1 /*
2 * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
30 #include "gc_implementation/parallelScavenge/generationSizer.hpp"
31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
32 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
37 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
39 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
40 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
41 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/gcCause.hpp"
44 #include "memory/gcLocker.inline.hpp"
45 #include "memory/referencePolicy.hpp"
46 #include "memory/referenceProcessor.hpp"
47 #include "oops/methodData.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "oops/oop.pcgc.inline.hpp"
50 #include "runtime/fprofiler.hpp"
51 #include "runtime/safepoint.hpp"
52 #include "runtime/vmThread.hpp"
53 #include "services/management.hpp"
54 #include "services/memoryService.hpp"
55 #include "services/memTracker.hpp"
56 #include "utilities/events.hpp"
57 #include "utilities/stack.inline.hpp"
58
59 #include <math.h>
60
61 // All sizes are in HeapWords.
62 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words
63 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
64 const size_t ParallelCompactData::RegionSizeBytes =
65 RegionSize << LogHeapWordSize;
66 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
67 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
68 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
69
70 const ParallelCompactData::RegionData::region_sz_t
71 ParallelCompactData::RegionData::dc_shift = 27;
72
73 const ParallelCompactData::RegionData::region_sz_t
74 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
75
76 const ParallelCompactData::RegionData::region_sz_t
77 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
78
79 const ParallelCompactData::RegionData::region_sz_t
80 ParallelCompactData::RegionData::los_mask = ~dc_mask;
81
82 const ParallelCompactData::RegionData::region_sz_t
83 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
84
85 const ParallelCompactData::RegionData::region_sz_t
86 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
87
88 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
89 bool PSParallelCompact::_print_phases = false;
90
91 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
92 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL;
93
94 double PSParallelCompact::_dwl_mean;
95 double PSParallelCompact::_dwl_std_dev;
96 double PSParallelCompact::_dwl_first_term;
97 double PSParallelCompact::_dwl_adjustment;
98 #ifdef ASSERT
99 bool PSParallelCompact::_dwl_initialized = false;
100 #endif // #ifdef ASSERT
101
102 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
103 HeapWord* destination)
104 {
105 assert(src_region_idx != 0, "invalid src_region_idx");
106 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
107 assert(destination != NULL, "invalid destination argument");
108
109 _src_region_idx = src_region_idx;
110 _partial_obj_size = partial_obj_size;
111 _destination = destination;
112
113 // These fields may not be updated below, so make sure they're clear.
114 assert(_dest_region_addr == NULL, "should have been cleared");
115 assert(_first_src_addr == NULL, "should have been cleared");
116
117 // Determine the number of destination regions for the partial object.
118 HeapWord* const last_word = destination + partial_obj_size - 1;
119 const ParallelCompactData& sd = PSParallelCompact::summary_data();
120 HeapWord* const beg_region_addr = sd.region_align_down(destination);
121 HeapWord* const end_region_addr = sd.region_align_down(last_word);
122
123 if (beg_region_addr == end_region_addr) {
124 // One destination region.
125 _destination_count = 1;
126 if (end_region_addr == destination) {
127 // The destination falls on a region boundary, thus the first word of the
128 // partial object will be the first word copied to the destination region.
129 _dest_region_addr = end_region_addr;
130 _first_src_addr = sd.region_to_addr(src_region_idx);
131 }
132 } else {
133 // Two destination regions. When copied, the partial object will cross a
134 // destination region boundary, so a word somewhere within the partial
135 // object will be the first word copied to the second destination region.
136 _destination_count = 2;
137 _dest_region_addr = end_region_addr;
138 const size_t ofs = pointer_delta(end_region_addr, destination);
139 assert(ofs < _partial_obj_size, "sanity");
140 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
141 }
142 }
143
144 void SplitInfo::clear()
145 {
146 _src_region_idx = 0;
147 _partial_obj_size = 0;
148 _destination = NULL;
149 _destination_count = 0;
150 _dest_region_addr = NULL;
151 _first_src_addr = NULL;
152 assert(!is_valid(), "sanity");
153 }
154
155 #ifdef ASSERT
156 void SplitInfo::verify_clear()
157 {
158 assert(_src_region_idx == 0, "not clear");
159 assert(_partial_obj_size == 0, "not clear");
160 assert(_destination == NULL, "not clear");
161 assert(_destination_count == 0, "not clear");
162 assert(_dest_region_addr == NULL, "not clear");
163 assert(_first_src_addr == NULL, "not clear");
164 }
165 #endif // #ifdef ASSERT
166
167
168 void PSParallelCompact::print_on_error(outputStream* st) {
169 _mark_bitmap.print_on_error(st);
170 }
171
172 #ifndef PRODUCT
173 const char* PSParallelCompact::space_names[] = {
174 "old ", "eden", "from", "to "
175 };
176
177 void PSParallelCompact::print_region_ranges()
178 {
179 tty->print_cr("space bottom top end new_top");
180 tty->print_cr("------ ---------- ---------- ---------- ----------");
181
182 for (unsigned int id = 0; id < last_space_id; ++id) {
183 const MutableSpace* space = _space_info[id].space();
184 tty->print_cr("%u %s "
185 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
186 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
187 id, space_names[id],
188 summary_data().addr_to_region_idx(space->bottom()),
189 summary_data().addr_to_region_idx(space->top()),
190 summary_data().addr_to_region_idx(space->end()),
191 summary_data().addr_to_region_idx(_space_info[id].new_top()));
192 }
193 }
194
195 void
196 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
197 {
198 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
199 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
200
201 ParallelCompactData& sd = PSParallelCompact::summary_data();
202 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
203 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
204 REGION_IDX_FORMAT " " PTR_FORMAT " "
205 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
206 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
207 i, c->data_location(), dci, c->destination(),
208 c->partial_obj_size(), c->live_obj_size(),
209 c->data_size(), c->source_region(), c->destination_count());
210
211 #undef REGION_IDX_FORMAT
212 #undef REGION_DATA_FORMAT
213 }
214
215 void
216 print_generic_summary_data(ParallelCompactData& summary_data,
217 HeapWord* const beg_addr,
218 HeapWord* const end_addr)
219 {
220 size_t total_words = 0;
221 size_t i = summary_data.addr_to_region_idx(beg_addr);
222 const size_t last = summary_data.addr_to_region_idx(end_addr);
223 HeapWord* pdest = 0;
224
225 while (i <= last) {
226 ParallelCompactData::RegionData* c = summary_data.region(i);
227 if (c->data_size() != 0 || c->destination() != pdest) {
228 print_generic_summary_region(i, c);
229 total_words += c->data_size();
230 pdest = c->destination();
231 }
232 ++i;
233 }
234
235 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
236 }
237
238 void
239 print_generic_summary_data(ParallelCompactData& summary_data,
240 SpaceInfo* space_info)
241 {
242 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
243 const MutableSpace* space = space_info[id].space();
244 print_generic_summary_data(summary_data, space->bottom(),
245 MAX2(space->top(), space_info[id].new_top()));
246 }
247 }
248
249 void
250 print_initial_summary_region(size_t i,
251 const ParallelCompactData::RegionData* c,
252 bool newline = true)
253 {
254 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
255 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
256 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
257 i, c->destination(),
258 c->partial_obj_size(), c->live_obj_size(),
259 c->data_size(), c->source_region(), c->destination_count());
260 if (newline) tty->cr();
261 }
262
263 void
264 print_initial_summary_data(ParallelCompactData& summary_data,
265 const MutableSpace* space) {
266 if (space->top() == space->bottom()) {
267 return;
268 }
269
270 const size_t region_size = ParallelCompactData::RegionSize;
271 typedef ParallelCompactData::RegionData RegionData;
272 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
273 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
274 const RegionData* c = summary_data.region(end_region - 1);
275 HeapWord* end_addr = c->destination() + c->data_size();
276 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
277
278 // Print (and count) the full regions at the beginning of the space.
279 size_t full_region_count = 0;
280 size_t i = summary_data.addr_to_region_idx(space->bottom());
281 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
282 print_initial_summary_region(i, summary_data.region(i));
283 ++full_region_count;
284 ++i;
285 }
286
287 size_t live_to_right = live_in_space - full_region_count * region_size;
288
289 double max_reclaimed_ratio = 0.0;
290 size_t max_reclaimed_ratio_region = 0;
291 size_t max_dead_to_right = 0;
292 size_t max_live_to_right = 0;
293
294 // Print the 'reclaimed ratio' for regions while there is something live in
295 // the region or to the right of it. The remaining regions are empty (and
296 // uninteresting), and computing the ratio will result in division by 0.
297 while (i < end_region && live_to_right > 0) {
298 c = summary_data.region(i);
299 HeapWord* const region_addr = summary_data.region_to_addr(i);
300 const size_t used_to_right = pointer_delta(space->top(), region_addr);
301 const size_t dead_to_right = used_to_right - live_to_right;
302 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
303
304 if (reclaimed_ratio > max_reclaimed_ratio) {
305 max_reclaimed_ratio = reclaimed_ratio;
306 max_reclaimed_ratio_region = i;
307 max_dead_to_right = dead_to_right;
308 max_live_to_right = live_to_right;
309 }
310
311 print_initial_summary_region(i, c, false);
312 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
313 reclaimed_ratio, dead_to_right, live_to_right);
314
315 live_to_right -= c->data_size();
316 ++i;
317 }
318
319 // Any remaining regions are empty. Print one more if there is one.
320 if (i < end_region) {
321 print_initial_summary_region(i, summary_data.region(i));
322 }
323
324 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
325 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
326 max_reclaimed_ratio_region, max_dead_to_right,
327 max_live_to_right, max_reclaimed_ratio);
328 }
329
330 void
331 print_initial_summary_data(ParallelCompactData& summary_data,
332 SpaceInfo* space_info) {
333 unsigned int id = PSParallelCompact::old_space_id;
334 const MutableSpace* space;
335 do {
336 space = space_info[id].space();
337 print_initial_summary_data(summary_data, space);
338 } while (++id < PSParallelCompact::eden_space_id);
339
340 do {
341 space = space_info[id].space();
342 print_generic_summary_data(summary_data, space->bottom(), space->top());
343 } while (++id < PSParallelCompact::last_space_id);
344 }
345 #endif // #ifndef PRODUCT
346
347 #ifdef ASSERT
348 size_t add_obj_count;
349 size_t add_obj_size;
350 size_t mark_bitmap_count;
351 size_t mark_bitmap_size;
352 #endif // #ifdef ASSERT
353
354 ParallelCompactData::ParallelCompactData()
355 {
356 _region_start = 0;
357
358 _region_vspace = 0;
359 _reserved_byte_size = 0;
360 _region_data = 0;
361 _region_count = 0;
362 }
363
364 bool ParallelCompactData::initialize(MemRegion covered_region)
365 {
366 _region_start = covered_region.start();
367 const size_t region_size = covered_region.word_size();
368 DEBUG_ONLY(_region_end = _region_start + region_size;)
369
370 assert(region_align_down(_region_start) == _region_start,
371 "region start not aligned");
372 assert((region_size & RegionSizeOffsetMask) == 0,
373 "region size not a multiple of RegionSize");
374
375 bool result = initialize_region_data(region_size);
376
377 return result;
378 }
379
380 PSVirtualSpace*
381 ParallelCompactData::create_vspace(size_t count, size_t element_size)
382 {
383 const size_t raw_bytes = count * element_size;
384 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
385 const size_t granularity = os::vm_allocation_granularity();
386 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
387
388 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
389 MAX2(page_sz, granularity);
390 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
391 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
392 rs.size());
393
394 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
395
396 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
397 if (vspace != 0) {
398 if (vspace->expand_by(_reserved_byte_size)) {
399 return vspace;
400 }
401 delete vspace;
402 // Release memory reserved in the space.
403 rs.release();
404 }
405
406 return 0;
407 }
408
409 bool ParallelCompactData::initialize_region_data(size_t region_size)
410 {
411 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
412 _region_vspace = create_vspace(count, sizeof(RegionData));
413 if (_region_vspace != 0) {
414 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
415 _region_count = count;
416 return true;
417 }
418 return false;
419 }
420
421 void ParallelCompactData::clear()
422 {
423 memset(_region_data, 0, _region_vspace->committed_size());
424 }
425
426 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
427 assert(beg_region <= _region_count, "beg_region out of range");
428 assert(end_region <= _region_count, "end_region out of range");
429
430 const size_t region_cnt = end_region - beg_region;
431 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
432 }
433
434 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
435 {
436 const RegionData* cur_cp = region(region_idx);
437 const RegionData* const end_cp = region(region_count() - 1);
438
439 HeapWord* result = region_to_addr(region_idx);
440 if (cur_cp < end_cp) {
441 do {
442 result += cur_cp->partial_obj_size();
443 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
444 }
445 return result;
446 }
447
448 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
449 {
450 const size_t obj_ofs = pointer_delta(addr, _region_start);
451 const size_t beg_region = obj_ofs >> Log2RegionSize;
452 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
453
454 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
455 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
456
457 if (beg_region == end_region) {
458 // All in one region.
459 _region_data[beg_region].add_live_obj(len);
460 return;
461 }
462
463 // First region.
464 const size_t beg_ofs = region_offset(addr);
465 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
466
467 Klass* klass = ((oop)addr)->klass();
468 // Middle regions--completely spanned by this object.
469 for (size_t region = beg_region + 1; region < end_region; ++region) {
470 _region_data[region].set_partial_obj_size(RegionSize);
471 _region_data[region].set_partial_obj_addr(addr);
472 }
473
474 // Last region.
475 const size_t end_ofs = region_offset(addr + len - 1);
476 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
477 _region_data[end_region].set_partial_obj_addr(addr);
478 }
479
480 void
481 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
482 {
483 assert(region_offset(beg) == 0, "not RegionSize aligned");
484 assert(region_offset(end) == 0, "not RegionSize aligned");
485
486 size_t cur_region = addr_to_region_idx(beg);
487 const size_t end_region = addr_to_region_idx(end);
488 HeapWord* addr = beg;
489 while (cur_region < end_region) {
490 _region_data[cur_region].set_destination(addr);
491 _region_data[cur_region].set_destination_count(0);
492 _region_data[cur_region].set_source_region(cur_region);
493 _region_data[cur_region].set_data_location(addr);
494
495 // Update live_obj_size so the region appears completely full.
496 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
497 _region_data[cur_region].set_live_obj_size(live_size);
498
499 ++cur_region;
500 addr += RegionSize;
501 }
502 }
503
504 // Find the point at which a space can be split and, if necessary, record the
505 // split point.
506 //
507 // If the current src region (which overflowed the destination space) doesn't
508 // have a partial object, the split point is at the beginning of the current src
509 // region (an "easy" split, no extra bookkeeping required).
510 //
511 // If the current src region has a partial object, the split point is in the
512 // region where that partial object starts (call it the split_region). If
513 // split_region has a partial object, then the split point is just after that
514 // partial object (a "hard" split where we have to record the split data and
515 // zero the partial_obj_size field). With a "hard" split, we know that the
516 // partial_obj ends within split_region because the partial object that caused
517 // the overflow starts in split_region. If split_region doesn't have a partial
518 // obj, then the split is at the beginning of split_region (another "easy"
519 // split).
520 HeapWord*
521 ParallelCompactData::summarize_split_space(size_t src_region,
522 SplitInfo& split_info,
523 HeapWord* destination,
524 HeapWord* target_end,
525 HeapWord** target_next)
526 {
527 assert(destination <= target_end, "sanity");
528 assert(destination + _region_data[src_region].data_size() > target_end,
529 "region should not fit into target space");
530 assert(is_region_aligned(target_end), "sanity");
531
532 size_t split_region = src_region;
533 HeapWord* split_destination = destination;
534 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
535
536 if (destination + partial_obj_size > target_end) {
537 // The split point is just after the partial object (if any) in the
538 // src_region that contains the start of the object that overflowed the
539 // destination space.
540 //
541 // Find the start of the "overflow" object and set split_region to the
542 // region containing it.
543 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
544 split_region = addr_to_region_idx(overflow_obj);
545
546 // Clear the source_region field of all destination regions whose first word
547 // came from data after the split point (a non-null source_region field
548 // implies a region must be filled).
549 //
550 // An alternative to the simple loop below: clear during post_compact(),
551 // which uses memcpy instead of individual stores, and is easy to
552 // parallelize. (The downside is that it clears the entire RegionData
553 // object as opposed to just one field.)
554 //
555 // post_compact() would have to clear the summary data up to the highest
556 // address that was written during the summary phase, which would be
557 //
558 // max(top, max(new_top, clear_top))
559 //
560 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
561 // to target_end.
562 const RegionData* const sr = region(split_region);
563 const size_t beg_idx =
564 addr_to_region_idx(region_align_up(sr->destination() +
565 sr->partial_obj_size()));
566 const size_t end_idx = addr_to_region_idx(target_end);
567
568 if (TraceParallelOldGCSummaryPhase) {
569 gclog_or_tty->print_cr("split: clearing source_region field in ["
570 SIZE_FORMAT ", " SIZE_FORMAT ")",
571 beg_idx, end_idx);
572 }
573 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
574 _region_data[idx].set_source_region(0);
575 }
576
577 // Set split_destination and partial_obj_size to reflect the split region.
578 split_destination = sr->destination();
579 partial_obj_size = sr->partial_obj_size();
580 }
581
582 // The split is recorded only if a partial object extends onto the region.
583 if (partial_obj_size != 0) {
584 _region_data[split_region].set_partial_obj_size(0);
585 split_info.record(split_region, partial_obj_size, split_destination);
586 }
587
588 // Setup the continuation addresses.
589 *target_next = split_destination + partial_obj_size;
590 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
591
592 if (TraceParallelOldGCSummaryPhase) {
593 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
594 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
595 " pos=" SIZE_FORMAT,
596 split_type, source_next, split_region,
597 partial_obj_size);
598 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
599 " tn=" PTR_FORMAT,
600 split_type, split_destination,
601 addr_to_region_idx(split_destination),
602 *target_next);
603
604 if (partial_obj_size != 0) {
605 HeapWord* const po_beg = split_info.destination();
606 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
607 gclog_or_tty->print_cr("%s split: "
608 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
609 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
610 split_type,
611 po_beg, addr_to_region_idx(po_beg),
612 po_end, addr_to_region_idx(po_end));
613 }
614 }
615
616 return source_next;
617 }
618
619 bool ParallelCompactData::summarize(SplitInfo& split_info,
620 HeapWord* source_beg, HeapWord* source_end,
621 HeapWord** source_next,
622 HeapWord* target_beg, HeapWord* target_end,
623 HeapWord** target_next)
624 {
625 if (TraceParallelOldGCSummaryPhase) {
626 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
627 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
628 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
629 source_beg, source_end, source_next_val,
630 target_beg, target_end, *target_next);
631 }
632
633 size_t cur_region = addr_to_region_idx(source_beg);
634 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
635
636 HeapWord *dest_addr = target_beg;
637 while (cur_region < end_region) {
638 // The destination must be set even if the region has no data.
639 _region_data[cur_region].set_destination(dest_addr);
640
641 size_t words = _region_data[cur_region].data_size();
642 if (words > 0) {
643 // If cur_region does not fit entirely into the target space, find a point
644 // at which the source space can be 'split' so that part is copied to the
645 // target space and the rest is copied elsewhere.
646 if (dest_addr + words > target_end) {
647 assert(source_next != NULL, "source_next is NULL when splitting");
648 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
649 target_end, target_next);
650 return false;
651 }
652
653 // Compute the destination_count for cur_region, and if necessary, update
654 // source_region for a destination region. The source_region field is
655 // updated if cur_region is the first (left-most) region to be copied to a
656 // destination region.
657 //
658 // The destination_count calculation is a bit subtle. A region that has
659 // data that compacts into itself does not count itself as a destination.
660 // This maintains the invariant that a zero count means the region is
661 // available and can be claimed and then filled.
662 uint destination_count = 0;
663 if (split_info.is_split(cur_region)) {
664 // The current region has been split: the partial object will be copied
665 // to one destination space and the remaining data will be copied to
666 // another destination space. Adjust the initial destination_count and,
667 // if necessary, set the source_region field if the partial object will
668 // cross a destination region boundary.
669 destination_count = split_info.destination_count();
670 if (destination_count == 2) {
671 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
672 _region_data[dest_idx].set_source_region(cur_region);
673 }
674 }
675
676 HeapWord* const last_addr = dest_addr + words - 1;
677 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
678 const size_t dest_region_2 = addr_to_region_idx(last_addr);
679
680 // Initially assume that the destination regions will be the same and
681 // adjust the value below if necessary. Under this assumption, if
682 // cur_region == dest_region_2, then cur_region will be compacted
683 // completely into itself.
684 destination_count += cur_region == dest_region_2 ? 0 : 1;
685 if (dest_region_1 != dest_region_2) {
686 // Destination regions differ; adjust destination_count.
687 destination_count += 1;
688 // Data from cur_region will be copied to the start of dest_region_2.
689 _region_data[dest_region_2].set_source_region(cur_region);
690 } else if (region_offset(dest_addr) == 0) {
691 // Data from cur_region will be copied to the start of the destination
692 // region.
693 _region_data[dest_region_1].set_source_region(cur_region);
694 }
695
696 _region_data[cur_region].set_destination_count(destination_count);
697 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
698 dest_addr += words;
699 }
700
701 ++cur_region;
702 }
703
704 *target_next = dest_addr;
705 return true;
706 }
707
708 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
709 assert(addr != NULL, "Should detect NULL oop earlier");
710 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
711 #ifdef ASSERT
712 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
713 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
714 }
715 #endif
716 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
717
718 // Region covering the object.
719 size_t region_index = addr_to_region_idx(addr);
720 const RegionData* const region_ptr = region(region_index);
721 HeapWord* const region_addr = region_align_down(addr);
722
723 assert(addr < region_addr + RegionSize, "Region does not cover object");
724 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
725
726 HeapWord* result = region_ptr->destination();
727
728 // If all the data in the region is live, then the new location of the object
729 // can be calculated from the destination of the region plus the offset of the
730 // object in the region.
731 if (region_ptr->data_size() == RegionSize) {
732 result += pointer_delta(addr, region_addr);
733 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
734 return result;
735 }
736
737 // The new location of the object is
738 // region destination +
739 // size of the partial object extending onto the region +
740 // sizes of the live objects in the Region that are to the left of addr
741 const size_t partial_obj_size = region_ptr->partial_obj_size();
742 HeapWord* const search_start = region_addr + partial_obj_size;
743
744 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
745 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
746
747 result += partial_obj_size + live_to_left;
748 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
749 return result;
750 }
751
752 #ifdef ASSERT
753 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
754 {
755 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
756 const size_t* const end = (const size_t*)vspace->committed_high_addr();
757 for (const size_t* p = beg; p < end; ++p) {
758 assert(*p == 0, "not zero");
759 }
760 }
761
762 void ParallelCompactData::verify_clear()
763 {
764 verify_clear(_region_vspace);
765 }
766 #endif // #ifdef ASSERT
767
768 #ifdef NOT_PRODUCT
769 ParallelCompactData::RegionData* debug_region(size_t region_index) {
770 ParallelCompactData& sd = PSParallelCompact::summary_data();
771 return sd.region(region_index);
772 }
773 #endif
774
775 elapsedTimer PSParallelCompact::_accumulated_time;
776 unsigned int PSParallelCompact::_total_invocations = 0;
777 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
778 jlong PSParallelCompact::_time_of_last_gc = 0;
779 CollectorCounters* PSParallelCompact::_counters = NULL;
780 ParMarkBitMap PSParallelCompact::_mark_bitmap;
781 ParallelCompactData PSParallelCompact::_summary_data;
782
783 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
784
785 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
786
787 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
788 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
789
790 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
791 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
792
793 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); }
794 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
795
796 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
797
798 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) {
799 mark_and_push(_compaction_manager, p);
800 }
801 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
802
803 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
804 klass->oops_do(_mark_and_push_closure);
805 }
806 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
807 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
808 }
809
810 void PSParallelCompact::post_initialize() {
811 ParallelScavengeHeap* heap = gc_heap();
812 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
813
814 MemRegion mr = heap->reserved_region();
815 _ref_processor =
816 new ReferenceProcessor(mr, // span
817 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
818 (int) ParallelGCThreads, // mt processing degree
819 true, // mt discovery
820 (int) ParallelGCThreads, // mt discovery degree
821 true, // atomic_discovery
822 &_is_alive_closure, // non-header is alive closure
823 false); // write barrier for next field updates
824 _counters = new CollectorCounters("PSParallelCompact", 1);
825
826 // Initialize static fields in ParCompactionManager.
827 ParCompactionManager::initialize(mark_bitmap());
828 }
829
830 bool PSParallelCompact::initialize() {
831 ParallelScavengeHeap* heap = gc_heap();
832 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
833 MemRegion mr = heap->reserved_region();
834
835 // Was the old gen get allocated successfully?
836 if (!heap->old_gen()->is_allocated()) {
837 return false;
838 }
839
840 initialize_space_info();
841 initialize_dead_wood_limiter();
842
843 if (!_mark_bitmap.initialize(mr)) {
844 vm_shutdown_during_initialization(
845 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
846 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
847 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
848 return false;
849 }
850
851 if (!_summary_data.initialize(mr)) {
852 vm_shutdown_during_initialization(
853 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
854 "garbage collection for the requested " SIZE_FORMAT "KB heap.",
855 _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
856 return false;
857 }
858
859 return true;
860 }
861
862 void PSParallelCompact::initialize_space_info()
863 {
864 memset(&_space_info, 0, sizeof(_space_info));
865
866 ParallelScavengeHeap* heap = gc_heap();
867 PSYoungGen* young_gen = heap->young_gen();
868
869 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
870 _space_info[eden_space_id].set_space(young_gen->eden_space());
871 _space_info[from_space_id].set_space(young_gen->from_space());
872 _space_info[to_space_id].set_space(young_gen->to_space());
873
874 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
875 }
876
877 void PSParallelCompact::initialize_dead_wood_limiter()
878 {
879 const size_t max = 100;
880 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
881 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
882 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
883 DEBUG_ONLY(_dwl_initialized = true;)
884 _dwl_adjustment = normal_distribution(1.0);
885 }
886
887 // Simple class for storing info about the heap at the start of GC, to be used
888 // after GC for comparison/printing.
889 class PreGCValues {
890 public:
891 PreGCValues() { }
892 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
893
894 void fill(ParallelScavengeHeap* heap) {
895 _heap_used = heap->used();
896 _young_gen_used = heap->young_gen()->used_in_bytes();
897 _old_gen_used = heap->old_gen()->used_in_bytes();
898 _metadata_used = MetaspaceAux::allocated_used_bytes();
899 };
900
901 size_t heap_used() const { return _heap_used; }
902 size_t young_gen_used() const { return _young_gen_used; }
903 size_t old_gen_used() const { return _old_gen_used; }
904 size_t metadata_used() const { return _metadata_used; }
905
906 private:
907 size_t _heap_used;
908 size_t _young_gen_used;
909 size_t _old_gen_used;
910 size_t _metadata_used;
911 };
912
913 void
914 PSParallelCompact::clear_data_covering_space(SpaceId id)
915 {
916 // At this point, top is the value before GC, new_top() is the value that will
917 // be set at the end of GC. The marking bitmap is cleared to top; nothing
918 // should be marked above top. The summary data is cleared to the larger of
919 // top & new_top.
920 MutableSpace* const space = _space_info[id].space();
921 HeapWord* const bot = space->bottom();
922 HeapWord* const top = space->top();
923 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
924
925 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
926 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
927 _mark_bitmap.clear_range(beg_bit, end_bit);
928
929 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
930 const size_t end_region =
931 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
932 _summary_data.clear_range(beg_region, end_region);
933
934 // Clear the data used to 'split' regions.
935 SplitInfo& split_info = _space_info[id].split_info();
936 if (split_info.is_valid()) {
937 split_info.clear();
938 }
939 DEBUG_ONLY(split_info.verify_clear();)
940 }
941
942 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
943 {
944 // Update the from & to space pointers in space_info, since they are swapped
945 // at each young gen gc. Do the update unconditionally (even though a
946 // promotion failure does not swap spaces) because an unknown number of minor
947 // collections will have swapped the spaces an unknown number of times.
948 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
949 ParallelScavengeHeap* heap = gc_heap();
950 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
951 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
952
953 pre_gc_values->fill(heap);
954
955 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
956 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
957
958 // Increment the invocation count
959 heap->increment_total_collections(true);
960
961 // We need to track unique mark sweep invocations as well.
962 _total_invocations++;
963
964 heap->print_heap_before_gc();
965
966 // Fill in TLABs
967 heap->accumulate_statistics_all_tlabs();
968 heap->ensure_parsability(true); // retire TLABs
969
970 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
971 HandleMark hm; // Discard invalid handles created during verification
972 Universe::verify(" VerifyBeforeGC:");
973 }
974
975 // Verify object start arrays
976 if (VerifyObjectStartArray &&
977 VerifyBeforeGC) {
978 heap->old_gen()->verify_object_start_array();
979 }
980
981 DEBUG_ONLY(mark_bitmap()->verify_clear();)
982 DEBUG_ONLY(summary_data().verify_clear();)
983
984 // Have worker threads release resources the next time they run a task.
985 gc_task_manager()->release_all_resources();
986 }
987
988 void PSParallelCompact::post_compact()
989 {
990 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
991
992 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
993 // Clear the marking bitmap, summary data and split info.
994 clear_data_covering_space(SpaceId(id));
995 // Update top(). Must be done after clearing the bitmap and summary data.
996 _space_info[id].publish_new_top();
997 }
998
999 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1000 MutableSpace* const from_space = _space_info[from_space_id].space();
1001 MutableSpace* const to_space = _space_info[to_space_id].space();
1002
1003 ParallelScavengeHeap* heap = gc_heap();
1004 bool eden_empty = eden_space->is_empty();
1005 if (!eden_empty) {
1006 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1007 heap->young_gen(), heap->old_gen());
1008 }
1009
1010 // Update heap occupancy information which is used as input to the soft ref
1011 // clearing policy at the next gc.
1012 Universe::update_heap_info_at_gc();
1013
1014 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1015 to_space->is_empty();
1016
1017 BarrierSet* bs = heap->barrier_set();
1018 if (bs->is_a(BarrierSet::ModRef)) {
1019 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1020 MemRegion old_mr = heap->old_gen()->reserved();
1021
1022 if (young_gen_empty) {
1023 modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1024 } else {
1025 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1026 }
1027 }
1028
1029 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1030 ClassLoaderDataGraph::purge();
1031 MetaspaceAux::verify_metrics();
1032
1033 Threads::gc_epilogue();
1034 CodeCache::gc_epilogue();
1035 JvmtiExport::gc_epilogue();
1036
1037 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1038
1039 ref_processor()->enqueue_discovered_references(NULL);
1040
1041 if (ZapUnusedHeapArea) {
1042 heap->gen_mangle_unused_area();
1043 }
1044
1045 // Update time of last GC
1046 reset_millis_since_last_gc();
1047 }
1048
1049 HeapWord*
1050 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1051 bool maximum_compaction)
1052 {
1053 const size_t region_size = ParallelCompactData::RegionSize;
1054 const ParallelCompactData& sd = summary_data();
1055
1056 const MutableSpace* const space = _space_info[id].space();
1057 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1058 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1059 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1060
1061 // Skip full regions at the beginning of the space--they are necessarily part
1062 // of the dense prefix.
1063 size_t full_count = 0;
1064 const RegionData* cp;
1065 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1066 ++full_count;
1067 }
1068
1069 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1070 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1071 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1072 if (maximum_compaction || cp == end_cp || interval_ended) {
1073 _maximum_compaction_gc_num = total_invocations();
1074 return sd.region_to_addr(cp);
1075 }
1076
1077 HeapWord* const new_top = _space_info[id].new_top();
1078 const size_t space_live = pointer_delta(new_top, space->bottom());
1079 const size_t space_used = space->used_in_words();
1080 const size_t space_capacity = space->capacity_in_words();
1081
1082 const double cur_density = double(space_live) / space_capacity;
1083 const double deadwood_density =
1084 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1085 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1086
1087 if (TraceParallelOldGCDensePrefix) {
1088 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1089 cur_density, deadwood_density, deadwood_goal);
1090 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1091 "space_cap=" SIZE_FORMAT,
1092 space_live, space_used,
1093 space_capacity);
1094 }
1095
1096 // XXX - Use binary search?
1097 HeapWord* dense_prefix = sd.region_to_addr(cp);
1098 const RegionData* full_cp = cp;
1099 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1100 while (cp < end_cp) {
1101 HeapWord* region_destination = cp->destination();
1102 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1103 if (TraceParallelOldGCDensePrefix && Verbose) {
1104 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1105 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1106 sd.region(cp), region_destination,
1107 dense_prefix, cur_deadwood);
1108 }
1109
1110 if (cur_deadwood >= deadwood_goal) {
1111 // Found the region that has the correct amount of deadwood to the left.
1112 // This typically occurs after crossing a fairly sparse set of regions, so
1113 // iterate backwards over those sparse regions, looking for the region
1114 // that has the lowest density of live objects 'to the right.'
1115 size_t space_to_left = sd.region(cp) * region_size;
1116 size_t live_to_left = space_to_left - cur_deadwood;
1117 size_t space_to_right = space_capacity - space_to_left;
1118 size_t live_to_right = space_live - live_to_left;
1119 double density_to_right = double(live_to_right) / space_to_right;
1120 while (cp > full_cp) {
1121 --cp;
1122 const size_t prev_region_live_to_right = live_to_right -
1123 cp->data_size();
1124 const size_t prev_region_space_to_right = space_to_right + region_size;
1125 double prev_region_density_to_right =
1126 double(prev_region_live_to_right) / prev_region_space_to_right;
1127 if (density_to_right <= prev_region_density_to_right) {
1128 return dense_prefix;
1129 }
1130 if (TraceParallelOldGCDensePrefix && Verbose) {
1131 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1132 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1133 prev_region_density_to_right);
1134 }
1135 dense_prefix -= region_size;
1136 live_to_right = prev_region_live_to_right;
1137 space_to_right = prev_region_space_to_right;
1138 density_to_right = prev_region_density_to_right;
1139 }
1140 return dense_prefix;
1141 }
1142
1143 dense_prefix += region_size;
1144 ++cp;
1145 }
1146
1147 return dense_prefix;
1148 }
1149
1150 #ifndef PRODUCT
1151 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1152 const SpaceId id,
1153 const bool maximum_compaction,
1154 HeapWord* const addr)
1155 {
1156 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1157 RegionData* const cp = summary_data().region(region_idx);
1158 const MutableSpace* const space = _space_info[id].space();
1159 HeapWord* const new_top = _space_info[id].new_top();
1160
1161 const size_t space_live = pointer_delta(new_top, space->bottom());
1162 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1163 const size_t space_cap = space->capacity_in_words();
1164 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1165 const size_t live_to_right = new_top - cp->destination();
1166 const size_t dead_to_right = space->top() - addr - live_to_right;
1167
1168 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1169 "spl=" SIZE_FORMAT " "
1170 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1171 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1172 " ratio=%10.8f",
1173 algorithm, addr, region_idx,
1174 space_live,
1175 dead_to_left, dead_to_left_pct,
1176 dead_to_right, live_to_right,
1177 double(dead_to_right) / live_to_right);
1178 }
1179 #endif // #ifndef PRODUCT
1180
1181 // Return a fraction indicating how much of the generation can be treated as
1182 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1183 // based on the density of live objects in the generation to determine a limit,
1184 // which is then adjusted so the return value is min_percent when the density is
1185 // 1.
1186 //
1187 // The following table shows some return values for a different values of the
1188 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1189 // min_percent is 1.
1190 //
1191 // fraction allowed as dead wood
1192 // -----------------------------------------------------------------
1193 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1194 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1195 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1196 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1197 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1198 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1199 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1200 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1201 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1202 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1203 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1204 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1205 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1206 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1207 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1208 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1209 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1210 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1211 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1212 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1213 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1214 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1215 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1216
1217 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1218 {
1219 assert(_dwl_initialized, "uninitialized");
1220
1221 // The raw limit is the value of the normal distribution at x = density.
1222 const double raw_limit = normal_distribution(density);
1223
1224 // Adjust the raw limit so it becomes the minimum when the density is 1.
1225 //
1226 // First subtract the adjustment value (which is simply the precomputed value
1227 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1228 // Then add the minimum value, so the minimum is returned when the density is
1229 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1230 const double min = double(min_percent) / 100.0;
1231 const double limit = raw_limit - _dwl_adjustment + min;
1232 return MAX2(limit, 0.0);
1233 }
1234
1235 ParallelCompactData::RegionData*
1236 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1237 const RegionData* end)
1238 {
1239 const size_t region_size = ParallelCompactData::RegionSize;
1240 ParallelCompactData& sd = summary_data();
1241 size_t left = sd.region(beg);
1242 size_t right = end > beg ? sd.region(end) - 1 : left;
1243
1244 // Binary search.
1245 while (left < right) {
1246 // Equivalent to (left + right) / 2, but does not overflow.
1247 const size_t middle = left + (right - left) / 2;
1248 RegionData* const middle_ptr = sd.region(middle);
1249 HeapWord* const dest = middle_ptr->destination();
1250 HeapWord* const addr = sd.region_to_addr(middle);
1251 assert(dest != NULL, "sanity");
1252 assert(dest <= addr, "must move left");
1253
1254 if (middle > left && dest < addr) {
1255 right = middle - 1;
1256 } else if (middle < right && middle_ptr->data_size() == region_size) {
1257 left = middle + 1;
1258 } else {
1259 return middle_ptr;
1260 }
1261 }
1262 return sd.region(left);
1263 }
1264
1265 ParallelCompactData::RegionData*
1266 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1267 const RegionData* end,
1268 size_t dead_words)
1269 {
1270 ParallelCompactData& sd = summary_data();
1271 size_t left = sd.region(beg);
1272 size_t right = end > beg ? sd.region(end) - 1 : left;
1273
1274 // Binary search.
1275 while (left < right) {
1276 // Equivalent to (left + right) / 2, but does not overflow.
1277 const size_t middle = left + (right - left) / 2;
1278 RegionData* const middle_ptr = sd.region(middle);
1279 HeapWord* const dest = middle_ptr->destination();
1280 HeapWord* const addr = sd.region_to_addr(middle);
1281 assert(dest != NULL, "sanity");
1282 assert(dest <= addr, "must move left");
1283
1284 const size_t dead_to_left = pointer_delta(addr, dest);
1285 if (middle > left && dead_to_left > dead_words) {
1286 right = middle - 1;
1287 } else if (middle < right && dead_to_left < dead_words) {
1288 left = middle + 1;
1289 } else {
1290 return middle_ptr;
1291 }
1292 }
1293 return sd.region(left);
1294 }
1295
1296 // The result is valid during the summary phase, after the initial summarization
1297 // of each space into itself, and before final summarization.
1298 inline double
1299 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1300 HeapWord* const bottom,
1301 HeapWord* const top,
1302 HeapWord* const new_top)
1303 {
1304 ParallelCompactData& sd = summary_data();
1305
1306 assert(cp != NULL, "sanity");
1307 assert(bottom != NULL, "sanity");
1308 assert(top != NULL, "sanity");
1309 assert(new_top != NULL, "sanity");
1310 assert(top >= new_top, "summary data problem?");
1311 assert(new_top > bottom, "space is empty; should not be here");
1312 assert(new_top >= cp->destination(), "sanity");
1313 assert(top >= sd.region_to_addr(cp), "sanity");
1314
1315 HeapWord* const destination = cp->destination();
1316 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1317 const size_t compacted_region_live = pointer_delta(new_top, destination);
1318 const size_t compacted_region_used = pointer_delta(top,
1319 sd.region_to_addr(cp));
1320 const size_t reclaimable = compacted_region_used - compacted_region_live;
1321
1322 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1323 return double(reclaimable) / divisor;
1324 }
1325
1326 // Return the address of the end of the dense prefix, a.k.a. the start of the
1327 // compacted region. The address is always on a region boundary.
1328 //
1329 // Completely full regions at the left are skipped, since no compaction can
1330 // occur in those regions. Then the maximum amount of dead wood to allow is
1331 // computed, based on the density (amount live / capacity) of the generation;
1332 // the region with approximately that amount of dead space to the left is
1333 // identified as the limit region. Regions between the last completely full
1334 // region and the limit region are scanned and the one that has the best
1335 // (maximum) reclaimed_ratio() is selected.
1336 HeapWord*
1337 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1338 bool maximum_compaction)
1339 {
1340 if (ParallelOldGCSplitALot) {
1341 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1342 // The value was chosen to provoke splitting a young gen space; use it.
1343 return _space_info[id].dense_prefix();
1344 }
1345 }
1346
1347 const size_t region_size = ParallelCompactData::RegionSize;
1348 const ParallelCompactData& sd = summary_data();
1349
1350 const MutableSpace* const space = _space_info[id].space();
1351 HeapWord* const top = space->top();
1352 HeapWord* const top_aligned_up = sd.region_align_up(top);
1353 HeapWord* const new_top = _space_info[id].new_top();
1354 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1355 HeapWord* const bottom = space->bottom();
1356 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1357 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1358 const RegionData* const new_top_cp =
1359 sd.addr_to_region_ptr(new_top_aligned_up);
1360
1361 // Skip full regions at the beginning of the space--they are necessarily part
1362 // of the dense prefix.
1363 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1364 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1365 space->is_empty(), "no dead space allowed to the left");
1366 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1367 "region must have dead space");
1368
1369 // The gc number is saved whenever a maximum compaction is done, and used to
1370 // determine when the maximum compaction interval has expired. This avoids
1371 // successive max compactions for different reasons.
1372 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1373 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1374 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1375 total_invocations() == HeapFirstMaximumCompactionCount;
1376 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1377 _maximum_compaction_gc_num = total_invocations();
1378 return sd.region_to_addr(full_cp);
1379 }
1380
1381 const size_t space_live = pointer_delta(new_top, bottom);
1382 const size_t space_used = space->used_in_words();
1383 const size_t space_capacity = space->capacity_in_words();
1384
1385 const double density = double(space_live) / double(space_capacity);
1386 const size_t min_percent_free = MarkSweepDeadRatio;
1387 const double limiter = dead_wood_limiter(density, min_percent_free);
1388 const size_t dead_wood_max = space_used - space_live;
1389 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1390 dead_wood_max);
1391
1392 if (TraceParallelOldGCDensePrefix) {
1393 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1394 "space_cap=" SIZE_FORMAT,
1395 space_live, space_used,
1396 space_capacity);
1397 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1398 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1399 density, min_percent_free, limiter,
1400 dead_wood_max, dead_wood_limit);
1401 }
1402
1403 // Locate the region with the desired amount of dead space to the left.
1404 const RegionData* const limit_cp =
1405 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1406
1407 // Scan from the first region with dead space to the limit region and find the
1408 // one with the best (largest) reclaimed ratio.
1409 double best_ratio = 0.0;
1410 const RegionData* best_cp = full_cp;
1411 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1412 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1413 if (tmp_ratio > best_ratio) {
1414 best_cp = cp;
1415 best_ratio = tmp_ratio;
1416 }
1417 }
1418
1419 #if 0
1420 // Something to consider: if the region with the best ratio is 'close to' the
1421 // first region w/free space, choose the first region with free space
1422 // ("first-free"). The first-free region is usually near the start of the
1423 // heap, which means we are copying most of the heap already, so copy a bit
1424 // more to get complete compaction.
1425 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1426 _maximum_compaction_gc_num = total_invocations();
1427 best_cp = full_cp;
1428 }
1429 #endif // #if 0
1430
1431 return sd.region_to_addr(best_cp);
1432 }
1433
1434 #ifndef PRODUCT
1435 void
1436 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1437 size_t words)
1438 {
1439 if (TraceParallelOldGCSummaryPhase) {
1440 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1441 SIZE_FORMAT, start, start + words, words);
1442 }
1443
1444 ObjectStartArray* const start_array = _space_info[id].start_array();
1445 CollectedHeap::fill_with_objects(start, words);
1446 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1447 _mark_bitmap.mark_obj(p, words);
1448 _summary_data.add_obj(p, words);
1449 start_array->allocate_block(p);
1450 }
1451 }
1452
1453 void
1454 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1455 {
1456 ParallelCompactData& sd = summary_data();
1457 MutableSpace* space = _space_info[id].space();
1458
1459 // Find the source and destination start addresses.
1460 HeapWord* const src_addr = sd.region_align_down(start);
1461 HeapWord* dst_addr;
1462 if (src_addr < start) {
1463 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1464 } else if (src_addr > space->bottom()) {
1465 // The start (the original top() value) is aligned to a region boundary so
1466 // the associated region does not have a destination. Compute the
1467 // destination from the previous region.
1468 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1469 dst_addr = cp->destination() + cp->data_size();
1470 } else {
1471 // Filling the entire space.
1472 dst_addr = space->bottom();
1473 }
1474 assert(dst_addr != NULL, "sanity");
1475
1476 // Update the summary data.
1477 bool result = _summary_data.summarize(_space_info[id].split_info(),
1478 src_addr, space->top(), NULL,
1479 dst_addr, space->end(),
1480 _space_info[id].new_top_addr());
1481 assert(result, "should not fail: bad filler object size");
1482 }
1483
1484 void
1485 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1486 {
1487 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1488 return;
1489 }
1490
1491 MutableSpace* const space = _space_info[id].space();
1492 if (space->is_empty()) {
1493 HeapWord* b = space->bottom();
1494 HeapWord* t = b + space->capacity_in_words() / 2;
1495 space->set_top(t);
1496 if (ZapUnusedHeapArea) {
1497 space->set_top_for_allocations();
1498 }
1499
1500 size_t min_size = CollectedHeap::min_fill_size();
1501 size_t obj_len = min_size;
1502 while (b + obj_len <= t) {
1503 CollectedHeap::fill_with_object(b, obj_len);
1504 mark_bitmap()->mark_obj(b, obj_len);
1505 summary_data().add_obj(b, obj_len);
1506 b += obj_len;
1507 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1508 }
1509 if (b < t) {
1510 // The loop didn't completely fill to t (top); adjust top downward.
1511 space->set_top(b);
1512 if (ZapUnusedHeapArea) {
1513 space->set_top_for_allocations();
1514 }
1515 }
1516
1517 HeapWord** nta = _space_info[id].new_top_addr();
1518 bool result = summary_data().summarize(_space_info[id].split_info(),
1519 space->bottom(), space->top(), NULL,
1520 space->bottom(), space->end(), nta);
1521 assert(result, "space must fit into itself");
1522 }
1523 }
1524
1525 void
1526 PSParallelCompact::provoke_split(bool & max_compaction)
1527 {
1528 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1529 return;
1530 }
1531
1532 const size_t region_size = ParallelCompactData::RegionSize;
1533 ParallelCompactData& sd = summary_data();
1534
1535 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1536 MutableSpace* const from_space = _space_info[from_space_id].space();
1537 const size_t eden_live = pointer_delta(eden_space->top(),
1538 _space_info[eden_space_id].new_top());
1539 const size_t from_live = pointer_delta(from_space->top(),
1540 _space_info[from_space_id].new_top());
1541
1542 const size_t min_fill_size = CollectedHeap::min_fill_size();
1543 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1544 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1545 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1546 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1547
1548 // Choose the space to split; need at least 2 regions live (or fillable).
1549 SpaceId id;
1550 MutableSpace* space;
1551 size_t live_words;
1552 size_t fill_words;
1553 if (eden_live + eden_fillable >= region_size * 2) {
1554 id = eden_space_id;
1555 space = eden_space;
1556 live_words = eden_live;
1557 fill_words = eden_fillable;
1558 } else if (from_live + from_fillable >= region_size * 2) {
1559 id = from_space_id;
1560 space = from_space;
1561 live_words = from_live;
1562 fill_words = from_fillable;
1563 } else {
1564 return; // Give up.
1565 }
1566 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1567
1568 if (live_words < region_size * 2) {
1569 // Fill from top() to end() w/live objects of mixed sizes.
1570 HeapWord* const fill_start = space->top();
1571 live_words += fill_words;
1572
1573 space->set_top(fill_start + fill_words);
1574 if (ZapUnusedHeapArea) {
1575 space->set_top_for_allocations();
1576 }
1577
1578 HeapWord* cur_addr = fill_start;
1579 while (fill_words > 0) {
1580 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1581 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1582 if (fill_words - cur_size < min_fill_size) {
1583 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1584 }
1585
1586 CollectedHeap::fill_with_object(cur_addr, cur_size);
1587 mark_bitmap()->mark_obj(cur_addr, cur_size);
1588 sd.add_obj(cur_addr, cur_size);
1589
1590 cur_addr += cur_size;
1591 fill_words -= cur_size;
1592 }
1593
1594 summarize_new_objects(id, fill_start);
1595 }
1596
1597 max_compaction = false;
1598
1599 // Manipulate the old gen so that it has room for about half of the live data
1600 // in the target young gen space (live_words / 2).
1601 id = old_space_id;
1602 space = _space_info[id].space();
1603 const size_t free_at_end = space->free_in_words();
1604 const size_t free_target = align_object_size(live_words / 2);
1605 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1606
1607 if (free_at_end >= free_target + min_fill_size) {
1608 // Fill space above top() and set the dense prefix so everything survives.
1609 HeapWord* const fill_start = space->top();
1610 const size_t fill_size = free_at_end - free_target;
1611 space->set_top(space->top() + fill_size);
1612 if (ZapUnusedHeapArea) {
1613 space->set_top_for_allocations();
1614 }
1615 fill_with_live_objects(id, fill_start, fill_size);
1616 summarize_new_objects(id, fill_start);
1617 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1618 } else if (dead + free_at_end > free_target) {
1619 // Find a dense prefix that makes the right amount of space available.
1620 HeapWord* cur = sd.region_align_down(space->top());
1621 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1622 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1623 while (dead_to_right < free_target) {
1624 cur -= region_size;
1625 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1626 dead_to_right = pointer_delta(space->end(), cur_destination);
1627 }
1628 _space_info[id].set_dense_prefix(cur);
1629 }
1630 }
1631 #endif // #ifndef PRODUCT
1632
1633 void PSParallelCompact::summarize_spaces_quick()
1634 {
1635 for (unsigned int i = 0; i < last_space_id; ++i) {
1636 const MutableSpace* space = _space_info[i].space();
1637 HeapWord** nta = _space_info[i].new_top_addr();
1638 bool result = _summary_data.summarize(_space_info[i].split_info(),
1639 space->bottom(), space->top(), NULL,
1640 space->bottom(), space->end(), nta);
1641 assert(result, "space must fit into itself");
1642 _space_info[i].set_dense_prefix(space->bottom());
1643 }
1644
1645 #ifndef PRODUCT
1646 if (ParallelOldGCSplitALot) {
1647 provoke_split_fill_survivor(to_space_id);
1648 }
1649 #endif // #ifndef PRODUCT
1650 }
1651
1652 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1653 {
1654 HeapWord* const dense_prefix_end = dense_prefix(id);
1655 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1656 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1657 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1658 // Only enough dead space is filled so that any remaining dead space to the
1659 // left is larger than the minimum filler object. (The remainder is filled
1660 // during the copy/update phase.)
1661 //
1662 // The size of the dead space to the right of the boundary is not a
1663 // concern, since compaction will be able to use whatever space is
1664 // available.
1665 //
1666 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1667 // surrounds the space to be filled with an object.
1668 //
1669 // In the 32-bit VM, each bit represents two 32-bit words:
1670 // +---+
1671 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1672 // end_bits: ... x x x | 0 | || 0 x x ...
1673 // +---+
1674 //
1675 // In the 64-bit VM, each bit represents one 64-bit word:
1676 // +------------+
1677 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1678 // end_bits: ... x x 1 | 0 || 0 | x x ...
1679 // +------------+
1680 // +-------+
1681 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1682 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1683 // +-------+
1684 // +-----------+
1685 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1686 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1687 // +-----------+
1688 // +-------+
1689 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1690 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1691 // +-------+
1692
1693 // Initially assume case a, c or e will apply.
1694 size_t obj_len = CollectedHeap::min_fill_size();
1695 HeapWord* obj_beg = dense_prefix_end - obj_len;
1696
1697 #ifdef _LP64
1698 if (MinObjAlignment > 1) { // object alignment > heap word size
1699 // Cases a, c or e.
1700 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1701 // Case b above.
1702 obj_beg = dense_prefix_end - 1;
1703 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1704 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1705 // Case d above.
1706 obj_beg = dense_prefix_end - 3;
1707 obj_len = 3;
1708 }
1709 #endif // #ifdef _LP64
1710
1711 CollectedHeap::fill_with_object(obj_beg, obj_len);
1712 _mark_bitmap.mark_obj(obj_beg, obj_len);
1713 _summary_data.add_obj(obj_beg, obj_len);
1714 assert(start_array(id) != NULL, "sanity");
1715 start_array(id)->allocate_block(obj_beg);
1716 }
1717 }
1718
1719 void
1720 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1721 {
1722 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1723 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1724 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1725 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1726 cur->set_source_region(0);
1727 }
1728 }
1729
1730 void
1731 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1732 {
1733 assert(id < last_space_id, "id out of range");
1734 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1735 ParallelOldGCSplitALot && id == old_space_id,
1736 "should have been reset in summarize_spaces_quick()");
1737
1738 const MutableSpace* space = _space_info[id].space();
1739 if (_space_info[id].new_top() != space->bottom()) {
1740 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1741 _space_info[id].set_dense_prefix(dense_prefix_end);
1742
1743 #ifndef PRODUCT
1744 if (TraceParallelOldGCDensePrefix) {
1745 print_dense_prefix_stats("ratio", id, maximum_compaction,
1746 dense_prefix_end);
1747 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1748 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1749 }
1750 #endif // #ifndef PRODUCT
1751
1752 // Recompute the summary data, taking into account the dense prefix. If
1753 // every last byte will be reclaimed, then the existing summary data which
1754 // compacts everything can be left in place.
1755 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1756 // If dead space crosses the dense prefix boundary, it is (at least
1757 // partially) filled with a dummy object, marked live and added to the
1758 // summary data. This simplifies the copy/update phase and must be done
1759 // before the final locations of objects are determined, to prevent
1760 // leaving a fragment of dead space that is too small to fill.
1761 fill_dense_prefix_end(id);
1762
1763 // Compute the destination of each Region, and thus each object.
1764 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1765 _summary_data.summarize(_space_info[id].split_info(),
1766 dense_prefix_end, space->top(), NULL,
1767 dense_prefix_end, space->end(),
1768 _space_info[id].new_top_addr());
1769 }
1770 }
1771
1772 if (TraceParallelOldGCSummaryPhase) {
1773 const size_t region_size = ParallelCompactData::RegionSize;
1774 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1775 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1776 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1777 HeapWord* const new_top = _space_info[id].new_top();
1778 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1779 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1780 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1781 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1782 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1783 id, space->capacity_in_words(), dense_prefix_end,
1784 dp_region, dp_words / region_size,
1785 cr_words / region_size, new_top);
1786 }
1787 }
1788
1789 #ifndef PRODUCT
1790 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1791 HeapWord* dst_beg, HeapWord* dst_end,
1792 SpaceId src_space_id,
1793 HeapWord* src_beg, HeapWord* src_end)
1794 {
1795 if (TraceParallelOldGCSummaryPhase) {
1796 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1797 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1798 SIZE_FORMAT "-" SIZE_FORMAT " "
1799 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1800 SIZE_FORMAT "-" SIZE_FORMAT,
1801 src_space_id, space_names[src_space_id],
1802 dst_space_id, space_names[dst_space_id],
1803 src_beg, src_end,
1804 _summary_data.addr_to_region_idx(src_beg),
1805 _summary_data.addr_to_region_idx(src_end),
1806 dst_beg, dst_end,
1807 _summary_data.addr_to_region_idx(dst_beg),
1808 _summary_data.addr_to_region_idx(dst_end));
1809 }
1810 }
1811 #endif // #ifndef PRODUCT
1812
1813 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1814 bool maximum_compaction)
1815 {
1816 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1817 // trace("2");
1818
1819 #ifdef ASSERT
1820 if (TraceParallelOldGCMarkingPhase) {
1821 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1822 "add_obj_bytes=" SIZE_FORMAT,
1823 add_obj_count, add_obj_size * HeapWordSize);
1824 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1825 "mark_bitmap_bytes=" SIZE_FORMAT,
1826 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1827 }
1828 #endif // #ifdef ASSERT
1829
1830 // Quick summarization of each space into itself, to see how much is live.
1831 summarize_spaces_quick();
1832
1833 if (TraceParallelOldGCSummaryPhase) {
1834 tty->print_cr("summary_phase: after summarizing each space to self");
1835 Universe::print();
1836 NOT_PRODUCT(print_region_ranges());
1837 if (Verbose) {
1838 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1839 }
1840 }
1841
1842 // The amount of live data that will end up in old space (assuming it fits).
1843 size_t old_space_total_live = 0;
1844 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1845 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1846 _space_info[id].space()->bottom());
1847 }
1848
1849 MutableSpace* const old_space = _space_info[old_space_id].space();
1850 const size_t old_capacity = old_space->capacity_in_words();
1851 if (old_space_total_live > old_capacity) {
1852 // XXX - should also try to expand
1853 maximum_compaction = true;
1854 }
1855 #ifndef PRODUCT
1856 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1857 provoke_split(maximum_compaction);
1858 }
1859 #endif // #ifndef PRODUCT
1860
1861 // Old generations.
1862 summarize_space(old_space_id, maximum_compaction);
1863
1864 // Summarize the remaining spaces in the young gen. The initial target space
1865 // is the old gen. If a space does not fit entirely into the target, then the
1866 // remainder is compacted into the space itself and that space becomes the new
1867 // target.
1868 SpaceId dst_space_id = old_space_id;
1869 HeapWord* dst_space_end = old_space->end();
1870 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1871 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1872 const MutableSpace* space = _space_info[id].space();
1873 const size_t live = pointer_delta(_space_info[id].new_top(),
1874 space->bottom());
1875 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1876
1877 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1878 SpaceId(id), space->bottom(), space->top());)
1879 if (live > 0 && live <= available) {
1880 // All the live data will fit.
1881 bool done = _summary_data.summarize(_space_info[id].split_info(),
1882 space->bottom(), space->top(),
1883 NULL,
1884 *new_top_addr, dst_space_end,
1885 new_top_addr);
1886 assert(done, "space must fit into old gen");
1887
1888 // Reset the new_top value for the space.
1889 _space_info[id].set_new_top(space->bottom());
1890 } else if (live > 0) {
1891 // Attempt to fit part of the source space into the target space.
1892 HeapWord* next_src_addr = NULL;
1893 bool done = _summary_data.summarize(_space_info[id].split_info(),
1894 space->bottom(), space->top(),
1895 &next_src_addr,
1896 *new_top_addr, dst_space_end,
1897 new_top_addr);
1898 assert(!done, "space should not fit into old gen");
1899 assert(next_src_addr != NULL, "sanity");
1900
1901 // The source space becomes the new target, so the remainder is compacted
1902 // within the space itself.
1903 dst_space_id = SpaceId(id);
1904 dst_space_end = space->end();
1905 new_top_addr = _space_info[id].new_top_addr();
1906 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1907 space->bottom(), dst_space_end,
1908 SpaceId(id), next_src_addr, space->top());)
1909 done = _summary_data.summarize(_space_info[id].split_info(),
1910 next_src_addr, space->top(),
1911 NULL,
1912 space->bottom(), dst_space_end,
1913 new_top_addr);
1914 assert(done, "space must fit when compacted into itself");
1915 assert(*new_top_addr <= space->top(), "usage should not grow");
1916 }
1917 }
1918
1919 if (TraceParallelOldGCSummaryPhase) {
1920 tty->print_cr("summary_phase: after final summarization");
1921 Universe::print();
1922 NOT_PRODUCT(print_region_ranges());
1923 if (Verbose) {
1924 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1925 }
1926 }
1927 }
1928
1929 // This method should contain all heap-specific policy for invoking a full
1930 // collection. invoke_no_policy() will only attempt to compact the heap; it
1931 // will do nothing further. If we need to bail out for policy reasons, scavenge
1932 // before full gc, or any other specialized behavior, it needs to be added here.
1933 //
1934 // Note that this method should only be called from the vm_thread while at a
1935 // safepoint.
1936 //
1937 // Note that the all_soft_refs_clear flag in the collector policy
1938 // may be true because this method can be called without intervening
1939 // activity. For example when the heap space is tight and full measure
1940 // are being taken to free space.
1941 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1942 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1943 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1944 "should be in vm thread");
1945
1946 ParallelScavengeHeap* heap = gc_heap();
1947 GCCause::Cause gc_cause = heap->gc_cause();
1948 assert(!heap->is_gc_active(), "not reentrant");
1949
1950 PSAdaptiveSizePolicy* policy = heap->size_policy();
1951 IsGCActiveMark mark;
1952
1953 if (ScavengeBeforeFullGC) {
1954 PSScavenge::invoke_no_policy();
1955 }
1956
1957 const bool clear_all_soft_refs =
1958 heap->collector_policy()->should_clear_all_soft_refs();
1959
1960 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1961 maximum_heap_compaction);
1962 }
1963
1964 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1965 size_t addr_region_index = addr_to_region_idx(addr);
1966 return region_index == addr_region_index;
1967 }
1968
1969 // This method contains no policy. You should probably
1970 // be calling invoke() instead.
1971 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1972 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1973 assert(ref_processor() != NULL, "Sanity");
1974
1975 if (GC_locker::check_active_before_gc()) {
1976 return false;
1977 }
1978
1979 TimeStamp marking_start;
1980 TimeStamp compaction_start;
1981 TimeStamp collection_exit;
1982
1983 ParallelScavengeHeap* heap = gc_heap();
1984 GCCause::Cause gc_cause = heap->gc_cause();
1985 PSYoungGen* young_gen = heap->young_gen();
1986 PSOldGen* old_gen = heap->old_gen();
1987 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1988
1989 // The scope of casr should end after code that can change
1990 // CollectorPolicy::_should_clear_all_soft_refs.
1991 ClearedAllSoftRefs casr(maximum_heap_compaction,
1992 heap->collector_policy());
1993
1994 if (ZapUnusedHeapArea) {
1995 // Save information needed to minimize mangling
1996 heap->record_gen_tops_before_GC();
1997 }
1998
1999 heap->pre_full_gc_dump();
2000
2001 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2002
2003 // Make sure data structures are sane, make the heap parsable, and do other
2004 // miscellaneous bookkeeping.
2005 PreGCValues pre_gc_values;
2006 pre_compact(&pre_gc_values);
2007
2008 // Get the compaction manager reserved for the VM thread.
2009 ParCompactionManager* const vmthread_cm =
2010 ParCompactionManager::manager_array(gc_task_manager()->workers());
2011
2012 // Place after pre_compact() where the number of invocations is incremented.
2013 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2014
2015 {
2016 ResourceMark rm;
2017 HandleMark hm;
2018
2019 // Set the number of GC threads to be used in this collection
2020 gc_task_manager()->set_active_gang();
2021 gc_task_manager()->task_idle_workers();
2022 heap->set_par_threads(gc_task_manager()->active_workers());
2023
2024 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2025 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2026 TraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty);
2027 TraceCollectorStats tcs(counters());
2028 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2029
2030 if (TraceGen1Time) accumulated_time()->start();
2031
2032 // Let the size policy know we're starting
2033 size_policy->major_collection_begin();
2034
2035 CodeCache::gc_prologue();
2036 Threads::gc_prologue();
2037
2038 COMPILER2_PRESENT(DerivedPointerTable::clear());
2039
2040 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2041 ref_processor()->setup_policy(maximum_heap_compaction);
2042
2043 bool marked_for_unloading = false;
2044
2045 marking_start.update();
2046 marking_phase(vmthread_cm, maximum_heap_compaction);
2047
2048 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2049 && gc_cause == GCCause::_java_lang_system_gc;
2050 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2051
2052 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2053 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2054
2055 // adjust_roots() updates Universe::_intArrayKlassObj which is
2056 // needed by the compaction for filling holes in the dense prefix.
2057 adjust_roots();
2058
2059 compaction_start.update();
2060 compact();
2061
2062 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2063 // done before resizing.
2064 post_compact();
2065
2066 // Let the size policy know we're done
2067 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2068
2069 if (UseAdaptiveSizePolicy) {
2070 if (PrintAdaptiveSizePolicy) {
2071 gclog_or_tty->print("AdaptiveSizeStart: ");
2072 gclog_or_tty->stamp();
2073 gclog_or_tty->print_cr(" collection: %d ",
2074 heap->total_collections());
2075 if (Verbose) {
2076 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d",
2077 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2078 }
2079 }
2080
2081 // Don't check if the size_policy is ready here. Let
2082 // the size_policy check that internally.
2083 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2084 ((gc_cause != GCCause::_java_lang_system_gc) ||
2085 UseAdaptiveSizePolicyWithSystemGC)) {
2086 // Calculate optimal free space amounts
2087 assert(young_gen->max_size() >
2088 young_gen->from_space()->capacity_in_bytes() +
2089 young_gen->to_space()->capacity_in_bytes(),
2090 "Sizes of space in young gen are out-of-bounds");
2091
2092 size_t young_live = young_gen->used_in_bytes();
2093 size_t eden_live = young_gen->eden_space()->used_in_bytes();
2094 size_t old_live = old_gen->used_in_bytes();
2095 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2096 size_t max_old_gen_size = old_gen->max_gen_size();
2097 size_t max_eden_size = young_gen->max_size() -
2098 young_gen->from_space()->capacity_in_bytes() -
2099 young_gen->to_space()->capacity_in_bytes();
2100
2101 // Used for diagnostics
2102 size_policy->clear_generation_free_space_flags();
2103
2104 size_policy->compute_generation_free_space(young_live,
2105 eden_live,
2106 old_live,
2107 cur_eden,
2108 max_old_gen_size,
2109 max_eden_size,
2110 true /* full gc*/);
2111
2112 size_policy->check_gc_overhead_limit(young_live,
2113 eden_live,
2114 max_old_gen_size,
2115 max_eden_size,
2116 true /* full gc*/,
2117 gc_cause,
2118 heap->collector_policy());
2119
2120 size_policy->decay_supplemental_growth(true /* full gc*/);
2121
2122 heap->resize_old_gen(
2123 size_policy->calculated_old_free_size_in_bytes());
2124
2125 // Don't resize the young generation at an major collection. A
2126 // desired young generation size may have been calculated but
2127 // resizing the young generation complicates the code because the
2128 // resizing of the old generation may have moved the boundary
2129 // between the young generation and the old generation. Let the
2130 // young generation resizing happen at the minor collections.
2131 }
2132 if (PrintAdaptiveSizePolicy) {
2133 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2134 heap->total_collections());
2135 }
2136 }
2137
2138 if (UsePerfData) {
2139 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2140 counters->update_counters();
2141 counters->update_old_capacity(old_gen->capacity_in_bytes());
2142 counters->update_young_capacity(young_gen->capacity_in_bytes());
2143 }
2144
2145 heap->resize_all_tlabs();
2146
2147 // Resize the metaspace capactiy after a collection
2148 MetaspaceGC::compute_new_size();
2149
2150 if (TraceGen1Time) accumulated_time()->stop();
2151
2152 if (PrintGC) {
2153 if (PrintGCDetails) {
2154 // No GC timestamp here. This is after GC so it would be confusing.
2155 young_gen->print_used_change(pre_gc_values.young_gen_used());
2156 old_gen->print_used_change(pre_gc_values.old_gen_used());
2157 heap->print_heap_change(pre_gc_values.heap_used());
2158 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2159 } else {
2160 heap->print_heap_change(pre_gc_values.heap_used());
2161 }
2162 }
2163
2164 // Track memory usage and detect low memory
2165 MemoryService::track_memory_usage();
2166 heap->update_counters();
2167 gc_task_manager()->release_idle_workers();
2168 }
2169
2170 #ifdef ASSERT
2171 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2172 ParCompactionManager* const cm =
2173 ParCompactionManager::manager_array(int(i));
2174 assert(cm->marking_stack()->is_empty(), "should be empty");
2175 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2176 }
2177 #endif // ASSERT
2178
2179 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2180 HandleMark hm; // Discard invalid handles created during verification
2181 Universe::verify(" VerifyAfterGC:");
2182 }
2183
2184 // Re-verify object start arrays
2185 if (VerifyObjectStartArray &&
2186 VerifyAfterGC) {
2187 old_gen->verify_object_start_array();
2188 }
2189
2190 if (ZapUnusedHeapArea) {
2191 old_gen->object_space()->check_mangled_unused_area_complete();
2192 }
2193
2194 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2195
2196 collection_exit.update();
2197
2198 heap->print_heap_after_gc();
2199 if (PrintGCTaskTimeStamps) {
2200 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2201 INT64_FORMAT,
2202 marking_start.ticks(), compaction_start.ticks(),
2203 collection_exit.ticks());
2204 gc_task_manager()->print_task_time_stamps();
2205 }
2206
2207 heap->post_full_gc_dump();
2208
2209 #ifdef TRACESPINNING
2210 ParallelTaskTerminator::print_termination_counts();
2211 #endif
2212
2213 return true;
2214 }
2215
2216 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2217 PSYoungGen* young_gen,
2218 PSOldGen* old_gen) {
2219 MutableSpace* const eden_space = young_gen->eden_space();
2220 assert(!eden_space->is_empty(), "eden must be non-empty");
2221 assert(young_gen->virtual_space()->alignment() ==
2222 old_gen->virtual_space()->alignment(), "alignments do not match");
2223
2224 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2225 return false;
2226 }
2227
2228 // Both generations must be completely committed.
2229 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2230 return false;
2231 }
2232 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2233 return false;
2234 }
2235
2236 // Figure out how much to take from eden. Include the average amount promoted
2237 // in the total; otherwise the next young gen GC will simply bail out to a
2238 // full GC.
2239 const size_t alignment = old_gen->virtual_space()->alignment();
2240 const size_t eden_used = eden_space->used_in_bytes();
2241 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2242 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2243 const size_t eden_capacity = eden_space->capacity_in_bytes();
2244
2245 if (absorb_size >= eden_capacity) {
2246 return false; // Must leave some space in eden.
2247 }
2248
2249 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2250 if (new_young_size < young_gen->min_gen_size()) {
2251 return false; // Respect young gen minimum size.
2252 }
2253
2254 if (TraceAdaptiveGCBoundary && Verbose) {
2255 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2256 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2257 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2258 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2259 absorb_size / K,
2260 eden_capacity / K, (eden_capacity - absorb_size) / K,
2261 young_gen->from_space()->used_in_bytes() / K,
2262 young_gen->to_space()->used_in_bytes() / K,
2263 young_gen->capacity_in_bytes() / K, new_young_size / K);
2264 }
2265
2266 // Fill the unused part of the old gen.
2267 MutableSpace* const old_space = old_gen->object_space();
2268 HeapWord* const unused_start = old_space->top();
2269 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2270
2271 if (unused_words > 0) {
2272 if (unused_words < CollectedHeap::min_fill_size()) {
2273 return false; // If the old gen cannot be filled, must give up.
2274 }
2275 CollectedHeap::fill_with_objects(unused_start, unused_words);
2276 }
2277
2278 // Take the live data from eden and set both top and end in the old gen to
2279 // eden top. (Need to set end because reset_after_change() mangles the region
2280 // from end to virtual_space->high() in debug builds).
2281 HeapWord* const new_top = eden_space->top();
2282 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2283 absorb_size);
2284 young_gen->reset_after_change();
2285 old_space->set_top(new_top);
2286 old_space->set_end(new_top);
2287 old_gen->reset_after_change();
2288
2289 // Update the object start array for the filler object and the data from eden.
2290 ObjectStartArray* const start_array = old_gen->start_array();
2291 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2292 start_array->allocate_block(p);
2293 }
2294
2295 // Could update the promoted average here, but it is not typically updated at
2296 // full GCs and the value to use is unclear. Something like
2297 //
2298 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2299
2300 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2301 return true;
2302 }
2303
2304 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2305 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2306 "shouldn't return NULL");
2307 return ParallelScavengeHeap::gc_task_manager();
2308 }
2309
2310 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2311 bool maximum_heap_compaction) {
2312 // Recursively traverse all live objects and mark them
2313 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2314
2315 ParallelScavengeHeap* heap = gc_heap();
2316 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2317 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2318 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2319 ParallelTaskTerminator terminator(active_gc_threads, qset);
2320
2321 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2322 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2323
2324 // Need new claim bits before marking starts.
2325 ClassLoaderDataGraph::clear_claimed_marks();
2326
2327 {
2328 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2329 ParallelScavengeHeap::ParStrongRootsScope psrs;
2330
2331 GCTaskQueue* q = GCTaskQueue::create();
2332
2333 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2334 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2335 // We scan the thread roots in parallel
2336 Threads::create_thread_roots_marking_tasks(q);
2337 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2338 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2339 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2340 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2341 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2342 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2343 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2344
2345 if (active_gc_threads > 1) {
2346 for (uint j = 0; j < active_gc_threads; j++) {
2347 q->enqueue(new StealMarkingTask(&terminator));
2348 }
2349 }
2350
2351 gc_task_manager()->execute_and_wait(q);
2352 }
2353
2354 // Process reference objects found during marking
2355 {
2356 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2357 if (ref_processor()->processing_is_mt()) {
2358 RefProcTaskExecutor task_executor;
2359 ref_processor()->process_discovered_references(
2360 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2361 &task_executor);
2362 } else {
2363 ref_processor()->process_discovered_references(
2364 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2365 }
2366 }
2367
2368 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2369
2370 // This is the point where the entire marking should have completed.
2371 assert(cm->marking_stacks_empty(), "Marking should have completed");
2372
2373 // Follow system dictionary roots and unload classes.
2374 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2375
2376 // Unload nmethods.
2377 CodeCache::do_unloading(is_alive_closure(), purged_class);
2378
2379 // Prune dead klasses from subklass/sibling/implementor lists.
2380 Klass::clean_weak_klass_links(is_alive_closure());
2381
2382 // Delete entries for dead interned strings.
2383 StringTable::unlink(is_alive_closure());
2384
2385 // Clean up unreferenced symbols in symbol table.
2386 SymbolTable::unlink();
2387 }
2388
2389 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) {
2390 ClassLoaderData* cld = klass->class_loader_data();
2391 // The actual processing of the klass is done when we
2392 // traverse the list of Klasses in the class loader data.
2393 PSParallelCompact::follow_class_loader(cm, cld);
2394 }
2395
2396 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) {
2397 ClassLoaderData* cld = klass->class_loader_data();
2398 // The actual processing of the klass is done when we
2399 // traverse the list of Klasses in the class loader data.
2400 PSParallelCompact::adjust_class_loader(cm, cld);
2401 }
2402
2403 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2404 ClassLoaderData* cld) {
2405 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2406 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2407
2408 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2409 }
2410
2411 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm,
2412 ClassLoaderData* cld) {
2413 cld->oops_do(PSParallelCompact::adjust_pointer_closure(),
2414 PSParallelCompact::adjust_klass_closure(),
2415 true);
2416 }
2417
2418 // This should be moved to the shared markSweep code!
2419 class PSAlwaysTrueClosure: public BoolObjectClosure {
2420 public:
2421 bool do_object_b(oop p) { return true; }
2422 };
2423 static PSAlwaysTrueClosure always_true;
2424
2425 void PSParallelCompact::adjust_roots() {
2426 // Adjust the pointers to reflect the new locations
2427 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2428
2429 // Need new claim bits when tracing through and adjusting pointers.
2430 ClassLoaderDataGraph::clear_claimed_marks();
2431
2432 // General strong roots.
2433 Universe::oops_do(adjust_pointer_closure());
2434 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles
2435 CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2436 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2437 ObjectSynchronizer::oops_do(adjust_pointer_closure());
2438 FlatProfiler::oops_do(adjust_pointer_closure());
2439 Management::oops_do(adjust_pointer_closure());
2440 JvmtiExport::oops_do(adjust_pointer_closure());
2441 // SO_AllClasses
2442 SystemDictionary::oops_do(adjust_pointer_closure());
2443 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2444
2445 // Now adjust pointers in remaining weak roots. (All of which should
2446 // have been cleared if they pointed to non-surviving objects.)
2447 // Global (weak) JNI handles
2448 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2449
2450 CodeCache::oops_do(adjust_pointer_closure());
2451 StringTable::oops_do(adjust_pointer_closure());
2452 ref_processor()->weak_oops_do(adjust_pointer_closure());
2453 // Roots were visited so references into the young gen in roots
2454 // may have been scanned. Process them also.
2455 // Should the reference processor have a span that excludes
2456 // young gen objects?
2457 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2458 }
2459
2460 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2461 uint parallel_gc_threads)
2462 {
2463 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2464
2465 // Find the threads that are active
2466 unsigned int which = 0;
2467
2468 const uint task_count = MAX2(parallel_gc_threads, 1U);
2469 for (uint j = 0; j < task_count; j++) {
2470 q->enqueue(new DrainStacksCompactionTask(j));
2471 ParCompactionManager::verify_region_list_empty(j);
2472 // Set the region stacks variables to "no" region stack values
2473 // so that they will be recognized and needing a region stack
2474 // in the stealing tasks if they do not get one by executing
2475 // a draining stack.
2476 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2477 cm->set_region_stack(NULL);
2478 cm->set_region_stack_index((uint)max_uintx);
2479 }
2480 ParCompactionManager::reset_recycled_stack_index();
2481
2482 // Find all regions that are available (can be filled immediately) and
2483 // distribute them to the thread stacks. The iteration is done in reverse
2484 // order (high to low) so the regions will be removed in ascending order.
2485
2486 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2487
2488 size_t fillable_regions = 0; // A count for diagnostic purposes.
2489 // A region index which corresponds to the tasks created above.
2490 // "which" must be 0 <= which < task_count
2491
2492 which = 0;
2493 // id + 1 is used to test termination so unsigned can
2494 // be used with an old_space_id == 0.
2495 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2496 SpaceInfo* const space_info = _space_info + id;
2497 MutableSpace* const space = space_info->space();
2498 HeapWord* const new_top = space_info->new_top();
2499
2500 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2501 const size_t end_region =
2502 sd.addr_to_region_idx(sd.region_align_up(new_top));
2503
2504 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2505 if (sd.region(cur)->claim_unsafe()) {
2506 ParCompactionManager::region_list_push(which, cur);
2507
2508 if (TraceParallelOldGCCompactionPhase && Verbose) {
2509 const size_t count_mod_8 = fillable_regions & 7;
2510 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2511 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2512 if (count_mod_8 == 7) gclog_or_tty->cr();
2513 }
2514
2515 NOT_PRODUCT(++fillable_regions;)
2516
2517 // Assign regions to tasks in round-robin fashion.
2518 if (++which == task_count) {
2519 assert(which <= parallel_gc_threads,
2520 "Inconsistent number of workers");
2521 which = 0;
2522 }
2523 }
2524 }
2525 }
2526
2527 if (TraceParallelOldGCCompactionPhase) {
2528 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2529 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2530 }
2531 }
2532
2533 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2534
2535 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2536 uint parallel_gc_threads) {
2537 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2538
2539 ParallelCompactData& sd = PSParallelCompact::summary_data();
2540
2541 // Iterate over all the spaces adding tasks for updating
2542 // regions in the dense prefix. Assume that 1 gc thread
2543 // will work on opening the gaps and the remaining gc threads
2544 // will work on the dense prefix.
2545 unsigned int space_id;
2546 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2547 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2548 const MutableSpace* const space = _space_info[space_id].space();
2549
2550 if (dense_prefix_end == space->bottom()) {
2551 // There is no dense prefix for this space.
2552 continue;
2553 }
2554
2555 // The dense prefix is before this region.
2556 size_t region_index_end_dense_prefix =
2557 sd.addr_to_region_idx(dense_prefix_end);
2558 RegionData* const dense_prefix_cp =
2559 sd.region(region_index_end_dense_prefix);
2560 assert(dense_prefix_end == space->end() ||
2561 dense_prefix_cp->available() ||
2562 dense_prefix_cp->claimed(),
2563 "The region after the dense prefix should always be ready to fill");
2564
2565 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2566
2567 // Is there dense prefix work?
2568 size_t total_dense_prefix_regions =
2569 region_index_end_dense_prefix - region_index_start;
2570 // How many regions of the dense prefix should be given to
2571 // each thread?
2572 if (total_dense_prefix_regions > 0) {
2573 uint tasks_for_dense_prefix = 1;
2574 if (total_dense_prefix_regions <=
2575 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2576 // Don't over partition. This assumes that
2577 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2578 // so there are not many regions to process.
2579 tasks_for_dense_prefix = parallel_gc_threads;
2580 } else {
2581 // Over partition
2582 tasks_for_dense_prefix = parallel_gc_threads *
2583 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2584 }
2585 size_t regions_per_thread = total_dense_prefix_regions /
2586 tasks_for_dense_prefix;
2587 // Give each thread at least 1 region.
2588 if (regions_per_thread == 0) {
2589 regions_per_thread = 1;
2590 }
2591
2592 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2593 if (region_index_start >= region_index_end_dense_prefix) {
2594 break;
2595 }
2596 // region_index_end is not processed
2597 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2598 region_index_end_dense_prefix);
2599 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2600 region_index_start,
2601 region_index_end));
2602 region_index_start = region_index_end;
2603 }
2604 }
2605 // This gets any part of the dense prefix that did not
2606 // fit evenly.
2607 if (region_index_start < region_index_end_dense_prefix) {
2608 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2609 region_index_start,
2610 region_index_end_dense_prefix));
2611 }
2612 }
2613 }
2614
2615 void PSParallelCompact::enqueue_region_stealing_tasks(
2616 GCTaskQueue* q,
2617 ParallelTaskTerminator* terminator_ptr,
2618 uint parallel_gc_threads) {
2619 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2620
2621 // Once a thread has drained it's stack, it should try to steal regions from
2622 // other threads.
2623 if (parallel_gc_threads > 1) {
2624 for (uint j = 0; j < parallel_gc_threads; j++) {
2625 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2626 }
2627 }
2628 }
2629
2630 void PSParallelCompact::compact() {
2631 // trace("5");
2632 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2633
2634 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2635 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2636 PSOldGen* old_gen = heap->old_gen();
2637 old_gen->start_array()->reset();
2638 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2639 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2640 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2641 ParallelTaskTerminator terminator(active_gc_threads, qset);
2642
2643 GCTaskQueue* q = GCTaskQueue::create();
2644 enqueue_region_draining_tasks(q, active_gc_threads);
2645 enqueue_dense_prefix_tasks(q, active_gc_threads);
2646 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2647
2648 {
2649 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2650
2651 gc_task_manager()->execute_and_wait(q);
2652
2653 #ifdef ASSERT
2654 // Verify that all regions have been processed before the deferred updates.
2655 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2656 verify_complete(SpaceId(id));
2657 }
2658 #endif
2659 }
2660
2661 {
2662 // Update the deferred objects, if any. Any compaction manager can be used.
2663 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2664 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2665 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2666 update_deferred_objects(cm, SpaceId(id));
2667 }
2668 }
2669 }
2670
2671 #ifdef ASSERT
2672 void PSParallelCompact::verify_complete(SpaceId space_id) {
2673 // All Regions between space bottom() to new_top() should be marked as filled
2674 // and all Regions between new_top() and top() should be available (i.e.,
2675 // should have been emptied).
2676 ParallelCompactData& sd = summary_data();
2677 SpaceInfo si = _space_info[space_id];
2678 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2679 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2680 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2681 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2682 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2683
2684 bool issued_a_warning = false;
2685
2686 size_t cur_region;
2687 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2688 const RegionData* const c = sd.region(cur_region);
2689 if (!c->completed()) {
2690 warning("region " SIZE_FORMAT " not filled: "
2691 "destination_count=" SIZE_FORMAT,
2692 cur_region, c->destination_count());
2693 issued_a_warning = true;
2694 }
2695 }
2696
2697 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2698 const RegionData* const c = sd.region(cur_region);
2699 if (!c->available()) {
2700 warning("region " SIZE_FORMAT " not empty: "
2701 "destination_count=" SIZE_FORMAT,
2702 cur_region, c->destination_count());
2703 issued_a_warning = true;
2704 }
2705 }
2706
2707 if (issued_a_warning) {
2708 print_region_ranges();
2709 }
2710 }
2711 #endif // #ifdef ASSERT
2712
2713 // Update interior oops in the ranges of regions [beg_region, end_region).
2714 void
2715 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2716 SpaceId space_id,
2717 size_t beg_region,
2718 size_t end_region) {
2719 ParallelCompactData& sd = summary_data();
2720 ParMarkBitMap* const mbm = mark_bitmap();
2721
2722 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2723 HeapWord* const end_addr = sd.region_to_addr(end_region);
2724 assert(beg_region <= end_region, "bad region range");
2725 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2726
2727 #ifdef ASSERT
2728 // Claim the regions to avoid triggering an assert when they are marked as
2729 // filled.
2730 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2731 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2732 }
2733 #endif // #ifdef ASSERT
2734
2735 if (beg_addr != space(space_id)->bottom()) {
2736 // Find the first live object or block of dead space that *starts* in this
2737 // range of regions. If a partial object crosses onto the region, skip it;
2738 // it will be marked for 'deferred update' when the object head is
2739 // processed. If dead space crosses onto the region, it is also skipped; it
2740 // will be filled when the prior region is processed. If neither of those
2741 // apply, the first word in the region is the start of a live object or dead
2742 // space.
2743 assert(beg_addr > space(space_id)->bottom(), "sanity");
2744 const RegionData* const cp = sd.region(beg_region);
2745 if (cp->partial_obj_size() != 0) {
2746 beg_addr = sd.partial_obj_end(beg_region);
2747 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2748 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2749 }
2750 }
2751
2752 if (beg_addr < end_addr) {
2753 // A live object or block of dead space starts in this range of Regions.
2754 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2755
2756 // Create closures and iterate.
2757 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2758 FillClosure fill_closure(cm, space_id);
2759 ParMarkBitMap::IterationStatus status;
2760 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2761 dense_prefix_end);
2762 if (status == ParMarkBitMap::incomplete) {
2763 update_closure.do_addr(update_closure.source());
2764 }
2765 }
2766
2767 // Mark the regions as filled.
2768 RegionData* const beg_cp = sd.region(beg_region);
2769 RegionData* const end_cp = sd.region(end_region);
2770 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2771 cp->set_completed();
2772 }
2773 }
2774
2775 // Return the SpaceId for the space containing addr. If addr is not in the
2776 // heap, last_space_id is returned. In debug mode it expects the address to be
2777 // in the heap and asserts such.
2778 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2779 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2780
2781 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2782 if (_space_info[id].space()->contains(addr)) {
2783 return SpaceId(id);
2784 }
2785 }
2786
2787 assert(false, "no space contains the addr");
2788 return last_space_id;
2789 }
2790
2791 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2792 SpaceId id) {
2793 assert(id < last_space_id, "bad space id");
2794
2795 ParallelCompactData& sd = summary_data();
2796 const SpaceInfo* const space_info = _space_info + id;
2797 ObjectStartArray* const start_array = space_info->start_array();
2798
2799 const MutableSpace* const space = space_info->space();
2800 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2801 HeapWord* const beg_addr = space_info->dense_prefix();
2802 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2803
2804 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2805 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2806 const RegionData* cur_region;
2807 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2808 HeapWord* const addr = cur_region->deferred_obj_addr();
2809 if (addr != NULL) {
2810 if (start_array != NULL) {
2811 start_array->allocate_block(addr);
2812 }
2813 oop(addr)->update_contents(cm);
2814 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2815 }
2816 }
2817 }
2818
2819 // Skip over count live words starting from beg, and return the address of the
2820 // next live word. Unless marked, the word corresponding to beg is assumed to
2821 // be dead. Callers must either ensure beg does not correspond to the middle of
2822 // an object, or account for those live words in some other way. Callers must
2823 // also ensure that there are enough live words in the range [beg, end) to skip.
2824 HeapWord*
2825 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2826 {
2827 assert(count > 0, "sanity");
2828
2829 ParMarkBitMap* m = mark_bitmap();
2830 idx_t bits_to_skip = m->words_to_bits(count);
2831 idx_t cur_beg = m->addr_to_bit(beg);
2832 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2833
2834 do {
2835 cur_beg = m->find_obj_beg(cur_beg, search_end);
2836 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2837 const size_t obj_bits = cur_end - cur_beg + 1;
2838 if (obj_bits > bits_to_skip) {
2839 return m->bit_to_addr(cur_beg + bits_to_skip);
2840 }
2841 bits_to_skip -= obj_bits;
2842 cur_beg = cur_end + 1;
2843 } while (bits_to_skip > 0);
2844
2845 // Skipping the desired number of words landed just past the end of an object.
2846 // Find the start of the next object.
2847 cur_beg = m->find_obj_beg(cur_beg, search_end);
2848 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2849 return m->bit_to_addr(cur_beg);
2850 }
2851
2852 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2853 SpaceId src_space_id,
2854 size_t src_region_idx)
2855 {
2856 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2857
2858 const SplitInfo& split_info = _space_info[src_space_id].split_info();
2859 if (split_info.dest_region_addr() == dest_addr) {
2860 // The partial object ending at the split point contains the first word to
2861 // be copied to dest_addr.
2862 return split_info.first_src_addr();
2863 }
2864
2865 const ParallelCompactData& sd = summary_data();
2866 ParMarkBitMap* const bitmap = mark_bitmap();
2867 const size_t RegionSize = ParallelCompactData::RegionSize;
2868
2869 assert(sd.is_region_aligned(dest_addr), "not aligned");
2870 const RegionData* const src_region_ptr = sd.region(src_region_idx);
2871 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2872 HeapWord* const src_region_destination = src_region_ptr->destination();
2873
2874 assert(dest_addr >= src_region_destination, "wrong src region");
2875 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2876
2877 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2878 HeapWord* const src_region_end = src_region_beg + RegionSize;
2879
2880 HeapWord* addr = src_region_beg;
2881 if (dest_addr == src_region_destination) {
2882 // Return the first live word in the source region.
2883 if (partial_obj_size == 0) {
2884 addr = bitmap->find_obj_beg(addr, src_region_end);
2885 assert(addr < src_region_end, "no objects start in src region");
2886 }
2887 return addr;
2888 }
2889
2890 // Must skip some live data.
2891 size_t words_to_skip = dest_addr - src_region_destination;
2892 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2893
2894 if (partial_obj_size >= words_to_skip) {
2895 // All the live words to skip are part of the partial object.
2896 addr += words_to_skip;
2897 if (partial_obj_size == words_to_skip) {
2898 // Find the first live word past the partial object.
2899 addr = bitmap->find_obj_beg(addr, src_region_end);
2900 assert(addr < src_region_end, "wrong src region");
2901 }
2902 return addr;
2903 }
2904
2905 // Skip over the partial object (if any).
2906 if (partial_obj_size != 0) {
2907 words_to_skip -= partial_obj_size;
2908 addr += partial_obj_size;
2909 }
2910
2911 // Skip over live words due to objects that start in the region.
2912 addr = skip_live_words(addr, src_region_end, words_to_skip);
2913 assert(addr < src_region_end, "wrong src region");
2914 return addr;
2915 }
2916
2917 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2918 SpaceId src_space_id,
2919 size_t beg_region,
2920 HeapWord* end_addr)
2921 {
2922 ParallelCompactData& sd = summary_data();
2923
2924 #ifdef ASSERT
2925 MutableSpace* const src_space = _space_info[src_space_id].space();
2926 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2927 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2928 "src_space_id does not match beg_addr");
2929 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2930 "src_space_id does not match end_addr");
2931 #endif // #ifdef ASSERT
2932
2933 RegionData* const beg = sd.region(beg_region);
2934 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2935
2936 // Regions up to new_top() are enqueued if they become available.
2937 HeapWord* const new_top = _space_info[src_space_id].new_top();
2938 RegionData* const enqueue_end =
2939 sd.addr_to_region_ptr(sd.region_align_up(new_top));
2940
2941 for (RegionData* cur = beg; cur < end; ++cur) {
2942 assert(cur->data_size() > 0, "region must have live data");
2943 cur->decrement_destination_count();
2944 if (cur < enqueue_end && cur->available() && cur->claim()) {
2945 cm->push_region(sd.region(cur));
2946 }
2947 }
2948 }
2949
2950 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2951 SpaceId& src_space_id,
2952 HeapWord*& src_space_top,
2953 HeapWord* end_addr)
2954 {
2955 typedef ParallelCompactData::RegionData RegionData;
2956
2957 ParallelCompactData& sd = PSParallelCompact::summary_data();
2958 const size_t region_size = ParallelCompactData::RegionSize;
2959
2960 size_t src_region_idx = 0;
2961
2962 // Skip empty regions (if any) up to the top of the space.
2963 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2964 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2965 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2966 const RegionData* const top_region_ptr =
2967 sd.addr_to_region_ptr(top_aligned_up);
2968 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2969 ++src_region_ptr;
2970 }
2971
2972 if (src_region_ptr < top_region_ptr) {
2973 // The next source region is in the current space. Update src_region_idx
2974 // and the source address to match src_region_ptr.
2975 src_region_idx = sd.region(src_region_ptr);
2976 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2977 if (src_region_addr > closure.source()) {
2978 closure.set_source(src_region_addr);
2979 }
2980 return src_region_idx;
2981 }
2982
2983 // Switch to a new source space and find the first non-empty region.
2984 unsigned int space_id = src_space_id + 1;
2985 assert(space_id < last_space_id, "not enough spaces");
2986
2987 HeapWord* const destination = closure.destination();
2988
2989 do {
2990 MutableSpace* space = _space_info[space_id].space();
2991 HeapWord* const bottom = space->bottom();
2992 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2993
2994 // Iterate over the spaces that do not compact into themselves.
2995 if (bottom_cp->destination() != bottom) {
2996 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2997 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2998
2999 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3000 if (src_cp->live_obj_size() > 0) {
3001 // Found it.
3002 assert(src_cp->destination() == destination,
3003 "first live obj in the space must match the destination");
3004 assert(src_cp->partial_obj_size() == 0,
3005 "a space cannot begin with a partial obj");
3006
3007 src_space_id = SpaceId(space_id);
3008 src_space_top = space->top();
3009 const size_t src_region_idx = sd.region(src_cp);
3010 closure.set_source(sd.region_to_addr(src_region_idx));
3011 return src_region_idx;
3012 } else {
3013 assert(src_cp->data_size() == 0, "sanity");
3014 }
3015 }
3016 }
3017 } while (++space_id < last_space_id);
3018
3019 assert(false, "no source region was found");
3020 return 0;
3021 }
3022
3023 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3024 {
3025 typedef ParMarkBitMap::IterationStatus IterationStatus;
3026 const size_t RegionSize = ParallelCompactData::RegionSize;
3027 ParMarkBitMap* const bitmap = mark_bitmap();
3028 ParallelCompactData& sd = summary_data();
3029 RegionData* const region_ptr = sd.region(region_idx);
3030
3031 // Get the items needed to construct the closure.
3032 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3033 SpaceId dest_space_id = space_id(dest_addr);
3034 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3035 HeapWord* new_top = _space_info[dest_space_id].new_top();
3036 assert(dest_addr < new_top, "sanity");
3037 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3038
3039 // Get the source region and related info.
3040 size_t src_region_idx = region_ptr->source_region();
3041 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3042 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3043
3044 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3045 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3046
3047 // Adjust src_region_idx to prepare for decrementing destination counts (the
3048 // destination count is not decremented when a region is copied to itself).
3049 if (src_region_idx == region_idx) {
3050 src_region_idx += 1;
3051 }
3052
3053 if (bitmap->is_unmarked(closure.source())) {
3054 // The first source word is in the middle of an object; copy the remainder
3055 // of the object or as much as will fit. The fact that pointer updates were
3056 // deferred will be noted when the object header is processed.
3057 HeapWord* const old_src_addr = closure.source();
3058 closure.copy_partial_obj();
3059 if (closure.is_full()) {
3060 decrement_destination_counts(cm, src_space_id, src_region_idx,
3061 closure.source());
3062 region_ptr->set_deferred_obj_addr(NULL);
3063 region_ptr->set_completed();
3064 return;
3065 }
3066
3067 HeapWord* const end_addr = sd.region_align_down(closure.source());
3068 if (sd.region_align_down(old_src_addr) != end_addr) {
3069 // The partial object was copied from more than one source region.
3070 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3071
3072 // Move to the next source region, possibly switching spaces as well. All
3073 // args except end_addr may be modified.
3074 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3075 end_addr);
3076 }
3077 }
3078
3079 do {
3080 HeapWord* const cur_addr = closure.source();
3081 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3082 src_space_top);
3083 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3084
3085 if (status == ParMarkBitMap::incomplete) {
3086 // The last obj that starts in the source region does not end in the
3087 // region.
3088 assert(closure.source() < end_addr, "sanity");
3089 HeapWord* const obj_beg = closure.source();
3090 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3091 src_space_top);
3092 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3093 if (obj_end < range_end) {
3094 // The end was found; the entire object will fit.
3095 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3096 assert(status != ParMarkBitMap::would_overflow, "sanity");
3097 } else {
3098 // The end was not found; the object will not fit.
3099 assert(range_end < src_space_top, "obj cannot cross space boundary");
3100 status = ParMarkBitMap::would_overflow;
3101 }
3102 }
3103
3104 if (status == ParMarkBitMap::would_overflow) {
3105 // The last object did not fit. Note that interior oop updates were
3106 // deferred, then copy enough of the object to fill the region.
3107 region_ptr->set_deferred_obj_addr(closure.destination());
3108 status = closure.copy_until_full(); // copies from closure.source()
3109
3110 decrement_destination_counts(cm, src_space_id, src_region_idx,
3111 closure.source());
3112 region_ptr->set_completed();
3113 return;
3114 }
3115
3116 if (status == ParMarkBitMap::full) {
3117 decrement_destination_counts(cm, src_space_id, src_region_idx,
3118 closure.source());
3119 region_ptr->set_deferred_obj_addr(NULL);
3120 region_ptr->set_completed();
3121 return;
3122 }
3123
3124 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3125
3126 // Move to the next source region, possibly switching spaces as well. All
3127 // args except end_addr may be modified.
3128 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3129 end_addr);
3130 } while (true);
3131 }
3132
3133 void
3134 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3135 const MutableSpace* sp = space(space_id);
3136 if (sp->is_empty()) {
3137 return;
3138 }
3139
3140 ParallelCompactData& sd = PSParallelCompact::summary_data();
3141 ParMarkBitMap* const bitmap = mark_bitmap();
3142 HeapWord* const dp_addr = dense_prefix(space_id);
3143 HeapWord* beg_addr = sp->bottom();
3144 HeapWord* end_addr = sp->top();
3145
3146 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3147
3148 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3149 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3150 if (beg_region < dp_region) {
3151 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3152 }
3153
3154 // The destination of the first live object that starts in the region is one
3155 // past the end of the partial object entering the region (if any).
3156 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3157 HeapWord* const new_top = _space_info[space_id].new_top();
3158 assert(new_top >= dest_addr, "bad new_top value");
3159 const size_t words = pointer_delta(new_top, dest_addr);
3160
3161 if (words > 0) {
3162 ObjectStartArray* start_array = _space_info[space_id].start_array();
3163 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3164
3165 ParMarkBitMap::IterationStatus status;
3166 status = bitmap->iterate(&closure, dest_addr, end_addr);
3167 assert(status == ParMarkBitMap::full, "iteration not complete");
3168 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3169 "live objects skipped because closure is full");
3170 }
3171 }
3172
3173 jlong PSParallelCompact::millis_since_last_gc() {
3174 // We need a monotonically non-deccreasing time in ms but
3175 // os::javaTimeMillis() does not guarantee monotonicity.
3176 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3177 jlong ret_val = now - _time_of_last_gc;
3178 // XXX See note in genCollectedHeap::millis_since_last_gc().
3179 if (ret_val < 0) {
3180 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3181 return 0;
3182 }
3183 return ret_val;
3184 }
3185
3186 void PSParallelCompact::reset_millis_since_last_gc() {
3187 // We need a monotonically non-deccreasing time in ms but
3188 // os::javaTimeMillis() does not guarantee monotonicity.
3189 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3190 }
3191
3192 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3193 {
3194 if (source() != destination()) {
3195 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3196 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3197 }
3198 update_state(words_remaining());
3199 assert(is_full(), "sanity");
3200 return ParMarkBitMap::full;
3201 }
3202
3203 void MoveAndUpdateClosure::copy_partial_obj()
3204 {
3205 size_t words = words_remaining();
3206
3207 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3208 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3209 if (end_addr < range_end) {
3210 words = bitmap()->obj_size(source(), end_addr);
3211 }
3212
3213 // This test is necessary; if omitted, the pointer updates to a partial object
3214 // that crosses the dense prefix boundary could be overwritten.
3215 if (source() != destination()) {
3216 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3217 Copy::aligned_conjoint_words(source(), destination(), words);
3218 }
3219 update_state(words);
3220 }
3221
3222 ParMarkBitMapClosure::IterationStatus
3223 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3224 assert(destination() != NULL, "sanity");
3225 assert(bitmap()->obj_size(addr) == words, "bad size");
3226
3227 _source = addr;
3228 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3229 destination(), "wrong destination");
3230
3231 if (words > words_remaining()) {
3232 return ParMarkBitMap::would_overflow;
3233 }
3234
3235 // The start_array must be updated even if the object is not moving.
3236 if (_start_array != NULL) {
3237 _start_array->allocate_block(destination());
3238 }
3239
3240 if (destination() != source()) {
3241 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3242 Copy::aligned_conjoint_words(source(), destination(), words);
3243 }
3244
3245 oop moved_oop = (oop) destination();
3246 moved_oop->update_contents(compaction_manager());
3247 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3248
3249 update_state(words);
3250 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3251 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3252 }
3253
3254 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3255 ParCompactionManager* cm,
3256 PSParallelCompact::SpaceId space_id) :
3257 ParMarkBitMapClosure(mbm, cm),
3258 _space_id(space_id),
3259 _start_array(PSParallelCompact::start_array(space_id))
3260 {
3261 }
3262
3263 // Updates the references in the object to their new values.
3264 ParMarkBitMapClosure::IterationStatus
3265 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3266 do_addr(addr);
3267 return ParMarkBitMap::incomplete;
3268 }