1 /*
2 * Copyright (c) 2005, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
30 #include "gc_implementation/parallelScavenge/generationSizer.hpp"
31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
32 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
37 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
39 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
40 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
41 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/gcCause.hpp"
44 #include "memory/gcLocker.inline.hpp"
45 #include "memory/referencePolicy.hpp"
46 #include "memory/referenceProcessor.hpp"
47 #include "oops/methodData.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "oops/oop.pcgc.inline.hpp"
50 #include "runtime/fprofiler.hpp"
51 #include "runtime/safepoint.hpp"
52 #include "runtime/vmThread.hpp"
53 #include "services/management.hpp"
54 #include "services/memoryService.hpp"
55 #include "services/memTracker.hpp"
56 #include "utilities/events.hpp"
57 #include "utilities/stack.inline.hpp"
58
59 #include <math.h>
60
61 // All sizes are in HeapWords.
62 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words
63 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
64 const size_t ParallelCompactData::RegionSizeBytes =
65 RegionSize << LogHeapWordSize;
66 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
67 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
68 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
69
70 const ParallelCompactData::RegionData::region_sz_t
71 ParallelCompactData::RegionData::dc_shift = 27;
72
73 const ParallelCompactData::RegionData::region_sz_t
74 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
75
76 const ParallelCompactData::RegionData::region_sz_t
77 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
78
79 const ParallelCompactData::RegionData::region_sz_t
80 ParallelCompactData::RegionData::los_mask = ~dc_mask;
81
82 const ParallelCompactData::RegionData::region_sz_t
83 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
84
85 const ParallelCompactData::RegionData::region_sz_t
86 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
87
88 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
89 bool PSParallelCompact::_print_phases = false;
90
91 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
92 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL;
93
94 double PSParallelCompact::_dwl_mean;
95 double PSParallelCompact::_dwl_std_dev;
96 double PSParallelCompact::_dwl_first_term;
97 double PSParallelCompact::_dwl_adjustment;
98 #ifdef ASSERT
99 bool PSParallelCompact::_dwl_initialized = false;
100 #endif // #ifdef ASSERT
101
102 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
103 HeapWord* destination)
104 {
105 assert(src_region_idx != 0, "invalid src_region_idx");
106 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
107 assert(destination != NULL, "invalid destination argument");
108
109 _src_region_idx = src_region_idx;
110 _partial_obj_size = partial_obj_size;
111 _destination = destination;
112
113 // These fields may not be updated below, so make sure they're clear.
114 assert(_dest_region_addr == NULL, "should have been cleared");
115 assert(_first_src_addr == NULL, "should have been cleared");
116
117 // Determine the number of destination regions for the partial object.
118 HeapWord* const last_word = destination + partial_obj_size - 1;
119 const ParallelCompactData& sd = PSParallelCompact::summary_data();
120 HeapWord* const beg_region_addr = sd.region_align_down(destination);
121 HeapWord* const end_region_addr = sd.region_align_down(last_word);
122
123 if (beg_region_addr == end_region_addr) {
124 // One destination region.
125 _destination_count = 1;
126 if (end_region_addr == destination) {
127 // The destination falls on a region boundary, thus the first word of the
128 // partial object will be the first word copied to the destination region.
129 _dest_region_addr = end_region_addr;
130 _first_src_addr = sd.region_to_addr(src_region_idx);
131 }
132 } else {
133 // Two destination regions. When copied, the partial object will cross a
134 // destination region boundary, so a word somewhere within the partial
135 // object will be the first word copied to the second destination region.
136 _destination_count = 2;
137 _dest_region_addr = end_region_addr;
138 const size_t ofs = pointer_delta(end_region_addr, destination);
139 assert(ofs < _partial_obj_size, "sanity");
140 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
141 }
142 }
143
144 void SplitInfo::clear()
145 {
146 _src_region_idx = 0;
147 _partial_obj_size = 0;
148 _destination = NULL;
149 _destination_count = 0;
150 _dest_region_addr = NULL;
151 _first_src_addr = NULL;
152 assert(!is_valid(), "sanity");
153 }
154
155 #ifdef ASSERT
156 void SplitInfo::verify_clear()
157 {
158 assert(_src_region_idx == 0, "not clear");
159 assert(_partial_obj_size == 0, "not clear");
160 assert(_destination == NULL, "not clear");
161 assert(_destination_count == 0, "not clear");
162 assert(_dest_region_addr == NULL, "not clear");
163 assert(_first_src_addr == NULL, "not clear");
164 }
165 #endif // #ifdef ASSERT
166
167
168 void PSParallelCompact::print_on_error(outputStream* st) {
169 _mark_bitmap.print_on_error(st);
170 }
171
172 #ifndef PRODUCT
173 const char* PSParallelCompact::space_names[] = {
174 "old ", "eden", "from", "to "
175 };
176
177 void PSParallelCompact::print_region_ranges()
178 {
179 tty->print_cr("space bottom top end new_top");
180 tty->print_cr("------ ---------- ---------- ---------- ----------");
181
182 for (unsigned int id = 0; id < last_space_id; ++id) {
183 const MutableSpace* space = _space_info[id].space();
184 tty->print_cr("%u %s "
185 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
186 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
187 id, space_names[id],
188 summary_data().addr_to_region_idx(space->bottom()),
189 summary_data().addr_to_region_idx(space->top()),
190 summary_data().addr_to_region_idx(space->end()),
191 summary_data().addr_to_region_idx(_space_info[id].new_top()));
192 }
193 }
194
195 void
196 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
197 {
198 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
199 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
200
201 ParallelCompactData& sd = PSParallelCompact::summary_data();
202 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
203 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
204 REGION_IDX_FORMAT " " PTR_FORMAT " "
205 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
206 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
207 i, c->data_location(), dci, c->destination(),
208 c->partial_obj_size(), c->live_obj_size(),
209 c->data_size(), c->source_region(), c->destination_count());
210
211 #undef REGION_IDX_FORMAT
212 #undef REGION_DATA_FORMAT
213 }
214
215 void
216 print_generic_summary_data(ParallelCompactData& summary_data,
217 HeapWord* const beg_addr,
218 HeapWord* const end_addr)
219 {
220 size_t total_words = 0;
221 size_t i = summary_data.addr_to_region_idx(beg_addr);
222 const size_t last = summary_data.addr_to_region_idx(end_addr);
223 HeapWord* pdest = 0;
224
225 while (i <= last) {
226 ParallelCompactData::RegionData* c = summary_data.region(i);
227 if (c->data_size() != 0 || c->destination() != pdest) {
228 print_generic_summary_region(i, c);
229 total_words += c->data_size();
230 pdest = c->destination();
231 }
232 ++i;
233 }
234
235 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
236 }
237
238 void
239 print_generic_summary_data(ParallelCompactData& summary_data,
240 SpaceInfo* space_info)
241 {
242 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
243 const MutableSpace* space = space_info[id].space();
244 print_generic_summary_data(summary_data, space->bottom(),
245 MAX2(space->top(), space_info[id].new_top()));
246 }
247 }
248
249 void
250 print_initial_summary_region(size_t i,
251 const ParallelCompactData::RegionData* c,
252 bool newline = true)
253 {
254 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
255 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
256 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
257 i, c->destination(),
258 c->partial_obj_size(), c->live_obj_size(),
259 c->data_size(), c->source_region(), c->destination_count());
260 if (newline) tty->cr();
261 }
262
263 void
264 print_initial_summary_data(ParallelCompactData& summary_data,
265 const MutableSpace* space) {
266 if (space->top() == space->bottom()) {
267 return;
268 }
269
270 const size_t region_size = ParallelCompactData::RegionSize;
271 typedef ParallelCompactData::RegionData RegionData;
272 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
273 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
274 const RegionData* c = summary_data.region(end_region - 1);
275 HeapWord* end_addr = c->destination() + c->data_size();
276 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
277
278 // Print (and count) the full regions at the beginning of the space.
279 size_t full_region_count = 0;
280 size_t i = summary_data.addr_to_region_idx(space->bottom());
281 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
282 print_initial_summary_region(i, summary_data.region(i));
283 ++full_region_count;
284 ++i;
285 }
286
287 size_t live_to_right = live_in_space - full_region_count * region_size;
288
289 double max_reclaimed_ratio = 0.0;
290 size_t max_reclaimed_ratio_region = 0;
291 size_t max_dead_to_right = 0;
292 size_t max_live_to_right = 0;
293
294 // Print the 'reclaimed ratio' for regions while there is something live in
295 // the region or to the right of it. The remaining regions are empty (and
296 // uninteresting), and computing the ratio will result in division by 0.
297 while (i < end_region && live_to_right > 0) {
298 c = summary_data.region(i);
299 HeapWord* const region_addr = summary_data.region_to_addr(i);
300 const size_t used_to_right = pointer_delta(space->top(), region_addr);
301 const size_t dead_to_right = used_to_right - live_to_right;
302 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
303
304 if (reclaimed_ratio > max_reclaimed_ratio) {
305 max_reclaimed_ratio = reclaimed_ratio;
306 max_reclaimed_ratio_region = i;
307 max_dead_to_right = dead_to_right;
308 max_live_to_right = live_to_right;
309 }
310
311 print_initial_summary_region(i, c, false);
312 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
313 reclaimed_ratio, dead_to_right, live_to_right);
314
315 live_to_right -= c->data_size();
316 ++i;
317 }
318
319 // Any remaining regions are empty. Print one more if there is one.
320 if (i < end_region) {
321 print_initial_summary_region(i, summary_data.region(i));
322 }
323
324 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
325 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
326 max_reclaimed_ratio_region, max_dead_to_right,
327 max_live_to_right, max_reclaimed_ratio);
328 }
329
330 void
331 print_initial_summary_data(ParallelCompactData& summary_data,
332 SpaceInfo* space_info) {
333 unsigned int id = PSParallelCompact::old_space_id;
334 const MutableSpace* space;
335 do {
336 space = space_info[id].space();
337 print_initial_summary_data(summary_data, space);
338 } while (++id < PSParallelCompact::eden_space_id);
339
340 do {
341 space = space_info[id].space();
342 print_generic_summary_data(summary_data, space->bottom(), space->top());
343 } while (++id < PSParallelCompact::last_space_id);
344 }
345 #endif // #ifndef PRODUCT
346
347 #ifdef ASSERT
348 size_t add_obj_count;
349 size_t add_obj_size;
350 size_t mark_bitmap_count;
351 size_t mark_bitmap_size;
352 #endif // #ifdef ASSERT
353
354 ParallelCompactData::ParallelCompactData()
355 {
356 _region_start = 0;
357
358 _region_vspace = 0;
359 _reserved_byte_size = 0;
360 _region_data = 0;
361 _region_count = 0;
362 }
363
364 bool ParallelCompactData::initialize(MemRegion covered_region)
365 {
366 _region_start = covered_region.start();
367 const size_t region_size = covered_region.word_size();
368 DEBUG_ONLY(_region_end = _region_start + region_size;)
369
370 assert(region_align_down(_region_start) == _region_start,
371 "region start not aligned");
372 assert((region_size & RegionSizeOffsetMask) == 0,
373 "region size not a multiple of RegionSize");
374
375 bool result = initialize_region_data(region_size);
376
377 return result;
378 }
379
380 PSVirtualSpace*
381 ParallelCompactData::create_vspace(size_t count, size_t element_size)
382 {
383 const size_t raw_bytes = count * element_size;
384 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
385 const size_t granularity = os::vm_allocation_granularity();
386 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
387
388 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
389 MAX2(page_sz, granularity);
390 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
391 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
392 rs.size());
393
394 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
395
396 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
397 if (vspace != 0) {
398 if (vspace->expand_by(_reserved_byte_size)) {
399 return vspace;
400 }
401 delete vspace;
402 // Release memory reserved in the space.
403 rs.release();
404 }
405
406 return 0;
407 }
408
409 bool ParallelCompactData::initialize_region_data(size_t region_size)
410 {
411 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
412 _region_vspace = create_vspace(count, sizeof(RegionData));
413 if (_region_vspace != 0) {
414 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
415 _region_count = count;
416 return true;
417 }
418 return false;
419 }
420
421 void ParallelCompactData::clear()
422 {
423 memset(_region_data, 0, _region_vspace->committed_size());
424 }
425
426 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
427 assert(beg_region <= _region_count, "beg_region out of range");
428 assert(end_region <= _region_count, "end_region out of range");
429
430 const size_t region_cnt = end_region - beg_region;
431 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
432 }
433
434 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
435 {
436 const RegionData* cur_cp = region(region_idx);
437 const RegionData* const end_cp = region(region_count() - 1);
438
439 HeapWord* result = region_to_addr(region_idx);
440 if (cur_cp < end_cp) {
441 do {
442 result += cur_cp->partial_obj_size();
443 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
444 }
445 return result;
446 }
447
448 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
449 {
450 const size_t obj_ofs = pointer_delta(addr, _region_start);
451 const size_t beg_region = obj_ofs >> Log2RegionSize;
452 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
453
454 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
455 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
456
457 if (beg_region == end_region) {
458 // All in one region.
459 _region_data[beg_region].add_live_obj(len);
460 return;
461 }
462
463 // First region.
464 const size_t beg_ofs = region_offset(addr);
465 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
466
467 Klass* klass = ((oop)addr)->klass();
468 // Middle regions--completely spanned by this object.
469 for (size_t region = beg_region + 1; region < end_region; ++region) {
470 _region_data[region].set_partial_obj_size(RegionSize);
471 _region_data[region].set_partial_obj_addr(addr);
472 }
473
474 // Last region.
475 const size_t end_ofs = region_offset(addr + len - 1);
476 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
477 _region_data[end_region].set_partial_obj_addr(addr);
478 }
479
480 void
481 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
482 {
483 assert(region_offset(beg) == 0, "not RegionSize aligned");
484 assert(region_offset(end) == 0, "not RegionSize aligned");
485
486 size_t cur_region = addr_to_region_idx(beg);
487 const size_t end_region = addr_to_region_idx(end);
488 HeapWord* addr = beg;
489 while (cur_region < end_region) {
490 _region_data[cur_region].set_destination(addr);
491 _region_data[cur_region].set_destination_count(0);
492 _region_data[cur_region].set_source_region(cur_region);
493 _region_data[cur_region].set_data_location(addr);
494
495 // Update live_obj_size so the region appears completely full.
496 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
497 _region_data[cur_region].set_live_obj_size(live_size);
498
499 ++cur_region;
500 addr += RegionSize;
501 }
502 }
503
504 // Find the point at which a space can be split and, if necessary, record the
505 // split point.
506 //
507 // If the current src region (which overflowed the destination space) doesn't
508 // have a partial object, the split point is at the beginning of the current src
509 // region (an "easy" split, no extra bookkeeping required).
510 //
511 // If the current src region has a partial object, the split point is in the
512 // region where that partial object starts (call it the split_region). If
513 // split_region has a partial object, then the split point is just after that
514 // partial object (a "hard" split where we have to record the split data and
515 // zero the partial_obj_size field). With a "hard" split, we know that the
516 // partial_obj ends within split_region because the partial object that caused
517 // the overflow starts in split_region. If split_region doesn't have a partial
518 // obj, then the split is at the beginning of split_region (another "easy"
519 // split).
520 HeapWord*
521 ParallelCompactData::summarize_split_space(size_t src_region,
522 SplitInfo& split_info,
523 HeapWord* destination,
524 HeapWord* target_end,
525 HeapWord** target_next)
526 {
527 assert(destination <= target_end, "sanity");
528 assert(destination + _region_data[src_region].data_size() > target_end,
529 "region should not fit into target space");
530 assert(is_region_aligned(target_end), "sanity");
531
532 size_t split_region = src_region;
533 HeapWord* split_destination = destination;
534 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
535
536 if (destination + partial_obj_size > target_end) {
537 // The split point is just after the partial object (if any) in the
538 // src_region that contains the start of the object that overflowed the
539 // destination space.
540 //
541 // Find the start of the "overflow" object and set split_region to the
542 // region containing it.
543 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
544 split_region = addr_to_region_idx(overflow_obj);
545
546 // Clear the source_region field of all destination regions whose first word
547 // came from data after the split point (a non-null source_region field
548 // implies a region must be filled).
549 //
550 // An alternative to the simple loop below: clear during post_compact(),
551 // which uses memcpy instead of individual stores, and is easy to
552 // parallelize. (The downside is that it clears the entire RegionData
553 // object as opposed to just one field.)
554 //
555 // post_compact() would have to clear the summary data up to the highest
556 // address that was written during the summary phase, which would be
557 //
558 // max(top, max(new_top, clear_top))
559 //
560 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
561 // to target_end.
562 const RegionData* const sr = region(split_region);
563 const size_t beg_idx =
564 addr_to_region_idx(region_align_up(sr->destination() +
565 sr->partial_obj_size()));
566 const size_t end_idx = addr_to_region_idx(target_end);
567
568 if (TraceParallelOldGCSummaryPhase) {
569 gclog_or_tty->print_cr("split: clearing source_region field in ["
570 SIZE_FORMAT ", " SIZE_FORMAT ")",
571 beg_idx, end_idx);
572 }
573 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
574 _region_data[idx].set_source_region(0);
575 }
576
577 // Set split_destination and partial_obj_size to reflect the split region.
578 split_destination = sr->destination();
579 partial_obj_size = sr->partial_obj_size();
580 }
581
582 // The split is recorded only if a partial object extends onto the region.
583 if (partial_obj_size != 0) {
584 _region_data[split_region].set_partial_obj_size(0);
585 split_info.record(split_region, partial_obj_size, split_destination);
586 }
587
588 // Setup the continuation addresses.
589 *target_next = split_destination + partial_obj_size;
590 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
591
592 if (TraceParallelOldGCSummaryPhase) {
593 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
594 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
595 " pos=" SIZE_FORMAT,
596 split_type, source_next, split_region,
597 partial_obj_size);
598 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
599 " tn=" PTR_FORMAT,
600 split_type, split_destination,
601 addr_to_region_idx(split_destination),
602 *target_next);
603
604 if (partial_obj_size != 0) {
605 HeapWord* const po_beg = split_info.destination();
606 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
607 gclog_or_tty->print_cr("%s split: "
608 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
609 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
610 split_type,
611 po_beg, addr_to_region_idx(po_beg),
612 po_end, addr_to_region_idx(po_end));
613 }
614 }
615
616 return source_next;
617 }
618
619 bool ParallelCompactData::summarize(SplitInfo& split_info,
620 HeapWord* source_beg, HeapWord* source_end,
621 HeapWord** source_next,
622 HeapWord* target_beg, HeapWord* target_end,
623 HeapWord** target_next)
624 {
625 if (TraceParallelOldGCSummaryPhase) {
626 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
627 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
628 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
629 source_beg, source_end, source_next_val,
630 target_beg, target_end, *target_next);
631 }
632
633 size_t cur_region = addr_to_region_idx(source_beg);
634 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
635
636 HeapWord *dest_addr = target_beg;
637 while (cur_region < end_region) {
638 // The destination must be set even if the region has no data.
639 _region_data[cur_region].set_destination(dest_addr);
640
641 size_t words = _region_data[cur_region].data_size();
642 if (words > 0) {
643 // If cur_region does not fit entirely into the target space, find a point
644 // at which the source space can be 'split' so that part is copied to the
645 // target space and the rest is copied elsewhere.
646 if (dest_addr + words > target_end) {
647 assert(source_next != NULL, "source_next is NULL when splitting");
648 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
649 target_end, target_next);
650 return false;
651 }
652
653 // Compute the destination_count for cur_region, and if necessary, update
654 // source_region for a destination region. The source_region field is
655 // updated if cur_region is the first (left-most) region to be copied to a
656 // destination region.
657 //
658 // The destination_count calculation is a bit subtle. A region that has
659 // data that compacts into itself does not count itself as a destination.
660 // This maintains the invariant that a zero count means the region is
661 // available and can be claimed and then filled.
662 uint destination_count = 0;
663 if (split_info.is_split(cur_region)) {
664 // The current region has been split: the partial object will be copied
665 // to one destination space and the remaining data will be copied to
666 // another destination space. Adjust the initial destination_count and,
667 // if necessary, set the source_region field if the partial object will
668 // cross a destination region boundary.
669 destination_count = split_info.destination_count();
670 if (destination_count == 2) {
671 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
672 _region_data[dest_idx].set_source_region(cur_region);
673 }
674 }
675
676 HeapWord* const last_addr = dest_addr + words - 1;
677 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
678 const size_t dest_region_2 = addr_to_region_idx(last_addr);
679
680 // Initially assume that the destination regions will be the same and
681 // adjust the value below if necessary. Under this assumption, if
682 // cur_region == dest_region_2, then cur_region will be compacted
683 // completely into itself.
684 destination_count += cur_region == dest_region_2 ? 0 : 1;
685 if (dest_region_1 != dest_region_2) {
686 // Destination regions differ; adjust destination_count.
687 destination_count += 1;
688 // Data from cur_region will be copied to the start of dest_region_2.
689 _region_data[dest_region_2].set_source_region(cur_region);
690 } else if (region_offset(dest_addr) == 0) {
691 // Data from cur_region will be copied to the start of the destination
692 // region.
693 _region_data[dest_region_1].set_source_region(cur_region);
694 }
695
696 _region_data[cur_region].set_destination_count(destination_count);
697 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
698 dest_addr += words;
699 }
700
701 ++cur_region;
702 }
703
704 *target_next = dest_addr;
705 return true;
706 }
707
708 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
709 assert(addr != NULL, "Should detect NULL oop earlier");
710 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
711 #ifdef ASSERT
712 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
713 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
714 }
715 #endif
716 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
717
718 // Region covering the object.
719 size_t region_index = addr_to_region_idx(addr);
720 const RegionData* const region_ptr = region(region_index);
721 HeapWord* const region_addr = region_align_down(addr);
722
723 assert(addr < region_addr + RegionSize, "Region does not cover object");
724 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
725
726 HeapWord* result = region_ptr->destination();
727
728 // If all the data in the region is live, then the new location of the object
729 // can be calculated from the destination of the region plus the offset of the
730 // object in the region.
731 if (region_ptr->data_size() == RegionSize) {
732 result += pointer_delta(addr, region_addr);
733 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
734 return result;
735 }
736
737 // The new location of the object is
738 // region destination +
739 // size of the partial object extending onto the region +
740 // sizes of the live objects in the Region that are to the left of addr
741 const size_t partial_obj_size = region_ptr->partial_obj_size();
742 HeapWord* const search_start = region_addr + partial_obj_size;
743
744 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
745 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
746
747 result += partial_obj_size + live_to_left;
748 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
749 return result;
750 }
751
752 #ifdef ASSERT
753 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
754 {
755 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
756 const size_t* const end = (const size_t*)vspace->committed_high_addr();
757 for (const size_t* p = beg; p < end; ++p) {
758 assert(*p == 0, "not zero");
759 }
760 }
761
762 void ParallelCompactData::verify_clear()
763 {
764 verify_clear(_region_vspace);
765 }
766 #endif // #ifdef ASSERT
767
768 #ifdef NOT_PRODUCT
769 ParallelCompactData::RegionData* debug_region(size_t region_index) {
770 ParallelCompactData& sd = PSParallelCompact::summary_data();
771 return sd.region(region_index);
772 }
773 #endif
774
775 elapsedTimer PSParallelCompact::_accumulated_time;
776 unsigned int PSParallelCompact::_total_invocations = 0;
777 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
778 jlong PSParallelCompact::_time_of_last_gc = 0;
779 CollectorCounters* PSParallelCompact::_counters = NULL;
780 ParMarkBitMap PSParallelCompact::_mark_bitmap;
781 ParallelCompactData PSParallelCompact::_summary_data;
782
783 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
784
785 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
786 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
787
788 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
789 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
790
791 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
792 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
793
794 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); }
795 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
796
797 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
798
799 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) {
800 mark_and_push(_compaction_manager, p);
801 }
802 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
803
804 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
805 klass->oops_do(_mark_and_push_closure);
806 }
807 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
808 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
809 }
810
811 void PSParallelCompact::post_initialize() {
812 ParallelScavengeHeap* heap = gc_heap();
813 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
814
815 MemRegion mr = heap->reserved_region();
816 _ref_processor =
817 new ReferenceProcessor(mr, // span
818 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
819 (int) ParallelGCThreads, // mt processing degree
820 true, // mt discovery
821 (int) ParallelGCThreads, // mt discovery degree
822 true, // atomic_discovery
823 &_is_alive_closure, // non-header is alive closure
824 false); // write barrier for next field updates
825 _counters = new CollectorCounters("PSParallelCompact", 1);
826
827 // Initialize static fields in ParCompactionManager.
828 ParCompactionManager::initialize(mark_bitmap());
829 }
830
831 bool PSParallelCompact::initialize() {
832 ParallelScavengeHeap* heap = gc_heap();
833 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
834 MemRegion mr = heap->reserved_region();
835
836 // Was the old gen get allocated successfully?
837 if (!heap->old_gen()->is_allocated()) {
838 return false;
839 }
840
841 initialize_space_info();
842 initialize_dead_wood_limiter();
843
844 if (!_mark_bitmap.initialize(mr)) {
845 vm_shutdown_during_initialization(
846 err_msg("Unable to allocate " SIZE_FORMAT " byte bitmap "
847 "for parallel garbage collection for the requested " SIZE_FORMAT " byte heap size.",
848 _mark_bitmap.reserved_byte_size(), mr.byte_size()));
849 return false;
850 }
851
852 if (!_summary_data.initialize(mr)) {
853 vm_shutdown_during_initialization(
854 err_msg("Unable to allocate " SIZE_FORMAT " byte tables "
855 "for parallel garbage collection for the requested " SIZE_FORMAT " byte heap size.",
856 _summary_data.reserved_byte_size(), mr.byte_size()));
857 return false;
858 }
859
860 return true;
861 }
862
863 void PSParallelCompact::initialize_space_info()
864 {
865 memset(&_space_info, 0, sizeof(_space_info));
866
867 ParallelScavengeHeap* heap = gc_heap();
868 PSYoungGen* young_gen = heap->young_gen();
869
870 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
871 _space_info[eden_space_id].set_space(young_gen->eden_space());
872 _space_info[from_space_id].set_space(young_gen->from_space());
873 _space_info[to_space_id].set_space(young_gen->to_space());
874
875 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
876 }
877
878 void PSParallelCompact::initialize_dead_wood_limiter()
879 {
880 const size_t max = 100;
881 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
882 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
883 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
884 DEBUG_ONLY(_dwl_initialized = true;)
885 _dwl_adjustment = normal_distribution(1.0);
886 }
887
888 // Simple class for storing info about the heap at the start of GC, to be used
889 // after GC for comparison/printing.
890 class PreGCValues {
891 public:
892 PreGCValues() { }
893 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
894
895 void fill(ParallelScavengeHeap* heap) {
896 _heap_used = heap->used();
897 _young_gen_used = heap->young_gen()->used_in_bytes();
898 _old_gen_used = heap->old_gen()->used_in_bytes();
899 _metadata_used = MetaspaceAux::allocated_used_bytes();
900 };
901
902 size_t heap_used() const { return _heap_used; }
903 size_t young_gen_used() const { return _young_gen_used; }
904 size_t old_gen_used() const { return _old_gen_used; }
905 size_t metadata_used() const { return _metadata_used; }
906
907 private:
908 size_t _heap_used;
909 size_t _young_gen_used;
910 size_t _old_gen_used;
911 size_t _metadata_used;
912 };
913
914 void
915 PSParallelCompact::clear_data_covering_space(SpaceId id)
916 {
917 // At this point, top is the value before GC, new_top() is the value that will
918 // be set at the end of GC. The marking bitmap is cleared to top; nothing
919 // should be marked above top. The summary data is cleared to the larger of
920 // top & new_top.
921 MutableSpace* const space = _space_info[id].space();
922 HeapWord* const bot = space->bottom();
923 HeapWord* const top = space->top();
924 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
925
926 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
927 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
928 _mark_bitmap.clear_range(beg_bit, end_bit);
929
930 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
931 const size_t end_region =
932 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
933 _summary_data.clear_range(beg_region, end_region);
934
935 // Clear the data used to 'split' regions.
936 SplitInfo& split_info = _space_info[id].split_info();
937 if (split_info.is_valid()) {
938 split_info.clear();
939 }
940 DEBUG_ONLY(split_info.verify_clear();)
941 }
942
943 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
944 {
945 // Update the from & to space pointers in space_info, since they are swapped
946 // at each young gen gc. Do the update unconditionally (even though a
947 // promotion failure does not swap spaces) because an unknown number of minor
948 // collections will have swapped the spaces an unknown number of times.
949 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
950 ParallelScavengeHeap* heap = gc_heap();
951 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
952 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
953
954 pre_gc_values->fill(heap);
955
956 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
957 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
958
959 // Increment the invocation count
960 heap->increment_total_collections(true);
961
962 // We need to track unique mark sweep invocations as well.
963 _total_invocations++;
964
965 heap->print_heap_before_gc();
966
967 // Fill in TLABs
968 heap->accumulate_statistics_all_tlabs();
969 heap->ensure_parsability(true); // retire TLABs
970
971 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
972 HandleMark hm; // Discard invalid handles created during verification
973 Universe::verify(" VerifyBeforeGC:");
974 }
975
976 // Verify object start arrays
977 if (VerifyObjectStartArray &&
978 VerifyBeforeGC) {
979 heap->old_gen()->verify_object_start_array();
980 }
981
982 DEBUG_ONLY(mark_bitmap()->verify_clear();)
983 DEBUG_ONLY(summary_data().verify_clear();)
984
985 // Have worker threads release resources the next time they run a task.
986 gc_task_manager()->release_all_resources();
987 }
988
989 void PSParallelCompact::post_compact()
990 {
991 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
992
993 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
994 // Clear the marking bitmap, summary data and split info.
995 clear_data_covering_space(SpaceId(id));
996 // Update top(). Must be done after clearing the bitmap and summary data.
997 _space_info[id].publish_new_top();
998 }
999
1000 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1001 MutableSpace* const from_space = _space_info[from_space_id].space();
1002 MutableSpace* const to_space = _space_info[to_space_id].space();
1003
1004 ParallelScavengeHeap* heap = gc_heap();
1005 bool eden_empty = eden_space->is_empty();
1006 if (!eden_empty) {
1007 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1008 heap->young_gen(), heap->old_gen());
1009 }
1010
1011 // Update heap occupancy information which is used as input to the soft ref
1012 // clearing policy at the next gc.
1013 Universe::update_heap_info_at_gc();
1014
1015 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1016 to_space->is_empty();
1017
1018 BarrierSet* bs = heap->barrier_set();
1019 if (bs->is_a(BarrierSet::ModRef)) {
1020 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1021 MemRegion old_mr = heap->old_gen()->reserved();
1022
1023 if (young_gen_empty) {
1024 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1025 } else {
1026 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1027 }
1028 }
1029
1030 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1031 ClassLoaderDataGraph::purge();
1032 MetaspaceAux::verify_metrics();
1033
1034 Threads::gc_epilogue();
1035 CodeCache::gc_epilogue();
1036 JvmtiExport::gc_epilogue();
1037
1038 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1039
1040 ref_processor()->enqueue_discovered_references(NULL);
1041
1042 if (ZapUnusedHeapArea) {
1043 heap->gen_mangle_unused_area();
1044 }
1045
1046 // Update time of last GC
1047 reset_millis_since_last_gc();
1048 }
1049
1050 HeapWord*
1051 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1052 bool maximum_compaction)
1053 {
1054 const size_t region_size = ParallelCompactData::RegionSize;
1055 const ParallelCompactData& sd = summary_data();
1056
1057 const MutableSpace* const space = _space_info[id].space();
1058 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1059 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1060 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1061
1062 // Skip full regions at the beginning of the space--they are necessarily part
1063 // of the dense prefix.
1064 size_t full_count = 0;
1065 const RegionData* cp;
1066 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1067 ++full_count;
1068 }
1069
1070 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1071 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1072 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1073 if (maximum_compaction || cp == end_cp || interval_ended) {
1074 _maximum_compaction_gc_num = total_invocations();
1075 return sd.region_to_addr(cp);
1076 }
1077
1078 HeapWord* const new_top = _space_info[id].new_top();
1079 const size_t space_live = pointer_delta(new_top, space->bottom());
1080 const size_t space_used = space->used_in_words();
1081 const size_t space_capacity = space->capacity_in_words();
1082
1083 const double cur_density = double(space_live) / space_capacity;
1084 const double deadwood_density =
1085 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1086 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1087
1088 if (TraceParallelOldGCDensePrefix) {
1089 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1090 cur_density, deadwood_density, deadwood_goal);
1091 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1092 "space_cap=" SIZE_FORMAT,
1093 space_live, space_used,
1094 space_capacity);
1095 }
1096
1097 // XXX - Use binary search?
1098 HeapWord* dense_prefix = sd.region_to_addr(cp);
1099 const RegionData* full_cp = cp;
1100 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1101 while (cp < end_cp) {
1102 HeapWord* region_destination = cp->destination();
1103 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1104 if (TraceParallelOldGCDensePrefix && Verbose) {
1105 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1106 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1107 sd.region(cp), region_destination,
1108 dense_prefix, cur_deadwood);
1109 }
1110
1111 if (cur_deadwood >= deadwood_goal) {
1112 // Found the region that has the correct amount of deadwood to the left.
1113 // This typically occurs after crossing a fairly sparse set of regions, so
1114 // iterate backwards over those sparse regions, looking for the region
1115 // that has the lowest density of live objects 'to the right.'
1116 size_t space_to_left = sd.region(cp) * region_size;
1117 size_t live_to_left = space_to_left - cur_deadwood;
1118 size_t space_to_right = space_capacity - space_to_left;
1119 size_t live_to_right = space_live - live_to_left;
1120 double density_to_right = double(live_to_right) / space_to_right;
1121 while (cp > full_cp) {
1122 --cp;
1123 const size_t prev_region_live_to_right = live_to_right -
1124 cp->data_size();
1125 const size_t prev_region_space_to_right = space_to_right + region_size;
1126 double prev_region_density_to_right =
1127 double(prev_region_live_to_right) / prev_region_space_to_right;
1128 if (density_to_right <= prev_region_density_to_right) {
1129 return dense_prefix;
1130 }
1131 if (TraceParallelOldGCDensePrefix && Verbose) {
1132 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1133 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1134 prev_region_density_to_right);
1135 }
1136 dense_prefix -= region_size;
1137 live_to_right = prev_region_live_to_right;
1138 space_to_right = prev_region_space_to_right;
1139 density_to_right = prev_region_density_to_right;
1140 }
1141 return dense_prefix;
1142 }
1143
1144 dense_prefix += region_size;
1145 ++cp;
1146 }
1147
1148 return dense_prefix;
1149 }
1150
1151 #ifndef PRODUCT
1152 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1153 const SpaceId id,
1154 const bool maximum_compaction,
1155 HeapWord* const addr)
1156 {
1157 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1158 RegionData* const cp = summary_data().region(region_idx);
1159 const MutableSpace* const space = _space_info[id].space();
1160 HeapWord* const new_top = _space_info[id].new_top();
1161
1162 const size_t space_live = pointer_delta(new_top, space->bottom());
1163 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1164 const size_t space_cap = space->capacity_in_words();
1165 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1166 const size_t live_to_right = new_top - cp->destination();
1167 const size_t dead_to_right = space->top() - addr - live_to_right;
1168
1169 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1170 "spl=" SIZE_FORMAT " "
1171 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1172 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1173 " ratio=%10.8f",
1174 algorithm, addr, region_idx,
1175 space_live,
1176 dead_to_left, dead_to_left_pct,
1177 dead_to_right, live_to_right,
1178 double(dead_to_right) / live_to_right);
1179 }
1180 #endif // #ifndef PRODUCT
1181
1182 // Return a fraction indicating how much of the generation can be treated as
1183 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1184 // based on the density of live objects in the generation to determine a limit,
1185 // which is then adjusted so the return value is min_percent when the density is
1186 // 1.
1187 //
1188 // The following table shows some return values for a different values of the
1189 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1190 // min_percent is 1.
1191 //
1192 // fraction allowed as dead wood
1193 // -----------------------------------------------------------------
1194 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1195 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1196 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1197 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1198 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1199 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1200 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1201 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1202 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1203 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1204 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1205 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1206 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1207 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1208 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1209 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1210 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1211 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1212 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1213 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1214 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1215 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1216 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1217
1218 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1219 {
1220 assert(_dwl_initialized, "uninitialized");
1221
1222 // The raw limit is the value of the normal distribution at x = density.
1223 const double raw_limit = normal_distribution(density);
1224
1225 // Adjust the raw limit so it becomes the minimum when the density is 1.
1226 //
1227 // First subtract the adjustment value (which is simply the precomputed value
1228 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1229 // Then add the minimum value, so the minimum is returned when the density is
1230 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1231 const double min = double(min_percent) / 100.0;
1232 const double limit = raw_limit - _dwl_adjustment + min;
1233 return MAX2(limit, 0.0);
1234 }
1235
1236 ParallelCompactData::RegionData*
1237 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1238 const RegionData* end)
1239 {
1240 const size_t region_size = ParallelCompactData::RegionSize;
1241 ParallelCompactData& sd = summary_data();
1242 size_t left = sd.region(beg);
1243 size_t right = end > beg ? sd.region(end) - 1 : left;
1244
1245 // Binary search.
1246 while (left < right) {
1247 // Equivalent to (left + right) / 2, but does not overflow.
1248 const size_t middle = left + (right - left) / 2;
1249 RegionData* const middle_ptr = sd.region(middle);
1250 HeapWord* const dest = middle_ptr->destination();
1251 HeapWord* const addr = sd.region_to_addr(middle);
1252 assert(dest != NULL, "sanity");
1253 assert(dest <= addr, "must move left");
1254
1255 if (middle > left && dest < addr) {
1256 right = middle - 1;
1257 } else if (middle < right && middle_ptr->data_size() == region_size) {
1258 left = middle + 1;
1259 } else {
1260 return middle_ptr;
1261 }
1262 }
1263 return sd.region(left);
1264 }
1265
1266 ParallelCompactData::RegionData*
1267 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1268 const RegionData* end,
1269 size_t dead_words)
1270 {
1271 ParallelCompactData& sd = summary_data();
1272 size_t left = sd.region(beg);
1273 size_t right = end > beg ? sd.region(end) - 1 : left;
1274
1275 // Binary search.
1276 while (left < right) {
1277 // Equivalent to (left + right) / 2, but does not overflow.
1278 const size_t middle = left + (right - left) / 2;
1279 RegionData* const middle_ptr = sd.region(middle);
1280 HeapWord* const dest = middle_ptr->destination();
1281 HeapWord* const addr = sd.region_to_addr(middle);
1282 assert(dest != NULL, "sanity");
1283 assert(dest <= addr, "must move left");
1284
1285 const size_t dead_to_left = pointer_delta(addr, dest);
1286 if (middle > left && dead_to_left > dead_words) {
1287 right = middle - 1;
1288 } else if (middle < right && dead_to_left < dead_words) {
1289 left = middle + 1;
1290 } else {
1291 return middle_ptr;
1292 }
1293 }
1294 return sd.region(left);
1295 }
1296
1297 // The result is valid during the summary phase, after the initial summarization
1298 // of each space into itself, and before final summarization.
1299 inline double
1300 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1301 HeapWord* const bottom,
1302 HeapWord* const top,
1303 HeapWord* const new_top)
1304 {
1305 ParallelCompactData& sd = summary_data();
1306
1307 assert(cp != NULL, "sanity");
1308 assert(bottom != NULL, "sanity");
1309 assert(top != NULL, "sanity");
1310 assert(new_top != NULL, "sanity");
1311 assert(top >= new_top, "summary data problem?");
1312 assert(new_top > bottom, "space is empty; should not be here");
1313 assert(new_top >= cp->destination(), "sanity");
1314 assert(top >= sd.region_to_addr(cp), "sanity");
1315
1316 HeapWord* const destination = cp->destination();
1317 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1318 const size_t compacted_region_live = pointer_delta(new_top, destination);
1319 const size_t compacted_region_used = pointer_delta(top,
1320 sd.region_to_addr(cp));
1321 const size_t reclaimable = compacted_region_used - compacted_region_live;
1322
1323 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1324 return double(reclaimable) / divisor;
1325 }
1326
1327 // Return the address of the end of the dense prefix, a.k.a. the start of the
1328 // compacted region. The address is always on a region boundary.
1329 //
1330 // Completely full regions at the left are skipped, since no compaction can
1331 // occur in those regions. Then the maximum amount of dead wood to allow is
1332 // computed, based on the density (amount live / capacity) of the generation;
1333 // the region with approximately that amount of dead space to the left is
1334 // identified as the limit region. Regions between the last completely full
1335 // region and the limit region are scanned and the one that has the best
1336 // (maximum) reclaimed_ratio() is selected.
1337 HeapWord*
1338 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1339 bool maximum_compaction)
1340 {
1341 if (ParallelOldGCSplitALot) {
1342 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1343 // The value was chosen to provoke splitting a young gen space; use it.
1344 return _space_info[id].dense_prefix();
1345 }
1346 }
1347
1348 const size_t region_size = ParallelCompactData::RegionSize;
1349 const ParallelCompactData& sd = summary_data();
1350
1351 const MutableSpace* const space = _space_info[id].space();
1352 HeapWord* const top = space->top();
1353 HeapWord* const top_aligned_up = sd.region_align_up(top);
1354 HeapWord* const new_top = _space_info[id].new_top();
1355 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1356 HeapWord* const bottom = space->bottom();
1357 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1358 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1359 const RegionData* const new_top_cp =
1360 sd.addr_to_region_ptr(new_top_aligned_up);
1361
1362 // Skip full regions at the beginning of the space--they are necessarily part
1363 // of the dense prefix.
1364 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1365 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1366 space->is_empty(), "no dead space allowed to the left");
1367 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1368 "region must have dead space");
1369
1370 // The gc number is saved whenever a maximum compaction is done, and used to
1371 // determine when the maximum compaction interval has expired. This avoids
1372 // successive max compactions for different reasons.
1373 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1374 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1375 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1376 total_invocations() == HeapFirstMaximumCompactionCount;
1377 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1378 _maximum_compaction_gc_num = total_invocations();
1379 return sd.region_to_addr(full_cp);
1380 }
1381
1382 const size_t space_live = pointer_delta(new_top, bottom);
1383 const size_t space_used = space->used_in_words();
1384 const size_t space_capacity = space->capacity_in_words();
1385
1386 const double density = double(space_live) / double(space_capacity);
1387 const size_t min_percent_free = MarkSweepDeadRatio;
1388 const double limiter = dead_wood_limiter(density, min_percent_free);
1389 const size_t dead_wood_max = space_used - space_live;
1390 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1391 dead_wood_max);
1392
1393 if (TraceParallelOldGCDensePrefix) {
1394 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1395 "space_cap=" SIZE_FORMAT,
1396 space_live, space_used,
1397 space_capacity);
1398 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1399 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1400 density, min_percent_free, limiter,
1401 dead_wood_max, dead_wood_limit);
1402 }
1403
1404 // Locate the region with the desired amount of dead space to the left.
1405 const RegionData* const limit_cp =
1406 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1407
1408 // Scan from the first region with dead space to the limit region and find the
1409 // one with the best (largest) reclaimed ratio.
1410 double best_ratio = 0.0;
1411 const RegionData* best_cp = full_cp;
1412 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1413 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1414 if (tmp_ratio > best_ratio) {
1415 best_cp = cp;
1416 best_ratio = tmp_ratio;
1417 }
1418 }
1419
1420 #if 0
1421 // Something to consider: if the region with the best ratio is 'close to' the
1422 // first region w/free space, choose the first region with free space
1423 // ("first-free"). The first-free region is usually near the start of the
1424 // heap, which means we are copying most of the heap already, so copy a bit
1425 // more to get complete compaction.
1426 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1427 _maximum_compaction_gc_num = total_invocations();
1428 best_cp = full_cp;
1429 }
1430 #endif // #if 0
1431
1432 return sd.region_to_addr(best_cp);
1433 }
1434
1435 #ifndef PRODUCT
1436 void
1437 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1438 size_t words)
1439 {
1440 if (TraceParallelOldGCSummaryPhase) {
1441 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1442 SIZE_FORMAT, start, start + words, words);
1443 }
1444
1445 ObjectStartArray* const start_array = _space_info[id].start_array();
1446 CollectedHeap::fill_with_objects(start, words);
1447 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1448 _mark_bitmap.mark_obj(p, words);
1449 _summary_data.add_obj(p, words);
1450 start_array->allocate_block(p);
1451 }
1452 }
1453
1454 void
1455 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1456 {
1457 ParallelCompactData& sd = summary_data();
1458 MutableSpace* space = _space_info[id].space();
1459
1460 // Find the source and destination start addresses.
1461 HeapWord* const src_addr = sd.region_align_down(start);
1462 HeapWord* dst_addr;
1463 if (src_addr < start) {
1464 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1465 } else if (src_addr > space->bottom()) {
1466 // The start (the original top() value) is aligned to a region boundary so
1467 // the associated region does not have a destination. Compute the
1468 // destination from the previous region.
1469 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1470 dst_addr = cp->destination() + cp->data_size();
1471 } else {
1472 // Filling the entire space.
1473 dst_addr = space->bottom();
1474 }
1475 assert(dst_addr != NULL, "sanity");
1476
1477 // Update the summary data.
1478 bool result = _summary_data.summarize(_space_info[id].split_info(),
1479 src_addr, space->top(), NULL,
1480 dst_addr, space->end(),
1481 _space_info[id].new_top_addr());
1482 assert(result, "should not fail: bad filler object size");
1483 }
1484
1485 void
1486 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1487 {
1488 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1489 return;
1490 }
1491
1492 MutableSpace* const space = _space_info[id].space();
1493 if (space->is_empty()) {
1494 HeapWord* b = space->bottom();
1495 HeapWord* t = b + space->capacity_in_words() / 2;
1496 space->set_top(t);
1497 if (ZapUnusedHeapArea) {
1498 space->set_top_for_allocations();
1499 }
1500
1501 size_t min_size = CollectedHeap::min_fill_size();
1502 size_t obj_len = min_size;
1503 while (b + obj_len <= t) {
1504 CollectedHeap::fill_with_object(b, obj_len);
1505 mark_bitmap()->mark_obj(b, obj_len);
1506 summary_data().add_obj(b, obj_len);
1507 b += obj_len;
1508 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1509 }
1510 if (b < t) {
1511 // The loop didn't completely fill to t (top); adjust top downward.
1512 space->set_top(b);
1513 if (ZapUnusedHeapArea) {
1514 space->set_top_for_allocations();
1515 }
1516 }
1517
1518 HeapWord** nta = _space_info[id].new_top_addr();
1519 bool result = summary_data().summarize(_space_info[id].split_info(),
1520 space->bottom(), space->top(), NULL,
1521 space->bottom(), space->end(), nta);
1522 assert(result, "space must fit into itself");
1523 }
1524 }
1525
1526 void
1527 PSParallelCompact::provoke_split(bool & max_compaction)
1528 {
1529 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1530 return;
1531 }
1532
1533 const size_t region_size = ParallelCompactData::RegionSize;
1534 ParallelCompactData& sd = summary_data();
1535
1536 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1537 MutableSpace* const from_space = _space_info[from_space_id].space();
1538 const size_t eden_live = pointer_delta(eden_space->top(),
1539 _space_info[eden_space_id].new_top());
1540 const size_t from_live = pointer_delta(from_space->top(),
1541 _space_info[from_space_id].new_top());
1542
1543 const size_t min_fill_size = CollectedHeap::min_fill_size();
1544 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1545 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1546 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1547 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1548
1549 // Choose the space to split; need at least 2 regions live (or fillable).
1550 SpaceId id;
1551 MutableSpace* space;
1552 size_t live_words;
1553 size_t fill_words;
1554 if (eden_live + eden_fillable >= region_size * 2) {
1555 id = eden_space_id;
1556 space = eden_space;
1557 live_words = eden_live;
1558 fill_words = eden_fillable;
1559 } else if (from_live + from_fillable >= region_size * 2) {
1560 id = from_space_id;
1561 space = from_space;
1562 live_words = from_live;
1563 fill_words = from_fillable;
1564 } else {
1565 return; // Give up.
1566 }
1567 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1568
1569 if (live_words < region_size * 2) {
1570 // Fill from top() to end() w/live objects of mixed sizes.
1571 HeapWord* const fill_start = space->top();
1572 live_words += fill_words;
1573
1574 space->set_top(fill_start + fill_words);
1575 if (ZapUnusedHeapArea) {
1576 space->set_top_for_allocations();
1577 }
1578
1579 HeapWord* cur_addr = fill_start;
1580 while (fill_words > 0) {
1581 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1582 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1583 if (fill_words - cur_size < min_fill_size) {
1584 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1585 }
1586
1587 CollectedHeap::fill_with_object(cur_addr, cur_size);
1588 mark_bitmap()->mark_obj(cur_addr, cur_size);
1589 sd.add_obj(cur_addr, cur_size);
1590
1591 cur_addr += cur_size;
1592 fill_words -= cur_size;
1593 }
1594
1595 summarize_new_objects(id, fill_start);
1596 }
1597
1598 max_compaction = false;
1599
1600 // Manipulate the old gen so that it has room for about half of the live data
1601 // in the target young gen space (live_words / 2).
1602 id = old_space_id;
1603 space = _space_info[id].space();
1604 const size_t free_at_end = space->free_in_words();
1605 const size_t free_target = align_object_size(live_words / 2);
1606 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1607
1608 if (free_at_end >= free_target + min_fill_size) {
1609 // Fill space above top() and set the dense prefix so everything survives.
1610 HeapWord* const fill_start = space->top();
1611 const size_t fill_size = free_at_end - free_target;
1612 space->set_top(space->top() + fill_size);
1613 if (ZapUnusedHeapArea) {
1614 space->set_top_for_allocations();
1615 }
1616 fill_with_live_objects(id, fill_start, fill_size);
1617 summarize_new_objects(id, fill_start);
1618 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1619 } else if (dead + free_at_end > free_target) {
1620 // Find a dense prefix that makes the right amount of space available.
1621 HeapWord* cur = sd.region_align_down(space->top());
1622 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1623 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1624 while (dead_to_right < free_target) {
1625 cur -= region_size;
1626 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1627 dead_to_right = pointer_delta(space->end(), cur_destination);
1628 }
1629 _space_info[id].set_dense_prefix(cur);
1630 }
1631 }
1632 #endif // #ifndef PRODUCT
1633
1634 void PSParallelCompact::summarize_spaces_quick()
1635 {
1636 for (unsigned int i = 0; i < last_space_id; ++i) {
1637 const MutableSpace* space = _space_info[i].space();
1638 HeapWord** nta = _space_info[i].new_top_addr();
1639 bool result = _summary_data.summarize(_space_info[i].split_info(),
1640 space->bottom(), space->top(), NULL,
1641 space->bottom(), space->end(), nta);
1642 assert(result, "space must fit into itself");
1643 _space_info[i].set_dense_prefix(space->bottom());
1644 }
1645
1646 #ifndef PRODUCT
1647 if (ParallelOldGCSplitALot) {
1648 provoke_split_fill_survivor(to_space_id);
1649 }
1650 #endif // #ifndef PRODUCT
1651 }
1652
1653 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1654 {
1655 HeapWord* const dense_prefix_end = dense_prefix(id);
1656 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1657 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1658 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1659 // Only enough dead space is filled so that any remaining dead space to the
1660 // left is larger than the minimum filler object. (The remainder is filled
1661 // during the copy/update phase.)
1662 //
1663 // The size of the dead space to the right of the boundary is not a
1664 // concern, since compaction will be able to use whatever space is
1665 // available.
1666 //
1667 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1668 // surrounds the space to be filled with an object.
1669 //
1670 // In the 32-bit VM, each bit represents two 32-bit words:
1671 // +---+
1672 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1673 // end_bits: ... x x x | 0 | || 0 x x ...
1674 // +---+
1675 //
1676 // In the 64-bit VM, each bit represents one 64-bit word:
1677 // +------------+
1678 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1679 // end_bits: ... x x 1 | 0 || 0 | x x ...
1680 // +------------+
1681 // +-------+
1682 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1683 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1684 // +-------+
1685 // +-----------+
1686 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1687 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1688 // +-----------+
1689 // +-------+
1690 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1691 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1692 // +-------+
1693
1694 // Initially assume case a, c or e will apply.
1695 size_t obj_len = CollectedHeap::min_fill_size();
1696 HeapWord* obj_beg = dense_prefix_end - obj_len;
1697
1698 #ifdef _LP64
1699 if (MinObjAlignment > 1) { // object alignment > heap word size
1700 // Cases a, c or e.
1701 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1702 // Case b above.
1703 obj_beg = dense_prefix_end - 1;
1704 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1705 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1706 // Case d above.
1707 obj_beg = dense_prefix_end - 3;
1708 obj_len = 3;
1709 }
1710 #endif // #ifdef _LP64
1711
1712 CollectedHeap::fill_with_object(obj_beg, obj_len);
1713 _mark_bitmap.mark_obj(obj_beg, obj_len);
1714 _summary_data.add_obj(obj_beg, obj_len);
1715 assert(start_array(id) != NULL, "sanity");
1716 start_array(id)->allocate_block(obj_beg);
1717 }
1718 }
1719
1720 void
1721 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1722 {
1723 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1724 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1725 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1726 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1727 cur->set_source_region(0);
1728 }
1729 }
1730
1731 void
1732 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1733 {
1734 assert(id < last_space_id, "id out of range");
1735 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1736 ParallelOldGCSplitALot && id == old_space_id,
1737 "should have been reset in summarize_spaces_quick()");
1738
1739 const MutableSpace* space = _space_info[id].space();
1740 if (_space_info[id].new_top() != space->bottom()) {
1741 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1742 _space_info[id].set_dense_prefix(dense_prefix_end);
1743
1744 #ifndef PRODUCT
1745 if (TraceParallelOldGCDensePrefix) {
1746 print_dense_prefix_stats("ratio", id, maximum_compaction,
1747 dense_prefix_end);
1748 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1749 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1750 }
1751 #endif // #ifndef PRODUCT
1752
1753 // Recompute the summary data, taking into account the dense prefix. If
1754 // every last byte will be reclaimed, then the existing summary data which
1755 // compacts everything can be left in place.
1756 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1757 // If dead space crosses the dense prefix boundary, it is (at least
1758 // partially) filled with a dummy object, marked live and added to the
1759 // summary data. This simplifies the copy/update phase and must be done
1760 // before the final locations of objects are determined, to prevent
1761 // leaving a fragment of dead space that is too small to fill.
1762 fill_dense_prefix_end(id);
1763
1764 // Compute the destination of each Region, and thus each object.
1765 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1766 _summary_data.summarize(_space_info[id].split_info(),
1767 dense_prefix_end, space->top(), NULL,
1768 dense_prefix_end, space->end(),
1769 _space_info[id].new_top_addr());
1770 }
1771 }
1772
1773 if (TraceParallelOldGCSummaryPhase) {
1774 const size_t region_size = ParallelCompactData::RegionSize;
1775 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1776 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1777 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1778 HeapWord* const new_top = _space_info[id].new_top();
1779 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1780 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1781 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1782 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1783 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1784 id, space->capacity_in_words(), dense_prefix_end,
1785 dp_region, dp_words / region_size,
1786 cr_words / region_size, new_top);
1787 }
1788 }
1789
1790 #ifndef PRODUCT
1791 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1792 HeapWord* dst_beg, HeapWord* dst_end,
1793 SpaceId src_space_id,
1794 HeapWord* src_beg, HeapWord* src_end)
1795 {
1796 if (TraceParallelOldGCSummaryPhase) {
1797 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1798 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1799 SIZE_FORMAT "-" SIZE_FORMAT " "
1800 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1801 SIZE_FORMAT "-" SIZE_FORMAT,
1802 src_space_id, space_names[src_space_id],
1803 dst_space_id, space_names[dst_space_id],
1804 src_beg, src_end,
1805 _summary_data.addr_to_region_idx(src_beg),
1806 _summary_data.addr_to_region_idx(src_end),
1807 dst_beg, dst_end,
1808 _summary_data.addr_to_region_idx(dst_beg),
1809 _summary_data.addr_to_region_idx(dst_end));
1810 }
1811 }
1812 #endif // #ifndef PRODUCT
1813
1814 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1815 bool maximum_compaction)
1816 {
1817 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1818 // trace("2");
1819
1820 #ifdef ASSERT
1821 if (TraceParallelOldGCMarkingPhase) {
1822 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1823 "add_obj_bytes=" SIZE_FORMAT,
1824 add_obj_count, add_obj_size * HeapWordSize);
1825 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1826 "mark_bitmap_bytes=" SIZE_FORMAT,
1827 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1828 }
1829 #endif // #ifdef ASSERT
1830
1831 // Quick summarization of each space into itself, to see how much is live.
1832 summarize_spaces_quick();
1833
1834 if (TraceParallelOldGCSummaryPhase) {
1835 tty->print_cr("summary_phase: after summarizing each space to self");
1836 Universe::print();
1837 NOT_PRODUCT(print_region_ranges());
1838 if (Verbose) {
1839 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1840 }
1841 }
1842
1843 // The amount of live data that will end up in old space (assuming it fits).
1844 size_t old_space_total_live = 0;
1845 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1846 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1847 _space_info[id].space()->bottom());
1848 }
1849
1850 MutableSpace* const old_space = _space_info[old_space_id].space();
1851 const size_t old_capacity = old_space->capacity_in_words();
1852 if (old_space_total_live > old_capacity) {
1853 // XXX - should also try to expand
1854 maximum_compaction = true;
1855 }
1856 #ifndef PRODUCT
1857 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1858 provoke_split(maximum_compaction);
1859 }
1860 #endif // #ifndef PRODUCT
1861
1862 // Old generations.
1863 summarize_space(old_space_id, maximum_compaction);
1864
1865 // Summarize the remaining spaces in the young gen. The initial target space
1866 // is the old gen. If a space does not fit entirely into the target, then the
1867 // remainder is compacted into the space itself and that space becomes the new
1868 // target.
1869 SpaceId dst_space_id = old_space_id;
1870 HeapWord* dst_space_end = old_space->end();
1871 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1872 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1873 const MutableSpace* space = _space_info[id].space();
1874 const size_t live = pointer_delta(_space_info[id].new_top(),
1875 space->bottom());
1876 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1877
1878 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1879 SpaceId(id), space->bottom(), space->top());)
1880 if (live > 0 && live <= available) {
1881 // All the live data will fit.
1882 bool done = _summary_data.summarize(_space_info[id].split_info(),
1883 space->bottom(), space->top(),
1884 NULL,
1885 *new_top_addr, dst_space_end,
1886 new_top_addr);
1887 assert(done, "space must fit into old gen");
1888
1889 // Reset the new_top value for the space.
1890 _space_info[id].set_new_top(space->bottom());
1891 } else if (live > 0) {
1892 // Attempt to fit part of the source space into the target space.
1893 HeapWord* next_src_addr = NULL;
1894 bool done = _summary_data.summarize(_space_info[id].split_info(),
1895 space->bottom(), space->top(),
1896 &next_src_addr,
1897 *new_top_addr, dst_space_end,
1898 new_top_addr);
1899 assert(!done, "space should not fit into old gen");
1900 assert(next_src_addr != NULL, "sanity");
1901
1902 // The source space becomes the new target, so the remainder is compacted
1903 // within the space itself.
1904 dst_space_id = SpaceId(id);
1905 dst_space_end = space->end();
1906 new_top_addr = _space_info[id].new_top_addr();
1907 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1908 space->bottom(), dst_space_end,
1909 SpaceId(id), next_src_addr, space->top());)
1910 done = _summary_data.summarize(_space_info[id].split_info(),
1911 next_src_addr, space->top(),
1912 NULL,
1913 space->bottom(), dst_space_end,
1914 new_top_addr);
1915 assert(done, "space must fit when compacted into itself");
1916 assert(*new_top_addr <= space->top(), "usage should not grow");
1917 }
1918 }
1919
1920 if (TraceParallelOldGCSummaryPhase) {
1921 tty->print_cr("summary_phase: after final summarization");
1922 Universe::print();
1923 NOT_PRODUCT(print_region_ranges());
1924 if (Verbose) {
1925 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1926 }
1927 }
1928 }
1929
1930 // This method should contain all heap-specific policy for invoking a full
1931 // collection. invoke_no_policy() will only attempt to compact the heap; it
1932 // will do nothing further. If we need to bail out for policy reasons, scavenge
1933 // before full gc, or any other specialized behavior, it needs to be added here.
1934 //
1935 // Note that this method should only be called from the vm_thread while at a
1936 // safepoint.
1937 //
1938 // Note that the all_soft_refs_clear flag in the collector policy
1939 // may be true because this method can be called without intervening
1940 // activity. For example when the heap space is tight and full measure
1941 // are being taken to free space.
1942 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1943 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1944 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1945 "should be in vm thread");
1946
1947 ParallelScavengeHeap* heap = gc_heap();
1948 GCCause::Cause gc_cause = heap->gc_cause();
1949 assert(!heap->is_gc_active(), "not reentrant");
1950
1951 PSAdaptiveSizePolicy* policy = heap->size_policy();
1952 IsGCActiveMark mark;
1953
1954 if (ScavengeBeforeFullGC) {
1955 PSScavenge::invoke_no_policy();
1956 }
1957
1958 const bool clear_all_soft_refs =
1959 heap->collector_policy()->should_clear_all_soft_refs();
1960
1961 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1962 maximum_heap_compaction);
1963 }
1964
1965 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1966 size_t addr_region_index = addr_to_region_idx(addr);
1967 return region_index == addr_region_index;
1968 }
1969
1970 // This method contains no policy. You should probably
1971 // be calling invoke() instead.
1972 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1973 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1974 assert(ref_processor() != NULL, "Sanity");
1975
1976 if (GC_locker::check_active_before_gc()) {
1977 return false;
1978 }
1979
1980 TimeStamp marking_start;
1981 TimeStamp compaction_start;
1982 TimeStamp collection_exit;
1983
1984 ParallelScavengeHeap* heap = gc_heap();
1985 GCCause::Cause gc_cause = heap->gc_cause();
1986 PSYoungGen* young_gen = heap->young_gen();
1987 PSOldGen* old_gen = heap->old_gen();
1988 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1989
1990 // The scope of casr should end after code that can change
1991 // CollectorPolicy::_should_clear_all_soft_refs.
1992 ClearedAllSoftRefs casr(maximum_heap_compaction,
1993 heap->collector_policy());
1994
1995 if (ZapUnusedHeapArea) {
1996 // Save information needed to minimize mangling
1997 heap->record_gen_tops_before_GC();
1998 }
1999
2000 heap->pre_full_gc_dump();
2001
2002 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2003
2004 // Make sure data structures are sane, make the heap parsable, and do other
2005 // miscellaneous bookkeeping.
2006 PreGCValues pre_gc_values;
2007 pre_compact(&pre_gc_values);
2008
2009 // Get the compaction manager reserved for the VM thread.
2010 ParCompactionManager* const vmthread_cm =
2011 ParCompactionManager::manager_array(gc_task_manager()->workers());
2012
2013 // Place after pre_compact() where the number of invocations is incremented.
2014 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2015
2016 {
2017 ResourceMark rm;
2018 HandleMark hm;
2019
2020 // Set the number of GC threads to be used in this collection
2021 gc_task_manager()->set_active_gang();
2022 gc_task_manager()->task_idle_workers();
2023 heap->set_par_threads(gc_task_manager()->active_workers());
2024
2025 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2026 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2027 TraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty);
2028 TraceCollectorStats tcs(counters());
2029 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2030
2031 if (TraceGen1Time) accumulated_time()->start();
2032
2033 // Let the size policy know we're starting
2034 size_policy->major_collection_begin();
2035
2036 CodeCache::gc_prologue();
2037 Threads::gc_prologue();
2038
2039 COMPILER2_PRESENT(DerivedPointerTable::clear());
2040
2041 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2042 ref_processor()->setup_policy(maximum_heap_compaction);
2043
2044 bool marked_for_unloading = false;
2045
2046 marking_start.update();
2047 marking_phase(vmthread_cm, maximum_heap_compaction);
2048
2049 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2050 && gc_cause == GCCause::_java_lang_system_gc;
2051 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2052
2053 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2054 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2055
2056 // adjust_roots() updates Universe::_intArrayKlassObj which is
2057 // needed by the compaction for filling holes in the dense prefix.
2058 adjust_roots();
2059
2060 compaction_start.update();
2061 compact();
2062
2063 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2064 // done before resizing.
2065 post_compact();
2066
2067 // Let the size policy know we're done
2068 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2069
2070 if (UseAdaptiveSizePolicy) {
2071 if (PrintAdaptiveSizePolicy) {
2072 gclog_or_tty->print("AdaptiveSizeStart: ");
2073 gclog_or_tty->stamp();
2074 gclog_or_tty->print_cr(" collection: %d ",
2075 heap->total_collections());
2076 if (Verbose) {
2077 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d",
2078 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2079 }
2080 }
2081
2082 // Don't check if the size_policy is ready here. Let
2083 // the size_policy check that internally.
2084 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2085 ((gc_cause != GCCause::_java_lang_system_gc) ||
2086 UseAdaptiveSizePolicyWithSystemGC)) {
2087 // Calculate optimal free space amounts
2088 assert(young_gen->max_size() >
2089 young_gen->from_space()->capacity_in_bytes() +
2090 young_gen->to_space()->capacity_in_bytes(),
2091 "Sizes of space in young gen are out-of-bounds");
2092 size_t max_eden_size = young_gen->max_size() -
2093 young_gen->from_space()->capacity_in_bytes() -
2094 young_gen->to_space()->capacity_in_bytes();
2095 size_policy->compute_generation_free_space(
2096 young_gen->used_in_bytes(),
2097 young_gen->eden_space()->used_in_bytes(),
2098 old_gen->used_in_bytes(),
2099 young_gen->eden_space()->capacity_in_bytes(),
2100 old_gen->max_gen_size(),
2101 max_eden_size,
2102 true /* full gc*/,
2103 gc_cause,
2104 heap->collector_policy());
2105
2106 heap->resize_old_gen(
2107 size_policy->calculated_old_free_size_in_bytes());
2108
2109 // Don't resize the young generation at an major collection. A
2110 // desired young generation size may have been calculated but
2111 // resizing the young generation complicates the code because the
2112 // resizing of the old generation may have moved the boundary
2113 // between the young generation and the old generation. Let the
2114 // young generation resizing happen at the minor collections.
2115 }
2116 if (PrintAdaptiveSizePolicy) {
2117 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2118 heap->total_collections());
2119 }
2120 }
2121
2122 if (UsePerfData) {
2123 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2124 counters->update_counters();
2125 counters->update_old_capacity(old_gen->capacity_in_bytes());
2126 counters->update_young_capacity(young_gen->capacity_in_bytes());
2127 }
2128
2129 heap->resize_all_tlabs();
2130
2131 // Resize the metaspace capactiy after a collection
2132 MetaspaceGC::compute_new_size();
2133
2134 if (TraceGen1Time) accumulated_time()->stop();
2135
2136 if (PrintGC) {
2137 if (PrintGCDetails) {
2138 // No GC timestamp here. This is after GC so it would be confusing.
2139 young_gen->print_used_change(pre_gc_values.young_gen_used());
2140 old_gen->print_used_change(pre_gc_values.old_gen_used());
2141 heap->print_heap_change(pre_gc_values.heap_used());
2142 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2143 } else {
2144 heap->print_heap_change(pre_gc_values.heap_used());
2145 }
2146 }
2147
2148 // Track memory usage and detect low memory
2149 MemoryService::track_memory_usage();
2150 heap->update_counters();
2151 gc_task_manager()->release_idle_workers();
2152 }
2153
2154 #ifdef ASSERT
2155 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2156 ParCompactionManager* const cm =
2157 ParCompactionManager::manager_array(int(i));
2158 assert(cm->marking_stack()->is_empty(), "should be empty");
2159 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2160 }
2161 #endif // ASSERT
2162
2163 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2164 HandleMark hm; // Discard invalid handles created during verification
2165 Universe::verify(" VerifyAfterGC:");
2166 }
2167
2168 // Re-verify object start arrays
2169 if (VerifyObjectStartArray &&
2170 VerifyAfterGC) {
2171 old_gen->verify_object_start_array();
2172 }
2173
2174 if (ZapUnusedHeapArea) {
2175 old_gen->object_space()->check_mangled_unused_area_complete();
2176 }
2177
2178 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2179
2180 collection_exit.update();
2181
2182 heap->print_heap_after_gc();
2183 if (PrintGCTaskTimeStamps) {
2184 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2185 INT64_FORMAT,
2186 marking_start.ticks(), compaction_start.ticks(),
2187 collection_exit.ticks());
2188 gc_task_manager()->print_task_time_stamps();
2189 }
2190
2191 heap->post_full_gc_dump();
2192
2193 #ifdef TRACESPINNING
2194 ParallelTaskTerminator::print_termination_counts();
2195 #endif
2196
2197 return true;
2198 }
2199
2200 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2201 PSYoungGen* young_gen,
2202 PSOldGen* old_gen) {
2203 MutableSpace* const eden_space = young_gen->eden_space();
2204 assert(!eden_space->is_empty(), "eden must be non-empty");
2205 assert(young_gen->virtual_space()->alignment() ==
2206 old_gen->virtual_space()->alignment(), "alignments do not match");
2207
2208 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2209 return false;
2210 }
2211
2212 // Both generations must be completely committed.
2213 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2214 return false;
2215 }
2216 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2217 return false;
2218 }
2219
2220 // Figure out how much to take from eden. Include the average amount promoted
2221 // in the total; otherwise the next young gen GC will simply bail out to a
2222 // full GC.
2223 const size_t alignment = old_gen->virtual_space()->alignment();
2224 const size_t eden_used = eden_space->used_in_bytes();
2225 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2226 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2227 const size_t eden_capacity = eden_space->capacity_in_bytes();
2228
2229 if (absorb_size >= eden_capacity) {
2230 return false; // Must leave some space in eden.
2231 }
2232
2233 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2234 if (new_young_size < young_gen->min_gen_size()) {
2235 return false; // Respect young gen minimum size.
2236 }
2237
2238 if (TraceAdaptiveGCBoundary && Verbose) {
2239 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2240 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2241 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2242 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2243 absorb_size / K,
2244 eden_capacity / K, (eden_capacity - absorb_size) / K,
2245 young_gen->from_space()->used_in_bytes() / K,
2246 young_gen->to_space()->used_in_bytes() / K,
2247 young_gen->capacity_in_bytes() / K, new_young_size / K);
2248 }
2249
2250 // Fill the unused part of the old gen.
2251 MutableSpace* const old_space = old_gen->object_space();
2252 HeapWord* const unused_start = old_space->top();
2253 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2254
2255 if (unused_words > 0) {
2256 if (unused_words < CollectedHeap::min_fill_size()) {
2257 return false; // If the old gen cannot be filled, must give up.
2258 }
2259 CollectedHeap::fill_with_objects(unused_start, unused_words);
2260 }
2261
2262 // Take the live data from eden and set both top and end in the old gen to
2263 // eden top. (Need to set end because reset_after_change() mangles the region
2264 // from end to virtual_space->high() in debug builds).
2265 HeapWord* const new_top = eden_space->top();
2266 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2267 absorb_size);
2268 young_gen->reset_after_change();
2269 old_space->set_top(new_top);
2270 old_space->set_end(new_top);
2271 old_gen->reset_after_change();
2272
2273 // Update the object start array for the filler object and the data from eden.
2274 ObjectStartArray* const start_array = old_gen->start_array();
2275 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2276 start_array->allocate_block(p);
2277 }
2278
2279 // Could update the promoted average here, but it is not typically updated at
2280 // full GCs and the value to use is unclear. Something like
2281 //
2282 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2283
2284 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2285 return true;
2286 }
2287
2288 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2289 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2290 "shouldn't return NULL");
2291 return ParallelScavengeHeap::gc_task_manager();
2292 }
2293
2294 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2295 bool maximum_heap_compaction) {
2296 // Recursively traverse all live objects and mark them
2297 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2298
2299 ParallelScavengeHeap* heap = gc_heap();
2300 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2301 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2302 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2303 ParallelTaskTerminator terminator(active_gc_threads, qset);
2304
2305 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2306 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2307
2308 // Need new claim bits before marking starts.
2309 ClassLoaderDataGraph::clear_claimed_marks();
2310
2311 {
2312 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2313 ParallelScavengeHeap::ParStrongRootsScope psrs;
2314
2315 GCTaskQueue* q = GCTaskQueue::create();
2316
2317 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2318 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2319 // We scan the thread roots in parallel
2320 Threads::create_thread_roots_marking_tasks(q);
2321 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2322 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2323 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2324 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2325 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2326 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2327
2328 if (active_gc_threads > 1) {
2329 for (uint j = 0; j < active_gc_threads; j++) {
2330 q->enqueue(new StealMarkingTask(&terminator));
2331 }
2332 }
2333
2334 gc_task_manager()->execute_and_wait(q);
2335 }
2336
2337 // Process reference objects found during marking
2338 {
2339 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2340 if (ref_processor()->processing_is_mt()) {
2341 RefProcTaskExecutor task_executor;
2342 ref_processor()->process_discovered_references(
2343 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2344 &task_executor);
2345 } else {
2346 ref_processor()->process_discovered_references(
2347 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2348 }
2349 }
2350
2351 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2352
2353 // This is the point where the entire marking should have completed.
2354 assert(cm->marking_stacks_empty(), "Marking should have completed");
2355
2356 // Follow system dictionary roots and unload classes.
2357 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2358
2359 // Unload nmethods.
2360 CodeCache::do_unloading(is_alive_closure(), purged_class);
2361
2362 // Prune dead klasses from subklass/sibling/implementor lists.
2363 Klass::clean_weak_klass_links(is_alive_closure());
2364
2365 // Delete entries for dead interned strings.
2366 StringTable::unlink(is_alive_closure());
2367
2368 // Clean up unreferenced symbols in symbol table.
2369 SymbolTable::unlink();
2370 }
2371
2372 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) {
2373 ClassLoaderData* cld = klass->class_loader_data();
2374 // The actual processing of the klass is done when we
2375 // traverse the list of Klasses in the class loader data.
2376 PSParallelCompact::follow_class_loader(cm, cld);
2377 }
2378
2379 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) {
2380 ClassLoaderData* cld = klass->class_loader_data();
2381 // The actual processing of the klass is done when we
2382 // traverse the list of Klasses in the class loader data.
2383 PSParallelCompact::adjust_class_loader(cm, cld);
2384 }
2385
2386 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2387 ClassLoaderData* cld) {
2388 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2389 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2390
2391 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2392 }
2393
2394 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm,
2395 ClassLoaderData* cld) {
2396 cld->oops_do(PSParallelCompact::adjust_pointer_closure(),
2397 PSParallelCompact::adjust_klass_closure(),
2398 true);
2399 }
2400
2401 // This should be moved to the shared markSweep code!
2402 class PSAlwaysTrueClosure: public BoolObjectClosure {
2403 public:
2404 void do_object(oop p) { ShouldNotReachHere(); }
2405 bool do_object_b(oop p) { return true; }
2406 };
2407 static PSAlwaysTrueClosure always_true;
2408
2409 void PSParallelCompact::adjust_roots() {
2410 // Adjust the pointers to reflect the new locations
2411 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2412
2413 // Need new claim bits when tracing through and adjusting pointers.
2414 ClassLoaderDataGraph::clear_claimed_marks();
2415
2416 // General strong roots.
2417 Universe::oops_do(adjust_pointer_closure());
2418 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles
2419 CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2420 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2421 ObjectSynchronizer::oops_do(adjust_pointer_closure());
2422 FlatProfiler::oops_do(adjust_pointer_closure());
2423 Management::oops_do(adjust_pointer_closure());
2424 JvmtiExport::oops_do(adjust_pointer_closure());
2425 // SO_AllClasses
2426 SystemDictionary::oops_do(adjust_pointer_closure());
2427 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2428
2429 // Now adjust pointers in remaining weak roots. (All of which should
2430 // have been cleared if they pointed to non-surviving objects.)
2431 // Global (weak) JNI handles
2432 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2433
2434 CodeCache::oops_do(adjust_pointer_closure());
2435 StringTable::oops_do(adjust_pointer_closure());
2436 ref_processor()->weak_oops_do(adjust_pointer_closure());
2437 // Roots were visited so references into the young gen in roots
2438 // may have been scanned. Process them also.
2439 // Should the reference processor have a span that excludes
2440 // young gen objects?
2441 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2442 }
2443
2444 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2445 uint parallel_gc_threads)
2446 {
2447 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2448
2449 // Find the threads that are active
2450 unsigned int which = 0;
2451
2452 const uint task_count = MAX2(parallel_gc_threads, 1U);
2453 for (uint j = 0; j < task_count; j++) {
2454 q->enqueue(new DrainStacksCompactionTask(j));
2455 ParCompactionManager::verify_region_list_empty(j);
2456 // Set the region stacks variables to "no" region stack values
2457 // so that they will be recognized and needing a region stack
2458 // in the stealing tasks if they do not get one by executing
2459 // a draining stack.
2460 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2461 cm->set_region_stack(NULL);
2462 cm->set_region_stack_index((uint)max_uintx);
2463 }
2464 ParCompactionManager::reset_recycled_stack_index();
2465
2466 // Find all regions that are available (can be filled immediately) and
2467 // distribute them to the thread stacks. The iteration is done in reverse
2468 // order (high to low) so the regions will be removed in ascending order.
2469
2470 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2471
2472 size_t fillable_regions = 0; // A count for diagnostic purposes.
2473 // A region index which corresponds to the tasks created above.
2474 // "which" must be 0 <= which < task_count
2475
2476 which = 0;
2477 // id + 1 is used to test termination so unsigned can
2478 // be used with an old_space_id == 0.
2479 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2480 SpaceInfo* const space_info = _space_info + id;
2481 MutableSpace* const space = space_info->space();
2482 HeapWord* const new_top = space_info->new_top();
2483
2484 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2485 const size_t end_region =
2486 sd.addr_to_region_idx(sd.region_align_up(new_top));
2487
2488 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2489 if (sd.region(cur)->claim_unsafe()) {
2490 ParCompactionManager::region_list_push(which, cur);
2491
2492 if (TraceParallelOldGCCompactionPhase && Verbose) {
2493 const size_t count_mod_8 = fillable_regions & 7;
2494 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2495 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2496 if (count_mod_8 == 7) gclog_or_tty->cr();
2497 }
2498
2499 NOT_PRODUCT(++fillable_regions;)
2500
2501 // Assign regions to tasks in round-robin fashion.
2502 if (++which == task_count) {
2503 assert(which <= parallel_gc_threads,
2504 "Inconsistent number of workers");
2505 which = 0;
2506 }
2507 }
2508 }
2509 }
2510
2511 if (TraceParallelOldGCCompactionPhase) {
2512 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2513 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2514 }
2515 }
2516
2517 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2518
2519 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2520 uint parallel_gc_threads) {
2521 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2522
2523 ParallelCompactData& sd = PSParallelCompact::summary_data();
2524
2525 // Iterate over all the spaces adding tasks for updating
2526 // regions in the dense prefix. Assume that 1 gc thread
2527 // will work on opening the gaps and the remaining gc threads
2528 // will work on the dense prefix.
2529 unsigned int space_id;
2530 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2531 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2532 const MutableSpace* const space = _space_info[space_id].space();
2533
2534 if (dense_prefix_end == space->bottom()) {
2535 // There is no dense prefix for this space.
2536 continue;
2537 }
2538
2539 // The dense prefix is before this region.
2540 size_t region_index_end_dense_prefix =
2541 sd.addr_to_region_idx(dense_prefix_end);
2542 RegionData* const dense_prefix_cp =
2543 sd.region(region_index_end_dense_prefix);
2544 assert(dense_prefix_end == space->end() ||
2545 dense_prefix_cp->available() ||
2546 dense_prefix_cp->claimed(),
2547 "The region after the dense prefix should always be ready to fill");
2548
2549 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2550
2551 // Is there dense prefix work?
2552 size_t total_dense_prefix_regions =
2553 region_index_end_dense_prefix - region_index_start;
2554 // How many regions of the dense prefix should be given to
2555 // each thread?
2556 if (total_dense_prefix_regions > 0) {
2557 uint tasks_for_dense_prefix = 1;
2558 if (total_dense_prefix_regions <=
2559 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2560 // Don't over partition. This assumes that
2561 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2562 // so there are not many regions to process.
2563 tasks_for_dense_prefix = parallel_gc_threads;
2564 } else {
2565 // Over partition
2566 tasks_for_dense_prefix = parallel_gc_threads *
2567 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2568 }
2569 size_t regions_per_thread = total_dense_prefix_regions /
2570 tasks_for_dense_prefix;
2571 // Give each thread at least 1 region.
2572 if (regions_per_thread == 0) {
2573 regions_per_thread = 1;
2574 }
2575
2576 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2577 if (region_index_start >= region_index_end_dense_prefix) {
2578 break;
2579 }
2580 // region_index_end is not processed
2581 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2582 region_index_end_dense_prefix);
2583 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2584 region_index_start,
2585 region_index_end));
2586 region_index_start = region_index_end;
2587 }
2588 }
2589 // This gets any part of the dense prefix that did not
2590 // fit evenly.
2591 if (region_index_start < region_index_end_dense_prefix) {
2592 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2593 region_index_start,
2594 region_index_end_dense_prefix));
2595 }
2596 }
2597 }
2598
2599 void PSParallelCompact::enqueue_region_stealing_tasks(
2600 GCTaskQueue* q,
2601 ParallelTaskTerminator* terminator_ptr,
2602 uint parallel_gc_threads) {
2603 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2604
2605 // Once a thread has drained it's stack, it should try to steal regions from
2606 // other threads.
2607 if (parallel_gc_threads > 1) {
2608 for (uint j = 0; j < parallel_gc_threads; j++) {
2609 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2610 }
2611 }
2612 }
2613
2614 void PSParallelCompact::compact() {
2615 // trace("5");
2616 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2617
2618 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2619 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2620 PSOldGen* old_gen = heap->old_gen();
2621 old_gen->start_array()->reset();
2622 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2623 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2624 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2625 ParallelTaskTerminator terminator(active_gc_threads, qset);
2626
2627 GCTaskQueue* q = GCTaskQueue::create();
2628 enqueue_region_draining_tasks(q, active_gc_threads);
2629 enqueue_dense_prefix_tasks(q, active_gc_threads);
2630 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2631
2632 {
2633 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2634
2635 gc_task_manager()->execute_and_wait(q);
2636
2637 #ifdef ASSERT
2638 // Verify that all regions have been processed before the deferred updates.
2639 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2640 verify_complete(SpaceId(id));
2641 }
2642 #endif
2643 }
2644
2645 {
2646 // Update the deferred objects, if any. Any compaction manager can be used.
2647 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2648 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2649 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2650 update_deferred_objects(cm, SpaceId(id));
2651 }
2652 }
2653 }
2654
2655 #ifdef ASSERT
2656 void PSParallelCompact::verify_complete(SpaceId space_id) {
2657 // All Regions between space bottom() to new_top() should be marked as filled
2658 // and all Regions between new_top() and top() should be available (i.e.,
2659 // should have been emptied).
2660 ParallelCompactData& sd = summary_data();
2661 SpaceInfo si = _space_info[space_id];
2662 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2663 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2664 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2665 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2666 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2667
2668 bool issued_a_warning = false;
2669
2670 size_t cur_region;
2671 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2672 const RegionData* const c = sd.region(cur_region);
2673 if (!c->completed()) {
2674 warning("region " SIZE_FORMAT " not filled: "
2675 "destination_count=" SIZE_FORMAT,
2676 cur_region, c->destination_count());
2677 issued_a_warning = true;
2678 }
2679 }
2680
2681 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2682 const RegionData* const c = sd.region(cur_region);
2683 if (!c->available()) {
2684 warning("region " SIZE_FORMAT " not empty: "
2685 "destination_count=" SIZE_FORMAT,
2686 cur_region, c->destination_count());
2687 issued_a_warning = true;
2688 }
2689 }
2690
2691 if (issued_a_warning) {
2692 print_region_ranges();
2693 }
2694 }
2695 #endif // #ifdef ASSERT
2696
2697 // Update interior oops in the ranges of regions [beg_region, end_region).
2698 void
2699 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2700 SpaceId space_id,
2701 size_t beg_region,
2702 size_t end_region) {
2703 ParallelCompactData& sd = summary_data();
2704 ParMarkBitMap* const mbm = mark_bitmap();
2705
2706 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2707 HeapWord* const end_addr = sd.region_to_addr(end_region);
2708 assert(beg_region <= end_region, "bad region range");
2709 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2710
2711 #ifdef ASSERT
2712 // Claim the regions to avoid triggering an assert when they are marked as
2713 // filled.
2714 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2715 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2716 }
2717 #endif // #ifdef ASSERT
2718
2719 if (beg_addr != space(space_id)->bottom()) {
2720 // Find the first live object or block of dead space that *starts* in this
2721 // range of regions. If a partial object crosses onto the region, skip it;
2722 // it will be marked for 'deferred update' when the object head is
2723 // processed. If dead space crosses onto the region, it is also skipped; it
2724 // will be filled when the prior region is processed. If neither of those
2725 // apply, the first word in the region is the start of a live object or dead
2726 // space.
2727 assert(beg_addr > space(space_id)->bottom(), "sanity");
2728 const RegionData* const cp = sd.region(beg_region);
2729 if (cp->partial_obj_size() != 0) {
2730 beg_addr = sd.partial_obj_end(beg_region);
2731 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2732 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2733 }
2734 }
2735
2736 if (beg_addr < end_addr) {
2737 // A live object or block of dead space starts in this range of Regions.
2738 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2739
2740 // Create closures and iterate.
2741 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2742 FillClosure fill_closure(cm, space_id);
2743 ParMarkBitMap::IterationStatus status;
2744 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2745 dense_prefix_end);
2746 if (status == ParMarkBitMap::incomplete) {
2747 update_closure.do_addr(update_closure.source());
2748 }
2749 }
2750
2751 // Mark the regions as filled.
2752 RegionData* const beg_cp = sd.region(beg_region);
2753 RegionData* const end_cp = sd.region(end_region);
2754 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2755 cp->set_completed();
2756 }
2757 }
2758
2759 // Return the SpaceId for the space containing addr. If addr is not in the
2760 // heap, last_space_id is returned. In debug mode it expects the address to be
2761 // in the heap and asserts such.
2762 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2763 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2764
2765 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2766 if (_space_info[id].space()->contains(addr)) {
2767 return SpaceId(id);
2768 }
2769 }
2770
2771 assert(false, "no space contains the addr");
2772 return last_space_id;
2773 }
2774
2775 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2776 SpaceId id) {
2777 assert(id < last_space_id, "bad space id");
2778
2779 ParallelCompactData& sd = summary_data();
2780 const SpaceInfo* const space_info = _space_info + id;
2781 ObjectStartArray* const start_array = space_info->start_array();
2782
2783 const MutableSpace* const space = space_info->space();
2784 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2785 HeapWord* const beg_addr = space_info->dense_prefix();
2786 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2787
2788 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2789 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2790 const RegionData* cur_region;
2791 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2792 HeapWord* const addr = cur_region->deferred_obj_addr();
2793 if (addr != NULL) {
2794 if (start_array != NULL) {
2795 start_array->allocate_block(addr);
2796 }
2797 oop(addr)->update_contents(cm);
2798 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2799 }
2800 }
2801 }
2802
2803 // Skip over count live words starting from beg, and return the address of the
2804 // next live word. Unless marked, the word corresponding to beg is assumed to
2805 // be dead. Callers must either ensure beg does not correspond to the middle of
2806 // an object, or account for those live words in some other way. Callers must
2807 // also ensure that there are enough live words in the range [beg, end) to skip.
2808 HeapWord*
2809 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2810 {
2811 assert(count > 0, "sanity");
2812
2813 ParMarkBitMap* m = mark_bitmap();
2814 idx_t bits_to_skip = m->words_to_bits(count);
2815 idx_t cur_beg = m->addr_to_bit(beg);
2816 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2817
2818 do {
2819 cur_beg = m->find_obj_beg(cur_beg, search_end);
2820 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2821 const size_t obj_bits = cur_end - cur_beg + 1;
2822 if (obj_bits > bits_to_skip) {
2823 return m->bit_to_addr(cur_beg + bits_to_skip);
2824 }
2825 bits_to_skip -= obj_bits;
2826 cur_beg = cur_end + 1;
2827 } while (bits_to_skip > 0);
2828
2829 // Skipping the desired number of words landed just past the end of an object.
2830 // Find the start of the next object.
2831 cur_beg = m->find_obj_beg(cur_beg, search_end);
2832 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2833 return m->bit_to_addr(cur_beg);
2834 }
2835
2836 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2837 SpaceId src_space_id,
2838 size_t src_region_idx)
2839 {
2840 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2841
2842 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2843 if (split_info.dest_region_addr() == dest_addr) {
2844 // The partial object ending at the split point contains the first word to
2845 // be copied to dest_addr.
2846 return split_info.first_src_addr();
2847 }
2848
2849 const ParallelCompactData& sd = summary_data();
2850 ParMarkBitMap* const bitmap = mark_bitmap();
2851 const size_t RegionSize = ParallelCompactData::RegionSize;
2852
2853 assert(sd.is_region_aligned(dest_addr), "not aligned");
2854 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2855 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2856 HeapWord* const src_region_destination = src_region_ptr->destination();
2857
2858 assert(dest_addr >= src_region_destination, "wrong src region");
2859 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2860
2861 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2862 HeapWord* const src_region_end = src_region_beg + RegionSize;
2863
2864 HeapWord* addr = src_region_beg;
2865 if (dest_addr == src_region_destination) {
2866 // Return the first live word in the source region.
2867 if (partial_obj_size == 0) {
2868 addr = bitmap->find_obj_beg(addr, src_region_end);
2869 assert(addr < src_region_end, "no objects start in src region");
2870 }
2871 return addr;
2872 }
2873
2874 // Must skip some live data.
2875 size_t words_to_skip = dest_addr - src_region_destination;
2876 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2877
2878 if (partial_obj_size >= words_to_skip) {
2879 // All the live words to skip are part of the partial object.
2880 addr += words_to_skip;
2881 if (partial_obj_size == words_to_skip) {
2882 // Find the first live word past the partial object.
2883 addr = bitmap->find_obj_beg(addr, src_region_end);
2884 assert(addr < src_region_end, "wrong src region");
2885 }
2886 return addr;
2887 }
2888
2889 // Skip over the partial object (if any).
2890 if (partial_obj_size != 0) {
2891 words_to_skip -= partial_obj_size;
2892 addr += partial_obj_size;
2893 }
2894
2895 // Skip over live words due to objects that start in the region.
2896 addr = skip_live_words(addr, src_region_end, words_to_skip);
2897 assert(addr < src_region_end, "wrong src region");
2898 return addr;
2899 }
2900
2901 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2902 SpaceId src_space_id,
2903 size_t beg_region,
2904 HeapWord* end_addr)
2905 {
2906 ParallelCompactData& sd = summary_data();
2907
2908 #ifdef ASSERT
2909 MutableSpace* const src_space = _space_info[src_space_id].space();
2910 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2911 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2912 "src_space_id does not match beg_addr");
2913 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2914 "src_space_id does not match end_addr");
2915 #endif // #ifdef ASSERT
2916
2917 RegionData* const beg = sd.region(beg_region);
2918 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2919
2920 // Regions up to new_top() are enqueued if they become available.
2921 HeapWord* const new_top = _space_info[src_space_id].new_top();
2922 RegionData* const enqueue_end =
2923 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2924
2925 for (RegionData* cur = beg; cur < end; ++cur) {
2926 assert(cur->data_size() > 0, "region must have live data");
2927 cur->decrement_destination_count();
2928 if (cur < enqueue_end && cur->available() && cur->claim()) {
2929 cm->push_region(sd.region(cur));
2930 }
2931 }
2932 }
2933
2934 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2935 SpaceId& src_space_id,
2936 HeapWord*& src_space_top,
2937 HeapWord* end_addr)
2938 {
2939 typedef ParallelCompactData::RegionData RegionData;
2940
2941 ParallelCompactData& sd = PSParallelCompact::summary_data();
2942 const size_t region_size = ParallelCompactData::RegionSize;
2943
2944 size_t src_region_idx = 0;
2945
2946 // Skip empty regions (if any) up to the top of the space.
2947 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2948 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2949 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2950 const RegionData* const top_region_ptr =
2951 sd.addr_to_region_ptr(top_aligned_up);
2952 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2953 ++src_region_ptr;
2954 }
2955
2956 if (src_region_ptr < top_region_ptr) {
2957 // The next source region is in the current space. Update src_region_idx
2958 // and the source address to match src_region_ptr.
2959 src_region_idx = sd.region(src_region_ptr);
2960 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2961 if (src_region_addr > closure.source()) {
2962 closure.set_source(src_region_addr);
2963 }
2964 return src_region_idx;
2965 }
2966
2967 // Switch to a new source space and find the first non-empty region.
2968 unsigned int space_id = src_space_id + 1;
2969 assert(space_id < last_space_id, "not enough spaces");
2970
2971 HeapWord* const destination = closure.destination();
2972
2973 do {
2974 MutableSpace* space = _space_info[space_id].space();
2975 HeapWord* const bottom = space->bottom();
2976 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2977
2978 // Iterate over the spaces that do not compact into themselves.
2979 if (bottom_cp->destination() != bottom) {
2980 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2981 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2982
2983 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2984 if (src_cp->live_obj_size() > 0) {
2985 // Found it.
2986 assert(src_cp->destination() == destination,
2987 "first live obj in the space must match the destination");
2988 assert(src_cp->partial_obj_size() == 0,
2989 "a space cannot begin with a partial obj");
2990
2991 src_space_id = SpaceId(space_id);
2992 src_space_top = space->top();
2993 const size_t src_region_idx = sd.region(src_cp);
2994 closure.set_source(sd.region_to_addr(src_region_idx));
2995 return src_region_idx;
2996 } else {
2997 assert(src_cp->data_size() == 0, "sanity");
2998 }
2999 }
3000 }
3001 } while (++space_id < last_space_id);
3002
3003 assert(false, "no source region was found");
3004 return 0;
3005 }
3006
3007 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3008 {
3009 typedef ParMarkBitMap::IterationStatus IterationStatus;
3010 const size_t RegionSize = ParallelCompactData::RegionSize;
3011 ParMarkBitMap* const bitmap = mark_bitmap();
3012 ParallelCompactData& sd = summary_data();
3013 RegionData* const region_ptr = sd.region(region_idx);
3014
3015 // Get the items needed to construct the closure.
3016 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3017 SpaceId dest_space_id = space_id(dest_addr);
3018 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3019 HeapWord* new_top = _space_info[dest_space_id].new_top();
3020 assert(dest_addr < new_top, "sanity");
3021 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3022
3023 // Get the source region and related info.
3024 size_t src_region_idx = region_ptr->source_region();
3025 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3026 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3027
3028 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3029 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3030
3031 // Adjust src_region_idx to prepare for decrementing destination counts (the
3032 // destination count is not decremented when a region is copied to itself).
3033 if (src_region_idx == region_idx) {
3034 src_region_idx += 1;
3035 }
3036
3037 if (bitmap->is_unmarked(closure.source())) {
3038 // The first source word is in the middle of an object; copy the remainder
3039 // of the object or as much as will fit. The fact that pointer updates were
3040 // deferred will be noted when the object header is processed.
3041 HeapWord* const old_src_addr = closure.source();
3042 closure.copy_partial_obj();
3043 if (closure.is_full()) {
3044 decrement_destination_counts(cm, src_space_id, src_region_idx,
3045 closure.source());
3046 region_ptr->set_deferred_obj_addr(NULL);
3047 region_ptr->set_completed();
3048 return;
3049 }
3050
3051 HeapWord* const end_addr = sd.region_align_down(closure.source());
3052 if (sd.region_align_down(old_src_addr) != end_addr) {
3053 // The partial object was copied from more than one source region.
3054 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3055
3056 // Move to the next source region, possibly switching spaces as well. All
3057 // args except end_addr may be modified.
3058 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3059 end_addr);
3060 }
3061 }
3062
3063 do {
3064 HeapWord* const cur_addr = closure.source();
3065 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3066 src_space_top);
3067 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3068
3069 if (status == ParMarkBitMap::incomplete) {
3070 // The last obj that starts in the source region does not end in the
3071 // region.
3072 assert(closure.source() < end_addr, "sanity");
3073 HeapWord* const obj_beg = closure.source();
3074 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3075 src_space_top);
3076 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3077 if (obj_end < range_end) {
3078 // The end was found; the entire object will fit.
3079 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3080 assert(status != ParMarkBitMap::would_overflow, "sanity");
3081 } else {
3082 // The end was not found; the object will not fit.
3083 assert(range_end < src_space_top, "obj cannot cross space boundary");
3084 status = ParMarkBitMap::would_overflow;
3085 }
3086 }
3087
3088 if (status == ParMarkBitMap::would_overflow) {
3089 // The last object did not fit. Note that interior oop updates were
3090 // deferred, then copy enough of the object to fill the region.
3091 region_ptr->set_deferred_obj_addr(closure.destination());
3092 status = closure.copy_until_full(); // copies from closure.source()
3093
3094 decrement_destination_counts(cm, src_space_id, src_region_idx,
3095 closure.source());
3096 region_ptr->set_completed();
3097 return;
3098 }
3099
3100 if (status == ParMarkBitMap::full) {
3101 decrement_destination_counts(cm, src_space_id, src_region_idx,
3102 closure.source());
3103 region_ptr->set_deferred_obj_addr(NULL);
3104 region_ptr->set_completed();
3105 return;
3106 }
3107
3108 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3109
3110 // Move to the next source region, possibly switching spaces as well. All
3111 // args except end_addr may be modified.
3112 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3113 end_addr);
3114 } while (true);
3115 }
3116
3117 void
3118 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3119 const MutableSpace* sp = space(space_id);
3120 if (sp->is_empty()) {
3121 return;
3122 }
3123
3124 ParallelCompactData& sd = PSParallelCompact::summary_data();
3125 ParMarkBitMap* const bitmap = mark_bitmap();
3126 HeapWord* const dp_addr = dense_prefix(space_id);
3127 HeapWord* beg_addr = sp->bottom();
3128 HeapWord* end_addr = sp->top();
3129
3130 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3131
3132 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3133 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3134 if (beg_region < dp_region) {
3135 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3136 }
3137
3138 // The destination of the first live object that starts in the region is one
3139 // past the end of the partial object entering the region (if any).
3140 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3141 HeapWord* const new_top = _space_info[space_id].new_top();
3142 assert(new_top >= dest_addr, "bad new_top value");
3143 const size_t words = pointer_delta(new_top, dest_addr);
3144
3145 if (words > 0) {
3146 ObjectStartArray* start_array = _space_info[space_id].start_array();
3147 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3148
3149 ParMarkBitMap::IterationStatus status;
3150 status = bitmap->iterate(&closure, dest_addr, end_addr);
3151 assert(status == ParMarkBitMap::full, "iteration not complete");
3152 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3153 "live objects skipped because closure is full");
3154 }
3155 }
3156
3157 jlong PSParallelCompact::millis_since_last_gc() {
3158 // We need a monotonically non-deccreasing time in ms but
3159 // os::javaTimeMillis() does not guarantee monotonicity.
3160 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3161 jlong ret_val = now - _time_of_last_gc;
3162 // XXX See note in genCollectedHeap::millis_since_last_gc().
3163 if (ret_val < 0) {
3164 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3165 return 0;
3166 }
3167 return ret_val;
3168 }
3169
3170 void PSParallelCompact::reset_millis_since_last_gc() {
3171 // We need a monotonically non-deccreasing time in ms but
3172 // os::javaTimeMillis() does not guarantee monotonicity.
3173 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3174 }
3175
3176 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3177 {
3178 if (source() != destination()) {
3179 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3180 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3181 }
3182 update_state(words_remaining());
3183 assert(is_full(), "sanity");
3184 return ParMarkBitMap::full;
3185 }
3186
3187 void MoveAndUpdateClosure::copy_partial_obj()
3188 {
3189 size_t words = words_remaining();
3190
3191 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3192 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3193 if (end_addr < range_end) {
3194 words = bitmap()->obj_size(source(), end_addr);
3195 }
3196
3197 // This test is necessary; if omitted, the pointer updates to a partial object
3198 // that crosses the dense prefix boundary could be overwritten.
3199 if (source() != destination()) {
3200 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3201 Copy::aligned_conjoint_words(source(), destination(), words);
3202 }
3203 update_state(words);
3204 }
3205
3206 ParMarkBitMapClosure::IterationStatus
3207 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3208 assert(destination() != NULL, "sanity");
3209 assert(bitmap()->obj_size(addr) == words, "bad size");
3210
3211 _source = addr;
3212 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3213 destination(), "wrong destination");
3214
3215 if (words > words_remaining()) {
3216 return ParMarkBitMap::would_overflow;
3217 }
3218
3219 // The start_array must be updated even if the object is not moving.
3220 if (_start_array != NULL) {
3221 _start_array->allocate_block(destination());
3222 }
3223
3224 if (destination() != source()) {
3225 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3226 Copy::aligned_conjoint_words(source(), destination(), words);
3227 }
3228
3229 oop moved_oop = (oop) destination();
3230 moved_oop->update_contents(compaction_manager());
3231 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3232
3233 update_state(words);
3234 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3235 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3236 }
3237
3238 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3239 ParCompactionManager* cm,
3240 PSParallelCompact::SpaceId space_id) :
3241 ParMarkBitMapClosure(mbm, cm),
3242 _space_id(space_id),
3243 _start_array(PSParallelCompact::start_array(space_id))
3244 {
3245 }
3246
3247 // Updates the references in the object to their new values.
3248 ParMarkBitMapClosure::IterationStatus
3249 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3250 do_addr(addr);
3251 return ParMarkBitMap::incomplete;
3252 }