1 /*
2 * Copyright (c) 2005, 2010, 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/psPermGen.hpp"
40 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
41 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
42 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
43 #include "gc_implementation/shared/isGCActiveMark.hpp"
44 #include "gc_interface/gcCause.hpp"
45 #include "memory/gcLocker.inline.hpp"
46 #include "memory/referencePolicy.hpp"
47 #include "memory/referenceProcessor.hpp"
48 #include "oops/methodDataOop.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "oops/oop.pcgc.inline.hpp"
51 #include "runtime/fprofiler.hpp"
52 #include "runtime/safepoint.hpp"
53 #include "runtime/vmThread.hpp"
54 #include "services/management.hpp"
55 #include "services/memoryService.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 klassOop 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 #ifdef VALIDATE_MARK_SWEEP
103 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL;
104 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL;
105 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL;
106 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL;
107 size_t PSParallelCompact::_live_oops_index = 0;
108 size_t PSParallelCompact::_live_oops_index_at_perm = 0;
109 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL;
110 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL;
111 bool PSParallelCompact::_pointer_tracking = false;
112 bool PSParallelCompact::_root_tracking = true;
113
114 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
115 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
116 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL;
117 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
118 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
119 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL;
120 #endif
121
122 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
123 HeapWord* destination)
124 {
125 assert(src_region_idx != 0, "invalid src_region_idx");
126 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
127 assert(destination != NULL, "invalid destination argument");
128
129 _src_region_idx = src_region_idx;
130 _partial_obj_size = partial_obj_size;
131 _destination = destination;
132
133 // These fields may not be updated below, so make sure they're clear.
134 assert(_dest_region_addr == NULL, "should have been cleared");
135 assert(_first_src_addr == NULL, "should have been cleared");
136
137 // Determine the number of destination regions for the partial object.
138 HeapWord* const last_word = destination + partial_obj_size - 1;
139 const ParallelCompactData& sd = PSParallelCompact::summary_data();
140 HeapWord* const beg_region_addr = sd.region_align_down(destination);
141 HeapWord* const end_region_addr = sd.region_align_down(last_word);
142
143 if (beg_region_addr == end_region_addr) {
144 // One destination region.
145 _destination_count = 1;
146 if (end_region_addr == destination) {
147 // The destination falls on a region boundary, thus the first word of the
148 // partial object will be the first word copied to the destination region.
149 _dest_region_addr = end_region_addr;
150 _first_src_addr = sd.region_to_addr(src_region_idx);
151 }
152 } else {
153 // Two destination regions. When copied, the partial object will cross a
154 // destination region boundary, so a word somewhere within the partial
155 // object will be the first word copied to the second destination region.
156 _destination_count = 2;
157 _dest_region_addr = end_region_addr;
158 const size_t ofs = pointer_delta(end_region_addr, destination);
159 assert(ofs < _partial_obj_size, "sanity");
160 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
161 }
162 }
163
164 void SplitInfo::clear()
165 {
166 _src_region_idx = 0;
167 _partial_obj_size = 0;
168 _destination = NULL;
169 _destination_count = 0;
170 _dest_region_addr = NULL;
171 _first_src_addr = NULL;
172 assert(!is_valid(), "sanity");
173 }
174
175 #ifdef ASSERT
176 void SplitInfo::verify_clear()
177 {
178 assert(_src_region_idx == 0, "not clear");
179 assert(_partial_obj_size == 0, "not clear");
180 assert(_destination == NULL, "not clear");
181 assert(_destination_count == 0, "not clear");
182 assert(_dest_region_addr == NULL, "not clear");
183 assert(_first_src_addr == NULL, "not clear");
184 }
185 #endif // #ifdef ASSERT
186
187
188 #ifndef PRODUCT
189 const char* PSParallelCompact::space_names[] = {
190 "perm", "old ", "eden", "from", "to "
191 };
192
193 void PSParallelCompact::print_region_ranges()
194 {
195 tty->print_cr("space bottom top end new_top");
196 tty->print_cr("------ ---------- ---------- ---------- ----------");
197
198 for (unsigned int id = 0; id < last_space_id; ++id) {
199 const MutableSpace* space = _space_info[id].space();
200 tty->print_cr("%u %s "
201 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
202 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
203 id, space_names[id],
204 summary_data().addr_to_region_idx(space->bottom()),
205 summary_data().addr_to_region_idx(space->top()),
206 summary_data().addr_to_region_idx(space->end()),
207 summary_data().addr_to_region_idx(_space_info[id].new_top()));
208 }
209 }
210
211 void
212 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
213 {
214 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
215 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
216
217 ParallelCompactData& sd = PSParallelCompact::summary_data();
218 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
219 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
220 REGION_IDX_FORMAT " " PTR_FORMAT " "
221 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
222 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
223 i, c->data_location(), dci, c->destination(),
224 c->partial_obj_size(), c->live_obj_size(),
225 c->data_size(), c->source_region(), c->destination_count());
226
227 #undef REGION_IDX_FORMAT
228 #undef REGION_DATA_FORMAT
229 }
230
231 void
232 print_generic_summary_data(ParallelCompactData& summary_data,
233 HeapWord* const beg_addr,
234 HeapWord* const end_addr)
235 {
236 size_t total_words = 0;
237 size_t i = summary_data.addr_to_region_idx(beg_addr);
238 const size_t last = summary_data.addr_to_region_idx(end_addr);
239 HeapWord* pdest = 0;
240
241 while (i <= last) {
242 ParallelCompactData::RegionData* c = summary_data.region(i);
243 if (c->data_size() != 0 || c->destination() != pdest) {
244 print_generic_summary_region(i, c);
245 total_words += c->data_size();
246 pdest = c->destination();
247 }
248 ++i;
249 }
250
251 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
252 }
253
254 void
255 print_generic_summary_data(ParallelCompactData& summary_data,
256 SpaceInfo* space_info)
257 {
258 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
259 const MutableSpace* space = space_info[id].space();
260 print_generic_summary_data(summary_data, space->bottom(),
261 MAX2(space->top(), space_info[id].new_top()));
262 }
263 }
264
265 void
266 print_initial_summary_region(size_t i,
267 const ParallelCompactData::RegionData* c,
268 bool newline = true)
269 {
270 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
271 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
272 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
273 i, c->destination(),
274 c->partial_obj_size(), c->live_obj_size(),
275 c->data_size(), c->source_region(), c->destination_count());
276 if (newline) tty->cr();
277 }
278
279 void
280 print_initial_summary_data(ParallelCompactData& summary_data,
281 const MutableSpace* space) {
282 if (space->top() == space->bottom()) {
283 return;
284 }
285
286 const size_t region_size = ParallelCompactData::RegionSize;
287 typedef ParallelCompactData::RegionData RegionData;
288 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
289 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
290 const RegionData* c = summary_data.region(end_region - 1);
291 HeapWord* end_addr = c->destination() + c->data_size();
292 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
293
294 // Print (and count) the full regions at the beginning of the space.
295 size_t full_region_count = 0;
296 size_t i = summary_data.addr_to_region_idx(space->bottom());
297 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
298 print_initial_summary_region(i, summary_data.region(i));
299 ++full_region_count;
300 ++i;
301 }
302
303 size_t live_to_right = live_in_space - full_region_count * region_size;
304
305 double max_reclaimed_ratio = 0.0;
306 size_t max_reclaimed_ratio_region = 0;
307 size_t max_dead_to_right = 0;
308 size_t max_live_to_right = 0;
309
310 // Print the 'reclaimed ratio' for regions while there is something live in
311 // the region or to the right of it. The remaining regions are empty (and
312 // uninteresting), and computing the ratio will result in division by 0.
313 while (i < end_region && live_to_right > 0) {
314 c = summary_data.region(i);
315 HeapWord* const region_addr = summary_data.region_to_addr(i);
316 const size_t used_to_right = pointer_delta(space->top(), region_addr);
317 const size_t dead_to_right = used_to_right - live_to_right;
318 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
319
320 if (reclaimed_ratio > max_reclaimed_ratio) {
321 max_reclaimed_ratio = reclaimed_ratio;
322 max_reclaimed_ratio_region = i;
323 max_dead_to_right = dead_to_right;
324 max_live_to_right = live_to_right;
325 }
326
327 print_initial_summary_region(i, c, false);
328 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
329 reclaimed_ratio, dead_to_right, live_to_right);
330
331 live_to_right -= c->data_size();
332 ++i;
333 }
334
335 // Any remaining regions are empty. Print one more if there is one.
336 if (i < end_region) {
337 print_initial_summary_region(i, summary_data.region(i));
338 }
339
340 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
341 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
342 max_reclaimed_ratio_region, max_dead_to_right,
343 max_live_to_right, max_reclaimed_ratio);
344 }
345
346 void
347 print_initial_summary_data(ParallelCompactData& summary_data,
348 SpaceInfo* space_info) {
349 unsigned int id = PSParallelCompact::perm_space_id;
350 const MutableSpace* space;
351 do {
352 space = space_info[id].space();
353 print_initial_summary_data(summary_data, space);
354 } while (++id < PSParallelCompact::eden_space_id);
355
356 do {
357 space = space_info[id].space();
358 print_generic_summary_data(summary_data, space->bottom(), space->top());
359 } while (++id < PSParallelCompact::last_space_id);
360 }
361 #endif // #ifndef PRODUCT
362
363 #ifdef ASSERT
364 size_t add_obj_count;
365 size_t add_obj_size;
366 size_t mark_bitmap_count;
367 size_t mark_bitmap_size;
368 #endif // #ifdef ASSERT
369
370 ParallelCompactData::ParallelCompactData()
371 {
372 _region_start = 0;
373
374 _region_vspace = 0;
375 _region_data = 0;
376 _region_count = 0;
377 }
378
379 bool ParallelCompactData::initialize(MemRegion covered_region)
380 {
381 _region_start = covered_region.start();
382 const size_t region_size = covered_region.word_size();
383 DEBUG_ONLY(_region_end = _region_start + region_size;)
384
385 assert(region_align_down(_region_start) == _region_start,
386 "region start not aligned");
387 assert((region_size & RegionSizeOffsetMask) == 0,
388 "region size not a multiple of RegionSize");
389
390 bool result = initialize_region_data(region_size);
391
392 return result;
393 }
394
395 PSVirtualSpace*
396 ParallelCompactData::create_vspace(size_t count, size_t element_size)
397 {
398 const size_t raw_bytes = count * element_size;
399 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
400 const size_t granularity = os::vm_allocation_granularity();
401 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
402
403 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
404 MAX2(page_sz, granularity);
405 ReservedSpace rs(bytes, rs_align, rs_align > 0);
406 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
407 rs.size());
408 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
409 if (vspace != 0) {
410 if (vspace->expand_by(bytes)) {
411 return vspace;
412 }
413 delete vspace;
414 // Release memory reserved in the space.
415 rs.release();
416 }
417
418 return 0;
419 }
420
421 bool ParallelCompactData::initialize_region_data(size_t region_size)
422 {
423 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
424 _region_vspace = create_vspace(count, sizeof(RegionData));
425 if (_region_vspace != 0) {
426 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
427 _region_count = count;
428 return true;
429 }
430 return false;
431 }
432
433 void ParallelCompactData::clear()
434 {
435 memset(_region_data, 0, _region_vspace->committed_size());
436 }
437
438 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
439 assert(beg_region <= _region_count, "beg_region out of range");
440 assert(end_region <= _region_count, "end_region out of range");
441
442 const size_t region_cnt = end_region - beg_region;
443 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
444 }
445
446 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
447 {
448 const RegionData* cur_cp = region(region_idx);
449 const RegionData* const end_cp = region(region_count() - 1);
450
451 HeapWord* result = region_to_addr(region_idx);
452 if (cur_cp < end_cp) {
453 do {
454 result += cur_cp->partial_obj_size();
455 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
456 }
457 return result;
458 }
459
460 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
461 {
462 const size_t obj_ofs = pointer_delta(addr, _region_start);
463 const size_t beg_region = obj_ofs >> Log2RegionSize;
464 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
465
466 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
467 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
468
469 if (beg_region == end_region) {
470 // All in one region.
471 _region_data[beg_region].add_live_obj(len);
472 return;
473 }
474
475 // First region.
476 const size_t beg_ofs = region_offset(addr);
477 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
478
479 klassOop klass = ((oop)addr)->klass();
480 // Middle regions--completely spanned by this object.
481 for (size_t region = beg_region + 1; region < end_region; ++region) {
482 _region_data[region].set_partial_obj_size(RegionSize);
483 _region_data[region].set_partial_obj_addr(addr);
484 }
485
486 // Last region.
487 const size_t end_ofs = region_offset(addr + len - 1);
488 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
489 _region_data[end_region].set_partial_obj_addr(addr);
490 }
491
492 void
493 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
494 {
495 assert(region_offset(beg) == 0, "not RegionSize aligned");
496 assert(region_offset(end) == 0, "not RegionSize aligned");
497
498 size_t cur_region = addr_to_region_idx(beg);
499 const size_t end_region = addr_to_region_idx(end);
500 HeapWord* addr = beg;
501 while (cur_region < end_region) {
502 _region_data[cur_region].set_destination(addr);
503 _region_data[cur_region].set_destination_count(0);
504 _region_data[cur_region].set_source_region(cur_region);
505 _region_data[cur_region].set_data_location(addr);
506
507 // Update live_obj_size so the region appears completely full.
508 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
509 _region_data[cur_region].set_live_obj_size(live_size);
510
511 ++cur_region;
512 addr += RegionSize;
513 }
514 }
515
516 // Find the point at which a space can be split and, if necessary, record the
517 // split point.
518 //
519 // If the current src region (which overflowed the destination space) doesn't
520 // have a partial object, the split point is at the beginning of the current src
521 // region (an "easy" split, no extra bookkeeping required).
522 //
523 // If the current src region has a partial object, the split point is in the
524 // region where that partial object starts (call it the split_region). If
525 // split_region has a partial object, then the split point is just after that
526 // partial object (a "hard" split where we have to record the split data and
527 // zero the partial_obj_size field). With a "hard" split, we know that the
528 // partial_obj ends within split_region because the partial object that caused
529 // the overflow starts in split_region. If split_region doesn't have a partial
530 // obj, then the split is at the beginning of split_region (another "easy"
531 // split).
532 HeapWord*
533 ParallelCompactData::summarize_split_space(size_t src_region,
534 SplitInfo& split_info,
535 HeapWord* destination,
536 HeapWord* target_end,
537 HeapWord** target_next)
538 {
539 assert(destination <= target_end, "sanity");
540 assert(destination + _region_data[src_region].data_size() > target_end,
541 "region should not fit into target space");
542 assert(is_region_aligned(target_end), "sanity");
543
544 size_t split_region = src_region;
545 HeapWord* split_destination = destination;
546 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
547
548 if (destination + partial_obj_size > target_end) {
549 // The split point is just after the partial object (if any) in the
550 // src_region that contains the start of the object that overflowed the
551 // destination space.
552 //
553 // Find the start of the "overflow" object and set split_region to the
554 // region containing it.
555 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
556 split_region = addr_to_region_idx(overflow_obj);
557
558 // Clear the source_region field of all destination regions whose first word
559 // came from data after the split point (a non-null source_region field
560 // implies a region must be filled).
561 //
562 // An alternative to the simple loop below: clear during post_compact(),
563 // which uses memcpy instead of individual stores, and is easy to
564 // parallelize. (The downside is that it clears the entire RegionData
565 // object as opposed to just one field.)
566 //
567 // post_compact() would have to clear the summary data up to the highest
568 // address that was written during the summary phase, which would be
569 //
570 // max(top, max(new_top, clear_top))
571 //
572 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
573 // to target_end.
574 const RegionData* const sr = region(split_region);
575 const size_t beg_idx =
576 addr_to_region_idx(region_align_up(sr->destination() +
577 sr->partial_obj_size()));
578 const size_t end_idx = addr_to_region_idx(target_end);
579
580 if (TraceParallelOldGCSummaryPhase) {
581 gclog_or_tty->print_cr("split: clearing source_region field in ["
582 SIZE_FORMAT ", " SIZE_FORMAT ")",
583 beg_idx, end_idx);
584 }
585 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
586 _region_data[idx].set_source_region(0);
587 }
588
589 // Set split_destination and partial_obj_size to reflect the split region.
590 split_destination = sr->destination();
591 partial_obj_size = sr->partial_obj_size();
592 }
593
594 // The split is recorded only if a partial object extends onto the region.
595 if (partial_obj_size != 0) {
596 _region_data[split_region].set_partial_obj_size(0);
597 split_info.record(split_region, partial_obj_size, split_destination);
598 }
599
600 // Setup the continuation addresses.
601 *target_next = split_destination + partial_obj_size;
602 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
603
604 if (TraceParallelOldGCSummaryPhase) {
605 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
606 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
607 " pos=" SIZE_FORMAT,
608 split_type, source_next, split_region,
609 partial_obj_size);
610 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
611 " tn=" PTR_FORMAT,
612 split_type, split_destination,
613 addr_to_region_idx(split_destination),
614 *target_next);
615
616 if (partial_obj_size != 0) {
617 HeapWord* const po_beg = split_info.destination();
618 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
619 gclog_or_tty->print_cr("%s split: "
620 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
621 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
622 split_type,
623 po_beg, addr_to_region_idx(po_beg),
624 po_end, addr_to_region_idx(po_end));
625 }
626 }
627
628 return source_next;
629 }
630
631 bool ParallelCompactData::summarize(SplitInfo& split_info,
632 HeapWord* source_beg, HeapWord* source_end,
633 HeapWord** source_next,
634 HeapWord* target_beg, HeapWord* target_end,
635 HeapWord** target_next)
636 {
637 if (TraceParallelOldGCSummaryPhase) {
638 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
639 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
640 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
641 source_beg, source_end, source_next_val,
642 target_beg, target_end, *target_next);
643 }
644
645 size_t cur_region = addr_to_region_idx(source_beg);
646 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
647
648 HeapWord *dest_addr = target_beg;
649 while (cur_region < end_region) {
650 // The destination must be set even if the region has no data.
651 _region_data[cur_region].set_destination(dest_addr);
652
653 size_t words = _region_data[cur_region].data_size();
654 if (words > 0) {
655 // If cur_region does not fit entirely into the target space, find a point
656 // at which the source space can be 'split' so that part is copied to the
657 // target space and the rest is copied elsewhere.
658 if (dest_addr + words > target_end) {
659 assert(source_next != NULL, "source_next is NULL when splitting");
660 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
661 target_end, target_next);
662 return false;
663 }
664
665 // Compute the destination_count for cur_region, and if necessary, update
666 // source_region for a destination region. The source_region field is
667 // updated if cur_region is the first (left-most) region to be copied to a
668 // destination region.
669 //
670 // The destination_count calculation is a bit subtle. A region that has
671 // data that compacts into itself does not count itself as a destination.
672 // This maintains the invariant that a zero count means the region is
673 // available and can be claimed and then filled.
674 uint destination_count = 0;
675 if (split_info.is_split(cur_region)) {
676 // The current region has been split: the partial object will be copied
677 // to one destination space and the remaining data will be copied to
678 // another destination space. Adjust the initial destination_count and,
679 // if necessary, set the source_region field if the partial object will
680 // cross a destination region boundary.
681 destination_count = split_info.destination_count();
682 if (destination_count == 2) {
683 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
684 _region_data[dest_idx].set_source_region(cur_region);
685 }
686 }
687
688 HeapWord* const last_addr = dest_addr + words - 1;
689 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
690 const size_t dest_region_2 = addr_to_region_idx(last_addr);
691
692 // Initially assume that the destination regions will be the same and
693 // adjust the value below if necessary. Under this assumption, if
694 // cur_region == dest_region_2, then cur_region will be compacted
695 // completely into itself.
696 destination_count += cur_region == dest_region_2 ? 0 : 1;
697 if (dest_region_1 != dest_region_2) {
698 // Destination regions differ; adjust destination_count.
699 destination_count += 1;
700 // Data from cur_region will be copied to the start of dest_region_2.
701 _region_data[dest_region_2].set_source_region(cur_region);
702 } else if (region_offset(dest_addr) == 0) {
703 // Data from cur_region will be copied to the start of the destination
704 // region.
705 _region_data[dest_region_1].set_source_region(cur_region);
706 }
707
708 _region_data[cur_region].set_destination_count(destination_count);
709 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
710 dest_addr += words;
711 }
712
713 ++cur_region;
714 }
715
716 *target_next = dest_addr;
717 return true;
718 }
719
720 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
721 assert(addr != NULL, "Should detect NULL oop earlier");
722 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
723 #ifdef ASSERT
724 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
725 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
726 }
727 #endif
728 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
729
730 // Region covering the object.
731 size_t region_index = addr_to_region_idx(addr);
732 const RegionData* const region_ptr = region(region_index);
733 HeapWord* const region_addr = region_align_down(addr);
734
735 assert(addr < region_addr + RegionSize, "Region does not cover object");
736 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
737
738 HeapWord* result = region_ptr->destination();
739
740 // If all the data in the region is live, then the new location of the object
741 // can be calculated from the destination of the region plus the offset of the
742 // object in the region.
743 if (region_ptr->data_size() == RegionSize) {
744 result += pointer_delta(addr, region_addr);
745 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
746 return result;
747 }
748
749 // The new location of the object is
750 // region destination +
751 // size of the partial object extending onto the region +
752 // sizes of the live objects in the Region that are to the left of addr
753 const size_t partial_obj_size = region_ptr->partial_obj_size();
754 HeapWord* const search_start = region_addr + partial_obj_size;
755
756 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
757 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
758
759 result += partial_obj_size + live_to_left;
760 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
761 return result;
762 }
763
764 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
765 klassOop updated_klass;
766 if (PSParallelCompact::should_update_klass(old_klass)) {
767 updated_klass = (klassOop) calc_new_pointer(old_klass);
768 } else {
769 updated_klass = old_klass;
770 }
771
772 return updated_klass;
773 }
774
775 #ifdef ASSERT
776 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
777 {
778 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
779 const size_t* const end = (const size_t*)vspace->committed_high_addr();
780 for (const size_t* p = beg; p < end; ++p) {
781 assert(*p == 0, "not zero");
782 }
783 }
784
785 void ParallelCompactData::verify_clear()
786 {
787 verify_clear(_region_vspace);
788 }
789 #endif // #ifdef ASSERT
790
791 #ifdef NOT_PRODUCT
792 ParallelCompactData::RegionData* debug_region(size_t region_index) {
793 ParallelCompactData& sd = PSParallelCompact::summary_data();
794 return sd.region(region_index);
795 }
796 #endif
797
798 elapsedTimer PSParallelCompact::_accumulated_time;
799 unsigned int PSParallelCompact::_total_invocations = 0;
800 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
801 jlong PSParallelCompact::_time_of_last_gc = 0;
802 CollectorCounters* PSParallelCompact::_counters = NULL;
803 ParMarkBitMap PSParallelCompact::_mark_bitmap;
804 ParallelCompactData PSParallelCompact::_summary_data;
805
806 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
807
808 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
809 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
810
811 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
812 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
813
814 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
815 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
816
817 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); }
818 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
819
820 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
821
822 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); }
823 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
824
825 void PSParallelCompact::post_initialize() {
826 ParallelScavengeHeap* heap = gc_heap();
827 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
828
829 MemRegion mr = heap->reserved_region();
830 _ref_processor = ReferenceProcessor::create_ref_processor(
831 mr, // span
832 true, // atomic_discovery
833 true, // mt_discovery
834 &_is_alive_closure,
835 ParallelGCThreads,
836 ParallelRefProcEnabled);
837 _counters = new CollectorCounters("PSParallelCompact", 1);
838
839 // Initialize static fields in ParCompactionManager.
840 ParCompactionManager::initialize(mark_bitmap());
841 }
842
843 bool PSParallelCompact::initialize() {
844 ParallelScavengeHeap* heap = gc_heap();
845 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
846 MemRegion mr = heap->reserved_region();
847
848 // Was the old gen get allocated successfully?
849 if (!heap->old_gen()->is_allocated()) {
850 return false;
851 }
852
853 initialize_space_info();
854 initialize_dead_wood_limiter();
855
856 if (!_mark_bitmap.initialize(mr)) {
857 vm_shutdown_during_initialization("Unable to allocate bit map for "
858 "parallel garbage collection for the requested heap size.");
859 return false;
860 }
861
862 if (!_summary_data.initialize(mr)) {
863 vm_shutdown_during_initialization("Unable to allocate tables for "
864 "parallel garbage collection for the requested heap size.");
865 return false;
866 }
867
868 return true;
869 }
870
871 void PSParallelCompact::initialize_space_info()
872 {
873 memset(&_space_info, 0, sizeof(_space_info));
874
875 ParallelScavengeHeap* heap = gc_heap();
876 PSYoungGen* young_gen = heap->young_gen();
877 MutableSpace* perm_space = heap->perm_gen()->object_space();
878
879 _space_info[perm_space_id].set_space(perm_space);
880 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
881 _space_info[eden_space_id].set_space(young_gen->eden_space());
882 _space_info[from_space_id].set_space(young_gen->from_space());
883 _space_info[to_space_id].set_space(young_gen->to_space());
884
885 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
886 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
887
888 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
889 if (TraceParallelOldGCDensePrefix) {
890 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
891 _space_info[perm_space_id].min_dense_prefix());
892 }
893 }
894
895 void PSParallelCompact::initialize_dead_wood_limiter()
896 {
897 const size_t max = 100;
898 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
899 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
900 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
901 DEBUG_ONLY(_dwl_initialized = true;)
902 _dwl_adjustment = normal_distribution(1.0);
903 }
904
905 // Simple class for storing info about the heap at the start of GC, to be used
906 // after GC for comparison/printing.
907 class PreGCValues {
908 public:
909 PreGCValues() { }
910 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
911
912 void fill(ParallelScavengeHeap* heap) {
913 _heap_used = heap->used();
914 _young_gen_used = heap->young_gen()->used_in_bytes();
915 _old_gen_used = heap->old_gen()->used_in_bytes();
916 _perm_gen_used = heap->perm_gen()->used_in_bytes();
917 };
918
919 size_t heap_used() const { return _heap_used; }
920 size_t young_gen_used() const { return _young_gen_used; }
921 size_t old_gen_used() const { return _old_gen_used; }
922 size_t perm_gen_used() const { return _perm_gen_used; }
923
924 private:
925 size_t _heap_used;
926 size_t _young_gen_used;
927 size_t _old_gen_used;
928 size_t _perm_gen_used;
929 };
930
931 void
932 PSParallelCompact::clear_data_covering_space(SpaceId id)
933 {
934 // At this point, top is the value before GC, new_top() is the value that will
935 // be set at the end of GC. The marking bitmap is cleared to top; nothing
936 // should be marked above top. The summary data is cleared to the larger of
937 // top & new_top.
938 MutableSpace* const space = _space_info[id].space();
939 HeapWord* const bot = space->bottom();
940 HeapWord* const top = space->top();
941 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
942
943 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
944 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
945 _mark_bitmap.clear_range(beg_bit, end_bit);
946
947 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
948 const size_t end_region =
949 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
950 _summary_data.clear_range(beg_region, end_region);
951
952 // Clear the data used to 'split' regions.
953 SplitInfo& split_info = _space_info[id].split_info();
954 if (split_info.is_valid()) {
955 split_info.clear();
956 }
957 DEBUG_ONLY(split_info.verify_clear();)
958 }
959
960 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
961 {
962 // Update the from & to space pointers in space_info, since they are swapped
963 // at each young gen gc. Do the update unconditionally (even though a
964 // promotion failure does not swap spaces) because an unknown number of minor
965 // collections will have swapped the spaces an unknown number of times.
966 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
967 ParallelScavengeHeap* heap = gc_heap();
968 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
969 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
970
971 pre_gc_values->fill(heap);
972
973 ParCompactionManager::reset();
974 NOT_PRODUCT(_mark_bitmap.reset_counters());
975 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
976 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
977
978 // Increment the invocation count
979 heap->increment_total_collections(true);
980
981 // We need to track unique mark sweep invocations as well.
982 _total_invocations++;
983
984 if (PrintHeapAtGC) {
985 Universe::print_heap_before_gc();
986 }
987
988 // Fill in TLABs
989 heap->accumulate_statistics_all_tlabs();
990 heap->ensure_parsability(true); // retire TLABs
991
992 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
993 HandleMark hm; // Discard invalid handles created during verification
994 gclog_or_tty->print(" VerifyBeforeGC:");
995 Universe::verify(true);
996 }
997
998 // Verify object start arrays
999 if (VerifyObjectStartArray &&
1000 VerifyBeforeGC) {
1001 heap->old_gen()->verify_object_start_array();
1002 heap->perm_gen()->verify_object_start_array();
1003 }
1004
1005 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1006 DEBUG_ONLY(summary_data().verify_clear();)
1007
1008 // Have worker threads release resources the next time they run a task.
1009 gc_task_manager()->release_all_resources();
1010 }
1011
1012 void PSParallelCompact::post_compact()
1013 {
1014 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
1015
1016 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1017 // Clear the marking bitmap, summary data and split info.
1018 clear_data_covering_space(SpaceId(id));
1019 // Update top(). Must be done after clearing the bitmap and summary data.
1020 _space_info[id].publish_new_top();
1021 }
1022
1023 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1024 MutableSpace* const from_space = _space_info[from_space_id].space();
1025 MutableSpace* const to_space = _space_info[to_space_id].space();
1026
1027 ParallelScavengeHeap* heap = gc_heap();
1028 bool eden_empty = eden_space->is_empty();
1029 if (!eden_empty) {
1030 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1031 heap->young_gen(), heap->old_gen());
1032 }
1033
1034 // Update heap occupancy information which is used as input to the soft ref
1035 // clearing policy at the next gc.
1036 Universe::update_heap_info_at_gc();
1037
1038 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1039 to_space->is_empty();
1040
1041 BarrierSet* bs = heap->barrier_set();
1042 if (bs->is_a(BarrierSet::ModRef)) {
1043 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1044 MemRegion old_mr = heap->old_gen()->reserved();
1045 MemRegion perm_mr = heap->perm_gen()->reserved();
1046 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1047
1048 if (young_gen_empty) {
1049 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1050 } else {
1051 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1052 }
1053 }
1054
1055 Threads::gc_epilogue();
1056 CodeCache::gc_epilogue();
1057
1058 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1059
1060 ref_processor()->enqueue_discovered_references(NULL);
1061
1062 if (ZapUnusedHeapArea) {
1063 heap->gen_mangle_unused_area();
1064 }
1065
1066 // Update time of last GC
1067 reset_millis_since_last_gc();
1068 }
1069
1070 HeapWord*
1071 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1072 bool maximum_compaction)
1073 {
1074 const size_t region_size = ParallelCompactData::RegionSize;
1075 const ParallelCompactData& sd = summary_data();
1076
1077 const MutableSpace* const space = _space_info[id].space();
1078 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1079 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1080 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1081
1082 // Skip full regions at the beginning of the space--they are necessarily part
1083 // of the dense prefix.
1084 size_t full_count = 0;
1085 const RegionData* cp;
1086 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1087 ++full_count;
1088 }
1089
1090 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1091 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1092 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1093 if (maximum_compaction || cp == end_cp || interval_ended) {
1094 _maximum_compaction_gc_num = total_invocations();
1095 return sd.region_to_addr(cp);
1096 }
1097
1098 HeapWord* const new_top = _space_info[id].new_top();
1099 const size_t space_live = pointer_delta(new_top, space->bottom());
1100 const size_t space_used = space->used_in_words();
1101 const size_t space_capacity = space->capacity_in_words();
1102
1103 const double cur_density = double(space_live) / space_capacity;
1104 const double deadwood_density =
1105 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1106 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1107
1108 if (TraceParallelOldGCDensePrefix) {
1109 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1110 cur_density, deadwood_density, deadwood_goal);
1111 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1112 "space_cap=" SIZE_FORMAT,
1113 space_live, space_used,
1114 space_capacity);
1115 }
1116
1117 // XXX - Use binary search?
1118 HeapWord* dense_prefix = sd.region_to_addr(cp);
1119 const RegionData* full_cp = cp;
1120 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1121 while (cp < end_cp) {
1122 HeapWord* region_destination = cp->destination();
1123 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1124 if (TraceParallelOldGCDensePrefix && Verbose) {
1125 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1126 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1127 sd.region(cp), region_destination,
1128 dense_prefix, cur_deadwood);
1129 }
1130
1131 if (cur_deadwood >= deadwood_goal) {
1132 // Found the region that has the correct amount of deadwood to the left.
1133 // This typically occurs after crossing a fairly sparse set of regions, so
1134 // iterate backwards over those sparse regions, looking for the region
1135 // that has the lowest density of live objects 'to the right.'
1136 size_t space_to_left = sd.region(cp) * region_size;
1137 size_t live_to_left = space_to_left - cur_deadwood;
1138 size_t space_to_right = space_capacity - space_to_left;
1139 size_t live_to_right = space_live - live_to_left;
1140 double density_to_right = double(live_to_right) / space_to_right;
1141 while (cp > full_cp) {
1142 --cp;
1143 const size_t prev_region_live_to_right = live_to_right -
1144 cp->data_size();
1145 const size_t prev_region_space_to_right = space_to_right + region_size;
1146 double prev_region_density_to_right =
1147 double(prev_region_live_to_right) / prev_region_space_to_right;
1148 if (density_to_right <= prev_region_density_to_right) {
1149 return dense_prefix;
1150 }
1151 if (TraceParallelOldGCDensePrefix && Verbose) {
1152 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1153 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1154 prev_region_density_to_right);
1155 }
1156 dense_prefix -= region_size;
1157 live_to_right = prev_region_live_to_right;
1158 space_to_right = prev_region_space_to_right;
1159 density_to_right = prev_region_density_to_right;
1160 }
1161 return dense_prefix;
1162 }
1163
1164 dense_prefix += region_size;
1165 ++cp;
1166 }
1167
1168 return dense_prefix;
1169 }
1170
1171 #ifndef PRODUCT
1172 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1173 const SpaceId id,
1174 const bool maximum_compaction,
1175 HeapWord* const addr)
1176 {
1177 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1178 RegionData* const cp = summary_data().region(region_idx);
1179 const MutableSpace* const space = _space_info[id].space();
1180 HeapWord* const new_top = _space_info[id].new_top();
1181
1182 const size_t space_live = pointer_delta(new_top, space->bottom());
1183 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1184 const size_t space_cap = space->capacity_in_words();
1185 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1186 const size_t live_to_right = new_top - cp->destination();
1187 const size_t dead_to_right = space->top() - addr - live_to_right;
1188
1189 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1190 "spl=" SIZE_FORMAT " "
1191 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1192 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1193 " ratio=%10.8f",
1194 algorithm, addr, region_idx,
1195 space_live,
1196 dead_to_left, dead_to_left_pct,
1197 dead_to_right, live_to_right,
1198 double(dead_to_right) / live_to_right);
1199 }
1200 #endif // #ifndef PRODUCT
1201
1202 // Return a fraction indicating how much of the generation can be treated as
1203 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1204 // based on the density of live objects in the generation to determine a limit,
1205 // which is then adjusted so the return value is min_percent when the density is
1206 // 1.
1207 //
1208 // The following table shows some return values for a different values of the
1209 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1210 // min_percent is 1.
1211 //
1212 // fraction allowed as dead wood
1213 // -----------------------------------------------------------------
1214 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1215 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1216 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1217 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1218 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1219 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1220 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1221 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1222 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1223 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1224 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1225 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1226 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1227 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1228 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1229 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1230 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1231 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1232 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1233 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1234 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1235 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1236 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1237
1238 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1239 {
1240 assert(_dwl_initialized, "uninitialized");
1241
1242 // The raw limit is the value of the normal distribution at x = density.
1243 const double raw_limit = normal_distribution(density);
1244
1245 // Adjust the raw limit so it becomes the minimum when the density is 1.
1246 //
1247 // First subtract the adjustment value (which is simply the precomputed value
1248 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1249 // Then add the minimum value, so the minimum is returned when the density is
1250 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1251 const double min = double(min_percent) / 100.0;
1252 const double limit = raw_limit - _dwl_adjustment + min;
1253 return MAX2(limit, 0.0);
1254 }
1255
1256 ParallelCompactData::RegionData*
1257 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1258 const RegionData* end)
1259 {
1260 const size_t region_size = ParallelCompactData::RegionSize;
1261 ParallelCompactData& sd = summary_data();
1262 size_t left = sd.region(beg);
1263 size_t right = end > beg ? sd.region(end) - 1 : left;
1264
1265 // Binary search.
1266 while (left < right) {
1267 // Equivalent to (left + right) / 2, but does not overflow.
1268 const size_t middle = left + (right - left) / 2;
1269 RegionData* const middle_ptr = sd.region(middle);
1270 HeapWord* const dest = middle_ptr->destination();
1271 HeapWord* const addr = sd.region_to_addr(middle);
1272 assert(dest != NULL, "sanity");
1273 assert(dest <= addr, "must move left");
1274
1275 if (middle > left && dest < addr) {
1276 right = middle - 1;
1277 } else if (middle < right && middle_ptr->data_size() == region_size) {
1278 left = middle + 1;
1279 } else {
1280 return middle_ptr;
1281 }
1282 }
1283 return sd.region(left);
1284 }
1285
1286 ParallelCompactData::RegionData*
1287 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1288 const RegionData* end,
1289 size_t dead_words)
1290 {
1291 ParallelCompactData& sd = summary_data();
1292 size_t left = sd.region(beg);
1293 size_t right = end > beg ? sd.region(end) - 1 : left;
1294
1295 // Binary search.
1296 while (left < right) {
1297 // Equivalent to (left + right) / 2, but does not overflow.
1298 const size_t middle = left + (right - left) / 2;
1299 RegionData* const middle_ptr = sd.region(middle);
1300 HeapWord* const dest = middle_ptr->destination();
1301 HeapWord* const addr = sd.region_to_addr(middle);
1302 assert(dest != NULL, "sanity");
1303 assert(dest <= addr, "must move left");
1304
1305 const size_t dead_to_left = pointer_delta(addr, dest);
1306 if (middle > left && dead_to_left > dead_words) {
1307 right = middle - 1;
1308 } else if (middle < right && dead_to_left < dead_words) {
1309 left = middle + 1;
1310 } else {
1311 return middle_ptr;
1312 }
1313 }
1314 return sd.region(left);
1315 }
1316
1317 // The result is valid during the summary phase, after the initial summarization
1318 // of each space into itself, and before final summarization.
1319 inline double
1320 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1321 HeapWord* const bottom,
1322 HeapWord* const top,
1323 HeapWord* const new_top)
1324 {
1325 ParallelCompactData& sd = summary_data();
1326
1327 assert(cp != NULL, "sanity");
1328 assert(bottom != NULL, "sanity");
1329 assert(top != NULL, "sanity");
1330 assert(new_top != NULL, "sanity");
1331 assert(top >= new_top, "summary data problem?");
1332 assert(new_top > bottom, "space is empty; should not be here");
1333 assert(new_top >= cp->destination(), "sanity");
1334 assert(top >= sd.region_to_addr(cp), "sanity");
1335
1336 HeapWord* const destination = cp->destination();
1337 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1338 const size_t compacted_region_live = pointer_delta(new_top, destination);
1339 const size_t compacted_region_used = pointer_delta(top,
1340 sd.region_to_addr(cp));
1341 const size_t reclaimable = compacted_region_used - compacted_region_live;
1342
1343 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1344 return double(reclaimable) / divisor;
1345 }
1346
1347 // Return the address of the end of the dense prefix, a.k.a. the start of the
1348 // compacted region. The address is always on a region boundary.
1349 //
1350 // Completely full regions at the left are skipped, since no compaction can
1351 // occur in those regions. Then the maximum amount of dead wood to allow is
1352 // computed, based on the density (amount live / capacity) of the generation;
1353 // the region with approximately that amount of dead space to the left is
1354 // identified as the limit region. Regions between the last completely full
1355 // region and the limit region are scanned and the one that has the best
1356 // (maximum) reclaimed_ratio() is selected.
1357 HeapWord*
1358 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1359 bool maximum_compaction)
1360 {
1361 if (ParallelOldGCSplitALot) {
1362 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1363 // The value was chosen to provoke splitting a young gen space; use it.
1364 return _space_info[id].dense_prefix();
1365 }
1366 }
1367
1368 const size_t region_size = ParallelCompactData::RegionSize;
1369 const ParallelCompactData& sd = summary_data();
1370
1371 const MutableSpace* const space = _space_info[id].space();
1372 HeapWord* const top = space->top();
1373 HeapWord* const top_aligned_up = sd.region_align_up(top);
1374 HeapWord* const new_top = _space_info[id].new_top();
1375 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1376 HeapWord* const bottom = space->bottom();
1377 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1378 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1379 const RegionData* const new_top_cp =
1380 sd.addr_to_region_ptr(new_top_aligned_up);
1381
1382 // Skip full regions at the beginning of the space--they are necessarily part
1383 // of the dense prefix.
1384 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1385 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1386 space->is_empty(), "no dead space allowed to the left");
1387 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1388 "region must have dead space");
1389
1390 // The gc number is saved whenever a maximum compaction is done, and used to
1391 // determine when the maximum compaction interval has expired. This avoids
1392 // successive max compactions for different reasons.
1393 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1394 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1395 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1396 total_invocations() == HeapFirstMaximumCompactionCount;
1397 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1398 _maximum_compaction_gc_num = total_invocations();
1399 return sd.region_to_addr(full_cp);
1400 }
1401
1402 const size_t space_live = pointer_delta(new_top, bottom);
1403 const size_t space_used = space->used_in_words();
1404 const size_t space_capacity = space->capacity_in_words();
1405
1406 const double density = double(space_live) / double(space_capacity);
1407 const size_t min_percent_free =
1408 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1409 const double limiter = dead_wood_limiter(density, min_percent_free);
1410 const size_t dead_wood_max = space_used - space_live;
1411 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1412 dead_wood_max);
1413
1414 if (TraceParallelOldGCDensePrefix) {
1415 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1416 "space_cap=" SIZE_FORMAT,
1417 space_live, space_used,
1418 space_capacity);
1419 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1420 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1421 density, min_percent_free, limiter,
1422 dead_wood_max, dead_wood_limit);
1423 }
1424
1425 // Locate the region with the desired amount of dead space to the left.
1426 const RegionData* const limit_cp =
1427 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1428
1429 // Scan from the first region with dead space to the limit region and find the
1430 // one with the best (largest) reclaimed ratio.
1431 double best_ratio = 0.0;
1432 const RegionData* best_cp = full_cp;
1433 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1434 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1435 if (tmp_ratio > best_ratio) {
1436 best_cp = cp;
1437 best_ratio = tmp_ratio;
1438 }
1439 }
1440
1441 #if 0
1442 // Something to consider: if the region with the best ratio is 'close to' the
1443 // first region w/free space, choose the first region with free space
1444 // ("first-free"). The first-free region is usually near the start of the
1445 // heap, which means we are copying most of the heap already, so copy a bit
1446 // more to get complete compaction.
1447 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1448 _maximum_compaction_gc_num = total_invocations();
1449 best_cp = full_cp;
1450 }
1451 #endif // #if 0
1452
1453 return sd.region_to_addr(best_cp);
1454 }
1455
1456 #ifndef PRODUCT
1457 void
1458 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1459 size_t words)
1460 {
1461 if (TraceParallelOldGCSummaryPhase) {
1462 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1463 SIZE_FORMAT, start, start + words, words);
1464 }
1465
1466 ObjectStartArray* const start_array = _space_info[id].start_array();
1467 CollectedHeap::fill_with_objects(start, words);
1468 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1469 _mark_bitmap.mark_obj(p, words);
1470 _summary_data.add_obj(p, words);
1471 start_array->allocate_block(p);
1472 }
1473 }
1474
1475 void
1476 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1477 {
1478 ParallelCompactData& sd = summary_data();
1479 MutableSpace* space = _space_info[id].space();
1480
1481 // Find the source and destination start addresses.
1482 HeapWord* const src_addr = sd.region_align_down(start);
1483 HeapWord* dst_addr;
1484 if (src_addr < start) {
1485 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1486 } else if (src_addr > space->bottom()) {
1487 // The start (the original top() value) is aligned to a region boundary so
1488 // the associated region does not have a destination. Compute the
1489 // destination from the previous region.
1490 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1491 dst_addr = cp->destination() + cp->data_size();
1492 } else {
1493 // Filling the entire space.
1494 dst_addr = space->bottom();
1495 }
1496 assert(dst_addr != NULL, "sanity");
1497
1498 // Update the summary data.
1499 bool result = _summary_data.summarize(_space_info[id].split_info(),
1500 src_addr, space->top(), NULL,
1501 dst_addr, space->end(),
1502 _space_info[id].new_top_addr());
1503 assert(result, "should not fail: bad filler object size");
1504 }
1505
1506 void
1507 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1508 {
1509 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1510 return;
1511 }
1512
1513 MutableSpace* const space = _space_info[id].space();
1514 if (space->is_empty()) {
1515 HeapWord* b = space->bottom();
1516 HeapWord* t = b + space->capacity_in_words() / 2;
1517 space->set_top(t);
1518 if (ZapUnusedHeapArea) {
1519 space->set_top_for_allocations();
1520 }
1521
1522 size_t min_size = CollectedHeap::min_fill_size();
1523 size_t obj_len = min_size;
1524 while (b + obj_len <= t) {
1525 CollectedHeap::fill_with_object(b, obj_len);
1526 mark_bitmap()->mark_obj(b, obj_len);
1527 summary_data().add_obj(b, obj_len);
1528 b += obj_len;
1529 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1530 }
1531 if (b < t) {
1532 // The loop didn't completely fill to t (top); adjust top downward.
1533 space->set_top(b);
1534 if (ZapUnusedHeapArea) {
1535 space->set_top_for_allocations();
1536 }
1537 }
1538
1539 HeapWord** nta = _space_info[id].new_top_addr();
1540 bool result = summary_data().summarize(_space_info[id].split_info(),
1541 space->bottom(), space->top(), NULL,
1542 space->bottom(), space->end(), nta);
1543 assert(result, "space must fit into itself");
1544 }
1545 }
1546
1547 void
1548 PSParallelCompact::provoke_split(bool & max_compaction)
1549 {
1550 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1551 return;
1552 }
1553
1554 const size_t region_size = ParallelCompactData::RegionSize;
1555 ParallelCompactData& sd = summary_data();
1556
1557 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1558 MutableSpace* const from_space = _space_info[from_space_id].space();
1559 const size_t eden_live = pointer_delta(eden_space->top(),
1560 _space_info[eden_space_id].new_top());
1561 const size_t from_live = pointer_delta(from_space->top(),
1562 _space_info[from_space_id].new_top());
1563
1564 const size_t min_fill_size = CollectedHeap::min_fill_size();
1565 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1566 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1567 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1568 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1569
1570 // Choose the space to split; need at least 2 regions live (or fillable).
1571 SpaceId id;
1572 MutableSpace* space;
1573 size_t live_words;
1574 size_t fill_words;
1575 if (eden_live + eden_fillable >= region_size * 2) {
1576 id = eden_space_id;
1577 space = eden_space;
1578 live_words = eden_live;
1579 fill_words = eden_fillable;
1580 } else if (from_live + from_fillable >= region_size * 2) {
1581 id = from_space_id;
1582 space = from_space;
1583 live_words = from_live;
1584 fill_words = from_fillable;
1585 } else {
1586 return; // Give up.
1587 }
1588 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1589
1590 if (live_words < region_size * 2) {
1591 // Fill from top() to end() w/live objects of mixed sizes.
1592 HeapWord* const fill_start = space->top();
1593 live_words += fill_words;
1594
1595 space->set_top(fill_start + fill_words);
1596 if (ZapUnusedHeapArea) {
1597 space->set_top_for_allocations();
1598 }
1599
1600 HeapWord* cur_addr = fill_start;
1601 while (fill_words > 0) {
1602 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1603 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1604 if (fill_words - cur_size < min_fill_size) {
1605 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1606 }
1607
1608 CollectedHeap::fill_with_object(cur_addr, cur_size);
1609 mark_bitmap()->mark_obj(cur_addr, cur_size);
1610 sd.add_obj(cur_addr, cur_size);
1611
1612 cur_addr += cur_size;
1613 fill_words -= cur_size;
1614 }
1615
1616 summarize_new_objects(id, fill_start);
1617 }
1618
1619 max_compaction = false;
1620
1621 // Manipulate the old gen so that it has room for about half of the live data
1622 // in the target young gen space (live_words / 2).
1623 id = old_space_id;
1624 space = _space_info[id].space();
1625 const size_t free_at_end = space->free_in_words();
1626 const size_t free_target = align_object_size(live_words / 2);
1627 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1628
1629 if (free_at_end >= free_target + min_fill_size) {
1630 // Fill space above top() and set the dense prefix so everything survives.
1631 HeapWord* const fill_start = space->top();
1632 const size_t fill_size = free_at_end - free_target;
1633 space->set_top(space->top() + fill_size);
1634 if (ZapUnusedHeapArea) {
1635 space->set_top_for_allocations();
1636 }
1637 fill_with_live_objects(id, fill_start, fill_size);
1638 summarize_new_objects(id, fill_start);
1639 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1640 } else if (dead + free_at_end > free_target) {
1641 // Find a dense prefix that makes the right amount of space available.
1642 HeapWord* cur = sd.region_align_down(space->top());
1643 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1644 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1645 while (dead_to_right < free_target) {
1646 cur -= region_size;
1647 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1648 dead_to_right = pointer_delta(space->end(), cur_destination);
1649 }
1650 _space_info[id].set_dense_prefix(cur);
1651 }
1652 }
1653 #endif // #ifndef PRODUCT
1654
1655 void PSParallelCompact::summarize_spaces_quick()
1656 {
1657 for (unsigned int i = 0; i < last_space_id; ++i) {
1658 const MutableSpace* space = _space_info[i].space();
1659 HeapWord** nta = _space_info[i].new_top_addr();
1660 bool result = _summary_data.summarize(_space_info[i].split_info(),
1661 space->bottom(), space->top(), NULL,
1662 space->bottom(), space->end(), nta);
1663 assert(result, "space must fit into itself");
1664 _space_info[i].set_dense_prefix(space->bottom());
1665 }
1666
1667 #ifndef PRODUCT
1668 if (ParallelOldGCSplitALot) {
1669 provoke_split_fill_survivor(to_space_id);
1670 }
1671 #endif // #ifndef PRODUCT
1672 }
1673
1674 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1675 {
1676 HeapWord* const dense_prefix_end = dense_prefix(id);
1677 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1678 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1679 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1680 // Only enough dead space is filled so that any remaining dead space to the
1681 // left is larger than the minimum filler object. (The remainder is filled
1682 // during the copy/update phase.)
1683 //
1684 // The size of the dead space to the right of the boundary is not a
1685 // concern, since compaction will be able to use whatever space is
1686 // available.
1687 //
1688 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1689 // surrounds the space to be filled with an object.
1690 //
1691 // In the 32-bit VM, each bit represents two 32-bit words:
1692 // +---+
1693 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1694 // end_bits: ... x x x | 0 | || 0 x x ...
1695 // +---+
1696 //
1697 // In the 64-bit VM, each bit represents one 64-bit word:
1698 // +------------+
1699 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1700 // end_bits: ... x x 1 | 0 || 0 | x x ...
1701 // +------------+
1702 // +-------+
1703 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1704 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1705 // +-------+
1706 // +-----------+
1707 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1708 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1709 // +-----------+
1710 // +-------+
1711 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1712 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1713 // +-------+
1714
1715 // Initially assume case a, c or e will apply.
1716 size_t obj_len = CollectedHeap::min_fill_size();
1717 HeapWord* obj_beg = dense_prefix_end - obj_len;
1718
1719 #ifdef _LP64
1720 if (MinObjAlignment > 1) { // object alignment > heap word size
1721 // Cases a, c or e.
1722 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1723 // Case b above.
1724 obj_beg = dense_prefix_end - 1;
1725 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1726 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1727 // Case d above.
1728 obj_beg = dense_prefix_end - 3;
1729 obj_len = 3;
1730 }
1731 #endif // #ifdef _LP64
1732
1733 CollectedHeap::fill_with_object(obj_beg, obj_len);
1734 _mark_bitmap.mark_obj(obj_beg, obj_len);
1735 _summary_data.add_obj(obj_beg, obj_len);
1736 assert(start_array(id) != NULL, "sanity");
1737 start_array(id)->allocate_block(obj_beg);
1738 }
1739 }
1740
1741 void
1742 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1743 {
1744 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1745 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1746 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1747 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1748 cur->set_source_region(0);
1749 }
1750 }
1751
1752 void
1753 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1754 {
1755 assert(id < last_space_id, "id out of range");
1756 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1757 ParallelOldGCSplitALot && id == old_space_id,
1758 "should have been reset in summarize_spaces_quick()");
1759
1760 const MutableSpace* space = _space_info[id].space();
1761 if (_space_info[id].new_top() != space->bottom()) {
1762 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1763 _space_info[id].set_dense_prefix(dense_prefix_end);
1764
1765 #ifndef PRODUCT
1766 if (TraceParallelOldGCDensePrefix) {
1767 print_dense_prefix_stats("ratio", id, maximum_compaction,
1768 dense_prefix_end);
1769 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1770 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1771 }
1772 #endif // #ifndef PRODUCT
1773
1774 // Recompute the summary data, taking into account the dense prefix. If
1775 // every last byte will be reclaimed, then the existing summary data which
1776 // compacts everything can be left in place.
1777 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1778 // If dead space crosses the dense prefix boundary, it is (at least
1779 // partially) filled with a dummy object, marked live and added to the
1780 // summary data. This simplifies the copy/update phase and must be done
1781 // before the final locations of objects are determined, to prevent
1782 // leaving a fragment of dead space that is too small to fill.
1783 fill_dense_prefix_end(id);
1784
1785 // Compute the destination of each Region, and thus each object.
1786 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1787 _summary_data.summarize(_space_info[id].split_info(),
1788 dense_prefix_end, space->top(), NULL,
1789 dense_prefix_end, space->end(),
1790 _space_info[id].new_top_addr());
1791 }
1792 }
1793
1794 if (TraceParallelOldGCSummaryPhase) {
1795 const size_t region_size = ParallelCompactData::RegionSize;
1796 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1797 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1798 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1799 HeapWord* const new_top = _space_info[id].new_top();
1800 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1801 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1802 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1803 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1804 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1805 id, space->capacity_in_words(), dense_prefix_end,
1806 dp_region, dp_words / region_size,
1807 cr_words / region_size, new_top);
1808 }
1809 }
1810
1811 #ifndef PRODUCT
1812 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1813 HeapWord* dst_beg, HeapWord* dst_end,
1814 SpaceId src_space_id,
1815 HeapWord* src_beg, HeapWord* src_end)
1816 {
1817 if (TraceParallelOldGCSummaryPhase) {
1818 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1819 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1820 SIZE_FORMAT "-" SIZE_FORMAT " "
1821 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1822 SIZE_FORMAT "-" SIZE_FORMAT,
1823 src_space_id, space_names[src_space_id],
1824 dst_space_id, space_names[dst_space_id],
1825 src_beg, src_end,
1826 _summary_data.addr_to_region_idx(src_beg),
1827 _summary_data.addr_to_region_idx(src_end),
1828 dst_beg, dst_end,
1829 _summary_data.addr_to_region_idx(dst_beg),
1830 _summary_data.addr_to_region_idx(dst_end));
1831 }
1832 }
1833 #endif // #ifndef PRODUCT
1834
1835 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1836 bool maximum_compaction)
1837 {
1838 EventMark m("2 summarize");
1839 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1840 // trace("2");
1841
1842 #ifdef ASSERT
1843 if (TraceParallelOldGCMarkingPhase) {
1844 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1845 "add_obj_bytes=" SIZE_FORMAT,
1846 add_obj_count, add_obj_size * HeapWordSize);
1847 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1848 "mark_bitmap_bytes=" SIZE_FORMAT,
1849 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1850 }
1851 #endif // #ifdef ASSERT
1852
1853 // Quick summarization of each space into itself, to see how much is live.
1854 summarize_spaces_quick();
1855
1856 if (TraceParallelOldGCSummaryPhase) {
1857 tty->print_cr("summary_phase: after summarizing each space to self");
1858 Universe::print();
1859 NOT_PRODUCT(print_region_ranges());
1860 if (Verbose) {
1861 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1862 }
1863 }
1864
1865 // The amount of live data that will end up in old space (assuming it fits).
1866 size_t old_space_total_live = 0;
1867 assert(perm_space_id < old_space_id, "should not count perm data here");
1868 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1869 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1870 _space_info[id].space()->bottom());
1871 }
1872
1873 MutableSpace* const old_space = _space_info[old_space_id].space();
1874 const size_t old_capacity = old_space->capacity_in_words();
1875 if (old_space_total_live > old_capacity) {
1876 // XXX - should also try to expand
1877 maximum_compaction = true;
1878 }
1879 #ifndef PRODUCT
1880 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1881 provoke_split(maximum_compaction);
1882 }
1883 #endif // #ifndef PRODUCT
1884
1885 // Permanent and Old generations.
1886 summarize_space(perm_space_id, maximum_compaction);
1887 summarize_space(old_space_id, maximum_compaction);
1888
1889 // Summarize the remaining spaces in the young gen. The initial target space
1890 // is the old gen. If a space does not fit entirely into the target, then the
1891 // remainder is compacted into the space itself and that space becomes the new
1892 // target.
1893 SpaceId dst_space_id = old_space_id;
1894 HeapWord* dst_space_end = old_space->end();
1895 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1896 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1897 const MutableSpace* space = _space_info[id].space();
1898 const size_t live = pointer_delta(_space_info[id].new_top(),
1899 space->bottom());
1900 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1901
1902 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1903 SpaceId(id), space->bottom(), space->top());)
1904 if (live > 0 && live <= available) {
1905 // All the live data will fit.
1906 bool done = _summary_data.summarize(_space_info[id].split_info(),
1907 space->bottom(), space->top(),
1908 NULL,
1909 *new_top_addr, dst_space_end,
1910 new_top_addr);
1911 assert(done, "space must fit into old gen");
1912
1913 // Reset the new_top value for the space.
1914 _space_info[id].set_new_top(space->bottom());
1915 } else if (live > 0) {
1916 // Attempt to fit part of the source space into the target space.
1917 HeapWord* next_src_addr = NULL;
1918 bool done = _summary_data.summarize(_space_info[id].split_info(),
1919 space->bottom(), space->top(),
1920 &next_src_addr,
1921 *new_top_addr, dst_space_end,
1922 new_top_addr);
1923 assert(!done, "space should not fit into old gen");
1924 assert(next_src_addr != NULL, "sanity");
1925
1926 // The source space becomes the new target, so the remainder is compacted
1927 // within the space itself.
1928 dst_space_id = SpaceId(id);
1929 dst_space_end = space->end();
1930 new_top_addr = _space_info[id].new_top_addr();
1931 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1932 space->bottom(), dst_space_end,
1933 SpaceId(id), next_src_addr, space->top());)
1934 done = _summary_data.summarize(_space_info[id].split_info(),
1935 next_src_addr, space->top(),
1936 NULL,
1937 space->bottom(), dst_space_end,
1938 new_top_addr);
1939 assert(done, "space must fit when compacted into itself");
1940 assert(*new_top_addr <= space->top(), "usage should not grow");
1941 }
1942 }
1943
1944 if (TraceParallelOldGCSummaryPhase) {
1945 tty->print_cr("summary_phase: after final summarization");
1946 Universe::print();
1947 NOT_PRODUCT(print_region_ranges());
1948 if (Verbose) {
1949 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1950 }
1951 }
1952 }
1953
1954 // This method should contain all heap-specific policy for invoking a full
1955 // collection. invoke_no_policy() will only attempt to compact the heap; it
1956 // will do nothing further. If we need to bail out for policy reasons, scavenge
1957 // before full gc, or any other specialized behavior, it needs to be added here.
1958 //
1959 // Note that this method should only be called from the vm_thread while at a
1960 // safepoint.
1961 //
1962 // Note that the all_soft_refs_clear flag in the collector policy
1963 // may be true because this method can be called without intervening
1964 // activity. For example when the heap space is tight and full measure
1965 // are being taken to free space.
1966 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1967 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1968 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1969 "should be in vm thread");
1970
1971 ParallelScavengeHeap* heap = gc_heap();
1972 GCCause::Cause gc_cause = heap->gc_cause();
1973 assert(!heap->is_gc_active(), "not reentrant");
1974
1975 PSAdaptiveSizePolicy* policy = heap->size_policy();
1976 IsGCActiveMark mark;
1977
1978 if (ScavengeBeforeFullGC) {
1979 PSScavenge::invoke_no_policy();
1980 }
1981
1982 const bool clear_all_soft_refs =
1983 heap->collector_policy()->should_clear_all_soft_refs();
1984
1985 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1986 maximum_heap_compaction);
1987 }
1988
1989 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1990 size_t addr_region_index = addr_to_region_idx(addr);
1991 return region_index == addr_region_index;
1992 }
1993
1994 // This method contains no policy. You should probably
1995 // be calling invoke() instead.
1996 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1997 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1998 assert(ref_processor() != NULL, "Sanity");
1999
2000 if (GC_locker::check_active_before_gc()) {
2001 return;
2002 }
2003
2004 TimeStamp marking_start;
2005 TimeStamp compaction_start;
2006 TimeStamp collection_exit;
2007
2008 ParallelScavengeHeap* heap = gc_heap();
2009 GCCause::Cause gc_cause = heap->gc_cause();
2010 PSYoungGen* young_gen = heap->young_gen();
2011 PSOldGen* old_gen = heap->old_gen();
2012 PSPermGen* perm_gen = heap->perm_gen();
2013 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2014
2015 // The scope of casr should end after code that can change
2016 // CollectorPolicy::_should_clear_all_soft_refs.
2017 ClearedAllSoftRefs casr(maximum_heap_compaction,
2018 heap->collector_policy());
2019
2020 if (ZapUnusedHeapArea) {
2021 // Save information needed to minimize mangling
2022 heap->record_gen_tops_before_GC();
2023 }
2024
2025 heap->pre_full_gc_dump();
2026
2027 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2028
2029 // Make sure data structures are sane, make the heap parsable, and do other
2030 // miscellaneous bookkeeping.
2031 PreGCValues pre_gc_values;
2032 pre_compact(&pre_gc_values);
2033
2034 // Get the compaction manager reserved for the VM thread.
2035 ParCompactionManager* const vmthread_cm =
2036 ParCompactionManager::manager_array(gc_task_manager()->workers());
2037
2038 // Place after pre_compact() where the number of invocations is incremented.
2039 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2040
2041 {
2042 ResourceMark rm;
2043 HandleMark hm;
2044
2045 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
2046
2047 // This is useful for debugging but don't change the output the
2048 // the customer sees.
2049 const char* gc_cause_str = "Full GC";
2050 if (is_system_gc && PrintGCDetails) {
2051 gc_cause_str = "Full GC (System)";
2052 }
2053 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2054 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2055 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
2056 TraceCollectorStats tcs(counters());
2057 TraceMemoryManagerStats tms(true /* Full GC */);
2058
2059 if (TraceGen1Time) accumulated_time()->start();
2060
2061 // Let the size policy know we're starting
2062 size_policy->major_collection_begin();
2063
2064 // When collecting the permanent generation methodOops may be moving,
2065 // so we either have to flush all bcp data or convert it into bci.
2066 CodeCache::gc_prologue();
2067 Threads::gc_prologue();
2068
2069 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2070 COMPILER2_PRESENT(DerivedPointerTable::clear());
2071
2072 ref_processor()->enable_discovery();
2073 ref_processor()->setup_policy(maximum_heap_compaction);
2074
2075 bool marked_for_unloading = false;
2076
2077 marking_start.update();
2078 marking_phase(vmthread_cm, maximum_heap_compaction);
2079
2080 #ifndef PRODUCT
2081 if (TraceParallelOldGCMarkingPhase) {
2082 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2083 "cas_by_another %d",
2084 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2085 mark_bitmap()->cas_by_another());
2086 }
2087 #endif // #ifndef PRODUCT
2088
2089 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
2090 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2091
2092 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2093 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2094
2095 // adjust_roots() updates Universe::_intArrayKlassObj which is
2096 // needed by the compaction for filling holes in the dense prefix.
2097 adjust_roots();
2098
2099 compaction_start.update();
2100 // Does the perm gen always have to be done serially because
2101 // klasses are used in the update of an object?
2102 compact_perm(vmthread_cm);
2103
2104 if (UseParallelOldGCCompacting) {
2105 compact();
2106 } else {
2107 compact_serial(vmthread_cm);
2108 }
2109
2110 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2111 // done before resizing.
2112 post_compact();
2113
2114 // Let the size policy know we're done
2115 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2116
2117 if (UseAdaptiveSizePolicy) {
2118 if (PrintAdaptiveSizePolicy) {
2119 gclog_or_tty->print("AdaptiveSizeStart: ");
2120 gclog_or_tty->stamp();
2121 gclog_or_tty->print_cr(" collection: %d ",
2122 heap->total_collections());
2123 if (Verbose) {
2124 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2125 " perm_gen_capacity: %d ",
2126 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2127 perm_gen->capacity_in_bytes());
2128 }
2129 }
2130
2131 // Don't check if the size_policy is ready here. Let
2132 // the size_policy check that internally.
2133 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2134 ((gc_cause != GCCause::_java_lang_system_gc) ||
2135 UseAdaptiveSizePolicyWithSystemGC)) {
2136 // Calculate optimal free space amounts
2137 assert(young_gen->max_size() >
2138 young_gen->from_space()->capacity_in_bytes() +
2139 young_gen->to_space()->capacity_in_bytes(),
2140 "Sizes of space in young gen are out-of-bounds");
2141 size_t max_eden_size = young_gen->max_size() -
2142 young_gen->from_space()->capacity_in_bytes() -
2143 young_gen->to_space()->capacity_in_bytes();
2144 size_policy->compute_generation_free_space(
2145 young_gen->used_in_bytes(),
2146 young_gen->eden_space()->used_in_bytes(),
2147 old_gen->used_in_bytes(),
2148 perm_gen->used_in_bytes(),
2149 young_gen->eden_space()->capacity_in_bytes(),
2150 old_gen->max_gen_size(),
2151 max_eden_size,
2152 true /* full gc*/,
2153 gc_cause,
2154 heap->collector_policy());
2155
2156 heap->resize_old_gen(
2157 size_policy->calculated_old_free_size_in_bytes());
2158
2159 // Don't resize the young generation at an major collection. A
2160 // desired young generation size may have been calculated but
2161 // resizing the young generation complicates the code because the
2162 // resizing of the old generation may have moved the boundary
2163 // between the young generation and the old generation. Let the
2164 // young generation resizing happen at the minor collections.
2165 }
2166 if (PrintAdaptiveSizePolicy) {
2167 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2168 heap->total_collections());
2169 }
2170 }
2171
2172 if (UsePerfData) {
2173 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2174 counters->update_counters();
2175 counters->update_old_capacity(old_gen->capacity_in_bytes());
2176 counters->update_young_capacity(young_gen->capacity_in_bytes());
2177 }
2178
2179 heap->resize_all_tlabs();
2180
2181 // We collected the perm gen, so we'll resize it here.
2182 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2183
2184 if (TraceGen1Time) accumulated_time()->stop();
2185
2186 if (PrintGC) {
2187 if (PrintGCDetails) {
2188 // No GC timestamp here. This is after GC so it would be confusing.
2189 young_gen->print_used_change(pre_gc_values.young_gen_used());
2190 old_gen->print_used_change(pre_gc_values.old_gen_used());
2191 heap->print_heap_change(pre_gc_values.heap_used());
2192 // Print perm gen last (print_heap_change() excludes the perm gen).
2193 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2194 } else {
2195 heap->print_heap_change(pre_gc_values.heap_used());
2196 }
2197 }
2198
2199 // Track memory usage and detect low memory
2200 MemoryService::track_memory_usage();
2201 heap->update_counters();
2202 }
2203
2204 #ifdef ASSERT
2205 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2206 ParCompactionManager* const cm =
2207 ParCompactionManager::manager_array(int(i));
2208 assert(cm->marking_stack()->is_empty(), "should be empty");
2209 assert(cm->region_stack()->is_empty(), "should be empty");
2210 assert(cm->revisit_klass_stack()->is_empty(), "should be empty");
2211 }
2212 #endif // ASSERT
2213
2214 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2215 HandleMark hm; // Discard invalid handles created during verification
2216 gclog_or_tty->print(" VerifyAfterGC:");
2217 Universe::verify(false);
2218 }
2219
2220 // Re-verify object start arrays
2221 if (VerifyObjectStartArray &&
2222 VerifyAfterGC) {
2223 old_gen->verify_object_start_array();
2224 perm_gen->verify_object_start_array();
2225 }
2226
2227 if (ZapUnusedHeapArea) {
2228 old_gen->object_space()->check_mangled_unused_area_complete();
2229 perm_gen->object_space()->check_mangled_unused_area_complete();
2230 }
2231
2232 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2233
2234 collection_exit.update();
2235
2236 if (PrintHeapAtGC) {
2237 Universe::print_heap_after_gc();
2238 }
2239 if (PrintGCTaskTimeStamps) {
2240 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2241 INT64_FORMAT,
2242 marking_start.ticks(), compaction_start.ticks(),
2243 collection_exit.ticks());
2244 gc_task_manager()->print_task_time_stamps();
2245 }
2246
2247 heap->post_full_gc_dump();
2248
2249 #ifdef TRACESPINNING
2250 ParallelTaskTerminator::print_termination_counts();
2251 #endif
2252 }
2253
2254 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2255 PSYoungGen* young_gen,
2256 PSOldGen* old_gen) {
2257 MutableSpace* const eden_space = young_gen->eden_space();
2258 assert(!eden_space->is_empty(), "eden must be non-empty");
2259 assert(young_gen->virtual_space()->alignment() ==
2260 old_gen->virtual_space()->alignment(), "alignments do not match");
2261
2262 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2263 return false;
2264 }
2265
2266 // Both generations must be completely committed.
2267 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2268 return false;
2269 }
2270 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2271 return false;
2272 }
2273
2274 // Figure out how much to take from eden. Include the average amount promoted
2275 // in the total; otherwise the next young gen GC will simply bail out to a
2276 // full GC.
2277 const size_t alignment = old_gen->virtual_space()->alignment();
2278 const size_t eden_used = eden_space->used_in_bytes();
2279 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2280 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2281 const size_t eden_capacity = eden_space->capacity_in_bytes();
2282
2283 if (absorb_size >= eden_capacity) {
2284 return false; // Must leave some space in eden.
2285 }
2286
2287 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2288 if (new_young_size < young_gen->min_gen_size()) {
2289 return false; // Respect young gen minimum size.
2290 }
2291
2292 if (TraceAdaptiveGCBoundary && Verbose) {
2293 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2294 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2295 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2296 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2297 absorb_size / K,
2298 eden_capacity / K, (eden_capacity - absorb_size) / K,
2299 young_gen->from_space()->used_in_bytes() / K,
2300 young_gen->to_space()->used_in_bytes() / K,
2301 young_gen->capacity_in_bytes() / K, new_young_size / K);
2302 }
2303
2304 // Fill the unused part of the old gen.
2305 MutableSpace* const old_space = old_gen->object_space();
2306 HeapWord* const unused_start = old_space->top();
2307 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2308
2309 if (unused_words > 0) {
2310 if (unused_words < CollectedHeap::min_fill_size()) {
2311 return false; // If the old gen cannot be filled, must give up.
2312 }
2313 CollectedHeap::fill_with_objects(unused_start, unused_words);
2314 }
2315
2316 // Take the live data from eden and set both top and end in the old gen to
2317 // eden top. (Need to set end because reset_after_change() mangles the region
2318 // from end to virtual_space->high() in debug builds).
2319 HeapWord* const new_top = eden_space->top();
2320 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2321 absorb_size);
2322 young_gen->reset_after_change();
2323 old_space->set_top(new_top);
2324 old_space->set_end(new_top);
2325 old_gen->reset_after_change();
2326
2327 // Update the object start array for the filler object and the data from eden.
2328 ObjectStartArray* const start_array = old_gen->start_array();
2329 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2330 start_array->allocate_block(p);
2331 }
2332
2333 // Could update the promoted average here, but it is not typically updated at
2334 // full GCs and the value to use is unclear. Something like
2335 //
2336 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2337
2338 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2339 return true;
2340 }
2341
2342 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2343 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2344 "shouldn't return NULL");
2345 return ParallelScavengeHeap::gc_task_manager();
2346 }
2347
2348 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2349 bool maximum_heap_compaction) {
2350 // Recursively traverse all live objects and mark them
2351 EventMark m("1 mark object");
2352 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2353
2354 ParallelScavengeHeap* heap = gc_heap();
2355 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2356 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2357 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2358
2359 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2360 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2361
2362 {
2363 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2364 ParallelScavengeHeap::ParStrongRootsScope psrs;
2365
2366 GCTaskQueue* q = GCTaskQueue::create();
2367
2368 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2369 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2370 // We scan the thread roots in parallel
2371 Threads::create_thread_roots_marking_tasks(q);
2372 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2373 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2374 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2375 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2376 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2377 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols));
2378 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2379
2380 if (parallel_gc_threads > 1) {
2381 for (uint j = 0; j < parallel_gc_threads; j++) {
2382 q->enqueue(new StealMarkingTask(&terminator));
2383 }
2384 }
2385
2386 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2387 q->enqueue(fin);
2388
2389 gc_task_manager()->add_list(q);
2390
2391 fin->wait_for();
2392
2393 // We have to release the barrier tasks!
2394 WaitForBarrierGCTask::destroy(fin);
2395 }
2396
2397 // Process reference objects found during marking
2398 {
2399 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2400 if (ref_processor()->processing_is_mt()) {
2401 RefProcTaskExecutor task_executor;
2402 ref_processor()->process_discovered_references(
2403 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2404 &task_executor);
2405 } else {
2406 ref_processor()->process_discovered_references(
2407 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2408 }
2409 }
2410
2411 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2412 // Follow system dictionary roots and unload classes.
2413 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2414
2415 // Follow code cache roots.
2416 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2417 purged_class);
2418 cm->follow_marking_stacks(); // Flush marking stack.
2419
2420 // Update subklass/sibling/implementor links of live klasses
2421 // revisit_klass_stack is used in follow_weak_klass_links().
2422 follow_weak_klass_links();
2423
2424 // Revisit memoized MDO's and clear any unmarked weak refs
2425 follow_mdo_weak_refs();
2426
2427 // Visit symbol and interned string tables and delete unmarked oops
2428 SymbolTable::unlink(is_alive_closure());
2429 StringTable::unlink(is_alive_closure());
2430
2431 assert(cm->marking_stacks_empty(), "marking stacks should be empty");
2432 }
2433
2434 // This should be moved to the shared markSweep code!
2435 class PSAlwaysTrueClosure: public BoolObjectClosure {
2436 public:
2437 void do_object(oop p) { ShouldNotReachHere(); }
2438 bool do_object_b(oop p) { return true; }
2439 };
2440 static PSAlwaysTrueClosure always_true;
2441
2442 void PSParallelCompact::adjust_roots() {
2443 // Adjust the pointers to reflect the new locations
2444 EventMark m("3 adjust roots");
2445 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2446
2447 // General strong roots.
2448 Universe::oops_do(adjust_root_pointer_closure());
2449 ReferenceProcessor::oops_do(adjust_root_pointer_closure());
2450 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
2451 Threads::oops_do(adjust_root_pointer_closure(), NULL);
2452 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2453 FlatProfiler::oops_do(adjust_root_pointer_closure());
2454 Management::oops_do(adjust_root_pointer_closure());
2455 JvmtiExport::oops_do(adjust_root_pointer_closure());
2456 // SO_AllClasses
2457 SystemDictionary::oops_do(adjust_root_pointer_closure());
2458 vmSymbols::oops_do(adjust_root_pointer_closure());
2459
2460 // Now adjust pointers in remaining weak roots. (All of which should
2461 // have been cleared if they pointed to non-surviving objects.)
2462 // Global (weak) JNI handles
2463 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2464
2465 CodeCache::oops_do(adjust_pointer_closure());
2466 SymbolTable::oops_do(adjust_root_pointer_closure());
2467 StringTable::oops_do(adjust_root_pointer_closure());
2468 ref_processor()->weak_oops_do(adjust_root_pointer_closure());
2469 // Roots were visited so references into the young gen in roots
2470 // may have been scanned. Process them also.
2471 // Should the reference processor have a span that excludes
2472 // young gen objects?
2473 PSScavenge::reference_processor()->weak_oops_do(
2474 adjust_root_pointer_closure());
2475 }
2476
2477 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2478 EventMark m("4 compact perm");
2479 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
2480 // trace("4");
2481
2482 gc_heap()->perm_gen()->start_array()->reset();
2483 move_and_update(cm, perm_space_id);
2484 }
2485
2486 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2487 uint parallel_gc_threads)
2488 {
2489 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2490
2491 const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
2492 for (unsigned int j = 0; j < task_count; j++) {
2493 q->enqueue(new DrainStacksCompactionTask(j));
2494 }
2495
2496 // Find all regions that are available (can be filled immediately) and
2497 // distribute them to the thread stacks. The iteration is done in reverse
2498 // order (high to low) so the regions will be removed in ascending order.
2499
2500 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2501
2502 size_t fillable_regions = 0; // A count for diagnostic purposes.
2503 unsigned int which = 0; // The worker thread number.
2504
2505 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2506 SpaceInfo* const space_info = _space_info + id;
2507 MutableSpace* const space = space_info->space();
2508 HeapWord* const new_top = space_info->new_top();
2509
2510 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2511 const size_t end_region =
2512 sd.addr_to_region_idx(sd.region_align_up(new_top));
2513 assert(end_region > 0, "perm gen cannot be empty");
2514
2515 for (size_t cur = end_region - 1; cur >= beg_region; --cur) {
2516 if (sd.region(cur)->claim_unsafe()) {
2517 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2518 cm->push_region(cur);
2519
2520 if (TraceParallelOldGCCompactionPhase && Verbose) {
2521 const size_t count_mod_8 = fillable_regions & 7;
2522 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2523 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2524 if (count_mod_8 == 7) gclog_or_tty->cr();
2525 }
2526
2527 NOT_PRODUCT(++fillable_regions;)
2528
2529 // Assign regions to threads in round-robin fashion.
2530 if (++which == task_count) {
2531 which = 0;
2532 }
2533 }
2534 }
2535 }
2536
2537 if (TraceParallelOldGCCompactionPhase) {
2538 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2539 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2540 }
2541 }
2542
2543 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2544
2545 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2546 uint parallel_gc_threads) {
2547 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2548
2549 ParallelCompactData& sd = PSParallelCompact::summary_data();
2550
2551 // Iterate over all the spaces adding tasks for updating
2552 // regions in the dense prefix. Assume that 1 gc thread
2553 // will work on opening the gaps and the remaining gc threads
2554 // will work on the dense prefix.
2555 unsigned int space_id;
2556 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2557 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2558 const MutableSpace* const space = _space_info[space_id].space();
2559
2560 if (dense_prefix_end == space->bottom()) {
2561 // There is no dense prefix for this space.
2562 continue;
2563 }
2564
2565 // The dense prefix is before this region.
2566 size_t region_index_end_dense_prefix =
2567 sd.addr_to_region_idx(dense_prefix_end);
2568 RegionData* const dense_prefix_cp =
2569 sd.region(region_index_end_dense_prefix);
2570 assert(dense_prefix_end == space->end() ||
2571 dense_prefix_cp->available() ||
2572 dense_prefix_cp->claimed(),
2573 "The region after the dense prefix should always be ready to fill");
2574
2575 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2576
2577 // Is there dense prefix work?
2578 size_t total_dense_prefix_regions =
2579 region_index_end_dense_prefix - region_index_start;
2580 // How many regions of the dense prefix should be given to
2581 // each thread?
2582 if (total_dense_prefix_regions > 0) {
2583 uint tasks_for_dense_prefix = 1;
2584 if (UseParallelDensePrefixUpdate) {
2585 if (total_dense_prefix_regions <=
2586 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2587 // Don't over partition. This assumes that
2588 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2589 // so there are not many regions to process.
2590 tasks_for_dense_prefix = parallel_gc_threads;
2591 } else {
2592 // Over partition
2593 tasks_for_dense_prefix = parallel_gc_threads *
2594 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2595 }
2596 }
2597 size_t regions_per_thread = total_dense_prefix_regions /
2598 tasks_for_dense_prefix;
2599 // Give each thread at least 1 region.
2600 if (regions_per_thread == 0) {
2601 regions_per_thread = 1;
2602 }
2603
2604 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2605 if (region_index_start >= region_index_end_dense_prefix) {
2606 break;
2607 }
2608 // region_index_end is not processed
2609 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2610 region_index_end_dense_prefix);
2611 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2612 region_index_start,
2613 region_index_end));
2614 region_index_start = region_index_end;
2615 }
2616 }
2617 // This gets any part of the dense prefix that did not
2618 // fit evenly.
2619 if (region_index_start < region_index_end_dense_prefix) {
2620 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2621 region_index_start,
2622 region_index_end_dense_prefix));
2623 }
2624 }
2625 }
2626
2627 void PSParallelCompact::enqueue_region_stealing_tasks(
2628 GCTaskQueue* q,
2629 ParallelTaskTerminator* terminator_ptr,
2630 uint parallel_gc_threads) {
2631 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2632
2633 // Once a thread has drained it's stack, it should try to steal regions from
2634 // other threads.
2635 if (parallel_gc_threads > 1) {
2636 for (uint j = 0; j < parallel_gc_threads; j++) {
2637 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2638 }
2639 }
2640 }
2641
2642 void PSParallelCompact::compact() {
2643 EventMark m("5 compact");
2644 // trace("5");
2645 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2646
2647 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2648 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2649 PSOldGen* old_gen = heap->old_gen();
2650 old_gen->start_array()->reset();
2651 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2652 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2653 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2654
2655 GCTaskQueue* q = GCTaskQueue::create();
2656 enqueue_region_draining_tasks(q, parallel_gc_threads);
2657 enqueue_dense_prefix_tasks(q, parallel_gc_threads);
2658 enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads);
2659
2660 {
2661 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2662
2663 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2664 q->enqueue(fin);
2665
2666 gc_task_manager()->add_list(q);
2667
2668 fin->wait_for();
2669
2670 // We have to release the barrier tasks!
2671 WaitForBarrierGCTask::destroy(fin);
2672
2673 #ifdef ASSERT
2674 // Verify that all regions have been processed before the deferred updates.
2675 // Note that perm_space_id is skipped; this type of verification is not
2676 // valid until the perm gen is compacted by regions.
2677 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2678 verify_complete(SpaceId(id));
2679 }
2680 #endif
2681 }
2682
2683 {
2684 // Update the deferred objects, if any. Any compaction manager can be used.
2685 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2686 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2687 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2688 update_deferred_objects(cm, SpaceId(id));
2689 }
2690 }
2691 }
2692
2693 #ifdef ASSERT
2694 void PSParallelCompact::verify_complete(SpaceId space_id) {
2695 // All Regions between space bottom() to new_top() should be marked as filled
2696 // and all Regions between new_top() and top() should be available (i.e.,
2697 // should have been emptied).
2698 ParallelCompactData& sd = summary_data();
2699 SpaceInfo si = _space_info[space_id];
2700 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2701 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2702 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2703 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2704 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2705
2706 bool issued_a_warning = false;
2707
2708 size_t cur_region;
2709 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2710 const RegionData* const c = sd.region(cur_region);
2711 if (!c->completed()) {
2712 warning("region " SIZE_FORMAT " not filled: "
2713 "destination_count=" SIZE_FORMAT,
2714 cur_region, c->destination_count());
2715 issued_a_warning = true;
2716 }
2717 }
2718
2719 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2720 const RegionData* const c = sd.region(cur_region);
2721 if (!c->available()) {
2722 warning("region " SIZE_FORMAT " not empty: "
2723 "destination_count=" SIZE_FORMAT,
2724 cur_region, c->destination_count());
2725 issued_a_warning = true;
2726 }
2727 }
2728
2729 if (issued_a_warning) {
2730 print_region_ranges();
2731 }
2732 }
2733 #endif // #ifdef ASSERT
2734
2735 void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
2736 EventMark m("5 compact serial");
2737 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty);
2738
2739 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2740 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2741
2742 PSYoungGen* young_gen = heap->young_gen();
2743 PSOldGen* old_gen = heap->old_gen();
2744
2745 old_gen->start_array()->reset();
2746 old_gen->move_and_update(cm);
2747 young_gen->move_and_update(cm);
2748 }
2749
2750 void
2751 PSParallelCompact::follow_weak_klass_links() {
2752 // All klasses on the revisit stack are marked at this point.
2753 // Update and follow all subklass, sibling and implementor links.
2754 if (PrintRevisitStats) {
2755 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2756 SystemDictionary::number_of_classes());
2757 }
2758 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2759 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2760 KeepAliveClosure keep_alive_closure(cm);
2761 Stack<Klass*>* const rks = cm->revisit_klass_stack();
2762 if (PrintRevisitStats) {
2763 gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT,
2764 i, rks->size());
2765 }
2766 while (!rks->is_empty()) {
2767 Klass* const k = rks->pop();
2768 k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure);
2769 }
2770
2771 cm->follow_marking_stacks();
2772 }
2773 }
2774
2775 void
2776 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2777 cm->revisit_klass_stack()->push(k);
2778 }
2779
2780 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) {
2781 cm->revisit_mdo_stack()->push(p);
2782 }
2783
2784 void PSParallelCompact::follow_mdo_weak_refs() {
2785 // All strongly reachable oops have been marked at this point;
2786 // we can visit and clear any weak references from MDO's which
2787 // we memoized during the strong marking phase.
2788 if (PrintRevisitStats) {
2789 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2790 SystemDictionary::number_of_classes());
2791 }
2792 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2793 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2794 Stack<DataLayout*>* rms = cm->revisit_mdo_stack();
2795 if (PrintRevisitStats) {
2796 gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT,
2797 i, rms->size());
2798 }
2799 while (!rms->is_empty()) {
2800 rms->pop()->follow_weak_refs(is_alive_closure());
2801 }
2802
2803 cm->follow_marking_stacks();
2804 }
2805 }
2806
2807
2808 #ifdef VALIDATE_MARK_SWEEP
2809
2810 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2811 if (!ValidateMarkSweep)
2812 return;
2813
2814 if (!isroot) {
2815 if (_pointer_tracking) {
2816 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2817 _adjusted_pointers->remove(p);
2818 }
2819 } else {
2820 ptrdiff_t index = _root_refs_stack->find(p);
2821 if (index != -1) {
2822 int l = _root_refs_stack->length();
2823 if (l > 0 && l - 1 != index) {
2824 void* last = _root_refs_stack->pop();
2825 assert(last != p, "should be different");
2826 _root_refs_stack->at_put(index, last);
2827 } else {
2828 _root_refs_stack->remove(p);
2829 }
2830 }
2831 }
2832 }
2833
2834
2835 void PSParallelCompact::check_adjust_pointer(void* p) {
2836 _adjusted_pointers->push(p);
2837 }
2838
2839
2840 class AdjusterTracker: public OopClosure {
2841 public:
2842 AdjusterTracker() {};
2843 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
2844 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); }
2845 };
2846
2847
2848 void PSParallelCompact::track_interior_pointers(oop obj) {
2849 if (ValidateMarkSweep) {
2850 _adjusted_pointers->clear();
2851 _pointer_tracking = true;
2852
2853 AdjusterTracker checker;
2854 obj->oop_iterate(&checker);
2855 }
2856 }
2857
2858
2859 void PSParallelCompact::check_interior_pointers() {
2860 if (ValidateMarkSweep) {
2861 _pointer_tracking = false;
2862 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2863 }
2864 }
2865
2866
2867 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2868 if (ValidateMarkSweep) {
2869 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2870 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2871 }
2872 }
2873
2874
2875 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2876 if (ValidateMarkSweep) {
2877 _live_oops->push(p);
2878 _live_oops_size->push(size);
2879 _live_oops_index++;
2880 }
2881 }
2882
2883 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2884 if (ValidateMarkSweep) {
2885 oop obj = _live_oops->at((int)_live_oops_index);
2886 guarantee(obj == p, "should be the same object");
2887 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2888 _live_oops_index++;
2889 }
2890 }
2891
2892 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2893 HeapWord* compaction_top) {
2894 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2895 "should be moved to forwarded location");
2896 if (ValidateMarkSweep) {
2897 PSParallelCompact::validate_live_oop(oop(q), size);
2898 _live_oops_moved_to->push(oop(compaction_top));
2899 }
2900 if (RecordMarkSweepCompaction) {
2901 _cur_gc_live_oops->push(q);
2902 _cur_gc_live_oops_moved_to->push(compaction_top);
2903 _cur_gc_live_oops_size->push(size);
2904 }
2905 }
2906
2907
2908 void PSParallelCompact::compaction_complete() {
2909 if (RecordMarkSweepCompaction) {
2910 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
2911 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2912 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
2913
2914 _cur_gc_live_oops = _last_gc_live_oops;
2915 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
2916 _cur_gc_live_oops_size = _last_gc_live_oops_size;
2917 _last_gc_live_oops = _tmp_live_oops;
2918 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2919 _last_gc_live_oops_size = _tmp_live_oops_size;
2920 }
2921 }
2922
2923
2924 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2925 if (!RecordMarkSweepCompaction) {
2926 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2927 return;
2928 }
2929
2930 if (_last_gc_live_oops == NULL) {
2931 tty->print_cr("No compaction information gathered yet");
2932 return;
2933 }
2934
2935 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2936 HeapWord* old_oop = _last_gc_live_oops->at(i);
2937 size_t sz = _last_gc_live_oops_size->at(i);
2938 if (old_oop <= q && q < (old_oop + sz)) {
2939 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2940 size_t offset = (q - old_oop);
2941 tty->print_cr("Address " PTR_FORMAT, q);
2942 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2943 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2944 return;
2945 }
2946 }
2947
2948 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2949 }
2950 #endif //VALIDATE_MARK_SWEEP
2951
2952 // Update interior oops in the ranges of regions [beg_region, end_region).
2953 void
2954 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2955 SpaceId space_id,
2956 size_t beg_region,
2957 size_t end_region) {
2958 ParallelCompactData& sd = summary_data();
2959 ParMarkBitMap* const mbm = mark_bitmap();
2960
2961 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2962 HeapWord* const end_addr = sd.region_to_addr(end_region);
2963 assert(beg_region <= end_region, "bad region range");
2964 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2965
2966 #ifdef ASSERT
2967 // Claim the regions to avoid triggering an assert when they are marked as
2968 // filled.
2969 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2970 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2971 }
2972 #endif // #ifdef ASSERT
2973
2974 if (beg_addr != space(space_id)->bottom()) {
2975 // Find the first live object or block of dead space that *starts* in this
2976 // range of regions. If a partial object crosses onto the region, skip it;
2977 // it will be marked for 'deferred update' when the object head is
2978 // processed. If dead space crosses onto the region, it is also skipped; it
2979 // will be filled when the prior region is processed. If neither of those
2980 // apply, the first word in the region is the start of a live object or dead
2981 // space.
2982 assert(beg_addr > space(space_id)->bottom(), "sanity");
2983 const RegionData* const cp = sd.region(beg_region);
2984 if (cp->partial_obj_size() != 0) {
2985 beg_addr = sd.partial_obj_end(beg_region);
2986 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2987 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2988 }
2989 }
2990
2991 if (beg_addr < end_addr) {
2992 // A live object or block of dead space starts in this range of Regions.
2993 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2994
2995 // Create closures and iterate.
2996 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2997 FillClosure fill_closure(cm, space_id);
2998 ParMarkBitMap::IterationStatus status;
2999 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
3000 dense_prefix_end);
3001 if (status == ParMarkBitMap::incomplete) {
3002 update_closure.do_addr(update_closure.source());
3003 }
3004 }
3005
3006 // Mark the regions as filled.
3007 RegionData* const beg_cp = sd.region(beg_region);
3008 RegionData* const end_cp = sd.region(end_region);
3009 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
3010 cp->set_completed();
3011 }
3012 }
3013
3014 // Return the SpaceId for the space containing addr. If addr is not in the
3015 // heap, last_space_id is returned. In debug mode it expects the address to be
3016 // in the heap and asserts such.
3017 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
3018 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
3019
3020 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
3021 if (_space_info[id].space()->contains(addr)) {
3022 return SpaceId(id);
3023 }
3024 }
3025
3026 assert(false, "no space contains the addr");
3027 return last_space_id;
3028 }
3029
3030 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
3031 SpaceId id) {
3032 assert(id < last_space_id, "bad space id");
3033
3034 ParallelCompactData& sd = summary_data();
3035 const SpaceInfo* const space_info = _space_info + id;
3036 ObjectStartArray* const start_array = space_info->start_array();
3037
3038 const MutableSpace* const space = space_info->space();
3039 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
3040 HeapWord* const beg_addr = space_info->dense_prefix();
3041 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
3042
3043 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
3044 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
3045 const RegionData* cur_region;
3046 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
3047 HeapWord* const addr = cur_region->deferred_obj_addr();
3048 if (addr != NULL) {
3049 if (start_array != NULL) {
3050 start_array->allocate_block(addr);
3051 }
3052 oop(addr)->update_contents(cm);
3053 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3054 }
3055 }
3056 }
3057
3058 // Skip over count live words starting from beg, and return the address of the
3059 // next live word. Unless marked, the word corresponding to beg is assumed to
3060 // be dead. Callers must either ensure beg does not correspond to the middle of
3061 // an object, or account for those live words in some other way. Callers must
3062 // also ensure that there are enough live words in the range [beg, end) to skip.
3063 HeapWord*
3064 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3065 {
3066 assert(count > 0, "sanity");
3067
3068 ParMarkBitMap* m = mark_bitmap();
3069 idx_t bits_to_skip = m->words_to_bits(count);
3070 idx_t cur_beg = m->addr_to_bit(beg);
3071 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3072
3073 do {
3074 cur_beg = m->find_obj_beg(cur_beg, search_end);
3075 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3076 const size_t obj_bits = cur_end - cur_beg + 1;
3077 if (obj_bits > bits_to_skip) {
3078 return m->bit_to_addr(cur_beg + bits_to_skip);
3079 }
3080 bits_to_skip -= obj_bits;
3081 cur_beg = cur_end + 1;
3082 } while (bits_to_skip > 0);
3083
3084 // Skipping the desired number of words landed just past the end of an object.
3085 // Find the start of the next object.
3086 cur_beg = m->find_obj_beg(cur_beg, search_end);
3087 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3088 return m->bit_to_addr(cur_beg);
3089 }
3090
3091 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3092 SpaceId src_space_id,
3093 size_t src_region_idx)
3094 {
3095 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
3096
3097 const SplitInfo& split_info = _space_info[src_space_id].split_info();
3098 if (split_info.dest_region_addr() == dest_addr) {
3099 // The partial object ending at the split point contains the first word to
3100 // be copied to dest_addr.
3101 return split_info.first_src_addr();
3102 }
3103
3104 const ParallelCompactData& sd = summary_data();
3105 ParMarkBitMap* const bitmap = mark_bitmap();
3106 const size_t RegionSize = ParallelCompactData::RegionSize;
3107
3108 assert(sd.is_region_aligned(dest_addr), "not aligned");
3109 const RegionData* const src_region_ptr = sd.region(src_region_idx);
3110 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
3111 HeapWord* const src_region_destination = src_region_ptr->destination();
3112
3113 assert(dest_addr >= src_region_destination, "wrong src region");
3114 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
3115
3116 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
3117 HeapWord* const src_region_end = src_region_beg + RegionSize;
3118
3119 HeapWord* addr = src_region_beg;
3120 if (dest_addr == src_region_destination) {
3121 // Return the first live word in the source region.
3122 if (partial_obj_size == 0) {
3123 addr = bitmap->find_obj_beg(addr, src_region_end);
3124 assert(addr < src_region_end, "no objects start in src region");
3125 }
3126 return addr;
3127 }
3128
3129 // Must skip some live data.
3130 size_t words_to_skip = dest_addr - src_region_destination;
3131 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
3132
3133 if (partial_obj_size >= words_to_skip) {
3134 // All the live words to skip are part of the partial object.
3135 addr += words_to_skip;
3136 if (partial_obj_size == words_to_skip) {
3137 // Find the first live word past the partial object.
3138 addr = bitmap->find_obj_beg(addr, src_region_end);
3139 assert(addr < src_region_end, "wrong src region");
3140 }
3141 return addr;
3142 }
3143
3144 // Skip over the partial object (if any).
3145 if (partial_obj_size != 0) {
3146 words_to_skip -= partial_obj_size;
3147 addr += partial_obj_size;
3148 }
3149
3150 // Skip over live words due to objects that start in the region.
3151 addr = skip_live_words(addr, src_region_end, words_to_skip);
3152 assert(addr < src_region_end, "wrong src region");
3153 return addr;
3154 }
3155
3156 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3157 SpaceId src_space_id,
3158 size_t beg_region,
3159 HeapWord* end_addr)
3160 {
3161 ParallelCompactData& sd = summary_data();
3162
3163 #ifdef ASSERT
3164 MutableSpace* const src_space = _space_info[src_space_id].space();
3165 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3166 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3167 "src_space_id does not match beg_addr");
3168 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3169 "src_space_id does not match end_addr");
3170 #endif // #ifdef ASSERT
3171
3172 RegionData* const beg = sd.region(beg_region);
3173 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3174
3175 // Regions up to new_top() are enqueued if they become available.
3176 HeapWord* const new_top = _space_info[src_space_id].new_top();
3177 RegionData* const enqueue_end =
3178 sd.addr_to_region_ptr(sd.region_align_up(new_top));
3179
3180 for (RegionData* cur = beg; cur < end; ++cur) {
3181 assert(cur->data_size() > 0, "region must have live data");
3182 cur->decrement_destination_count();
3183 if (cur < enqueue_end && cur->available() && cur->claim()) {
3184 cm->push_region(sd.region(cur));
3185 }
3186 }
3187 }
3188
3189 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3190 SpaceId& src_space_id,
3191 HeapWord*& src_space_top,
3192 HeapWord* end_addr)
3193 {
3194 typedef ParallelCompactData::RegionData RegionData;
3195
3196 ParallelCompactData& sd = PSParallelCompact::summary_data();
3197 const size_t region_size = ParallelCompactData::RegionSize;
3198
3199 size_t src_region_idx = 0;
3200
3201 // Skip empty regions (if any) up to the top of the space.
3202 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3203 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3204 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3205 const RegionData* const top_region_ptr =
3206 sd.addr_to_region_ptr(top_aligned_up);
3207 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3208 ++src_region_ptr;
3209 }
3210
3211 if (src_region_ptr < top_region_ptr) {
3212 // The next source region is in the current space. Update src_region_idx
3213 // and the source address to match src_region_ptr.
3214 src_region_idx = sd.region(src_region_ptr);
3215 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3216 if (src_region_addr > closure.source()) {
3217 closure.set_source(src_region_addr);
3218 }
3219 return src_region_idx;
3220 }
3221
3222 // Switch to a new source space and find the first non-empty region.
3223 unsigned int space_id = src_space_id + 1;
3224 assert(space_id < last_space_id, "not enough spaces");
3225
3226 HeapWord* const destination = closure.destination();
3227
3228 do {
3229 MutableSpace* space = _space_info[space_id].space();
3230 HeapWord* const bottom = space->bottom();
3231 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3232
3233 // Iterate over the spaces that do not compact into themselves.
3234 if (bottom_cp->destination() != bottom) {
3235 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3236 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3237
3238 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3239 if (src_cp->live_obj_size() > 0) {
3240 // Found it.
3241 assert(src_cp->destination() == destination,
3242 "first live obj in the space must match the destination");
3243 assert(src_cp->partial_obj_size() == 0,
3244 "a space cannot begin with a partial obj");
3245
3246 src_space_id = SpaceId(space_id);
3247 src_space_top = space->top();
3248 const size_t src_region_idx = sd.region(src_cp);
3249 closure.set_source(sd.region_to_addr(src_region_idx));
3250 return src_region_idx;
3251 } else {
3252 assert(src_cp->data_size() == 0, "sanity");
3253 }
3254 }
3255 }
3256 } while (++space_id < last_space_id);
3257
3258 assert(false, "no source region was found");
3259 return 0;
3260 }
3261
3262 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3263 {
3264 typedef ParMarkBitMap::IterationStatus IterationStatus;
3265 const size_t RegionSize = ParallelCompactData::RegionSize;
3266 ParMarkBitMap* const bitmap = mark_bitmap();
3267 ParallelCompactData& sd = summary_data();
3268 RegionData* const region_ptr = sd.region(region_idx);
3269
3270 // Get the items needed to construct the closure.
3271 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3272 SpaceId dest_space_id = space_id(dest_addr);
3273 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3274 HeapWord* new_top = _space_info[dest_space_id].new_top();
3275 assert(dest_addr < new_top, "sanity");
3276 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3277
3278 // Get the source region and related info.
3279 size_t src_region_idx = region_ptr->source_region();
3280 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3281 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3282
3283 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3284 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3285
3286 // Adjust src_region_idx to prepare for decrementing destination counts (the
3287 // destination count is not decremented when a region is copied to itself).
3288 if (src_region_idx == region_idx) {
3289 src_region_idx += 1;
3290 }
3291
3292 if (bitmap->is_unmarked(closure.source())) {
3293 // The first source word is in the middle of an object; copy the remainder
3294 // of the object or as much as will fit. The fact that pointer updates were
3295 // deferred will be noted when the object header is processed.
3296 HeapWord* const old_src_addr = closure.source();
3297 closure.copy_partial_obj();
3298 if (closure.is_full()) {
3299 decrement_destination_counts(cm, src_space_id, src_region_idx,
3300 closure.source());
3301 region_ptr->set_deferred_obj_addr(NULL);
3302 region_ptr->set_completed();
3303 return;
3304 }
3305
3306 HeapWord* const end_addr = sd.region_align_down(closure.source());
3307 if (sd.region_align_down(old_src_addr) != end_addr) {
3308 // The partial object was copied from more than one source region.
3309 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3310
3311 // Move to the next source region, possibly switching spaces as well. All
3312 // args except end_addr may be modified.
3313 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3314 end_addr);
3315 }
3316 }
3317
3318 do {
3319 HeapWord* const cur_addr = closure.source();
3320 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3321 src_space_top);
3322 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3323
3324 if (status == ParMarkBitMap::incomplete) {
3325 // The last obj that starts in the source region does not end in the
3326 // region.
3327 assert(closure.source() < end_addr, "sanity");
3328 HeapWord* const obj_beg = closure.source();
3329 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3330 src_space_top);
3331 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3332 if (obj_end < range_end) {
3333 // The end was found; the entire object will fit.
3334 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3335 assert(status != ParMarkBitMap::would_overflow, "sanity");
3336 } else {
3337 // The end was not found; the object will not fit.
3338 assert(range_end < src_space_top, "obj cannot cross space boundary");
3339 status = ParMarkBitMap::would_overflow;
3340 }
3341 }
3342
3343 if (status == ParMarkBitMap::would_overflow) {
3344 // The last object did not fit. Note that interior oop updates were
3345 // deferred, then copy enough of the object to fill the region.
3346 region_ptr->set_deferred_obj_addr(closure.destination());
3347 status = closure.copy_until_full(); // copies from closure.source()
3348
3349 decrement_destination_counts(cm, src_space_id, src_region_idx,
3350 closure.source());
3351 region_ptr->set_completed();
3352 return;
3353 }
3354
3355 if (status == ParMarkBitMap::full) {
3356 decrement_destination_counts(cm, src_space_id, src_region_idx,
3357 closure.source());
3358 region_ptr->set_deferred_obj_addr(NULL);
3359 region_ptr->set_completed();
3360 return;
3361 }
3362
3363 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3364
3365 // Move to the next source region, possibly switching spaces as well. All
3366 // args except end_addr may be modified.
3367 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3368 end_addr);
3369 } while (true);
3370 }
3371
3372 void
3373 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3374 const MutableSpace* sp = space(space_id);
3375 if (sp->is_empty()) {
3376 return;
3377 }
3378
3379 ParallelCompactData& sd = PSParallelCompact::summary_data();
3380 ParMarkBitMap* const bitmap = mark_bitmap();
3381 HeapWord* const dp_addr = dense_prefix(space_id);
3382 HeapWord* beg_addr = sp->bottom();
3383 HeapWord* end_addr = sp->top();
3384
3385 #ifdef ASSERT
3386 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3387 if (cm->should_verify_only()) {
3388 VerifyUpdateClosure verify_update(cm, sp);
3389 bitmap->iterate(&verify_update, beg_addr, end_addr);
3390 return;
3391 }
3392
3393 if (cm->should_reset_only()) {
3394 ResetObjectsClosure reset_objects(cm);
3395 bitmap->iterate(&reset_objects, beg_addr, end_addr);
3396 return;
3397 }
3398 #endif
3399
3400 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3401 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3402 if (beg_region < dp_region) {
3403 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3404 }
3405
3406 // The destination of the first live object that starts in the region is one
3407 // past the end of the partial object entering the region (if any).
3408 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3409 HeapWord* const new_top = _space_info[space_id].new_top();
3410 assert(new_top >= dest_addr, "bad new_top value");
3411 const size_t words = pointer_delta(new_top, dest_addr);
3412
3413 if (words > 0) {
3414 ObjectStartArray* start_array = _space_info[space_id].start_array();
3415 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3416
3417 ParMarkBitMap::IterationStatus status;
3418 status = bitmap->iterate(&closure, dest_addr, end_addr);
3419 assert(status == ParMarkBitMap::full, "iteration not complete");
3420 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3421 "live objects skipped because closure is full");
3422 }
3423 }
3424
3425 jlong PSParallelCompact::millis_since_last_gc() {
3426 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
3427 // XXX See note in genCollectedHeap::millis_since_last_gc().
3428 if (ret_val < 0) {
3429 NOT_PRODUCT(warning("time warp: %d", ret_val);)
3430 return 0;
3431 }
3432 return ret_val;
3433 }
3434
3435 void PSParallelCompact::reset_millis_since_last_gc() {
3436 _time_of_last_gc = os::javaTimeMillis();
3437 }
3438
3439 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3440 {
3441 if (source() != destination()) {
3442 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3443 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3444 }
3445 update_state(words_remaining());
3446 assert(is_full(), "sanity");
3447 return ParMarkBitMap::full;
3448 }
3449
3450 void MoveAndUpdateClosure::copy_partial_obj()
3451 {
3452 size_t words = words_remaining();
3453
3454 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3455 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3456 if (end_addr < range_end) {
3457 words = bitmap()->obj_size(source(), end_addr);
3458 }
3459
3460 // This test is necessary; if omitted, the pointer updates to a partial object
3461 // that crosses the dense prefix boundary could be overwritten.
3462 if (source() != destination()) {
3463 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3464 Copy::aligned_conjoint_words(source(), destination(), words);
3465 }
3466 update_state(words);
3467 }
3468
3469 ParMarkBitMapClosure::IterationStatus
3470 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3471 assert(destination() != NULL, "sanity");
3472 assert(bitmap()->obj_size(addr) == words, "bad size");
3473
3474 _source = addr;
3475 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3476 destination(), "wrong destination");
3477
3478 if (words > words_remaining()) {
3479 return ParMarkBitMap::would_overflow;
3480 }
3481
3482 // The start_array must be updated even if the object is not moving.
3483 if (_start_array != NULL) {
3484 _start_array->allocate_block(destination());
3485 }
3486
3487 if (destination() != source()) {
3488 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3489 Copy::aligned_conjoint_words(source(), destination(), words);
3490 }
3491
3492 oop moved_oop = (oop) destination();
3493 moved_oop->update_contents(compaction_manager());
3494 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3495
3496 update_state(words);
3497 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3498 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3499 }
3500
3501 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3502 ParCompactionManager* cm,
3503 PSParallelCompact::SpaceId space_id) :
3504 ParMarkBitMapClosure(mbm, cm),
3505 _space_id(space_id),
3506 _start_array(PSParallelCompact::start_array(space_id))
3507 {
3508 }
3509
3510 // Updates the references in the object to their new values.
3511 ParMarkBitMapClosure::IterationStatus
3512 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3513 do_addr(addr);
3514 return ParMarkBitMap::incomplete;
3515 }
3516
3517 // Verify the new location using the forwarding pointer
3518 // from MarkSweep::mark_sweep_phase2(). Set the mark_word
3519 // to the initial value.
3520 ParMarkBitMapClosure::IterationStatus
3521 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3522 // The second arg (words) is not used.
3523 oop obj = (oop) addr;
3524 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
3525 HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
3526 if (forwarding_ptr == NULL) {
3527 // The object is dead or not moving.
3528 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
3529 "Object liveness is wrong.");
3530 return ParMarkBitMap::incomplete;
3531 }
3532 assert(UseParallelOldGCDensePrefix ||
3533 (HeapMaximumCompactionInterval > 1) ||
3534 (MarkSweepAlwaysCompactCount > 1) ||
3535 (forwarding_ptr == new_pointer),
3536 "Calculation of new location is incorrect");
3537 return ParMarkBitMap::incomplete;
3538 }
3539
3540 // Reset objects modified for debug checking.
3541 ParMarkBitMapClosure::IterationStatus
3542 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
3543 // The second arg (words) is not used.
3544 oop obj = (oop) addr;
3545 obj->init_mark();
3546 return ParMarkBitMap::incomplete;
3547 }
3548
3549 // Prepare for compaction. This method is executed once
3550 // (i.e., by a single thread) before compaction.
3551 // Save the updated location of the intArrayKlassObj for
3552 // filling holes in the dense prefix.
3553 void PSParallelCompact::compact_prologue() {
3554 _updated_int_array_klass_obj = (klassOop)
3555 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3556 }
3557