_addr + words - 1;
744 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
745 const size_t dest_region_2 = addr_to_region_idx(last_addr);
746
747 // Initially assume that the destination regions will be the same and
748 // adjust the value below if necessary. Under this assumption, if
749 // cur_region == dest_region_2, then cur_region will be compacted
750 // completely into itself.
751 destination_count += cur_region == dest_region_2 ? 0 : 1;
752 if (dest_region_1 != dest_region_2) {
753 // Destination regions differ; adjust destination_count.
754 destination_count += 1;
755 // Data from cur_region will be copied to the start of dest_region_2.
756 _region_data[dest_region_2].set_source_region(cur_region);
757 } else if (region_offset(dest_addr) == 0) {
758 // Data from cur_region will be copied to the start of the destination
759 // region.
760 _region_data[dest_region_1].set_source_region(cur_region);
761 }
762
763 _region_data[cur_region].set_destination_count(destination_count);
764 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
765 dest_addr += words;
766 }
767
768 ++cur_region;
769 }
770
771 *target_next = dest_addr;
772 return true;
773 }
774
calc_new_pointer(HeapWord * addr,ParCompactionManager * cm)775 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
776 assert(addr != NULL, "Should detect NULL oop earlier");
777 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
778 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
779
780 // Region covering the object.
781 RegionData* const region_ptr = addr_to_region_ptr(addr);
782 HeapWord* result = region_ptr->destination();
783
784 // If the entire Region is live, the new location is region->destination + the
785 // offset of the object within in the Region.
786
787 // Run some performance tests to determine if this special case pays off. It
788 // is worth it for pointers into the dense prefix. If the optimization to
789 // avoid pointer updates in regions that only point to the dense prefix is
790 // ever implemented, this should be revisited.
791 if (region_ptr->data_size() == RegionSize) {
792 result += region_offset(addr);
793 return result;
794 }
795
796 // Otherwise, the new location is region->destination + block offset + the
797 // number of live words in the Block that are (a) to the left of addr and (b)
798 // due to objects that start in the Block.
799
800 // Fill in the block table if necessary. This is unsynchronized, so multiple
801 // threads may fill the block table for a region (harmless, since it is
802 // idempotent).
803 if (!region_ptr->blocks_filled()) {
804 PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
805 region_ptr->set_blocks_filled();
806 }
807
808 HeapWord* const search_start = block_align_down(addr);
809 const size_t block_offset = addr_to_block_ptr(addr)->offset();
810
811 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
812 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr));
813 result += block_offset + live;
814 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
815 return result;
816 }
817
818 #ifdef ASSERT
verify_clear(const PSVirtualSpace * vspace)819 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
820 {
821 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
822 const size_t* const end = (const size_t*)vspace->committed_high_addr();
823 for (const size_t* p = beg; p < end; ++p) {
824 assert(*p == 0, "not zero");
825 }
826 }
827
verify_clear()828 void ParallelCompactData::verify_clear()
829 {
830 verify_clear(_region_vspace);
831 verify_clear(_block_vspace);
832 }
833 #endif // #ifdef ASSERT
834
835 STWGCTimer PSParallelCompact::_gc_timer;
836 ParallelOldTracer PSParallelCompact::_gc_tracer;
837 elapsedTimer PSParallelCompact::_accumulated_time;
838 unsigned int PSParallelCompact::_total_invocations = 0;
839 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
840 jlong PSParallelCompact::_time_of_last_gc = 0;
841 CollectorCounters* PSParallelCompact::_counters = NULL;
842 ParMarkBitMap PSParallelCompact::_mark_bitmap;
843 ParallelCompactData PSParallelCompact::_summary_data;
844
845 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
846
do_object_b(oop p)847 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
848
849 class PCReferenceProcessor: public ReferenceProcessor {
850 public:
PCReferenceProcessor(BoolObjectClosure * is_subject_to_discovery,BoolObjectClosure * is_alive_non_header)851 PCReferenceProcessor(
852 BoolObjectClosure* is_subject_to_discovery,
853 BoolObjectClosure* is_alive_non_header) :
854 ReferenceProcessor(is_subject_to_discovery,
855 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
856 ParallelGCThreads, // mt processing degree
857 true, // mt discovery
858 ParallelGCThreads, // mt discovery degree
859 true, // atomic_discovery
860 is_alive_non_header) {
861 }
862
discover(oop obj,ReferenceType type)863 template<typename T> bool discover(oop obj, ReferenceType type) {
864 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
865 T heap_oop = RawAccess<>::oop_load(referent_addr);
866 oop referent = CompressedOops::decode_not_null(heap_oop);
867 return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
868 && ReferenceProcessor::discover_reference(obj, type);
869 }
discover_reference(oop obj,ReferenceType type)870 virtual bool discover_reference(oop obj, ReferenceType type) {
871 if (UseCompressedOops) {
872 return discover<narrowOop>(obj, type);
873 } else {
874 return discover<oop>(obj, type);
875 }
876 }
877 };
878
post_initialize()879 void PSParallelCompact::post_initialize() {
880 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
881 _span_based_discoverer.set_span(heap->reserved_region());
882 _ref_processor =
883 new PCReferenceProcessor(&_span_based_discoverer,
884 &_is_alive_closure); // non-header is alive closure
885
886 _counters = new CollectorCounters("PSParallelCompact", 1);
887
888 // Initialize static fields in ParCompactionManager.
889 ParCompactionManager::initialize(mark_bitmap());
890 }
891
initialize()892 bool PSParallelCompact::initialize() {
893 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
894 MemRegion mr = heap->reserved_region();
895
896 // Was the old gen get allocated successfully?
897 if (!heap->old_gen()->is_allocated()) {
898 return false;
899 }
900
901 initialize_space_info();
902 initialize_dead_wood_limiter();
903
904 if (!_mark_bitmap.initialize(mr)) {
905 vm_shutdown_during_initialization(
906 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
907 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
908 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
909 return false;
910 }
911
912 if (!_summary_data.initialize(mr)) {
913 vm_shutdown_during_initialization(
914 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
915 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
916 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
917 return false;
918 }
919
920 return true;
921 }
922
initialize_space_info()923 void PSParallelCompact::initialize_space_info()
924 {
925 memset(&_space_info, 0, sizeof(_space_info));
926
927 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
928 PSYoungGen* young_gen = heap->young_gen();
929
930 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
931 _space_info[eden_space_id].set_space(young_gen->eden_space());
932 _space_info[from_space_id].set_space(young_gen->from_space());
933 _space_info[to_space_id].set_space(young_gen->to_space());
934
935 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
936 }
937
initialize_dead_wood_limiter()938 void PSParallelCompact::initialize_dead_wood_limiter()
939 {
940 const size_t max = 100;
941 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
942 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
943 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
944 DEBUG_ONLY(_dwl_initialized = true;)
945 _dwl_adjustment = normal_distribution(1.0);
946 }
947
948 void
clear_data_covering_space(SpaceId id)949 PSParallelCompact::clear_data_covering_space(SpaceId id)
950 {
951 // At this point, top is the value before GC, new_top() is the value that will
952 // be set at the end of GC. The marking bitmap is cleared to top; nothing
953 // should be marked above top. The summary data is cleared to the larger of
954 // top & new_top.
955 MutableSpace* const space = _space_info[id].space();
956 HeapWord* const bot = space->bottom();
957 HeapWord* const top = space->top();
958 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
959
960 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
961 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
962 _mark_bitmap.clear_range(beg_bit, end_bit);
963
964 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
965 const size_t end_region =
966 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
967 _summary_data.clear_range(beg_region, end_region);
968
969 // Clear the data used to 'split' regions.
970 SplitInfo& split_info = _space_info[id].split_info();
971 if (split_info.is_valid()) {
972 split_info.clear();
973 }
974 DEBUG_ONLY(split_info.verify_clear();)
975 }
976
pre_compact()977 void PSParallelCompact::pre_compact()
978 {
979 // Update the from & to space pointers in space_info, since they are swapped
980 // at each young gen gc. Do the update unconditionally (even though a
981 // promotion failure does not swap spaces) because an unknown number of young
982 // collections will have swapped the spaces an unknown number of times.
983 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
984 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
985 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
986 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
987
988 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
989 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
990
991 // Increment the invocation count
992 heap->increment_total_collections(true);
993
994 // We need to track unique mark sweep invocations as well.
995 _total_invocations++;
996
997 heap->print_heap_before_gc();
998 heap->trace_heap_before_gc(&_gc_tracer);
999
1000 // Fill in TLABs
1001 heap->ensure_parsability(true); // retire TLABs
1002
1003 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1004 HandleMark hm; // Discard invalid handles created during verification
1005 Universe::verify("Before GC");
1006 }
1007
1008 // Verify object start arrays
1009 if (VerifyObjectStartArray &&
1010 VerifyBeforeGC) {
1011 heap->old_gen()->verify_object_start_array();
1012 }
1013
1014 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1015 DEBUG_ONLY(summary_data().verify_clear();)
1016
1017 // Have worker threads release resources the next time they run a task.
1018 gc_task_manager()->release_all_resources();
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
1027 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1028 // Clear the marking bitmap, summary data and split info.
1029 clear_data_covering_space(SpaceId(id));
1030 // Update top(). Must be done after clearing the bitmap and summary data.
1031 _space_info[id].publish_new_top();
1032 }
1033
1034 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1035 MutableSpace* const from_space = _space_info[from_space_id].space();
1036 MutableSpace* const to_space = _space_info[to_space_id].space();
1037
1038 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1039 bool eden_empty = eden_space->is_empty();
1040 if (!eden_empty) {
1041 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1042 heap->young_gen(), heap->old_gen());
1043 }
1044
1045 // Update heap occupancy information which is used as input to the soft ref
1046 // clearing policy at the next gc.
1047 Universe::update_heap_info_at_gc();
1048
1049 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1050 to_space->is_empty();
1051
1052 PSCardTable* ct = heap->card_table();
1053 MemRegion old_mr = heap->old_gen()->reserved();
1054 if (young_gen_empty) {
1055 ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1056 } else {
1057 ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1058 }
1059
1060 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1061 ClassLoaderDataGraph::purge();
1062 MetaspaceUtils::verify_metrics();
1063
1064 CodeCache::gc_epilogue();
1065 JvmtiExport::gc_epilogue();
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 collector 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 // CollectorPolicy::_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 PreGCValues pre_gc_values(heap);
1782
1783 // Get the compaction manager reserved for the VM thread.
1784 ParCompactionManager* const vmthread_cm =
1785 ParCompactionManager::manager_array(gc_task_manager()->workers());
1786
1787 {
1788 ResourceMark rm;
1789 HandleMark hm;
1790
1791 // Set the number of GC threads to be used in this collection
1792 gc_task_manager()->set_active_gang();
1793 gc_task_manager()->task_idle_workers();
1794
1795 GCTraceCPUTime tcpu;
1796 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1797
1798 heap->pre_full_gc_dump(&_gc_timer);
1799
1800 TraceCollectorStats tcs(counters());
1801 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1802
1803 if (log_is_enabled(Debug, gc, heap, exit)) {
1804 accumulated_time()->start();
1805 }
1806
1807 // Let the size policy know we're starting
1808 size_policy->major_collection_begin();
1809
1810 CodeCache::gc_prologue();
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(young_live,
1893 eden_live,
1894 max_old_gen_size,
1895 max_eden_size,
1896 true /* full gc*/,
1897 gc_cause,
1898 heap->soft_ref_policy());
1899
1900 size_policy->decay_supplemental_growth(true /* full gc*/);
1901
1902 heap->resize_old_gen(
1903 size_policy->calculated_old_free_size_in_bytes());
1904
1905 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1906 size_policy->calculated_survivor_size_in_bytes());
1907 }
1908
1909 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1910 }
1911
1912 if (UsePerfData) {
1913 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1914 counters->update_counters();
1915 counters->update_old_capacity(old_gen->capacity_in_bytes());
1916 counters->update_young_capacity(young_gen->capacity_in_bytes());
1917 }
1918
1919 heap->resize_all_tlabs();
1920
1921 // Resize the metaspace capacity after a collection
1922 MetaspaceGC::compute_new_size();
1923
1924 if (log_is_enabled(Debug, gc, heap, exit)) {
1925 accumulated_time()->stop();
1926 }
1927
1928 young_gen->print_used_change(pre_gc_values.young_gen_used());
1929 old_gen->print_used_change(pre_gc_values.old_gen_used());
1930 MetaspaceUtils::print_metaspace_change(pre_gc_values.metadata_used());
1931
1932 // Track memory usage and detect low memory
1933 MemoryService::track_memory_usage();
1934 heap->update_counters();
1935 gc_task_manager()->release_idle_workers();
1936
1937 heap->post_full_gc_dump(&_gc_timer);
1938 }
1939
1940 #ifdef ASSERT
1941 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1942 ParCompactionManager* const cm =
1943 ParCompactionManager::manager_array(int(i));
1944 assert(cm->marking_stack()->is_empty(), "should be empty");
1945 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i);
1946 }
1947 #endif // ASSERT
1948
1949 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1950 HandleMark hm; // Discard invalid handles created during verification
1951 Universe::verify("After GC");
1952 }
1953
1954 // Re-verify object start arrays
1955 if (VerifyObjectStartArray &&
1956 VerifyAfterGC) {
1957 old_gen->verify_object_start_array();
1958 }
1959
1960 if (ZapUnusedHeapArea) {
1961 old_gen->object_space()->check_mangled_unused_area_complete();
1962 }
1963
1964 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1965
1966 collection_exit.update();
1967
1968 heap->print_heap_after_gc();
1969 heap->trace_heap_after_gc(&_gc_tracer);
1970
1971 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1972 marking_start.ticks(), compaction_start.ticks(),
1973 collection_exit.ticks());
1974 gc_task_manager()->print_task_time_stamps();
1975
1976 #ifdef TRACESPINNING
1977 ParallelTaskTerminator::print_termination_counts();
1978 #endif
1979
1980 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1981
1982 _gc_timer.register_gc_end();
1983
1984 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1985 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1986
1987 return true;
1988 }
1989
absorb_live_data_from_eden(PSAdaptiveSizePolicy * size_policy,PSYoungGen * young_gen,PSOldGen * old_gen)1990 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1991 PSYoungGen* young_gen,
1992 PSOldGen* old_gen) {
1993 MutableSpace* const eden_space = young_gen->eden_space();
1994 assert(!eden_space->is_empty(), "eden must be non-empty");
1995 assert(young_gen->virtual_space()->alignment() ==
1996 old_gen->virtual_space()->alignment(), "alignments do not match");
1997
1998 // We also return false when it's a heterogenous heap because old generation cannot absorb data from eden
1999 // when it is allocated on different memory (example, nv-dimm) than young.
2000 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) ||
2001 ParallelScavengeHeap::heap()->ps_collector_policy()->is_hetero_heap()) {
2002 return false;
2003 }
2004
2005 // Both generations must be completely committed.
2006 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2007 return false;
2008 }
2009 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2010 return false;
2011 }
2012
2013 // Figure out how much to take from eden. Include the average amount promoted
2014 // in the total; otherwise the next young gen GC will simply bail out to a
2015 // full GC.
2016 const size_t alignment = old_gen->virtual_space()->alignment();
2017 const size_t eden_used = eden_space->used_in_bytes();
2018 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2019 const size_t absorb_size = align_up(eden_used + promoted, alignment);
2020 const size_t eden_capacity = eden_space->capacity_in_bytes();
2021
2022 if (absorb_size >= eden_capacity) {
2023 return false; // Must leave some space in eden.
2024 }
2025
2026 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2027 if (new_young_size < young_gen->min_gen_size()) {
2028 return false; // Respect young gen minimum size.
2029 }
2030
2031 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: "
2032 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2033 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2034 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2035 absorb_size / K,
2036 eden_capacity / K, (eden_capacity - absorb_size) / K,
2037 young_gen->from_space()->used_in_bytes() / K,
2038 young_gen->to_space()->used_in_bytes() / K,
2039 young_gen->capacity_in_bytes() / K, new_young_size / K);
2040
2041 // Fill the unused part of the old gen.
2042 MutableSpace* const old_space = old_gen->object_space();
2043 HeapWord* const unused_start = old_space->top();
2044 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2045
2046 if (unused_words > 0) {
2047 if (unused_words < CollectedHeap::min_fill_size()) {
2048 return false; // If the old gen cannot be filled, must give up.
2049 }
2050 CollectedHeap::fill_with_objects(unused_start, unused_words);
2051 }
2052
2053 // Take the live data from eden and set both top and end in the old gen to
2054 // eden top. (Need to set end because reset_after_change() mangles the region
2055 // from end to virtual_space->high() in debug builds).
2056 HeapWord* const new_top = eden_space->top();
2057 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2058 absorb_size);
2059 young_gen->reset_after_change();
2060 old_space->set_top(new_top);
2061 old_space->set_end(new_top);
2062 old_gen->reset_after_change();
2063
2064 // Update the object start array for the filler object and the data from eden.
2065 ObjectStartArray* const start_array = old_gen->start_array();
2066 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2067 start_array->allocate_block(p);
2068 }
2069
2070 // Could update the promoted average here, but it is not typically updated at
2071 // full GCs and the value to use is unclear. Something like
2072 //
2073 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2074
2075 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2076 return true;
2077 }
2078
gc_task_manager()2079 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2080 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2081 "shouldn't return NULL");
2082 return ParallelScavengeHeap::gc_task_manager();
2083 }
2084
2085 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
2086 private:
2087 GCTaskQueue* _q;
2088
2089 public:
PCAddThreadRootsMarkingTaskClosure(GCTaskQueue * q)2090 PCAddThreadRootsMarkingTaskClosure(GCTaskQueue* q) : _q(q) { }
do_thread(Thread * t)2091 void do_thread(Thread* t) {
2092 _q->enqueue(new ThreadRootsMarkingTask(t));
2093 }
2094 };
2095
marking_phase(ParCompactionManager * cm,bool maximum_heap_compaction,ParallelOldTracer * gc_tracer)2096 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2097 bool maximum_heap_compaction,
2098 ParallelOldTracer *gc_tracer) {
2099 // Recursively traverse all live objects and mark them
2100 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2101
2102 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2103 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2104 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2105 TaskQueueSetSuper* qset = ParCompactionManager::stack_array();
2106 TaskTerminator terminator(active_gc_threads, qset);
2107
2108 PCMarkAndPushClosure mark_and_push_closure(cm);
2109 ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2110
2111 // Need new claim bits before marking starts.
2112 ClassLoaderDataGraph::clear_claimed_marks();
2113
2114 {
2115 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2116
2117 ParallelScavengeHeap::ParStrongRootsScope psrs;
2118
2119 GCTaskQueue* q = GCTaskQueue::create();
2120
2121 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2122 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2123 // We scan the thread roots in parallel
2124 PCAddThreadRootsMarkingTaskClosure cl(q);
2125 Threads::java_threads_and_vm_thread_do(&cl);
2126 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2127 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2128 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2129 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2130 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2131 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2132
2133 if (active_gc_threads > 1) {
2134 for (uint j = 0; j < active_gc_threads; j++) {
2135 q->enqueue(new StealMarkingTask(terminator.terminator()));
2136 }
2137 }
2138
2139 gc_task_manager()->execute_and_wait(q);
2140 }
2141
2142 // Process reference objects found during marking
2143 {
2144 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2145
2146 ReferenceProcessorStats stats;
2147 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2148
2149 if (ref_processor()->processing_is_mt()) {
2150 ref_processor()->set_active_mt_degree(active_gc_threads);
2151
2152 RefProcTaskExecutor task_executor;
2153 stats = ref_processor()->process_discovered_references(
2154 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2155 &task_executor, &pt);
2156 } else {
2157 stats = ref_processor()->process_discovered_references(
2158 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2159 &pt);
2160 }
2161
2162 gc_tracer->report_gc_reference_stats(stats);
2163 pt.print_all_references();
2164 }
2165
2166 // This is the point where the entire marking should have completed.
2167 assert(cm->marking_stacks_empty(), "Marking should have completed");
2168
2169 {
2170 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2171 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl);
2172 }
2173
2174 {
2175 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2176
2177 // Follow system dictionary roots and unload classes.
2178 bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2179
2180 // Unload nmethods.
2181 CodeCache::do_unloading(is_alive_closure(), purged_class);
2182
2183 // Prune dead klasses from subklass/sibling/implementor lists.
2184 Klass::clean_weak_klass_links(purged_class);
2185 }
2186
2187 {
2188 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer);
2189 // Delete entries for dead interned strings.
2190 StringTable::unlink(is_alive_closure());
2191 }
2192
2193 {
2194 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer);
2195 // Clean up unreferenced symbols in symbol table.
2196 SymbolTable::unlink();
2197 }
2198
2199 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2200 }
2201
adjust_roots(ParCompactionManager * cm)2202 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2203 // Adjust the pointers to reflect the new locations
2204 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2205
2206 // Need new claim bits when tracing through and adjusting pointers.
2207 ClassLoaderDataGraph::clear_claimed_marks();
2208
2209 PCAdjustPointerClosure oop_closure(cm);
2210
2211 // General strong roots.
2212 Universe::oops_do(&oop_closure);
2213 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles
2214 Threads::oops_do(&oop_closure, NULL);
2215 ObjectSynchronizer::oops_do(&oop_closure);
2216 Management::oops_do(&oop_closure);
2217 JvmtiExport::oops_do(&oop_closure);
2218 SystemDictionary::oops_do(&oop_closure);
2219 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong);
2220 ClassLoaderDataGraph::cld_do(&cld_closure);
2221
2222 // Now adjust pointers in remaining weak roots. (All of which should
2223 // have been cleared if they pointed to non-surviving objects.)
2224 WeakProcessor::oops_do(&oop_closure);
2225
2226 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2227 CodeCache::blobs_do(&adjust_from_blobs);
2228 AOTLoader::oops_do(&oop_closure);
2229 StringTable::oops_do(&oop_closure);
2230 ref_processor()->weak_oops_do(&oop_closure);
2231 // Roots were visited so references into the young gen in roots
2232 // may have been scanned. Process them also.
2233 // Should the reference processor have a span that excludes
2234 // young gen objects?
2235 PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2236 }
2237
2238 // Helper class to print 8 region numbers per line and then print the total at the end.
2239 class FillableRegionLogger : public StackObj {
2240 private:
2241 Log(gc, compaction) log;
2242 static const int LineLength = 8;
2243 size_t _regions[LineLength];
2244 int _next_index;
2245 bool _enabled;
2246 size_t _total_regions;
2247 public:
FillableRegionLogger()2248 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
~FillableRegionLogger()2249 ~FillableRegionLogger() {
2250 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2251 }
2252
print_line()2253 void print_line() {
2254 if (!_enabled || _next_index == 0) {
2255 return;
2256 }
2257 FormatBuffer<> line("Fillable: ");
2258 for (int i = 0; i < _next_index; i++) {
2259 line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2260 }
2261 log.trace("%s", line.buffer());
2262 _next_index = 0;
2263 }
2264
handle(size_t region)2265 void handle(size_t region) {
2266 if (!_enabled) {
2267 return;
2268 }
2269 _regions[_next_index++] = region;
2270 if (_next_index == LineLength) {
2271 print_line();
2272 }
2273 _total_regions++;
2274 }
2275 };
2276
prepare_region_draining_tasks(GCTaskQueue * q,uint parallel_gc_threads)2277 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q,
2278 uint parallel_gc_threads)
2279 {
2280 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2281
2282 // Find the threads that are active
2283 unsigned int which = 0;
2284
2285 // Find all regions that are available (can be filled immediately) and
2286 // distribute them to the thread stacks. The iteration is done in reverse
2287 // order (high to low) so the regions will be removed in ascending order.
2288
2289 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2290
2291 which = 0;
2292 // id + 1 is used to test termination so unsigned can
2293 // be used with an old_space_id == 0.
2294 FillableRegionLogger region_logger;
2295 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2296 SpaceInfo* const space_info = _space_info + id;
2297 MutableSpace* const space = space_info->space();
2298 HeapWord* const new_top = space_info->new_top();
2299
2300 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2301 const size_t end_region =
2302 sd.addr_to_region_idx(sd.region_align_up(new_top));
2303
2304 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2305 if (sd.region(cur)->claim_unsafe()) {
2306 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2307 cm->region_stack()->push(cur);
2308 region_logger.handle(cur);
2309 // Assign regions to tasks in round-robin fashion.
2310 if (++which == parallel_gc_threads) {
2311 which = 0;
2312 }
2313 }
2314 }
2315 region_logger.print_line();
2316 }
2317 }
2318
2319 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2320
enqueue_dense_prefix_tasks(GCTaskQueue * q,uint parallel_gc_threads)2321 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2322 uint parallel_gc_threads) {
2323 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2324
2325 ParallelCompactData& sd = PSParallelCompact::summary_data();
2326
2327 // Iterate over all the spaces adding tasks for updating
2328 // regions in the dense prefix. Assume that 1 gc thread
2329 // will work on opening the gaps and the remaining gc threads
2330 // will work on the dense prefix.
2331 unsigned int space_id;
2332 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2333 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2334 const MutableSpace* const space = _space_info[space_id].space();
2335
2336 if (dense_prefix_end == space->bottom()) {
2337 // There is no dense prefix for this space.
2338 continue;
2339 }
2340
2341 // The dense prefix is before this region.
2342 size_t region_index_end_dense_prefix =
2343 sd.addr_to_region_idx(dense_prefix_end);
2344 RegionData* const dense_prefix_cp =
2345 sd.region(region_index_end_dense_prefix);
2346 assert(dense_prefix_end == space->end() ||
2347 dense_prefix_cp->available() ||
2348 dense_prefix_cp->claimed(),
2349 "The region after the dense prefix should always be ready to fill");
2350
2351 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2352
2353 // Is there dense prefix work?
2354 size_t total_dense_prefix_regions =
2355 region_index_end_dense_prefix - region_index_start;
2356 // How many regions of the dense prefix should be given to
2357 // each thread?
2358 if (total_dense_prefix_regions > 0) {
2359 uint tasks_for_dense_prefix = 1;
2360 if (total_dense_prefix_regions <=
2361 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2362 // Don't over partition. This assumes that
2363 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2364 // so there are not many regions to process.
2365 tasks_for_dense_prefix = parallel_gc_threads;
2366 } else {
2367 // Over partition
2368 tasks_for_dense_prefix = parallel_gc_threads *
2369 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2370 }
2371 size_t regions_per_thread = total_dense_prefix_regions /
2372 tasks_for_dense_prefix;
2373 // Give each thread at least 1 region.
2374 if (regions_per_thread == 0) {
2375 regions_per_thread = 1;
2376 }
2377
2378 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2379 if (region_index_start >= region_index_end_dense_prefix) {
2380 break;
2381 }
2382 // region_index_end is not processed
2383 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2384 region_index_end_dense_prefix);
2385 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2386 region_index_start,
2387 region_index_end));
2388 region_index_start = region_index_end;
2389 }
2390 }
2391 // This gets any part of the dense prefix that did not
2392 // fit evenly.
2393 if (region_index_start < region_index_end_dense_prefix) {
2394 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2395 region_index_start,
2396 region_index_end_dense_prefix));
2397 }
2398 }
2399 }
2400
enqueue_region_stealing_tasks(GCTaskQueue * q,ParallelTaskTerminator * terminator_ptr,uint parallel_gc_threads)2401 void PSParallelCompact::enqueue_region_stealing_tasks(
2402 GCTaskQueue* q,
2403 ParallelTaskTerminator* terminator_ptr,
2404 uint parallel_gc_threads) {
2405 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer);
2406
2407 // Once a thread has drained it's stack, it should try to steal regions from
2408 // other threads.
2409 for (uint j = 0; j < parallel_gc_threads; j++) {
2410 q->enqueue(new CompactionWithStealingTask(terminator_ptr));
2411 }
2412 }
2413
2414 #ifdef ASSERT
2415 // Write a histogram of the number of times the block table was filled for a
2416 // region.
write_block_fill_histogram()2417 void PSParallelCompact::write_block_fill_histogram()
2418 {
2419 if (!log_develop_is_enabled(Trace, gc, compaction)) {
2420 return;
2421 }
2422
2423 Log(gc, compaction) log;
2424 ResourceMark rm;
2425 LogStream ls(log.trace());
2426 outputStream* out = &ls;
2427
2428 typedef ParallelCompactData::RegionData rd_t;
2429 ParallelCompactData& sd = summary_data();
2430
2431 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2432 MutableSpace* const spc = _space_info[id].space();
2433 if (spc->bottom() != spc->top()) {
2434 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2435 HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2436 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2437
2438 size_t histo[5] = { 0, 0, 0, 0, 0 };
2439 const size_t histo_len = sizeof(histo) / sizeof(size_t);
2440 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2441
2442 for (const rd_t* cur = beg; cur < end; ++cur) {
2443 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2444 }
2445 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2446 for (size_t i = 0; i < histo_len; ++i) {
2447 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2448 histo[i], 100.0 * histo[i] / region_cnt);
2449 }
2450 out->cr();
2451 }
2452 }
2453 }
2454 #endif // #ifdef ASSERT
2455
compact()2456 void PSParallelCompact::compact() {
2457 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2458
2459 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2460 PSOldGen* old_gen = heap->old_gen();
2461 old_gen->start_array()->reset();
2462 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2463 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2464 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2465 TaskTerminator terminator(active_gc_threads, qset);
2466
2467 GCTaskQueue* q = GCTaskQueue::create();
2468 prepare_region_draining_tasks(q, active_gc_threads);
2469 enqueue_dense_prefix_tasks(q, active_gc_threads);
2470 enqueue_region_stealing_tasks(q, terminator.terminator(), active_gc_threads);
2471
2472 {
2473 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2474
2475 gc_task_manager()->execute_and_wait(q);
2476
2477 #ifdef ASSERT
2478 // Verify that all regions have been processed before the deferred updates.
2479 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2480 verify_complete(SpaceId(id));
2481 }
2482 #endif
2483 }
2484
2485 {
2486 // Update the deferred objects, if any. Any compaction manager can be used.
2487 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2488 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2489 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2490 update_deferred_objects(cm, SpaceId(id));
2491 }
2492 }
2493
2494 DEBUG_ONLY(write_block_fill_histogram());
2495 }
2496
2497 #ifdef ASSERT
verify_complete(SpaceId space_id)2498 void PSParallelCompact::verify_complete(SpaceId space_id) {
2499 // All Regions between space bottom() to new_top() should be marked as filled
2500 // and all Regions between new_top() and top() should be available (i.e.,
2501 // should have been emptied).
2502 ParallelCompactData& sd = summary_data();
2503 SpaceInfo si = _space_info[space_id];
2504 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2505 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2506 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2507 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2508 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2509
2510 bool issued_a_warning = false;
2511
2512 size_t cur_region;
2513 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2514 const RegionData* const c = sd.region(cur_region);
2515 if (!c->completed()) {
2516 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2517 cur_region, c->destination_count());
2518 issued_a_warning = true;
2519 }
2520 }
2521
2522 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2523 const RegionData* const c = sd.region(cur_region);
2524 if (!c->available()) {
2525 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2526 cur_region, c->destination_count());
2527 issued_a_warning = true;
2528 }
2529 }
2530
2531 if (issued_a_warning) {
2532 print_region_ranges();
2533 }
2534 }
2535 #endif // #ifdef ASSERT
2536
do_addr(HeapWord * addr)2537 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2538 _start_array->allocate_block(addr);
2539 compaction_manager()->update_contents(oop(addr));
2540 }
2541
2542 // Update interior oops in the ranges of regions [beg_region, end_region).
2543 void
update_and_deadwood_in_dense_prefix(ParCompactionManager * cm,SpaceId space_id,size_t beg_region,size_t end_region)2544 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2545 SpaceId space_id,
2546 size_t beg_region,
2547 size_t end_region) {
2548 ParallelCompactData& sd = summary_data();
2549 ParMarkBitMap* const mbm = mark_bitmap();
2550
2551 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2552 HeapWord* const end_addr = sd.region_to_addr(end_region);
2553 assert(beg_region <= end_region, "bad region range");
2554 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2555
2556 #ifdef ASSERT
2557 // Claim the regions to avoid triggering an assert when they are marked as
2558 // filled.
2559 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2560 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2561 }
2562 #endif // #ifdef ASSERT
2563
2564 if (beg_addr != space(space_id)->bottom()) {
2565 // Find the first live object or block of dead space that *starts* in this
2566 // range of regions. If a partial object crosses onto the region, skip it;
2567 // it will be marked for 'deferred update' when the object head is
2568 // processed. If dead space crosses onto the region, it is also skipped; it
2569 // will be filled when the prior region is processed. If neither of those
2570 // apply, the first word in the region is the start of a live object or dead
2571 // space.
2572 assert(beg_addr > space(space_id)->bottom(), "sanity");
2573 const RegionData* const cp = sd.region(beg_region);
2574 if (cp->partial_obj_size() != 0) {
2575 beg_addr = sd.partial_obj_end(beg_region);
2576 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2577 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2578 }
2579 }
2580
2581 if (beg_addr < end_addr) {
2582 // A live object or block of dead space starts in this range of Regions.
2583 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2584
2585 // Create closures and iterate.
2586 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2587 FillClosure fill_closure(cm, space_id);
2588 ParMarkBitMap::IterationStatus status;
2589 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2590 dense_prefix_end);
2591 if (status == ParMarkBitMap::incomplete) {
2592 update_closure.do_addr(update_closure.source());
2593 }
2594 }
2595
2596 // Mark the regions as filled.
2597 RegionData* const beg_cp = sd.region(beg_region);
2598 RegionData* const end_cp = sd.region(end_region);
2599 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2600 cp->set_completed();
2601 }
2602 }
2603
2604 // Return the SpaceId for the space containing addr. If addr is not in the
2605 // heap, last_space_id is returned. In debug mode it expects the address to be
2606 // in the heap and asserts such.
space_id(HeapWord * addr)2607 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2608 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2609
2610 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2611 if (_space_info[id].space()->contains(addr)) {
2612 return SpaceId(id);
2613 }
2614 }
2615
2616 assert(false, "no space contains the addr");
2617 return last_space_id;
2618 }
2619
update_deferred_objects(ParCompactionManager * cm,SpaceId id)2620 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2621 SpaceId id) {
2622 assert(id < last_space_id, "bad space id");
2623
2624 ParallelCompactData& sd = summary_data();
2625 const SpaceInfo* const space_info = _space_info + id;
2626 ObjectStartArray* const start_array = space_info->start_array();
2627
2628 const MutableSpace* const space = space_info->space();
2629 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2630 HeapWord* const beg_addr = space_info->dense_prefix();
2631 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2632
2633 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2634 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2635 const RegionData* cur_region;
2636 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2637 HeapWord* const addr = cur_region->deferred_obj_addr();
2638 if (addr != NULL) {
2639 if (start_array != NULL) {
2640 start_array->allocate_block(addr);
2641 }
2642 cm->update_contents(oop(addr));
2643 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2644 }
2645 }
2646 }
2647
2648 // Skip over count live words starting from beg, and return the address of the
2649 // next live word. Unless marked, the word corresponding to beg is assumed to
2650 // be dead. Callers must either ensure beg does not correspond to the middle of
2651 // an object, or account for those live words in some other way. Callers must
2652 // also ensure that there are enough live words in the range [beg, end) to skip.
2653 HeapWord*
skip_live_words(HeapWord * beg,HeapWord * end,size_t count)2654 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2655 {
2656 assert(count > 0, "sanity");
2657
2658 ParMarkBitMap* m = mark_bitmap();
2659 idx_t bits_to_skip = m->words_to_bits(count);
2660 idx_t cur_beg = m->addr_to_bit(beg);
2661 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2662
2663 do {
2664 cur_beg = m->find_obj_beg(cur_beg, search_end);
2665 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2666 const size_t obj_bits = cur_end - cur_beg + 1;
2667 if (obj_bits > bits_to_skip) {
2668 return m->bit_to_addr(cur_beg + bits_to_skip);
2669 }
2670 bits_to_skip -= obj_bits;
2671 cur_beg = cur_end + 1;
2672 } while (bits_to_skip > 0);
2673
2674 // Skipping the desired number of words landed just past the end of an object.
2675 // Find the start of the next object.
2676 cur_beg = m->find_obj_beg(cur_beg, search_end);
2677 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2678 return m->bit_to_addr(cur_beg);
2679 }
2680
first_src_addr(HeapWord * const dest_addr,SpaceId src_space_id,size_t src_region_idx)2681 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2682 SpaceId src_space_id,
2683 size_t src_region_idx)
2684 {
2685 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2686
2687 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2688 if (split_info.dest_region_addr() == dest_addr) {
2689 // The partial object ending at the split point contains the first word to
2690 // be copied to dest_addr.
2691 return split_info.first_src_addr();
2692 }
2693
2694 const ParallelCompactData& sd = summary_data();
2695 ParMarkBitMap* const bitmap = mark_bitmap();
2696 const size_t RegionSize = ParallelCompactData::RegionSize;
2697
2698 assert(sd.is_region_aligned(dest_addr), "not aligned");
2699 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2700 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2701 HeapWord* const src_region_destination = src_region_ptr->destination();
2702
2703 assert(dest_addr >= src_region_destination, "wrong src region");
2704 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2705
2706 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2707 HeapWord* const src_region_end = src_region_beg + RegionSize;
2708
2709 HeapWord* addr = src_region_beg;
2710 if (dest_addr == src_region_destination) {
2711 // Return the first live word in the source region.
2712 if (partial_obj_size == 0) {
2713 addr = bitmap->find_obj_beg(addr, src_region_end);
2714 assert(addr < src_region_end, "no objects start in src region");
2715 }
2716 return addr;
2717 }
2718
2719 // Must skip some live data.
2720 size_t words_to_skip = dest_addr - src_region_destination;
2721 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2722
2723 if (partial_obj_size >= words_to_skip) {
2724 // All the live words to skip are part of the partial object.
2725 addr += words_to_skip;
2726 if (partial_obj_size == words_to_skip) {
2727 // Find the first live word past the partial object.
2728 addr = bitmap->find_obj_beg(addr, src_region_end);
2729 assert(addr < src_region_end, "wrong src region");
2730 }
2731 return addr;
2732 }
2733
2734 // Skip over the partial object (if any).
2735 if (partial_obj_size != 0) {
2736 words_to_skip -= partial_obj_size;
2737 addr += partial_obj_size;
2738 }
2739
2740 // Skip over live words due to objects that start in the region.
2741 addr = skip_live_words(addr, src_region_end, words_to_skip);
2742 assert(addr < src_region_end, "wrong src region");
2743 return addr;
2744 }
2745
decrement_destination_counts(ParCompactionManager * cm,SpaceId src_space_id,size_t beg_region,HeapWord * end_addr)2746 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2747 SpaceId src_space_id,
2748 size_t beg_region,
2749 HeapWord* end_addr)
2750 {
2751 ParallelCompactData& sd = summary_data();
2752
2753 #ifdef ASSERT
2754 MutableSpace* const src_space = _space_info[src_space_id].space();
2755 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2756 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2757 "src_space_id does not match beg_addr");
2758 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2759 "src_space_id does not match end_addr");
2760 #endif // #ifdef ASSERT
2761
2762 RegionData* const beg = sd.region(beg_region);
2763 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2764
2765 // Regions up to new_top() are enqueued if they become available.
2766 HeapWord* const new_top = _space_info[src_space_id].new_top();
2767 RegionData* const enqueue_end =
2768 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2769
2770 for (RegionData* cur = beg; cur < end; ++cur) {
2771 assert(cur->data_size() > 0, "region must have live data");
2772 cur->decrement_destination_count();
2773 if (cur < enqueue_end && cur->available() && cur->claim()) {
2774 cm->push_region(sd.region(cur));
2775 }
2776 }
2777 }
2778
next_src_region(MoveAndUpdateClosure & closure,SpaceId & src_space_id,HeapWord * & src_space_top,HeapWord * end_addr)2779 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2780 SpaceId& src_space_id,
2781 HeapWord*& src_space_top,
2782 HeapWord* end_addr)
2783 {
2784 typedef ParallelCompactData::RegionData RegionData;
2785
2786 ParallelCompactData& sd = PSParallelCompact::summary_data();
2787 const size_t region_size = ParallelCompactData::RegionSize;
2788
2789 size_t src_region_idx = 0;
2790
2791 // Skip empty regions (if any) up to the top of the space.
2792 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2793 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2794 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2795 const RegionData* const top_region_ptr =
2796 sd.addr_to_region_ptr(top_aligned_up);
2797 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2798 ++src_region_ptr;
2799 }
2800
2801 if (src_region_ptr < top_region_ptr) {
2802 // The next source region is in the current space. Update src_region_idx
2803 // and the source address to match src_region_ptr.
2804 src_region_idx = sd.region(src_region_ptr);
2805 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2806 if (src_region_addr > closure.source()) {
2807 closure.set_source(src_region_addr);
2808 }
2809 return src_region_idx;
2810 }
2811
2812 // Switch to a new source space and find the first non-empty region.
2813 unsigned int space_id = src_space_id + 1;
2814 assert(space_id < last_space_id, "not enough spaces");
2815
2816 HeapWord* const destination = closure.destination();
2817
2818 do {
2819 MutableSpace* space = _space_info[space_id].space();
2820 HeapWord* const bottom = space->bottom();
2821 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2822
2823 // Iterate over the spaces that do not compact into themselves.
2824 if (bottom_cp->destination() != bottom) {
2825 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2826 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2827
2828 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2829 if (src_cp->live_obj_size() > 0) {
2830 // Found it.
2831 assert(src_cp->destination() == destination,
2832 "first live obj in the space must match the destination");
2833 assert(src_cp->partial_obj_size() == 0,
2834 "a space cannot begin with a partial obj");
2835
2836 src_space_id = SpaceId(space_id);
2837 src_space_top = space->top();
2838 const size_t src_region_idx = sd.region(src_cp);
2839 closure.set_source(sd.region_to_addr(src_region_idx));
2840 return src_region_idx;
2841 } else {
2842 assert(src_cp->data_size() == 0, "sanity");
2843 }
2844 }
2845 }
2846 } while (++space_id < last_space_id);
2847
2848 assert(false, "no source region was found");
2849 return 0;
2850 }
2851
fill_region(ParCompactionManager * cm,size_t region_idx)2852 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
2853 {
2854 typedef ParMarkBitMap::IterationStatus IterationStatus;
2855 const size_t RegionSize = ParallelCompactData::RegionSize;
2856 ParMarkBitMap* const bitmap = mark_bitmap();
2857 ParallelCompactData& sd = summary_data();
2858 RegionData* const region_ptr = sd.region(region_idx);
2859
2860 // Get the items needed to construct the closure.
2861 HeapWord* dest_addr = sd.region_to_addr(region_idx);
2862 SpaceId dest_space_id = space_id(dest_addr);
2863 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
2864 HeapWord* new_top = _space_info[dest_space_id].new_top();
2865 assert(dest_addr < new_top, "sanity");
2866 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
2867
2868 // Get the source region and related info.
2869 size_t src_region_idx = region_ptr->source_region();
2870 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2871 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2872
2873 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2874 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2875
2876 // Adjust src_region_idx to prepare for decrementing destination counts (the
2877 // destination count is not decremented when a region is copied to itself).
2878 if (src_region_idx == region_idx) {
2879 src_region_idx += 1;
2880 }
2881
2882 if (bitmap->is_unmarked(closure.source())) {
2883 // The first source word is in the middle of an object; copy the remainder
2884 // of the object or as much as will fit. The fact that pointer updates were
2885 // deferred will be noted when the object header is processed.
2886 HeapWord* const old_src_addr = closure.source();
2887 closure.copy_partial_obj();
2888 if (closure.is_full()) {
2889 decrement_destination_counts(cm, src_space_id, src_region_idx,
2890 closure.source());
2891 region_ptr->set_deferred_obj_addr(NULL);
2892 region_ptr->set_completed();
2893 return;
2894 }
2895
2896 HeapWord* const end_addr = sd.region_align_down(closure.source());
2897 if (sd.region_align_down(old_src_addr) != end_addr) {
2898 // The partial object was copied from more than one source region.
2899 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2900
2901 // Move to the next source region, possibly switching spaces as well. All
2902 // args except end_addr may be modified.
2903 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2904 end_addr);
2905 }
2906 }
2907
2908 do {
2909 HeapWord* const cur_addr = closure.source();
2910 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2911 src_space_top);
2912 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2913
2914 if (status == ParMarkBitMap::incomplete) {
2915 // The last obj that starts in the source region does not end in the
2916 // region.
2917 assert(closure.source() < end_addr, "sanity");
2918 HeapWord* const obj_beg = closure.source();
2919 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2920 src_space_top);
2921 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2922 if (obj_end < range_end) {
2923 // The end was found; the entire object will fit.
2924 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2925 assert(status != ParMarkBitMap::would_overflow, "sanity");
2926 } else {
2927 // The end was not found; the object will not fit.
2928 assert(range_end < src_space_top, "obj cannot cross space boundary");
2929 status = ParMarkBitMap::would_overflow;
2930 }
2931 }
2932
2933 if (status == ParMarkBitMap::would_overflow) {
2934 // The last object did not fit. Note that interior oop updates were
2935 // deferred, then copy enough of the object to fill the region.
2936 region_ptr->set_deferred_obj_addr(closure.destination());
2937 status = closure.copy_until_full(); // copies from closure.source()
2938
2939 decrement_destination_counts(cm, src_space_id, src_region_idx,
2940 closure.source());
2941 region_ptr->set_completed();
2942 return;
2943 }
2944
2945 if (status == ParMarkBitMap::full) {
2946 decrement_destination_counts(cm, src_space_id, src_region_idx,
2947 closure.source());
2948 region_ptr->set_deferred_obj_addr(NULL);
2949 region_ptr->set_completed();
2950 return;
2951 }
2952
2953 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2954
2955 // Move to the next source region, possibly switching spaces as well. All
2956 // args except end_addr may be modified.
2957 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2958 end_addr);
2959 } while (true);
2960 }
2961
fill_blocks(size_t region_idx)2962 void PSParallelCompact::fill_blocks(size_t region_idx)
2963 {
2964 // Fill in the block table elements for the specified region. Each block
2965 // table element holds the number of live words in the region that are to the
2966 // left of the first object that starts in the block. Thus only blocks in
2967 // which an object starts need to be filled.
2968 //
2969 // The algorithm scans the section of the bitmap that corresponds to the
2970 // region, keeping a running total of the live words. When an object start is
2971 // found, if it's the first to start in the block that contains it, the
2972 // current total is written to the block table element.
2973 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
2974 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
2975 const size_t RegionSize = ParallelCompactData::RegionSize;
2976
2977 ParallelCompactData& sd = summary_data();
2978 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
2979 if (partial_obj_size >= RegionSize) {
2980 return; // No objects start in this region.
2981 }
2982
2983 // Ensure the first loop iteration decides that the block has changed.
2984 size_t cur_block = sd.block_count();
2985
2986 const ParMarkBitMap* const bitmap = mark_bitmap();
2987
2988 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
2989 assert((size_t)1 << Log2BitsPerBlock ==
2990 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
2991
2992 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
2993 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
2994 size_t live_bits = bitmap->words_to_bits(partial_obj_size);
2995 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
2996 while (beg_bit < range_end) {
2997 const size_t new_block = beg_bit >> Log2BitsPerBlock;
2998 if (new_block != cur_block) {
2999 cur_block = new_block;
3000 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3001 }
3002
3003 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3004 if (end_bit < range_end - 1) {
3005 live_bits += end_bit - beg_bit + 1;
3006 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3007 } else {
3008 return;
3009 }
3010 }
3011 }
3012
3013 void
move_and_update(ParCompactionManager * cm,SpaceId space_id)3014 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3015 const MutableSpace* sp = space(space_id);
3016 if (sp->is_empty()) {
3017 return;
3018 }
3019
3020 ParallelCompactData& sd = PSParallelCompact::summary_data();
3021 ParMarkBitMap* const bitmap = mark_bitmap();
3022 HeapWord* const dp_addr = dense_prefix(space_id);
3023 HeapWord* beg_addr = sp->bottom();
3024 HeapWord* end_addr = sp->top();
3025
3026 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3027
3028 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3029 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3030 if (beg_region < dp_region) {
3031 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3032 }
3033
3034 // The destination of the first live object that starts in the region is one
3035 // past the end of the partial object entering the region (if any).
3036 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3037 HeapWord* const new_top = _space_info[space_id].new_top();
3038 assert(new_top >= dest_addr, "bad new_top value");
3039 const size_t words = pointer_delta(new_top, dest_addr);
3040
3041 if (words > 0) {
3042 ObjectStartArray* start_array = _space_info[space_id].start_array();
3043 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3044
3045 ParMarkBitMap::IterationStatus status;
3046 status = bitmap->iterate(&closure, dest_addr, end_addr);
3047 assert(status == ParMarkBitMap::full, "iteration not complete");
3048 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3049 "live objects skipped because closure is full");
3050 }
3051 }
3052
millis_since_last_gc()3053 jlong PSParallelCompact::millis_since_last_gc() {
3054 // We need a monotonically non-decreasing time in ms but
3055 // os::javaTimeMillis() does not guarantee monotonicity.
3056 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3057 jlong ret_val = now - _time_of_last_gc;
3058 // XXX See note in genCollectedHeap::millis_since_last_gc().
3059 if (ret_val < 0) {
3060 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);)
3061 return 0;
3062 }
3063 return ret_val;
3064 }
3065
reset_millis_since_last_gc()3066 void PSParallelCompact::reset_millis_since_last_gc() {
3067 // We need a monotonically non-decreasing time in ms but
3068 // os::javaTimeMillis() does not guarantee monotonicity.
3069 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3070 }
3071
copy_until_full()3072 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3073 {
3074 if (source() != destination()) {
3075 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3076 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3077 }
3078 update_state(words_remaining());
3079 assert(is_full(), "sanity");
3080 return ParMarkBitMap::full;
3081 }
3082
copy_partial_obj()3083 void MoveAndUpdateClosure::copy_partial_obj()
3084 {
3085 size_t words = words_remaining();
3086
3087 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3088 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3089 if (end_addr < range_end) {
3090 words = bitmap()->obj_size(source(), end_addr);
3091 }
3092
3093 // This test is necessary; if omitted, the pointer updates to a partial object
3094 // that crosses the dense prefix boundary could be overwritten.
3095 if (source() != destination()) {
3096 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3097 Copy::aligned_conjoint_words(source(), destination(), words);
3098 }
3099 update_state(words);
3100 }
3101
3102 ParMarkBitMapClosure::IterationStatus
do_addr(HeapWord * addr,size_t words)3103 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3104 assert(destination() != NULL, "sanity");
3105 assert(bitmap()->obj_size(addr) == words, "bad size");
3106
3107 _source = addr;
3108 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3109 destination(), "wrong destination");
3110
3111 if (words > words_remaining()) {
3112 return ParMarkBitMap::would_overflow;
3113 }
3114
3115 // The start_array must be updated even if the object is not moving.
3116 if (_start_array != NULL) {
3117 _start_array->allocate_block(destination());
3118 }
3119
3120 if (destination() != source()) {
3121 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3122 Copy::aligned_conjoint_words(source(), destination(), words);
3123 }
3124
3125 oop moved_oop = (oop) destination();
3126 compaction_manager()->update_contents(moved_oop);
3127 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3128
3129 update_state(words);
3130 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3131 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3132 }
3133
UpdateOnlyClosure(ParMarkBitMap * mbm,ParCompactionManager * cm,PSParallelCompact::SpaceId space_id)3134 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3135 ParCompactionManager* cm,
3136 PSParallelCompact::SpaceId space_id) :
3137 ParMarkBitMapClosure(mbm, cm),
3138 _space_id(space_id),
3139 _start_array(PSParallelCompact::start_array(space_id))
3140 {
3141 }
3142
3143 // Updates the references in the object to their new values.
3144 ParMarkBitMapClosure::IterationStatus
do_addr(HeapWord * addr,size_t words)3145 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3146 do_addr(addr);
3147 return ParMarkBitMap::incomplete;
3148 }
3149
FillClosure(ParCompactionManager * cm,PSParallelCompact::SpaceId space_id)3150 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3151 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3152 _start_array(PSParallelCompact::start_array(space_id))
3153 {
3154 assert(space_id == PSParallelCompact::old_space_id,
3155 "cannot use FillClosure in the young gen");
3156 }
3157
3158 ParMarkBitMapClosure::IterationStatus
do_addr(HeapWord * addr,size_t size)3159 FillClosure::do_addr(HeapWord* addr, size_t size) {
3160 CollectedHeap::fill_with_objects(addr, size);
3161 HeapWord* const end = addr + size;
3162 do {
3163 _start_array->allocate_block(addr);
3164 addr += oop(addr)->size();
3165 } while (addr < end);
3166 return ParMarkBitMap::incomplete;
3167 }
3168