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 _region_data = 0;
360 _region_count = 0;
361 }
362
363 bool ParallelCompactData::initialize(MemRegion covered_region)
364 {
365 _region_start = covered_region.start();
366 const size_t region_size = covered_region.word_size();
367 DEBUG_ONLY(_region_end = _region_start + region_size;)
368
369 assert(region_align_down(_region_start) == _region_start,
370 "region start not aligned");
371 assert((region_size & RegionSizeOffsetMask) == 0,
372 "region size not a multiple of RegionSize");
373
374 bool result = initialize_region_data(region_size);
375
376 return result;
377 }
378
379 PSVirtualSpace*
380 ParallelCompactData::create_vspace(size_t count, size_t element_size)
381 {
382 const size_t raw_bytes = count * element_size;
383 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
384 const size_t granularity = os::vm_allocation_granularity();
385 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
386
387 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
388 MAX2(page_sz, granularity);
389 ReservedSpace rs(bytes, rs_align, rs_align > 0);
390 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
391 rs.size());
392
393 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
394
395 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
396 if (vspace != 0) {
397 if (vspace->expand_by(bytes)) {
398 return vspace;
399 }
400 delete vspace;
401 // Release memory reserved in the space.
402 rs.release();
403 }
404
405 return 0;
406 }
407
408 bool ParallelCompactData::initialize_region_data(size_t region_size)
409 {
410 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
411 _region_vspace = create_vspace(count, sizeof(RegionData));
412 if (_region_vspace != 0) {
413 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
414 _region_count = count;
415 return true;
416 }
417 return false;
418 }
419
420 void ParallelCompactData::clear()
421 {
422 memset(_region_data, 0, _region_vspace->committed_size());
423 }
424
425 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
426 assert(beg_region <= _region_count, "beg_region out of range");
427 assert(end_region <= _region_count, "end_region out of range");
428
429 const size_t region_cnt = end_region - beg_region;
430 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
431 }
432
433 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
434 {
435 const RegionData* cur_cp = region(region_idx);
436 const RegionData* const end_cp = region(region_count() - 1);
437
438 HeapWord* result = region_to_addr(region_idx);
439 if (cur_cp < end_cp) {
440 do {
441 result += cur_cp->partial_obj_size();
442 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
443 }
444 return result;
445 }
446
447 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
448 {
449 const size_t obj_ofs = pointer_delta(addr, _region_start);
450 const size_t beg_region = obj_ofs >> Log2RegionSize;
451 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
452
453 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
454 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
455
456 if (beg_region == end_region) {
457 // All in one region.
458 _region_data[beg_region].add_live_obj(len);
459 return;
460 }
461
462 // First region.
463 const size_t beg_ofs = region_offset(addr);
464 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
465
466 Klass* klass = ((oop)addr)->klass();
467 // Middle regions--completely spanned by this object.
468 for (size_t region = beg_region + 1; region < end_region; ++region) {
469 _region_data[region].set_partial_obj_size(RegionSize);
470 _region_data[region].set_partial_obj_addr(addr);
471 }
472
473 // Last region.
474 const size_t end_ofs = region_offset(addr + len - 1);
475 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
476 _region_data[end_region].set_partial_obj_addr(addr);
477 }
478
479 void
480 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
481 {
482 assert(region_offset(beg) == 0, "not RegionSize aligned");
483 assert(region_offset(end) == 0, "not RegionSize aligned");
484
485 size_t cur_region = addr_to_region_idx(beg);
486 const size_t end_region = addr_to_region_idx(end);
487 HeapWord* addr = beg;
488 while (cur_region < end_region) {
489 _region_data[cur_region].set_destination(addr);
490 _region_data[cur_region].set_destination_count(0);
491 _region_data[cur_region].set_source_region(cur_region);
492 _region_data[cur_region].set_data_location(addr);
493
494 // Update live_obj_size so the region appears completely full.
495 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
496 _region_data[cur_region].set_live_obj_size(live_size);
497
498 ++cur_region;
499 addr += RegionSize;
500 }
501 }
502
503 // Find the point at which a space can be split and, if necessary, record the
504 // split point.
505 //
506 // If the current src region (which overflowed the destination space) doesn't
507 // have a partial object, the split point is at the beginning of the current src
508 // region (an "easy" split, no extra bookkeeping required).
509 //
510 // If the current src region has a partial object, the split point is in the
511 // region where that partial object starts (call it the split_region). If
512 // split_region has a partial object, then the split point is just after that
513 // partial object (a "hard" split where we have to record the split data and
514 // zero the partial_obj_size field). With a "hard" split, we know that the
515 // partial_obj ends within split_region because the partial object that caused
516 // the overflow starts in split_region. If split_region doesn't have a partial
517 // obj, then the split is at the beginning of split_region (another "easy"
518 // split).
519 HeapWord*
520 ParallelCompactData::summarize_split_space(size_t src_region,
521 SplitInfo& split_info,
522 HeapWord* destination,
523 HeapWord* target_end,
524 HeapWord** target_next)
525 {
526 assert(destination <= target_end, "sanity");
527 assert(destination + _region_data[src_region].data_size() > target_end,
528 "region should not fit into target space");
529 assert(is_region_aligned(target_end), "sanity");
530
531 size_t split_region = src_region;
532 HeapWord* split_destination = destination;
533 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
534
535 if (destination + partial_obj_size > target_end) {
536 // The split point is just after the partial object (if any) in the
537 // src_region that contains the start of the object that overflowed the
538 // destination space.
539 //
540 // Find the start of the "overflow" object and set split_region to the
541 // region containing it.
542 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
543 split_region = addr_to_region_idx(overflow_obj);
544
545 // Clear the source_region field of all destination regions whose first word
546 // came from data after the split point (a non-null source_region field
547 // implies a region must be filled).
548 //
549 // An alternative to the simple loop below: clear during post_compact(),
550 // which uses memcpy instead of individual stores, and is easy to
551 // parallelize. (The downside is that it clears the entire RegionData
552 // object as opposed to just one field.)
553 //
554 // post_compact() would have to clear the summary data up to the highest
555 // address that was written during the summary phase, which would be
556 //
557 // max(top, max(new_top, clear_top))
558 //
559 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
560 // to target_end.
561 const RegionData* const sr = region(split_region);
562 const size_t beg_idx =
563 addr_to_region_idx(region_align_up(sr->destination() +
564 sr->partial_obj_size()));
565 const size_t end_idx = addr_to_region_idx(target_end);
566
567 if (TraceParallelOldGCSummaryPhase) {
568 gclog_or_tty->print_cr("split: clearing source_region field in ["
569 SIZE_FORMAT ", " SIZE_FORMAT ")",
570 beg_idx, end_idx);
571 }
572 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
573 _region_data[idx].set_source_region(0);
574 }
575
576 // Set split_destination and partial_obj_size to reflect the split region.
577 split_destination = sr->destination();
578 partial_obj_size = sr->partial_obj_size();
579 }
580
581 // The split is recorded only if a partial object extends onto the region.
582 if (partial_obj_size != 0) {
583 _region_data[split_region].set_partial_obj_size(0);
584 split_info.record(split_region, partial_obj_size, split_destination);
585 }
586
587 // Setup the continuation addresses.
588 *target_next = split_destination + partial_obj_size;
589 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
590
591 if (TraceParallelOldGCSummaryPhase) {
592 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
593 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
594 " pos=" SIZE_FORMAT,
595 split_type, source_next, split_region,
596 partial_obj_size);
597 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
598 " tn=" PTR_FORMAT,
599 split_type, split_destination,
600 addr_to_region_idx(split_destination),
601 *target_next);
602
603 if (partial_obj_size != 0) {
604 HeapWord* const po_beg = split_info.destination();
605 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
606 gclog_or_tty->print_cr("%s split: "
607 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
608 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
609 split_type,
610 po_beg, addr_to_region_idx(po_beg),
611 po_end, addr_to_region_idx(po_end));
612 }
613 }
614
615 return source_next;
616 }
617
618 bool ParallelCompactData::summarize(SplitInfo& split_info,
619 HeapWord* source_beg, HeapWord* source_end,
620 HeapWord** source_next,
621 HeapWord* target_beg, HeapWord* target_end,
622 HeapWord** target_next)
623 {
624 if (TraceParallelOldGCSummaryPhase) {
625 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
626 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
627 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
628 source_beg, source_end, source_next_val,
629 target_beg, target_end, *target_next);
630 }
631
632 size_t cur_region = addr_to_region_idx(source_beg);
633 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
634
635 HeapWord *dest_addr = target_beg;
636 while (cur_region < end_region) {
637 // The destination must be set even if the region has no data.
638 _region_data[cur_region].set_destination(dest_addr);
639
640 size_t words = _region_data[cur_region].data_size();
641 if (words > 0) {
642 // If cur_region does not fit entirely into the target space, find a point
643 // at which the source space can be 'split' so that part is copied to the
644 // target space and the rest is copied elsewhere.
645 if (dest_addr + words > target_end) {
646 assert(source_next != NULL, "source_next is NULL when splitting");
647 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
648 target_end, target_next);
649 return false;
650 }
651
652 // Compute the destination_count for cur_region, and if necessary, update
653 // source_region for a destination region. The source_region field is
654 // updated if cur_region is the first (left-most) region to be copied to a
655 // destination region.
656 //
657 // The destination_count calculation is a bit subtle. A region that has
658 // data that compacts into itself does not count itself as a destination.
659 // This maintains the invariant that a zero count means the region is
660 // available and can be claimed and then filled.
661 uint destination_count = 0;
662 if (split_info.is_split(cur_region)) {
663 // The current region has been split: the partial object will be copied
664 // to one destination space and the remaining data will be copied to
665 // another destination space. Adjust the initial destination_count and,
666 // if necessary, set the source_region field if the partial object will
667 // cross a destination region boundary.
668 destination_count = split_info.destination_count();
669 if (destination_count == 2) {
670 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
671 _region_data[dest_idx].set_source_region(cur_region);
672 }
673 }
674
675 HeapWord* const last_addr = dest_addr + words - 1;
676 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
677 const size_t dest_region_2 = addr_to_region_idx(last_addr);
678
679 // Initially assume that the destination regions will be the same and
680 // adjust the value below if necessary. Under this assumption, if
681 // cur_region == dest_region_2, then cur_region will be compacted
682 // completely into itself.
683 destination_count += cur_region == dest_region_2 ? 0 : 1;
684 if (dest_region_1 != dest_region_2) {
685 // Destination regions differ; adjust destination_count.
686 destination_count += 1;
687 // Data from cur_region will be copied to the start of dest_region_2.
688 _region_data[dest_region_2].set_source_region(cur_region);
689 } else if (region_offset(dest_addr) == 0) {
690 // Data from cur_region will be copied to the start of the destination
691 // region.
692 _region_data[dest_region_1].set_source_region(cur_region);
693 }
694
695 _region_data[cur_region].set_destination_count(destination_count);
696 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
697 dest_addr += words;
698 }
699
700 ++cur_region;
701 }
702
703 *target_next = dest_addr;
704 return true;
705 }
706
707 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
708 assert(addr != NULL, "Should detect NULL oop earlier");
709 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
710 #ifdef ASSERT
711 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
712 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
713 }
714 #endif
715 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
716
717 // Region covering the object.
718 size_t region_index = addr_to_region_idx(addr);
719 const RegionData* const region_ptr = region(region_index);
720 HeapWord* const region_addr = region_align_down(addr);
721
722 assert(addr < region_addr + RegionSize, "Region does not cover object");
723 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
724
725 HeapWord* result = region_ptr->destination();
726
727 // If all the data in the region is live, then the new location of the object
728 // can be calculated from the destination of the region plus the offset of the
729 // object in the region.
730 if (region_ptr->data_size() == RegionSize) {
731 result += pointer_delta(addr, region_addr);
732 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
733 return result;
734 }
735
736 // The new location of the object is
737 // region destination +
738 // size of the partial object extending onto the region +
739 // sizes of the live objects in the Region that are to the left of addr
740 const size_t partial_obj_size = region_ptr->partial_obj_size();
741 HeapWord* const search_start = region_addr + partial_obj_size;
742
743 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
744 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
745
746 result += partial_obj_size + live_to_left;
747 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
748 return result;
749 }
750
751 #ifdef ASSERT
752 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
753 {
754 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
755 const size_t* const end = (const size_t*)vspace->committed_high_addr();
756 for (const size_t* p = beg; p < end; ++p) {
757 assert(*p == 0, "not zero");
758 }
759 }
760
761 void ParallelCompactData::verify_clear()
762 {
763 verify_clear(_region_vspace);
764 }
765 #endif // #ifdef ASSERT
766
767 #ifdef NOT_PRODUCT
768 ParallelCompactData::RegionData* debug_region(size_t region_index) {
769 ParallelCompactData& sd = PSParallelCompact::summary_data();
770 return sd.region(region_index);
771 }
772 #endif
773
774 elapsedTimer PSParallelCompact::_accumulated_time;
775 unsigned int PSParallelCompact::_total_invocations = 0;
776 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
777 jlong PSParallelCompact::_time_of_last_gc = 0;
778 CollectorCounters* PSParallelCompact::_counters = NULL;
779 ParMarkBitMap PSParallelCompact::_mark_bitmap;
780 ParallelCompactData PSParallelCompact::_summary_data;
781
782 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
783
784 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
785
786 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
787 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
788
789 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
790 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
791
792 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); }
793 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
794
795 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
796
797 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) {
798 mark_and_push(_compaction_manager, p);
799 }
800 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
801
802 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
803 klass->oops_do(_mark_and_push_closure);
804 }
805 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
806 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
807 }
808
809 void PSParallelCompact::post_initialize() {
810 ParallelScavengeHeap* heap = gc_heap();
811 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
812
813 MemRegion mr = heap->reserved_region();
814 _ref_processor =
815 new ReferenceProcessor(mr, // span
816 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
817 (int) ParallelGCThreads, // mt processing degree
818 true, // mt discovery
819 (int) ParallelGCThreads, // mt discovery degree
820 true, // atomic_discovery
821 &_is_alive_closure, // non-header is alive closure
822 false); // write barrier for next field updates
823 _counters = new CollectorCounters("PSParallelCompact", 1);
824
825 // Initialize static fields in ParCompactionManager.
826 ParCompactionManager::initialize(mark_bitmap());
827 }
828
829 bool PSParallelCompact::initialize() {
830 ParallelScavengeHeap* heap = gc_heap();
831 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
832 MemRegion mr = heap->reserved_region();
833
834 // Was the old gen get allocated successfully?
835 if (!heap->old_gen()->is_allocated()) {
836 return false;
837 }
838
839 initialize_space_info();
840 initialize_dead_wood_limiter();
841
842 if (!_mark_bitmap.initialize(mr)) {
843 vm_shutdown_during_initialization("Unable to allocate bit map for "
844 "parallel garbage collection for the requested heap size.");
845 return false;
846 }
847
848 if (!_summary_data.initialize(mr)) {
849 vm_shutdown_during_initialization("Unable to allocate tables for "
850 "parallel garbage collection for the requested heap size.");
851 return false;
852 }
853
854 return true;
855 }
856
857 void PSParallelCompact::initialize_space_info()
858 {
859 memset(&_space_info, 0, sizeof(_space_info));
860
861 ParallelScavengeHeap* heap = gc_heap();
862 PSYoungGen* young_gen = heap->young_gen();
863
864 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
865 _space_info[eden_space_id].set_space(young_gen->eden_space());
866 _space_info[from_space_id].set_space(young_gen->from_space());
867 _space_info[to_space_id].set_space(young_gen->to_space());
868
869 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
870 }
871
872 void PSParallelCompact::initialize_dead_wood_limiter()
873 {
874 const size_t max = 100;
875 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
876 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
877 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
878 DEBUG_ONLY(_dwl_initialized = true;)
879 _dwl_adjustment = normal_distribution(1.0);
880 }
881
882 // Simple class for storing info about the heap at the start of GC, to be used
883 // after GC for comparison/printing.
884 class PreGCValues {
885 public:
886 PreGCValues() { }
887 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
888
889 void fill(ParallelScavengeHeap* heap) {
890 _heap_used = heap->used();
891 _young_gen_used = heap->young_gen()->used_in_bytes();
892 _old_gen_used = heap->old_gen()->used_in_bytes();
893 _metadata_used = MetaspaceAux::allocated_used_bytes();
894 };
895
896 size_t heap_used() const { return _heap_used; }
897 size_t young_gen_used() const { return _young_gen_used; }
898 size_t old_gen_used() const { return _old_gen_used; }
899 size_t metadata_used() const { return _metadata_used; }
900
901 private:
902 size_t _heap_used;
903 size_t _young_gen_used;
904 size_t _old_gen_used;
905 size_t _metadata_used;
906 };
907
908 void
909 PSParallelCompact::clear_data_covering_space(SpaceId id)
910 {
911 // At this point, top is the value before GC, new_top() is the value that will
912 // be set at the end of GC. The marking bitmap is cleared to top; nothing
913 // should be marked above top. The summary data is cleared to the larger of
914 // top & new_top.
915 MutableSpace* const space = _space_info[id].space();
916 HeapWord* const bot = space->bottom();
917 HeapWord* const top = space->top();
918 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
919
920 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
921 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
922 _mark_bitmap.clear_range(beg_bit, end_bit);
923
924 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
925 const size_t end_region =
926 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
927 _summary_data.clear_range(beg_region, end_region);
928
929 // Clear the data used to 'split' regions.
930 SplitInfo& split_info = _space_info[id].split_info();
931 if (split_info.is_valid()) {
932 split_info.clear();
933 }
934 DEBUG_ONLY(split_info.verify_clear();)
935 }
936
937 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
938 {
939 // Update the from & to space pointers in space_info, since they are swapped
940 // at each young gen gc. Do the update unconditionally (even though a
941 // promotion failure does not swap spaces) because an unknown number of minor
942 // collections will have swapped the spaces an unknown number of times.
943 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
944 ParallelScavengeHeap* heap = gc_heap();
945 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
946 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
947
948 pre_gc_values->fill(heap);
949
950 NOT_PRODUCT(_mark_bitmap.reset_counters());
951 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
952 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
953
954 // Increment the invocation count
955 heap->increment_total_collections(true);
956
957 // We need to track unique mark sweep invocations as well.
958 _total_invocations++;
959
960 heap->print_heap_before_gc();
961
962 // Fill in TLABs
963 heap->accumulate_statistics_all_tlabs();
964 heap->ensure_parsability(true); // retire TLABs
965
966 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
967 HandleMark hm; // Discard invalid handles created during verification
968 Universe::verify(" VerifyBeforeGC:");
969 }
970
971 // Verify object start arrays
972 if (VerifyObjectStartArray &&
973 VerifyBeforeGC) {
974 heap->old_gen()->verify_object_start_array();
975 }
976
977 DEBUG_ONLY(mark_bitmap()->verify_clear();)
978 DEBUG_ONLY(summary_data().verify_clear();)
979
980 // Have worker threads release resources the next time they run a task.
981 gc_task_manager()->release_all_resources();
982 }
983
984 void PSParallelCompact::post_compact()
985 {
986 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
987
988 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
989 // Clear the marking bitmap, summary data and split info.
990 clear_data_covering_space(SpaceId(id));
991 // Update top(). Must be done after clearing the bitmap and summary data.
992 _space_info[id].publish_new_top();
993 }
994
995 MutableSpace* const eden_space = _space_info[eden_space_id].space();
996 MutableSpace* const from_space = _space_info[from_space_id].space();
997 MutableSpace* const to_space = _space_info[to_space_id].space();
998
999 ParallelScavengeHeap* heap = gc_heap();
1000 bool eden_empty = eden_space->is_empty();
1001 if (!eden_empty) {
1002 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1003 heap->young_gen(), heap->old_gen());
1004 }
1005
1006 // Update heap occupancy information which is used as input to the soft ref
1007 // clearing policy at the next gc.
1008 Universe::update_heap_info_at_gc();
1009
1010 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1011 to_space->is_empty();
1012
1013 BarrierSet* bs = heap->barrier_set();
1014 if (bs->is_a(BarrierSet::ModRef)) {
1015 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1016 MemRegion old_mr = heap->old_gen()->reserved();
1017
1018 if (young_gen_empty) {
1019 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1020 } else {
1021 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1022 }
1023 }
1024
1025 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1026 ClassLoaderDataGraph::purge();
1027 MetaspaceAux::verify_metrics();
1028
1029 Threads::gc_epilogue();
1030 CodeCache::gc_epilogue();
1031 JvmtiExport::gc_epilogue();
1032
1033 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1034
1035 ref_processor()->enqueue_discovered_references(NULL);
1036
1037 if (ZapUnusedHeapArea) {
1038 heap->gen_mangle_unused_area();
1039 }
1040
1041 // Update time of last GC
1042 reset_millis_since_last_gc();
1043 }
1044
1045 HeapWord*
1046 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1047 bool maximum_compaction)
1048 {
1049 const size_t region_size = ParallelCompactData::RegionSize;
1050 const ParallelCompactData& sd = summary_data();
1051
1052 const MutableSpace* const space = _space_info[id].space();
1053 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1054 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1055 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1056
1057 // Skip full regions at the beginning of the space--they are necessarily part
1058 // of the dense prefix.
1059 size_t full_count = 0;
1060 const RegionData* cp;
1061 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1062 ++full_count;
1063 }
1064
1065 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1066 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1067 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1068 if (maximum_compaction || cp == end_cp || interval_ended) {
1069 _maximum_compaction_gc_num = total_invocations();
1070 return sd.region_to_addr(cp);
1071 }
1072
1073 HeapWord* const new_top = _space_info[id].new_top();
1074 const size_t space_live = pointer_delta(new_top, space->bottom());
1075 const size_t space_used = space->used_in_words();
1076 const size_t space_capacity = space->capacity_in_words();
1077
1078 const double cur_density = double(space_live) / space_capacity;
1079 const double deadwood_density =
1080 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1081 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1082
1083 if (TraceParallelOldGCDensePrefix) {
1084 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1085 cur_density, deadwood_density, deadwood_goal);
1086 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1087 "space_cap=" SIZE_FORMAT,
1088 space_live, space_used,
1089 space_capacity);
1090 }
1091
1092 // XXX - Use binary search?
1093 HeapWord* dense_prefix = sd.region_to_addr(cp);
1094 const RegionData* full_cp = cp;
1095 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1096 while (cp < end_cp) {
1097 HeapWord* region_destination = cp->destination();
1098 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1099 if (TraceParallelOldGCDensePrefix && Verbose) {
1100 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1101 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1102 sd.region(cp), region_destination,
1103 dense_prefix, cur_deadwood);
1104 }
1105
1106 if (cur_deadwood >= deadwood_goal) {
1107 // Found the region that has the correct amount of deadwood to the left.
1108 // This typically occurs after crossing a fairly sparse set of regions, so
1109 // iterate backwards over those sparse regions, looking for the region
1110 // that has the lowest density of live objects 'to the right.'
1111 size_t space_to_left = sd.region(cp) * region_size;
1112 size_t live_to_left = space_to_left - cur_deadwood;
1113 size_t space_to_right = space_capacity - space_to_left;
1114 size_t live_to_right = space_live - live_to_left;
1115 double density_to_right = double(live_to_right) / space_to_right;
1116 while (cp > full_cp) {
1117 --cp;
1118 const size_t prev_region_live_to_right = live_to_right -
1119 cp->data_size();
1120 const size_t prev_region_space_to_right = space_to_right + region_size;
1121 double prev_region_density_to_right =
1122 double(prev_region_live_to_right) / prev_region_space_to_right;
1123 if (density_to_right <= prev_region_density_to_right) {
1124 return dense_prefix;
1125 }
1126 if (TraceParallelOldGCDensePrefix && Verbose) {
1127 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1128 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1129 prev_region_density_to_right);
1130 }
1131 dense_prefix -= region_size;
1132 live_to_right = prev_region_live_to_right;
1133 space_to_right = prev_region_space_to_right;
1134 density_to_right = prev_region_density_to_right;
1135 }
1136 return dense_prefix;
1137 }
1138
1139 dense_prefix += region_size;
1140 ++cp;
1141 }
1142
1143 return dense_prefix;
1144 }
1145
1146 #ifndef PRODUCT
1147 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1148 const SpaceId id,
1149 const bool maximum_compaction,
1150 HeapWord* const addr)
1151 {
1152 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1153 RegionData* const cp = summary_data().region(region_idx);
1154 const MutableSpace* const space = _space_info[id].space();
1155 HeapWord* const new_top = _space_info[id].new_top();
1156
1157 const size_t space_live = pointer_delta(new_top, space->bottom());
1158 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1159 const size_t space_cap = space->capacity_in_words();
1160 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1161 const size_t live_to_right = new_top - cp->destination();
1162 const size_t dead_to_right = space->top() - addr - live_to_right;
1163
1164 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1165 "spl=" SIZE_FORMAT " "
1166 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1167 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1168 " ratio=%10.8f",
1169 algorithm, addr, region_idx,
1170 space_live,
1171 dead_to_left, dead_to_left_pct,
1172 dead_to_right, live_to_right,
1173 double(dead_to_right) / live_to_right);
1174 }
1175 #endif // #ifndef PRODUCT
1176
1177 // Return a fraction indicating how much of the generation can be treated as
1178 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1179 // based on the density of live objects in the generation to determine a limit,
1180 // which is then adjusted so the return value is min_percent when the density is
1181 // 1.
1182 //
1183 // The following table shows some return values for a different values of the
1184 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1185 // min_percent is 1.
1186 //
1187 // fraction allowed as dead wood
1188 // -----------------------------------------------------------------
1189 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1190 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1191 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1192 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1193 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1194 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1195 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1196 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1197 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1198 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1199 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1200 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1201 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1202 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1203 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1204 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1205 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1206 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1207 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1208 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1209 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1210 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1211 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1212
1213 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1214 {
1215 assert(_dwl_initialized, "uninitialized");
1216
1217 // The raw limit is the value of the normal distribution at x = density.
1218 const double raw_limit = normal_distribution(density);
1219
1220 // Adjust the raw limit so it becomes the minimum when the density is 1.
1221 //
1222 // First subtract the adjustment value (which is simply the precomputed value
1223 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1224 // Then add the minimum value, so the minimum is returned when the density is
1225 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1226 const double min = double(min_percent) / 100.0;
1227 const double limit = raw_limit - _dwl_adjustment + min;
1228 return MAX2(limit, 0.0);
1229 }
1230
1231 ParallelCompactData::RegionData*
1232 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1233 const RegionData* end)
1234 {
1235 const size_t region_size = ParallelCompactData::RegionSize;
1236 ParallelCompactData& sd = summary_data();
1237 size_t left = sd.region(beg);
1238 size_t right = end > beg ? sd.region(end) - 1 : left;
1239
1240 // Binary search.
1241 while (left < right) {
1242 // Equivalent to (left + right) / 2, but does not overflow.
1243 const size_t middle = left + (right - left) / 2;
1244 RegionData* const middle_ptr = sd.region(middle);
1245 HeapWord* const dest = middle_ptr->destination();
1246 HeapWord* const addr = sd.region_to_addr(middle);
1247 assert(dest != NULL, "sanity");
1248 assert(dest <= addr, "must move left");
1249
1250 if (middle > left && dest < addr) {
1251 right = middle - 1;
1252 } else if (middle < right && middle_ptr->data_size() == region_size) {
1253 left = middle + 1;
1254 } else {
1255 return middle_ptr;
1256 }
1257 }
1258 return sd.region(left);
1259 }
1260
1261 ParallelCompactData::RegionData*
1262 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1263 const RegionData* end,
1264 size_t dead_words)
1265 {
1266 ParallelCompactData& sd = summary_data();
1267 size_t left = sd.region(beg);
1268 size_t right = end > beg ? sd.region(end) - 1 : left;
1269
1270 // Binary search.
1271 while (left < right) {
1272 // Equivalent to (left + right) / 2, but does not overflow.
1273 const size_t middle = left + (right - left) / 2;
1274 RegionData* const middle_ptr = sd.region(middle);
1275 HeapWord* const dest = middle_ptr->destination();
1276 HeapWord* const addr = sd.region_to_addr(middle);
1277 assert(dest != NULL, "sanity");
1278 assert(dest <= addr, "must move left");
1279
1280 const size_t dead_to_left = pointer_delta(addr, dest);
1281 if (middle > left && dead_to_left > dead_words) {
1282 right = middle - 1;
1283 } else if (middle < right && dead_to_left < dead_words) {
1284 left = middle + 1;
1285 } else {
1286 return middle_ptr;
1287 }
1288 }
1289 return sd.region(left);
1290 }
1291
1292 // The result is valid during the summary phase, after the initial summarization
1293 // of each space into itself, and before final summarization.
1294 inline double
1295 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1296 HeapWord* const bottom,
1297 HeapWord* const top,
1298 HeapWord* const new_top)
1299 {
1300 ParallelCompactData& sd = summary_data();
1301
1302 assert(cp != NULL, "sanity");
1303 assert(bottom != NULL, "sanity");
1304 assert(top != NULL, "sanity");
1305 assert(new_top != NULL, "sanity");
1306 assert(top >= new_top, "summary data problem?");
1307 assert(new_top > bottom, "space is empty; should not be here");
1308 assert(new_top >= cp->destination(), "sanity");
1309 assert(top >= sd.region_to_addr(cp), "sanity");
1310
1311 HeapWord* const destination = cp->destination();
1312 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1313 const size_t compacted_region_live = pointer_delta(new_top, destination);
1314 const size_t compacted_region_used = pointer_delta(top,
1315 sd.region_to_addr(cp));
1316 const size_t reclaimable = compacted_region_used - compacted_region_live;
1317
1318 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1319 return double(reclaimable) / divisor;
1320 }
1321
1322 // Return the address of the end of the dense prefix, a.k.a. the start of the
1323 // compacted region. The address is always on a region boundary.
1324 //
1325 // Completely full regions at the left are skipped, since no compaction can
1326 // occur in those regions. Then the maximum amount of dead wood to allow is
1327 // computed, based on the density (amount live / capacity) of the generation;
1328 // the region with approximately that amount of dead space to the left is
1329 // identified as the limit region. Regions between the last completely full
1330 // region and the limit region are scanned and the one that has the best
1331 // (maximum) reclaimed_ratio() is selected.
1332 HeapWord*
1333 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1334 bool maximum_compaction)
1335 {
1336 if (ParallelOldGCSplitALot) {
1337 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1338 // The value was chosen to provoke splitting a young gen space; use it.
1339 return _space_info[id].dense_prefix();
1340 }
1341 }
1342
1343 const size_t region_size = ParallelCompactData::RegionSize;
1344 const ParallelCompactData& sd = summary_data();
1345
1346 const MutableSpace* const space = _space_info[id].space();
1347 HeapWord* const top = space->top();
1348 HeapWord* const top_aligned_up = sd.region_align_up(top);
1349 HeapWord* const new_top = _space_info[id].new_top();
1350 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1351 HeapWord* const bottom = space->bottom();
1352 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1353 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1354 const RegionData* const new_top_cp =
1355 sd.addr_to_region_ptr(new_top_aligned_up);
1356
1357 // Skip full regions at the beginning of the space--they are necessarily part
1358 // of the dense prefix.
1359 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1360 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1361 space->is_empty(), "no dead space allowed to the left");
1362 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1363 "region must have dead space");
1364
1365 // The gc number is saved whenever a maximum compaction is done, and used to
1366 // determine when the maximum compaction interval has expired. This avoids
1367 // successive max compactions for different reasons.
1368 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1369 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1370 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1371 total_invocations() == HeapFirstMaximumCompactionCount;
1372 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1373 _maximum_compaction_gc_num = total_invocations();
1374 return sd.region_to_addr(full_cp);
1375 }
1376
1377 const size_t space_live = pointer_delta(new_top, bottom);
1378 const size_t space_used = space->used_in_words();
1379 const size_t space_capacity = space->capacity_in_words();
1380
1381 const double density = double(space_live) / double(space_capacity);
1382 const size_t min_percent_free = MarkSweepDeadRatio;
1383 const double limiter = dead_wood_limiter(density, min_percent_free);
1384 const size_t dead_wood_max = space_used - space_live;
1385 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1386 dead_wood_max);
1387
1388 if (TraceParallelOldGCDensePrefix) {
1389 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1390 "space_cap=" SIZE_FORMAT,
1391 space_live, space_used,
1392 space_capacity);
1393 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1394 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1395 density, min_percent_free, limiter,
1396 dead_wood_max, dead_wood_limit);
1397 }
1398
1399 // Locate the region with the desired amount of dead space to the left.
1400 const RegionData* const limit_cp =
1401 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1402
1403 // Scan from the first region with dead space to the limit region and find the
1404 // one with the best (largest) reclaimed ratio.
1405 double best_ratio = 0.0;
1406 const RegionData* best_cp = full_cp;
1407 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1408 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1409 if (tmp_ratio > best_ratio) {
1410 best_cp = cp;
1411 best_ratio = tmp_ratio;
1412 }
1413 }
1414
1415 #if 0
1416 // Something to consider: if the region with the best ratio is 'close to' the
1417 // first region w/free space, choose the first region with free space
1418 // ("first-free"). The first-free region is usually near the start of the
1419 // heap, which means we are copying most of the heap already, so copy a bit
1420 // more to get complete compaction.
1421 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1422 _maximum_compaction_gc_num = total_invocations();
1423 best_cp = full_cp;
1424 }
1425 #endif // #if 0
1426
1427 return sd.region_to_addr(best_cp);
1428 }
1429
1430 #ifndef PRODUCT
1431 void
1432 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1433 size_t words)
1434 {
1435 if (TraceParallelOldGCSummaryPhase) {
1436 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1437 SIZE_FORMAT, start, start + words, words);
1438 }
1439
1440 ObjectStartArray* const start_array = _space_info[id].start_array();
1441 CollectedHeap::fill_with_objects(start, words);
1442 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1443 _mark_bitmap.mark_obj(p, words);
1444 _summary_data.add_obj(p, words);
1445 start_array->allocate_block(p);
1446 }
1447 }
1448
1449 void
1450 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1451 {
1452 ParallelCompactData& sd = summary_data();
1453 MutableSpace* space = _space_info[id].space();
1454
1455 // Find the source and destination start addresses.
1456 HeapWord* const src_addr = sd.region_align_down(start);
1457 HeapWord* dst_addr;
1458 if (src_addr < start) {
1459 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1460 } else if (src_addr > space->bottom()) {
1461 // The start (the original top() value) is aligned to a region boundary so
1462 // the associated region does not have a destination. Compute the
1463 // destination from the previous region.
1464 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1465 dst_addr = cp->destination() + cp->data_size();
1466 } else {
1467 // Filling the entire space.
1468 dst_addr = space->bottom();
1469 }
1470 assert(dst_addr != NULL, "sanity");
1471
1472 // Update the summary data.
1473 bool result = _summary_data.summarize(_space_info[id].split_info(),
1474 src_addr, space->top(), NULL,
1475 dst_addr, space->end(),
1476 _space_info[id].new_top_addr());
1477 assert(result, "should not fail: bad filler object size");
1478 }
1479
1480 void
1481 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1482 {
1483 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1484 return;
1485 }
1486
1487 MutableSpace* const space = _space_info[id].space();
1488 if (space->is_empty()) {
1489 HeapWord* b = space->bottom();
1490 HeapWord* t = b + space->capacity_in_words() / 2;
1491 space->set_top(t);
1492 if (ZapUnusedHeapArea) {
1493 space->set_top_for_allocations();
1494 }
1495
1496 size_t min_size = CollectedHeap::min_fill_size();
1497 size_t obj_len = min_size;
1498 while (b + obj_len <= t) {
1499 CollectedHeap::fill_with_object(b, obj_len);
1500 mark_bitmap()->mark_obj(b, obj_len);
1501 summary_data().add_obj(b, obj_len);
1502 b += obj_len;
1503 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1504 }
1505 if (b < t) {
1506 // The loop didn't completely fill to t (top); adjust top downward.
1507 space->set_top(b);
1508 if (ZapUnusedHeapArea) {
1509 space->set_top_for_allocations();
1510 }
1511 }
1512
1513 HeapWord** nta = _space_info[id].new_top_addr();
1514 bool result = summary_data().summarize(_space_info[id].split_info(),
1515 space->bottom(), space->top(), NULL,
1516 space->bottom(), space->end(), nta);
1517 assert(result, "space must fit into itself");
1518 }
1519 }
1520
1521 void
1522 PSParallelCompact::provoke_split(bool & max_compaction)
1523 {
1524 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1525 return;
1526 }
1527
1528 const size_t region_size = ParallelCompactData::RegionSize;
1529 ParallelCompactData& sd = summary_data();
1530
1531 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1532 MutableSpace* const from_space = _space_info[from_space_id].space();
1533 const size_t eden_live = pointer_delta(eden_space->top(),
1534 _space_info[eden_space_id].new_top());
1535 const size_t from_live = pointer_delta(from_space->top(),
1536 _space_info[from_space_id].new_top());
1537
1538 const size_t min_fill_size = CollectedHeap::min_fill_size();
1539 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1540 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1541 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1542 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1543
1544 // Choose the space to split; need at least 2 regions live (or fillable).
1545 SpaceId id;
1546 MutableSpace* space;
1547 size_t live_words;
1548 size_t fill_words;
1549 if (eden_live + eden_fillable >= region_size * 2) {
1550 id = eden_space_id;
1551 space = eden_space;
1552 live_words = eden_live;
1553 fill_words = eden_fillable;
1554 } else if (from_live + from_fillable >= region_size * 2) {
1555 id = from_space_id;
1556 space = from_space;
1557 live_words = from_live;
1558 fill_words = from_fillable;
1559 } else {
1560 return; // Give up.
1561 }
1562 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1563
1564 if (live_words < region_size * 2) {
1565 // Fill from top() to end() w/live objects of mixed sizes.
1566 HeapWord* const fill_start = space->top();
1567 live_words += fill_words;
1568
1569 space->set_top(fill_start + fill_words);
1570 if (ZapUnusedHeapArea) {
1571 space->set_top_for_allocations();
1572 }
1573
1574 HeapWord* cur_addr = fill_start;
1575 while (fill_words > 0) {
1576 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1577 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1578 if (fill_words - cur_size < min_fill_size) {
1579 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1580 }
1581
1582 CollectedHeap::fill_with_object(cur_addr, cur_size);
1583 mark_bitmap()->mark_obj(cur_addr, cur_size);
1584 sd.add_obj(cur_addr, cur_size);
1585
1586 cur_addr += cur_size;
1587 fill_words -= cur_size;
1588 }
1589
1590 summarize_new_objects(id, fill_start);
1591 }
1592
1593 max_compaction = false;
1594
1595 // Manipulate the old gen so that it has room for about half of the live data
1596 // in the target young gen space (live_words / 2).
1597 id = old_space_id;
1598 space = _space_info[id].space();
1599 const size_t free_at_end = space->free_in_words();
1600 const size_t free_target = align_object_size(live_words / 2);
1601 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1602
1603 if (free_at_end >= free_target + min_fill_size) {
1604 // Fill space above top() and set the dense prefix so everything survives.
1605 HeapWord* const fill_start = space->top();
1606 const size_t fill_size = free_at_end - free_target;
1607 space->set_top(space->top() + fill_size);
1608 if (ZapUnusedHeapArea) {
1609 space->set_top_for_allocations();
1610 }
1611 fill_with_live_objects(id, fill_start, fill_size);
1612 summarize_new_objects(id, fill_start);
1613 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1614 } else if (dead + free_at_end > free_target) {
1615 // Find a dense prefix that makes the right amount of space available.
1616 HeapWord* cur = sd.region_align_down(space->top());
1617 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1618 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1619 while (dead_to_right < free_target) {
1620 cur -= region_size;
1621 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1622 dead_to_right = pointer_delta(space->end(), cur_destination);
1623 }
1624 _space_info[id].set_dense_prefix(cur);
1625 }
1626 }
1627 #endif // #ifndef PRODUCT
1628
1629 void PSParallelCompact::summarize_spaces_quick()
1630 {
1631 for (unsigned int i = 0; i < last_space_id; ++i) {
1632 const MutableSpace* space = _space_info[i].space();
1633 HeapWord** nta = _space_info[i].new_top_addr();
1634 bool result = _summary_data.summarize(_space_info[i].split_info(),
1635 space->bottom(), space->top(), NULL,
1636 space->bottom(), space->end(), nta);
1637 assert(result, "space must fit into itself");
1638 _space_info[i].set_dense_prefix(space->bottom());
1639 }
1640
1641 #ifndef PRODUCT
1642 if (ParallelOldGCSplitALot) {
1643 provoke_split_fill_survivor(to_space_id);
1644 }
1645 #endif // #ifndef PRODUCT
1646 }
1647
1648 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1649 {
1650 HeapWord* const dense_prefix_end = dense_prefix(id);
1651 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1652 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1653 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1654 // Only enough dead space is filled so that any remaining dead space to the
1655 // left is larger than the minimum filler object. (The remainder is filled
1656 // during the copy/update phase.)
1657 //
1658 // The size of the dead space to the right of the boundary is not a
1659 // concern, since compaction will be able to use whatever space is
1660 // available.
1661 //
1662 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1663 // surrounds the space to be filled with an object.
1664 //
1665 // In the 32-bit VM, each bit represents two 32-bit words:
1666 // +---+
1667 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1668 // end_bits: ... x x x | 0 | || 0 x x ...
1669 // +---+
1670 //
1671 // In the 64-bit VM, each bit represents one 64-bit word:
1672 // +------------+
1673 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1674 // end_bits: ... x x 1 | 0 || 0 | x x ...
1675 // +------------+
1676 // +-------+
1677 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1678 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1679 // +-------+
1680 // +-----------+
1681 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1682 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1683 // +-----------+
1684 // +-------+
1685 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1686 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1687 // +-------+
1688
1689 // Initially assume case a, c or e will apply.
1690 size_t obj_len = CollectedHeap::min_fill_size();
1691 HeapWord* obj_beg = dense_prefix_end - obj_len;
1692
1693 #ifdef _LP64
1694 if (MinObjAlignment > 1) { // object alignment > heap word size
1695 // Cases a, c or e.
1696 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1697 // Case b above.
1698 obj_beg = dense_prefix_end - 1;
1699 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1700 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1701 // Case d above.
1702 obj_beg = dense_prefix_end - 3;
1703 obj_len = 3;
1704 }
1705 #endif // #ifdef _LP64
1706
1707 CollectedHeap::fill_with_object(obj_beg, obj_len);
1708 _mark_bitmap.mark_obj(obj_beg, obj_len);
1709 _summary_data.add_obj(obj_beg, obj_len);
1710 assert(start_array(id) != NULL, "sanity");
1711 start_array(id)->allocate_block(obj_beg);
1712 }
1713 }
1714
1715 void
1716 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1717 {
1718 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1719 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1720 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1721 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1722 cur->set_source_region(0);
1723 }
1724 }
1725
1726 void
1727 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1728 {
1729 assert(id < last_space_id, "id out of range");
1730 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1731 ParallelOldGCSplitALot && id == old_space_id,
1732 "should have been reset in summarize_spaces_quick()");
1733
1734 const MutableSpace* space = _space_info[id].space();
1735 if (_space_info[id].new_top() != space->bottom()) {
1736 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1737 _space_info[id].set_dense_prefix(dense_prefix_end);
1738
1739 #ifndef PRODUCT
1740 if (TraceParallelOldGCDensePrefix) {
1741 print_dense_prefix_stats("ratio", id, maximum_compaction,
1742 dense_prefix_end);
1743 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1744 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1745 }
1746 #endif // #ifndef PRODUCT
1747
1748 // Recompute the summary data, taking into account the dense prefix. If
1749 // every last byte will be reclaimed, then the existing summary data which
1750 // compacts everything can be left in place.
1751 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1752 // If dead space crosses the dense prefix boundary, it is (at least
1753 // partially) filled with a dummy object, marked live and added to the
1754 // summary data. This simplifies the copy/update phase and must be done
1755 // before the final locations of objects are determined, to prevent
1756 // leaving a fragment of dead space that is too small to fill.
1757 fill_dense_prefix_end(id);
1758
1759 // Compute the destination of each Region, and thus each object.
1760 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1761 _summary_data.summarize(_space_info[id].split_info(),
1762 dense_prefix_end, space->top(), NULL,
1763 dense_prefix_end, space->end(),
1764 _space_info[id].new_top_addr());
1765 }
1766 }
1767
1768 if (TraceParallelOldGCSummaryPhase) {
1769 const size_t region_size = ParallelCompactData::RegionSize;
1770 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1771 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1772 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1773 HeapWord* const new_top = _space_info[id].new_top();
1774 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1775 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1776 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1777 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1778 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1779 id, space->capacity_in_words(), dense_prefix_end,
1780 dp_region, dp_words / region_size,
1781 cr_words / region_size, new_top);
1782 }
1783 }
1784
1785 #ifndef PRODUCT
1786 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1787 HeapWord* dst_beg, HeapWord* dst_end,
1788 SpaceId src_space_id,
1789 HeapWord* src_beg, HeapWord* src_end)
1790 {
1791 if (TraceParallelOldGCSummaryPhase) {
1792 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1793 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1794 SIZE_FORMAT "-" SIZE_FORMAT " "
1795 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1796 SIZE_FORMAT "-" SIZE_FORMAT,
1797 src_space_id, space_names[src_space_id],
1798 dst_space_id, space_names[dst_space_id],
1799 src_beg, src_end,
1800 _summary_data.addr_to_region_idx(src_beg),
1801 _summary_data.addr_to_region_idx(src_end),
1802 dst_beg, dst_end,
1803 _summary_data.addr_to_region_idx(dst_beg),
1804 _summary_data.addr_to_region_idx(dst_end));
1805 }
1806 }
1807 #endif // #ifndef PRODUCT
1808
1809 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1810 bool maximum_compaction)
1811 {
1812 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1813 // trace("2");
1814
1815 #ifdef ASSERT
1816 if (TraceParallelOldGCMarkingPhase) {
1817 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1818 "add_obj_bytes=" SIZE_FORMAT,
1819 add_obj_count, add_obj_size * HeapWordSize);
1820 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1821 "mark_bitmap_bytes=" SIZE_FORMAT,
1822 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1823 }
1824 #endif // #ifdef ASSERT
1825
1826 // Quick summarization of each space into itself, to see how much is live.
1827 summarize_spaces_quick();
1828
1829 if (TraceParallelOldGCSummaryPhase) {
1830 tty->print_cr("summary_phase: after summarizing each space to self");
1831 Universe::print();
1832 NOT_PRODUCT(print_region_ranges());
1833 if (Verbose) {
1834 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1835 }
1836 }
1837
1838 // The amount of live data that will end up in old space (assuming it fits).
1839 size_t old_space_total_live = 0;
1840 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1841 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1842 _space_info[id].space()->bottom());
1843 }
1844
1845 MutableSpace* const old_space = _space_info[old_space_id].space();
1846 const size_t old_capacity = old_space->capacity_in_words();
1847 if (old_space_total_live > old_capacity) {
1848 // XXX - should also try to expand
1849 maximum_compaction = true;
1850 }
1851 #ifndef PRODUCT
1852 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1853 provoke_split(maximum_compaction);
1854 }
1855 #endif // #ifndef PRODUCT
1856
1857 // Old generations.
1858 summarize_space(old_space_id, maximum_compaction);
1859
1860 // Summarize the remaining spaces in the young gen. The initial target space
1861 // is the old gen. If a space does not fit entirely into the target, then the
1862 // remainder is compacted into the space itself and that space becomes the new
1863 // target.
1864 SpaceId dst_space_id = old_space_id;
1865 HeapWord* dst_space_end = old_space->end();
1866 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1867 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1868 const MutableSpace* space = _space_info[id].space();
1869 const size_t live = pointer_delta(_space_info[id].new_top(),
1870 space->bottom());
1871 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1872
1873 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1874 SpaceId(id), space->bottom(), space->top());)
1875 if (live > 0 && live <= available) {
1876 // All the live data will fit.
1877 bool done = _summary_data.summarize(_space_info[id].split_info(),
1878 space->bottom(), space->top(),
1879 NULL,
1880 *new_top_addr, dst_space_end,
1881 new_top_addr);
1882 assert(done, "space must fit into old gen");
1883
1884 // Reset the new_top value for the space.
1885 _space_info[id].set_new_top(space->bottom());
1886 } else if (live > 0) {
1887 // Attempt to fit part of the source space into the target space.
1888 HeapWord* next_src_addr = NULL;
1889 bool done = _summary_data.summarize(_space_info[id].split_info(),
1890 space->bottom(), space->top(),
1891 &next_src_addr,
1892 *new_top_addr, dst_space_end,
1893 new_top_addr);
1894 assert(!done, "space should not fit into old gen");
1895 assert(next_src_addr != NULL, "sanity");
1896
1897 // The source space becomes the new target, so the remainder is compacted
1898 // within the space itself.
1899 dst_space_id = SpaceId(id);
1900 dst_space_end = space->end();
1901 new_top_addr = _space_info[id].new_top_addr();
1902 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1903 space->bottom(), dst_space_end,
1904 SpaceId(id), next_src_addr, space->top());)
1905 done = _summary_data.summarize(_space_info[id].split_info(),
1906 next_src_addr, space->top(),
1907 NULL,
1908 space->bottom(), dst_space_end,
1909 new_top_addr);
1910 assert(done, "space must fit when compacted into itself");
1911 assert(*new_top_addr <= space->top(), "usage should not grow");
1912 }
1913 }
1914
1915 if (TraceParallelOldGCSummaryPhase) {
1916 tty->print_cr("summary_phase: after final summarization");
1917 Universe::print();
1918 NOT_PRODUCT(print_region_ranges());
1919 if (Verbose) {
1920 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1921 }
1922 }
1923 }
1924
1925 // This method should contain all heap-specific policy for invoking a full
1926 // collection. invoke_no_policy() will only attempt to compact the heap; it
1927 // will do nothing further. If we need to bail out for policy reasons, scavenge
1928 // before full gc, or any other specialized behavior, it needs to be added here.
1929 //
1930 // Note that this method should only be called from the vm_thread while at a
1931 // safepoint.
1932 //
1933 // Note that the all_soft_refs_clear flag in the collector policy
1934 // may be true because this method can be called without intervening
1935 // activity. For example when the heap space is tight and full measure
1936 // are being taken to free space.
1937 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1938 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1939 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1940 "should be in vm thread");
1941
1942 ParallelScavengeHeap* heap = gc_heap();
1943 GCCause::Cause gc_cause = heap->gc_cause();
1944 assert(!heap->is_gc_active(), "not reentrant");
1945
1946 PSAdaptiveSizePolicy* policy = heap->size_policy();
1947 IsGCActiveMark mark;
1948
1949 if (ScavengeBeforeFullGC) {
1950 PSScavenge::invoke_no_policy();
1951 }
1952
1953 const bool clear_all_soft_refs =
1954 heap->collector_policy()->should_clear_all_soft_refs();
1955
1956 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1957 maximum_heap_compaction);
1958 }
1959
1960 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1961 size_t addr_region_index = addr_to_region_idx(addr);
1962 return region_index == addr_region_index;
1963 }
1964
1965 // This method contains no policy. You should probably
1966 // be calling invoke() instead.
1967 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1968 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1969 assert(ref_processor() != NULL, "Sanity");
1970
1971 if (GC_locker::check_active_before_gc()) {
1972 return false;
1973 }
1974
1975 TimeStamp marking_start;
1976 TimeStamp compaction_start;
1977 TimeStamp collection_exit;
1978
1979 ParallelScavengeHeap* heap = gc_heap();
1980 GCCause::Cause gc_cause = heap->gc_cause();
1981 PSYoungGen* young_gen = heap->young_gen();
1982 PSOldGen* old_gen = heap->old_gen();
1983 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1984
1985 // The scope of casr should end after code that can change
1986 // CollectorPolicy::_should_clear_all_soft_refs.
1987 ClearedAllSoftRefs casr(maximum_heap_compaction,
1988 heap->collector_policy());
1989
1990 if (ZapUnusedHeapArea) {
1991 // Save information needed to minimize mangling
1992 heap->record_gen_tops_before_GC();
1993 }
1994
1995 heap->pre_full_gc_dump();
1996
1997 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
1998
1999 // Make sure data structures are sane, make the heap parsable, and do other
2000 // miscellaneous bookkeeping.
2001 PreGCValues pre_gc_values;
2002 pre_compact(&pre_gc_values);
2003
2004 // Get the compaction manager reserved for the VM thread.
2005 ParCompactionManager* const vmthread_cm =
2006 ParCompactionManager::manager_array(gc_task_manager()->workers());
2007
2008 // Place after pre_compact() where the number of invocations is incremented.
2009 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2010
2011 {
2012 ResourceMark rm;
2013 HandleMark hm;
2014
2015 // Set the number of GC threads to be used in this collection
2016 gc_task_manager()->set_active_gang();
2017 gc_task_manager()->task_idle_workers();
2018 heap->set_par_threads(gc_task_manager()->active_workers());
2019
2020 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2021 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2022 TraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty);
2023 TraceCollectorStats tcs(counters());
2024 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2025
2026 if (TraceGen1Time) accumulated_time()->start();
2027
2028 // Let the size policy know we're starting
2029 size_policy->major_collection_begin();
2030
2031 CodeCache::gc_prologue();
2032 Threads::gc_prologue();
2033
2034 COMPILER2_PRESENT(DerivedPointerTable::clear());
2035
2036 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2037 ref_processor()->setup_policy(maximum_heap_compaction);
2038
2039 bool marked_for_unloading = false;
2040
2041 marking_start.update();
2042 marking_phase(vmthread_cm, maximum_heap_compaction);
2043
2044 #ifndef PRODUCT
2045 if (TraceParallelOldGCMarkingPhase) {
2046 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2047 "cas_by_another %d",
2048 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2049 mark_bitmap()->cas_by_another());
2050 }
2051 #endif // #ifndef PRODUCT
2052
2053 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2054 && gc_cause == GCCause::_java_lang_system_gc;
2055 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2056
2057 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2058 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2059
2060 // adjust_roots() updates Universe::_intArrayKlassObj which is
2061 // needed by the compaction for filling holes in the dense prefix.
2062 adjust_roots();
2063
2064 compaction_start.update();
2065 compact();
2066
2067 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2068 // done before resizing.
2069 post_compact();
2070
2071 // Let the size policy know we're done
2072 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2073
2074 if (UseAdaptiveSizePolicy) {
2075 if (PrintAdaptiveSizePolicy) {
2076 gclog_or_tty->print("AdaptiveSizeStart: ");
2077 gclog_or_tty->stamp();
2078 gclog_or_tty->print_cr(" collection: %d ",
2079 heap->total_collections());
2080 if (Verbose) {
2081 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d",
2082 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2083 }
2084 }
2085
2086 // Don't check if the size_policy is ready here. Let
2087 // the size_policy check that internally.
2088 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2089 ((gc_cause != GCCause::_java_lang_system_gc) ||
2090 UseAdaptiveSizePolicyWithSystemGC)) {
2091 // Calculate optimal free space amounts
2092 assert(young_gen->max_size() >
2093 young_gen->from_space()->capacity_in_bytes() +
2094 young_gen->to_space()->capacity_in_bytes(),
2095 "Sizes of space in young gen are out-of-bounds");
2096 size_t max_eden_size = young_gen->max_size() -
2097 young_gen->from_space()->capacity_in_bytes() -
2098 young_gen->to_space()->capacity_in_bytes();
2099 size_policy->compute_generation_free_space(
2100 young_gen->used_in_bytes(),
2101 young_gen->eden_space()->used_in_bytes(),
2102 old_gen->used_in_bytes(),
2103 young_gen->eden_space()->capacity_in_bytes(),
2104 old_gen->max_gen_size(),
2105 max_eden_size,
2106 true /* full gc*/,
2107 gc_cause,
2108 heap->collector_policy());
2109
2110 heap->resize_old_gen(
2111 size_policy->calculated_old_free_size_in_bytes());
2112
2113 // Don't resize the young generation at an major collection. A
2114 // desired young generation size may have been calculated but
2115 // resizing the young generation complicates the code because the
2116 // resizing of the old generation may have moved the boundary
2117 // between the young generation and the old generation. Let the
2118 // young generation resizing happen at the minor collections.
2119 }
2120 if (PrintAdaptiveSizePolicy) {
2121 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2122 heap->total_collections());
2123 }
2124 }
2125
2126 if (UsePerfData) {
2127 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2128 counters->update_counters();
2129 counters->update_old_capacity(old_gen->capacity_in_bytes());
2130 counters->update_young_capacity(young_gen->capacity_in_bytes());
2131 }
2132
2133 heap->resize_all_tlabs();
2134
2135 // Resize the metaspace capactiy after a collection
2136 MetaspaceGC::compute_new_size();
2137
2138 if (TraceGen1Time) accumulated_time()->stop();
2139
2140 if (PrintGC) {
2141 if (PrintGCDetails) {
2142 // No GC timestamp here. This is after GC so it would be confusing.
2143 young_gen->print_used_change(pre_gc_values.young_gen_used());
2144 old_gen->print_used_change(pre_gc_values.old_gen_used());
2145 heap->print_heap_change(pre_gc_values.heap_used());
2146 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2147 } else {
2148 heap->print_heap_change(pre_gc_values.heap_used());
2149 }
2150 }
2151
2152 // Track memory usage and detect low memory
2153 MemoryService::track_memory_usage();
2154 heap->update_counters();
2155 gc_task_manager()->release_idle_workers();
2156 }
2157
2158 #ifdef ASSERT
2159 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2160 ParCompactionManager* const cm =
2161 ParCompactionManager::manager_array(int(i));
2162 assert(cm->marking_stack()->is_empty(), "should be empty");
2163 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2164 }
2165 #endif // ASSERT
2166
2167 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2168 HandleMark hm; // Discard invalid handles created during verification
2169 Universe::verify(" VerifyAfterGC:");
2170 }
2171
2172 // Re-verify object start arrays
2173 if (VerifyObjectStartArray &&
2174 VerifyAfterGC) {
2175 old_gen->verify_object_start_array();
2176 }
2177
2178 if (ZapUnusedHeapArea) {
2179 old_gen->object_space()->check_mangled_unused_area_complete();
2180 }
2181
2182 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2183
2184 collection_exit.update();
2185
2186 heap->print_heap_after_gc();
2187 if (PrintGCTaskTimeStamps) {
2188 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2189 INT64_FORMAT,
2190 marking_start.ticks(), compaction_start.ticks(),
2191 collection_exit.ticks());
2192 gc_task_manager()->print_task_time_stamps();
2193 }
2194
2195 heap->post_full_gc_dump();
2196
2197 #ifdef TRACESPINNING
2198 ParallelTaskTerminator::print_termination_counts();
2199 #endif
2200
2201 return true;
2202 }
2203
2204 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2205 PSYoungGen* young_gen,
2206 PSOldGen* old_gen) {
2207 MutableSpace* const eden_space = young_gen->eden_space();
2208 assert(!eden_space->is_empty(), "eden must be non-empty");
2209 assert(young_gen->virtual_space()->alignment() ==
2210 old_gen->virtual_space()->alignment(), "alignments do not match");
2211
2212 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2213 return false;
2214 }
2215
2216 // Both generations must be completely committed.
2217 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2218 return false;
2219 }
2220 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2221 return false;
2222 }
2223
2224 // Figure out how much to take from eden. Include the average amount promoted
2225 // in the total; otherwise the next young gen GC will simply bail out to a
2226 // full GC.
2227 const size_t alignment = old_gen->virtual_space()->alignment();
2228 const size_t eden_used = eden_space->used_in_bytes();
2229 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2230 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2231 const size_t eden_capacity = eden_space->capacity_in_bytes();
2232
2233 if (absorb_size >= eden_capacity) {
2234 return false; // Must leave some space in eden.
2235 }
2236
2237 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2238 if (new_young_size < young_gen->min_gen_size()) {
2239 return false; // Respect young gen minimum size.
2240 }
2241
2242 if (TraceAdaptiveGCBoundary && Verbose) {
2243 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2244 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2245 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2246 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2247 absorb_size / K,
2248 eden_capacity / K, (eden_capacity - absorb_size) / K,
2249 young_gen->from_space()->used_in_bytes() / K,
2250 young_gen->to_space()->used_in_bytes() / K,
2251 young_gen->capacity_in_bytes() / K, new_young_size / K);
2252 }
2253
2254 // Fill the unused part of the old gen.
2255 MutableSpace* const old_space = old_gen->object_space();
2256 HeapWord* const unused_start = old_space->top();
2257 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2258
2259 if (unused_words > 0) {
2260 if (unused_words < CollectedHeap::min_fill_size()) {
2261 return false; // If the old gen cannot be filled, must give up.
2262 }
2263 CollectedHeap::fill_with_objects(unused_start, unused_words);
2264 }
2265
2266 // Take the live data from eden and set both top and end in the old gen to
2267 // eden top. (Need to set end because reset_after_change() mangles the region
2268 // from end to virtual_space->high() in debug builds).
2269 HeapWord* const new_top = eden_space->top();
2270 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2271 absorb_size);
2272 young_gen->reset_after_change();
2273 old_space->set_top(new_top);
2274 old_space->set_end(new_top);
2275 old_gen->reset_after_change();
2276
2277 // Update the object start array for the filler object and the data from eden.
2278 ObjectStartArray* const start_array = old_gen->start_array();
2279 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2280 start_array->allocate_block(p);
2281 }
2282
2283 // Could update the promoted average here, but it is not typically updated at
2284 // full GCs and the value to use is unclear. Something like
2285 //
2286 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2287
2288 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2289 return true;
2290 }
2291
2292 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2293 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2294 "shouldn't return NULL");
2295 return ParallelScavengeHeap::gc_task_manager();
2296 }
2297
2298 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2299 bool maximum_heap_compaction) {
2300 // Recursively traverse all live objects and mark them
2301 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2302
2303 ParallelScavengeHeap* heap = gc_heap();
2304 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2305 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2306 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2307 ParallelTaskTerminator terminator(active_gc_threads, qset);
2308
2309 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2310 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2311
2312 // Need new claim bits before marking starts.
2313 ClassLoaderDataGraph::clear_claimed_marks();
2314
2315 {
2316 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2317 ParallelScavengeHeap::ParStrongRootsScope psrs;
2318
2319 GCTaskQueue* q = GCTaskQueue::create();
2320
2321 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2322 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2323 // We scan the thread roots in parallel
2324 Threads::create_thread_roots_marking_tasks(q);
2325 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2326 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2327 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2328 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2329 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2330 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2331
2332 if (active_gc_threads > 1) {
2333 for (uint j = 0; j < active_gc_threads; j++) {
2334 q->enqueue(new StealMarkingTask(&terminator));
2335 }
2336 }
2337
2338 gc_task_manager()->execute_and_wait(q);
2339 }
2340
2341 // Process reference objects found during marking
2342 {
2343 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2344 if (ref_processor()->processing_is_mt()) {
2345 RefProcTaskExecutor task_executor;
2346 ref_processor()->process_discovered_references(
2347 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2348 &task_executor);
2349 } else {
2350 ref_processor()->process_discovered_references(
2351 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2352 }
2353 }
2354
2355 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2356
2357 // This is the point where the entire marking should have completed.
2358 assert(cm->marking_stacks_empty(), "Marking should have completed");
2359
2360 // Follow system dictionary roots and unload classes.
2361 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2362
2363 // Unload nmethods.
2364 CodeCache::do_unloading(is_alive_closure(), purged_class);
2365
2366 // Prune dead klasses from subklass/sibling/implementor lists.
2367 Klass::clean_weak_klass_links(is_alive_closure());
2368
2369 // Delete entries for dead interned strings.
2370 StringTable::unlink(is_alive_closure());
2371
2372 // Clean up unreferenced symbols in symbol table.
2373 SymbolTable::unlink();
2374 }
2375
2376 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) {
2377 ClassLoaderData* cld = klass->class_loader_data();
2378 // The actual processing of the klass is done when we
2379 // traverse the list of Klasses in the class loader data.
2380 PSParallelCompact::follow_class_loader(cm, cld);
2381 }
2382
2383 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) {
2384 ClassLoaderData* cld = klass->class_loader_data();
2385 // The actual processing of the klass is done when we
2386 // traverse the list of Klasses in the class loader data.
2387 PSParallelCompact::adjust_class_loader(cm, cld);
2388 }
2389
2390 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2391 ClassLoaderData* cld) {
2392 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2393 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2394
2395 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2396 }
2397
2398 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm,
2399 ClassLoaderData* cld) {
2400 cld->oops_do(PSParallelCompact::adjust_pointer_closure(),
2401 PSParallelCompact::adjust_klass_closure(),
2402 true);
2403 }
2404
2405 // This should be moved to the shared markSweep code!
2406 class PSAlwaysTrueClosure: public BoolObjectClosure {
2407 public:
2408 bool do_object_b(oop p) { return true; }
2409 };
2410 static PSAlwaysTrueClosure always_true;
2411
2412 void PSParallelCompact::adjust_roots() {
2413 // Adjust the pointers to reflect the new locations
2414 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2415
2416 // Need new claim bits when tracing through and adjusting pointers.
2417 ClassLoaderDataGraph::clear_claimed_marks();
2418
2419 // General strong roots.
2420 Universe::oops_do(adjust_pointer_closure());
2421 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles
2422 CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2423 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2424 ObjectSynchronizer::oops_do(adjust_pointer_closure());
2425 FlatProfiler::oops_do(adjust_pointer_closure());
2426 Management::oops_do(adjust_pointer_closure());
2427 JvmtiExport::oops_do(adjust_pointer_closure());
2428 // SO_AllClasses
2429 SystemDictionary::oops_do(adjust_pointer_closure());
2430 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2431
2432 // Now adjust pointers in remaining weak roots. (All of which should
2433 // have been cleared if they pointed to non-surviving objects.)
2434 // Global (weak) JNI handles
2435 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2436
2437 CodeCache::oops_do(adjust_pointer_closure());
2438 StringTable::oops_do(adjust_pointer_closure());
2439 ref_processor()->weak_oops_do(adjust_pointer_closure());
2440 // Roots were visited so references into the young gen in roots
2441 // may have been scanned. Process them also.
2442 // Should the reference processor have a span that excludes
2443 // young gen objects?
2444 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2445 }
2446
2447 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2448 uint parallel_gc_threads)
2449 {
2450 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2451
2452 // Find the threads that are active
2453 unsigned int which = 0;
2454
2455 const uint task_count = MAX2(parallel_gc_threads, 1U);
2456 for (uint j = 0; j < task_count; j++) {
2457 q->enqueue(new DrainStacksCompactionTask(j));
2458 ParCompactionManager::verify_region_list_empty(j);
2459 // Set the region stacks variables to "no" region stack values
2460 // so that they will be recognized and needing a region stack
2461 // in the stealing tasks if they do not get one by executing
2462 // a draining stack.
2463 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2464 cm->set_region_stack(NULL);
2465 cm->set_region_stack_index((uint)max_uintx);
2466 }
2467 ParCompactionManager::reset_recycled_stack_index();
2468
2469 // Find all regions that are available (can be filled immediately) and
2470 // distribute them to the thread stacks. The iteration is done in reverse
2471 // order (high to low) so the regions will be removed in ascending order.
2472
2473 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2474
2475 size_t fillable_regions = 0; // A count for diagnostic purposes.
2476 // A region index which corresponds to the tasks created above.
2477 // "which" must be 0 <= which < task_count
2478
2479 which = 0;
2480 // id + 1 is used to test termination so unsigned can
2481 // be used with an old_space_id == 0.
2482 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2483 SpaceInfo* const space_info = _space_info + id;
2484 MutableSpace* const space = space_info->space();
2485 HeapWord* const new_top = space_info->new_top();
2486
2487 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2488 const size_t end_region =
2489 sd.addr_to_region_idx(sd.region_align_up(new_top));
2490
2491 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2492 if (sd.region(cur)->claim_unsafe()) {
2493 ParCompactionManager::region_list_push(which, cur);
2494
2495 if (TraceParallelOldGCCompactionPhase && Verbose) {
2496 const size_t count_mod_8 = fillable_regions & 7;
2497 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2498 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2499 if (count_mod_8 == 7) gclog_or_tty->cr();
2500 }
2501
2502 NOT_PRODUCT(++fillable_regions;)
2503
2504 // Assign regions to tasks in round-robin fashion.
2505 if (++which == task_count) {
2506 assert(which <= parallel_gc_threads,
2507 "Inconsistent number of workers");
2508 which = 0;
2509 }
2510 }
2511 }
2512 }
2513
2514 if (TraceParallelOldGCCompactionPhase) {
2515 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2516 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2517 }
2518 }
2519
2520 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2521
2522 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2523 uint parallel_gc_threads) {
2524 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2525
2526 ParallelCompactData& sd = PSParallelCompact::summary_data();
2527
2528 // Iterate over all the spaces adding tasks for updating
2529 // regions in the dense prefix. Assume that 1 gc thread
2530 // will work on opening the gaps and the remaining gc threads
2531 // will work on the dense prefix.
2532 unsigned int space_id;
2533 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2534 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2535 const MutableSpace* const space = _space_info[space_id].space();
2536
2537 if (dense_prefix_end == space->bottom()) {
2538 // There is no dense prefix for this space.
2539 continue;
2540 }
2541
2542 // The dense prefix is before this region.
2543 size_t region_index_end_dense_prefix =
2544 sd.addr_to_region_idx(dense_prefix_end);
2545 RegionData* const dense_prefix_cp =
2546 sd.region(region_index_end_dense_prefix);
2547 assert(dense_prefix_end == space->end() ||
2548 dense_prefix_cp->available() ||
2549 dense_prefix_cp->claimed(),
2550 "The region after the dense prefix should always be ready to fill");
2551
2552 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2553
2554 // Is there dense prefix work?
2555 size_t total_dense_prefix_regions =
2556 region_index_end_dense_prefix - region_index_start;
2557 // How many regions of the dense prefix should be given to
2558 // each thread?
2559 if (total_dense_prefix_regions > 0) {
2560 uint tasks_for_dense_prefix = 1;
2561 if (total_dense_prefix_regions <=
2562 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2563 // Don't over partition. This assumes that
2564 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2565 // so there are not many regions to process.
2566 tasks_for_dense_prefix = parallel_gc_threads;
2567 } else {
2568 // Over partition
2569 tasks_for_dense_prefix = parallel_gc_threads *
2570 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2571 }
2572 size_t regions_per_thread = total_dense_prefix_regions /
2573 tasks_for_dense_prefix;
2574 // Give each thread at least 1 region.
2575 if (regions_per_thread == 0) {
2576 regions_per_thread = 1;
2577 }
2578
2579 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2580 if (region_index_start >= region_index_end_dense_prefix) {
2581 break;
2582 }
2583 // region_index_end is not processed
2584 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2585 region_index_end_dense_prefix);
2586 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2587 region_index_start,
2588 region_index_end));
2589 region_index_start = region_index_end;
2590 }
2591 }
2592 // This gets any part of the dense prefix that did not
2593 // fit evenly.
2594 if (region_index_start < region_index_end_dense_prefix) {
2595 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2596 region_index_start,
2597 region_index_end_dense_prefix));
2598 }
2599 }
2600 }
2601
2602 void PSParallelCompact::enqueue_region_stealing_tasks(
2603 GCTaskQueue* q,
2604 ParallelTaskTerminator* terminator_ptr,
2605 uint parallel_gc_threads) {
2606 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2607
2608 // Once a thread has drained it's stack, it should try to steal regions from
2609 // other threads.
2610 if (parallel_gc_threads > 1) {
2611 for (uint j = 0; j < parallel_gc_threads; j++) {
2612 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2613 }
2614 }
2615 }
2616
2617 void PSParallelCompact::compact() {
2618 // trace("5");
2619 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2620
2621 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2622 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2623 PSOldGen* old_gen = heap->old_gen();
2624 old_gen->start_array()->reset();
2625 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2626 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2627 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2628 ParallelTaskTerminator terminator(active_gc_threads, qset);
2629
2630 GCTaskQueue* q = GCTaskQueue::create();
2631 enqueue_region_draining_tasks(q, active_gc_threads);
2632 enqueue_dense_prefix_tasks(q, active_gc_threads);
2633 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2634
2635 {
2636 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2637
2638 gc_task_manager()->execute_and_wait(q);
2639
2640 #ifdef ASSERT
2641 // Verify that all regions have been processed before the deferred updates.
2642 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2643 verify_complete(SpaceId(id));
2644 }
2645 #endif
2646 }
2647
2648 {
2649 // Update the deferred objects, if any. Any compaction manager can be used.
2650 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2651 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2652 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2653 update_deferred_objects(cm, SpaceId(id));
2654 }
2655 }
2656 }
2657
2658 #ifdef ASSERT
2659 void PSParallelCompact::verify_complete(SpaceId space_id) {
2660 // All Regions between space bottom() to new_top() should be marked as filled
2661 // and all Regions between new_top() and top() should be available (i.e.,
2662 // should have been emptied).
2663 ParallelCompactData& sd = summary_data();
2664 SpaceInfo si = _space_info[space_id];
2665 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2666 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2667 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2668 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2669 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2670
2671 bool issued_a_warning = false;
2672
2673 size_t cur_region;
2674 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2675 const RegionData* const c = sd.region(cur_region);
2676 if (!c->completed()) {
2677 warning("region " SIZE_FORMAT " not filled: "
2678 "destination_count=" SIZE_FORMAT,
2679 cur_region, c->destination_count());
2680 issued_a_warning = true;
2681 }
2682 }
2683
2684 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2685 const RegionData* const c = sd.region(cur_region);
2686 if (!c->available()) {
2687 warning("region " SIZE_FORMAT " not empty: "
2688 "destination_count=" SIZE_FORMAT,
2689 cur_region, c->destination_count());
2690 issued_a_warning = true;
2691 }
2692 }
2693
2694 if (issued_a_warning) {
2695 print_region_ranges();
2696 }
2697 }
2698 #endif // #ifdef ASSERT
2699
2700 // Update interior oops in the ranges of regions [beg_region, end_region).
2701 void
2702 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2703 SpaceId space_id,
2704 size_t beg_region,
2705 size_t end_region) {
2706 ParallelCompactData& sd = summary_data();
2707 ParMarkBitMap* const mbm = mark_bitmap();
2708
2709 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2710 HeapWord* const end_addr = sd.region_to_addr(end_region);
2711 assert(beg_region <= end_region, "bad region range");
2712 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2713
2714 #ifdef ASSERT
2715 // Claim the regions to avoid triggering an assert when they are marked as
2716 // filled.
2717 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2718 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2719 }
2720 #endif // #ifdef ASSERT
2721
2722 if (beg_addr != space(space_id)->bottom()) {
2723 // Find the first live object or block of dead space that *starts* in this
2724 // range of regions. If a partial object crosses onto the region, skip it;
2725 // it will be marked for 'deferred update' when the object head is
2726 // processed. If dead space crosses onto the region, it is also skipped; it
2727 // will be filled when the prior region is processed. If neither of those
2728 // apply, the first word in the region is the start of a live object or dead
2729 // space.
2730 assert(beg_addr > space(space_id)->bottom(), "sanity");
2731 const RegionData* const cp = sd.region(beg_region);
2732 if (cp->partial_obj_size() != 0) {
2733 beg_addr = sd.partial_obj_end(beg_region);
2734 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2735 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2736 }
2737 }
2738
2739 if (beg_addr < end_addr) {
2740 // A live object or block of dead space starts in this range of Regions.
2741 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2742
2743 // Create closures and iterate.
2744 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2745 FillClosure fill_closure(cm, space_id);
2746 ParMarkBitMap::IterationStatus status;
2747 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2748 dense_prefix_end);
2749 if (status == ParMarkBitMap::incomplete) {
2750 update_closure.do_addr(update_closure.source());
2751 }
2752 }
2753
2754 // Mark the regions as filled.
2755 RegionData* const beg_cp = sd.region(beg_region);
2756 RegionData* const end_cp = sd.region(end_region);
2757 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2758 cp->set_completed();
2759 }
2760 }
2761
2762 // Return the SpaceId for the space containing addr. If addr is not in the
2763 // heap, last_space_id is returned. In debug mode it expects the address to be
2764 // in the heap and asserts such.
2765 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2766 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2767
2768 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2769 if (_space_info[id].space()->contains(addr)) {
2770 return SpaceId(id);
2771 }
2772 }
2773
2774 assert(false, "no space contains the addr");
2775 return last_space_id;
2776 }
2777
2778 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2779 SpaceId id) {
2780 assert(id < last_space_id, "bad space id");
2781
2782 ParallelCompactData& sd = summary_data();
2783 const SpaceInfo* const space_info = _space_info + id;
2784 ObjectStartArray* const start_array = space_info->start_array();
2785
2786 const MutableSpace* const space = space_info->space();
2787 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2788 HeapWord* const beg_addr = space_info->dense_prefix();
2789 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2790
2791 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2792 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2793 const RegionData* cur_region;
2794 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2795 HeapWord* const addr = cur_region->deferred_obj_addr();
2796 if (addr != NULL) {
2797 if (start_array != NULL) {
2798 start_array->allocate_block(addr);
2799 }
2800 oop(addr)->update_contents(cm);
2801 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2802 }
2803 }
2804 }
2805
2806 // Skip over count live words starting from beg, and return the address of the
2807 // next live word. Unless marked, the word corresponding to beg is assumed to
2808 // be dead. Callers must either ensure beg does not correspond to the middle of
2809 // an object, or account for those live words in some other way. Callers must
2810 // also ensure that there are enough live words in the range [beg, end) to skip.
2811 HeapWord*
2812 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2813 {
2814 assert(count > 0, "sanity");
2815
2816 ParMarkBitMap* m = mark_bitmap();
2817 idx_t bits_to_skip = m->words_to_bits(count);
2818 idx_t cur_beg = m->addr_to_bit(beg);
2819 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2820
2821 do {
2822 cur_beg = m->find_obj_beg(cur_beg, search_end);
2823 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2824 const size_t obj_bits = cur_end - cur_beg + 1;
2825 if (obj_bits > bits_to_skip) {
2826 return m->bit_to_addr(cur_beg + bits_to_skip);
2827 }
2828 bits_to_skip -= obj_bits;
2829 cur_beg = cur_end + 1;
2830 } while (bits_to_skip > 0);
2831
2832 // Skipping the desired number of words landed just past the end of an object.
2833 // Find the start of the next object.
2834 cur_beg = m->find_obj_beg(cur_beg, search_end);
2835 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2836 return m->bit_to_addr(cur_beg);
2837 }
2838
2839 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2840 SpaceId src_space_id,
2841 size_t src_region_idx)
2842 {
2843 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2844
2845 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2846 if (split_info.dest_region_addr() == dest_addr) {
2847 // The partial object ending at the split point contains the first word to
2848 // be copied to dest_addr.
2849 return split_info.first_src_addr();
2850 }
2851
2852 const ParallelCompactData& sd = summary_data();
2853 ParMarkBitMap* const bitmap = mark_bitmap();
2854 const size_t RegionSize = ParallelCompactData::RegionSize;
2855
2856 assert(sd.is_region_aligned(dest_addr), "not aligned");
2857 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2858 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2859 HeapWord* const src_region_destination = src_region_ptr->destination();
2860
2861 assert(dest_addr >= src_region_destination, "wrong src region");
2862 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2863
2864 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2865 HeapWord* const src_region_end = src_region_beg + RegionSize;
2866
2867 HeapWord* addr = src_region_beg;
2868 if (dest_addr == src_region_destination) {
2869 // Return the first live word in the source region.
2870 if (partial_obj_size == 0) {
2871 addr = bitmap->find_obj_beg(addr, src_region_end);
2872 assert(addr < src_region_end, "no objects start in src region");
2873 }
2874 return addr;
2875 }
2876
2877 // Must skip some live data.
2878 size_t words_to_skip = dest_addr - src_region_destination;
2879 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2880
2881 if (partial_obj_size >= words_to_skip) {
2882 // All the live words to skip are part of the partial object.
2883 addr += words_to_skip;
2884 if (partial_obj_size == words_to_skip) {
2885 // Find the first live word past the partial object.
2886 addr = bitmap->find_obj_beg(addr, src_region_end);
2887 assert(addr < src_region_end, "wrong src region");
2888 }
2889 return addr;
2890 }
2891
2892 // Skip over the partial object (if any).
2893 if (partial_obj_size != 0) {
2894 words_to_skip -= partial_obj_size;
2895 addr += partial_obj_size;
2896 }
2897
2898 // Skip over live words due to objects that start in the region.
2899 addr = skip_live_words(addr, src_region_end, words_to_skip);
2900 assert(addr < src_region_end, "wrong src region");
2901 return addr;
2902 }
2903
2904 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2905 SpaceId src_space_id,
2906 size_t beg_region,
2907 HeapWord* end_addr)
2908 {
2909 ParallelCompactData& sd = summary_data();
2910
2911 #ifdef ASSERT
2912 MutableSpace* const src_space = _space_info[src_space_id].space();
2913 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2914 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2915 "src_space_id does not match beg_addr");
2916 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2917 "src_space_id does not match end_addr");
2918 #endif // #ifdef ASSERT
2919
2920 RegionData* const beg = sd.region(beg_region);
2921 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2922
2923 // Regions up to new_top() are enqueued if they become available.
2924 HeapWord* const new_top = _space_info[src_space_id].new_top();
2925 RegionData* const enqueue_end =
2926 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2927
2928 for (RegionData* cur = beg; cur < end; ++cur) {
2929 assert(cur->data_size() > 0, "region must have live data");
2930 cur->decrement_destination_count();
2931 if (cur < enqueue_end && cur->available() && cur->claim()) {
2932 cm->push_region(sd.region(cur));
2933 }
2934 }
2935 }
2936
2937 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2938 SpaceId& src_space_id,
2939 HeapWord*& src_space_top,
2940 HeapWord* end_addr)
2941 {
2942 typedef ParallelCompactData::RegionData RegionData;
2943
2944 ParallelCompactData& sd = PSParallelCompact::summary_data();
2945 const size_t region_size = ParallelCompactData::RegionSize;
2946
2947 size_t src_region_idx = 0;
2948
2949 // Skip empty regions (if any) up to the top of the space.
2950 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2951 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2952 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2953 const RegionData* const top_region_ptr =
2954 sd.addr_to_region_ptr(top_aligned_up);
2955 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2956 ++src_region_ptr;
2957 }
2958
2959 if (src_region_ptr < top_region_ptr) {
2960 // The next source region is in the current space. Update src_region_idx
2961 // and the source address to match src_region_ptr.
2962 src_region_idx = sd.region(src_region_ptr);
2963 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2964 if (src_region_addr > closure.source()) {
2965 closure.set_source(src_region_addr);
2966 }
2967 return src_region_idx;
2968 }
2969
2970 // Switch to a new source space and find the first non-empty region.
2971 unsigned int space_id = src_space_id + 1;
2972 assert(space_id < last_space_id, "not enough spaces");
2973
2974 HeapWord* const destination = closure.destination();
2975
2976 do {
2977 MutableSpace* space = _space_info[space_id].space();
2978 HeapWord* const bottom = space->bottom();
2979 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2980
2981 // Iterate over the spaces that do not compact into themselves.
2982 if (bottom_cp->destination() != bottom) {
2983 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2984 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2985
2986 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2987 if (src_cp->live_obj_size() > 0) {
2988 // Found it.
2989 assert(src_cp->destination() == destination,
2990 "first live obj in the space must match the destination");
2991 assert(src_cp->partial_obj_size() == 0,
2992 "a space cannot begin with a partial obj");
2993
2994 src_space_id = SpaceId(space_id);
2995 src_space_top = space->top();
2996 const size_t src_region_idx = sd.region(src_cp);
2997 closure.set_source(sd.region_to_addr(src_region_idx));
2998 return src_region_idx;
2999 } else {
3000 assert(src_cp->data_size() == 0, "sanity");
3001 }
3002 }
3003 }
3004 } while (++space_id < last_space_id);
3005
3006 assert(false, "no source region was found");
3007 return 0;
3008 }
3009
3010 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3011 {
3012 typedef ParMarkBitMap::IterationStatus IterationStatus;
3013 const size_t RegionSize = ParallelCompactData::RegionSize;
3014 ParMarkBitMap* const bitmap = mark_bitmap();
3015 ParallelCompactData& sd = summary_data();
3016 RegionData* const region_ptr = sd.region(region_idx);
3017
3018 // Get the items needed to construct the closure.
3019 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3020 SpaceId dest_space_id = space_id(dest_addr);
3021 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3022 HeapWord* new_top = _space_info[dest_space_id].new_top();
3023 assert(dest_addr < new_top, "sanity");
3024 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3025
3026 // Get the source region and related info.
3027 size_t src_region_idx = region_ptr->source_region();
3028 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3029 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3030
3031 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3032 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3033
3034 // Adjust src_region_idx to prepare for decrementing destination counts (the
3035 // destination count is not decremented when a region is copied to itself).
3036 if (src_region_idx == region_idx) {
3037 src_region_idx += 1;
3038 }
3039
3040 if (bitmap->is_unmarked(closure.source())) {
3041 // The first source word is in the middle of an object; copy the remainder
3042 // of the object or as much as will fit. The fact that pointer updates were
3043 // deferred will be noted when the object header is processed.
3044 HeapWord* const old_src_addr = closure.source();
3045 closure.copy_partial_obj();
3046 if (closure.is_full()) {
3047 decrement_destination_counts(cm, src_space_id, src_region_idx,
3048 closure.source());
3049 region_ptr->set_deferred_obj_addr(NULL);
3050 region_ptr->set_completed();
3051 return;
3052 }
3053
3054 HeapWord* const end_addr = sd.region_align_down(closure.source());
3055 if (sd.region_align_down(old_src_addr) != end_addr) {
3056 // The partial object was copied from more than one source region.
3057 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3058
3059 // Move to the next source region, possibly switching spaces as well. All
3060 // args except end_addr may be modified.
3061 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3062 end_addr);
3063 }
3064 }
3065
3066 do {
3067 HeapWord* const cur_addr = closure.source();
3068 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3069 src_space_top);
3070 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3071
3072 if (status == ParMarkBitMap::incomplete) {
3073 // The last obj that starts in the source region does not end in the
3074 // region.
3075 assert(closure.source() < end_addr, "sanity");
3076 HeapWord* const obj_beg = closure.source();
3077 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3078 src_space_top);
3079 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3080 if (obj_end < range_end) {
3081 // The end was found; the entire object will fit.
3082 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3083 assert(status != ParMarkBitMap::would_overflow, "sanity");
3084 } else {
3085 // The end was not found; the object will not fit.
3086 assert(range_end < src_space_top, "obj cannot cross space boundary");
3087 status = ParMarkBitMap::would_overflow;
3088 }
3089 }
3090
3091 if (status == ParMarkBitMap::would_overflow) {
3092 // The last object did not fit. Note that interior oop updates were
3093 // deferred, then copy enough of the object to fill the region.
3094 region_ptr->set_deferred_obj_addr(closure.destination());
3095 status = closure.copy_until_full(); // copies from closure.source()
3096
3097 decrement_destination_counts(cm, src_space_id, src_region_idx,
3098 closure.source());
3099 region_ptr->set_completed();
3100 return;
3101 }
3102
3103 if (status == ParMarkBitMap::full) {
3104 decrement_destination_counts(cm, src_space_id, src_region_idx,
3105 closure.source());
3106 region_ptr->set_deferred_obj_addr(NULL);
3107 region_ptr->set_completed();
3108 return;
3109 }
3110
3111 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3112
3113 // Move to the next source region, possibly switching spaces as well. All
3114 // args except end_addr may be modified.
3115 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3116 end_addr);
3117 } while (true);
3118 }
3119
3120 void
3121 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3122 const MutableSpace* sp = space(space_id);
3123 if (sp->is_empty()) {
3124 return;
3125 }
3126
3127 ParallelCompactData& sd = PSParallelCompact::summary_data();
3128 ParMarkBitMap* const bitmap = mark_bitmap();
3129 HeapWord* const dp_addr = dense_prefix(space_id);
3130 HeapWord* beg_addr = sp->bottom();
3131 HeapWord* end_addr = sp->top();
3132
3133 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3134
3135 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3136 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3137 if (beg_region < dp_region) {
3138 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3139 }
3140
3141 // The destination of the first live object that starts in the region is one
3142 // past the end of the partial object entering the region (if any).
3143 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3144 HeapWord* const new_top = _space_info[space_id].new_top();
3145 assert(new_top >= dest_addr, "bad new_top value");
3146 const size_t words = pointer_delta(new_top, dest_addr);
3147
3148 if (words > 0) {
3149 ObjectStartArray* start_array = _space_info[space_id].start_array();
3150 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3151
3152 ParMarkBitMap::IterationStatus status;
3153 status = bitmap->iterate(&closure, dest_addr, end_addr);
3154 assert(status == ParMarkBitMap::full, "iteration not complete");
3155 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3156 "live objects skipped because closure is full");
3157 }
3158 }
3159
3160 jlong PSParallelCompact::millis_since_last_gc() {
3161 // We need a monotonically non-deccreasing time in ms but
3162 // os::javaTimeMillis() does not guarantee monotonicity.
3163 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3164 jlong ret_val = now - _time_of_last_gc;
3165 // XXX See note in genCollectedHeap::millis_since_last_gc().
3166 if (ret_val < 0) {
3167 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3168 return 0;
3169 }
3170 return ret_val;
3171 }
3172
3173 void PSParallelCompact::reset_millis_since_last_gc() {
3174 // We need a monotonically non-deccreasing time in ms but
3175 // os::javaTimeMillis() does not guarantee monotonicity.
3176 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3177 }
3178
3179 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3180 {
3181 if (source() != destination()) {
3182 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3183 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3184 }
3185 update_state(words_remaining());
3186 assert(is_full(), "sanity");
3187 return ParMarkBitMap::full;
3188 }
3189
3190 void MoveAndUpdateClosure::copy_partial_obj()
3191 {
3192 size_t words = words_remaining();
3193
3194 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3195 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3196 if (end_addr < range_end) {
3197 words = bitmap()->obj_size(source(), end_addr);
3198 }
3199
3200 // This test is necessary; if omitted, the pointer updates to a partial object
3201 // that crosses the dense prefix boundary could be overwritten.
3202 if (source() != destination()) {
3203 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3204 Copy::aligned_conjoint_words(source(), destination(), words);
3205 }
3206 update_state(words);
3207 }
3208
3209 ParMarkBitMapClosure::IterationStatus
3210 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3211 assert(destination() != NULL, "sanity");
3212 assert(bitmap()->obj_size(addr) == words, "bad size");
3213
3214 _source = addr;
3215 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3216 destination(), "wrong destination");
3217
3218 if (words > words_remaining()) {
3219 return ParMarkBitMap::would_overflow;
3220 }
3221
3222 // The start_array must be updated even if the object is not moving.
3223 if (_start_array != NULL) {
3224 _start_array->allocate_block(destination());
3225 }
3226
3227 if (destination() != source()) {
3228 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3229 Copy::aligned_conjoint_words(source(), destination(), words);
3230 }
3231
3232 oop moved_oop = (oop) destination();
3233 moved_oop->update_contents(compaction_manager());
3234 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3235
3236 update_state(words);
3237 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3238 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3239 }
3240
3241 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3242 ParCompactionManager* cm,
3243 PSParallelCompact::SpaceId space_id) :
3244 ParMarkBitMapClosure(mbm, cm),
3245 _space_id(space_id),
3246 _start_array(PSParallelCompact::start_array(space_id))
3247 {
3248 }
3249
3250 // Updates the references in the object to their new values.
3251 ParMarkBitMapClosure::IterationStatus
3252 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3253 do_addr(addr);
3254 return ParMarkBitMap::incomplete;
3255 }