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