Mon, 28 Jul 2008 15:30:23 -0700
Merge
1 /*
2 * Copyright 2005-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_psParallelCompact.cpp.incl"
28 #include <math.h>
30 // All sizes are in HeapWords.
31 const size_t ParallelCompactData::Log2ChunkSize = 9; // 512 words
32 const size_t ParallelCompactData::ChunkSize = (size_t)1 << Log2ChunkSize;
33 const size_t ParallelCompactData::ChunkSizeBytes = ChunkSize << LogHeapWordSize;
34 const size_t ParallelCompactData::ChunkSizeOffsetMask = ChunkSize - 1;
35 const size_t ParallelCompactData::ChunkAddrOffsetMask = ChunkSizeBytes - 1;
36 const size_t ParallelCompactData::ChunkAddrMask = ~ChunkAddrOffsetMask;
38 // 32-bit: 128 words covers 4 bitmap words
39 // 64-bit: 128 words covers 2 bitmap words
40 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
41 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
42 const size_t ParallelCompactData::BlockOffsetMask = BlockSize - 1;
43 const size_t ParallelCompactData::BlockMask = ~BlockOffsetMask;
45 const size_t ParallelCompactData::BlocksPerChunk = ChunkSize / BlockSize;
47 const ParallelCompactData::ChunkData::chunk_sz_t
48 ParallelCompactData::ChunkData::dc_shift = 27;
50 const ParallelCompactData::ChunkData::chunk_sz_t
51 ParallelCompactData::ChunkData::dc_mask = ~0U << dc_shift;
53 const ParallelCompactData::ChunkData::chunk_sz_t
54 ParallelCompactData::ChunkData::dc_one = 0x1U << dc_shift;
56 const ParallelCompactData::ChunkData::chunk_sz_t
57 ParallelCompactData::ChunkData::los_mask = ~dc_mask;
59 const ParallelCompactData::ChunkData::chunk_sz_t
60 ParallelCompactData::ChunkData::dc_claimed = 0x8U << dc_shift;
62 const ParallelCompactData::ChunkData::chunk_sz_t
63 ParallelCompactData::ChunkData::dc_completed = 0xcU << dc_shift;
65 #ifdef ASSERT
66 short ParallelCompactData::BlockData::_cur_phase = 0;
67 #endif
69 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
70 bool PSParallelCompact::_print_phases = false;
72 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
73 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL;
75 double PSParallelCompact::_dwl_mean;
76 double PSParallelCompact::_dwl_std_dev;
77 double PSParallelCompact::_dwl_first_term;
78 double PSParallelCompact::_dwl_adjustment;
79 #ifdef ASSERT
80 bool PSParallelCompact::_dwl_initialized = false;
81 #endif // #ifdef ASSERT
83 #ifdef VALIDATE_MARK_SWEEP
84 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL;
85 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL;
86 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL;
87 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL;
88 size_t PSParallelCompact::_live_oops_index = 0;
89 size_t PSParallelCompact::_live_oops_index_at_perm = 0;
90 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL;
91 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL;
92 bool PSParallelCompact::_pointer_tracking = false;
93 bool PSParallelCompact::_root_tracking = true;
95 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
96 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
97 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL;
98 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
99 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
100 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL;
101 #endif
103 // XXX beg - verification code; only works while we also mark in object headers
104 static void
105 verify_mark_bitmap(ParMarkBitMap& _mark_bitmap)
106 {
107 ParallelScavengeHeap* heap = PSParallelCompact::gc_heap();
109 PSPermGen* perm_gen = heap->perm_gen();
110 PSOldGen* old_gen = heap->old_gen();
111 PSYoungGen* young_gen = heap->young_gen();
113 MutableSpace* perm_space = perm_gen->object_space();
114 MutableSpace* old_space = old_gen->object_space();
115 MutableSpace* eden_space = young_gen->eden_space();
116 MutableSpace* from_space = young_gen->from_space();
117 MutableSpace* to_space = young_gen->to_space();
119 // 'from_space' here is the survivor space at the lower address.
120 if (to_space->bottom() < from_space->bottom()) {
121 from_space = to_space;
122 to_space = young_gen->from_space();
123 }
125 HeapWord* boundaries[12];
126 unsigned int bidx = 0;
127 const unsigned int bidx_max = sizeof(boundaries) / sizeof(boundaries[0]);
129 boundaries[0] = perm_space->bottom();
130 boundaries[1] = perm_space->top();
131 boundaries[2] = old_space->bottom();
132 boundaries[3] = old_space->top();
133 boundaries[4] = eden_space->bottom();
134 boundaries[5] = eden_space->top();
135 boundaries[6] = from_space->bottom();
136 boundaries[7] = from_space->top();
137 boundaries[8] = to_space->bottom();
138 boundaries[9] = to_space->top();
139 boundaries[10] = to_space->end();
140 boundaries[11] = to_space->end();
142 BitMap::idx_t beg_bit = 0;
143 BitMap::idx_t end_bit;
144 BitMap::idx_t tmp_bit;
145 const BitMap::idx_t last_bit = _mark_bitmap.size();
146 do {
147 HeapWord* addr = _mark_bitmap.bit_to_addr(beg_bit);
148 if (_mark_bitmap.is_marked(beg_bit)) {
149 oop obj = (oop)addr;
150 assert(obj->is_gc_marked(), "obj header is not marked");
151 end_bit = _mark_bitmap.find_obj_end(beg_bit, last_bit);
152 const size_t size = _mark_bitmap.obj_size(beg_bit, end_bit);
153 assert(size == (size_t)obj->size(), "end bit wrong?");
154 beg_bit = _mark_bitmap.find_obj_beg(beg_bit + 1, last_bit);
155 assert(beg_bit > end_bit, "bit set in middle of an obj");
156 } else {
157 if (addr >= boundaries[bidx] && addr < boundaries[bidx + 1]) {
158 // a dead object in the current space.
159 oop obj = (oop)addr;
160 end_bit = _mark_bitmap.addr_to_bit(addr + obj->size());
161 assert(!obj->is_gc_marked(), "obj marked in header, not in bitmap");
162 tmp_bit = beg_bit + 1;
163 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, end_bit);
164 assert(beg_bit == end_bit, "beg bit set in unmarked obj");
165 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, end_bit);
166 assert(beg_bit == end_bit, "end bit set in unmarked obj");
167 } else if (addr < boundaries[bidx + 2]) {
168 // addr is between top in the current space and bottom in the next.
169 end_bit = beg_bit + pointer_delta(boundaries[bidx + 2], addr);
170 tmp_bit = beg_bit;
171 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, end_bit);
172 assert(beg_bit == end_bit, "beg bit set above top");
173 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, end_bit);
174 assert(beg_bit == end_bit, "end bit set above top");
175 bidx += 2;
176 } else if (bidx < bidx_max - 2) {
177 bidx += 2; // ???
178 } else {
179 tmp_bit = beg_bit;
180 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, last_bit);
181 assert(beg_bit == last_bit, "beg bit set outside heap");
182 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, last_bit);
183 assert(beg_bit == last_bit, "end bit set outside heap");
184 }
185 }
186 } while (beg_bit < last_bit);
187 }
188 // XXX end - verification code; only works while we also mark in object headers
190 #ifndef PRODUCT
191 const char* PSParallelCompact::space_names[] = {
192 "perm", "old ", "eden", "from", "to "
193 };
195 void PSParallelCompact::print_chunk_ranges()
196 {
197 tty->print_cr("space bottom top end new_top");
198 tty->print_cr("------ ---------- ---------- ---------- ----------");
200 for (unsigned int id = 0; id < last_space_id; ++id) {
201 const MutableSpace* space = _space_info[id].space();
202 tty->print_cr("%u %s "
203 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
205 id, space_names[id],
206 summary_data().addr_to_chunk_idx(space->bottom()),
207 summary_data().addr_to_chunk_idx(space->top()),
208 summary_data().addr_to_chunk_idx(space->end()),
209 summary_data().addr_to_chunk_idx(_space_info[id].new_top()));
210 }
211 }
213 void
214 print_generic_summary_chunk(size_t i, const ParallelCompactData::ChunkData* c)
215 {
216 #define CHUNK_IDX_FORMAT SIZE_FORMAT_W(7)
217 #define CHUNK_DATA_FORMAT SIZE_FORMAT_W(5)
219 ParallelCompactData& sd = PSParallelCompact::summary_data();
220 size_t dci = c->destination() ? sd.addr_to_chunk_idx(c->destination()) : 0;
221 tty->print_cr(CHUNK_IDX_FORMAT " " PTR_FORMAT " "
222 CHUNK_IDX_FORMAT " " PTR_FORMAT " "
223 CHUNK_DATA_FORMAT " " CHUNK_DATA_FORMAT " "
224 CHUNK_DATA_FORMAT " " CHUNK_IDX_FORMAT " %d",
225 i, c->data_location(), dci, c->destination(),
226 c->partial_obj_size(), c->live_obj_size(),
227 c->data_size(), c->source_chunk(), c->destination_count());
229 #undef CHUNK_IDX_FORMAT
230 #undef CHUNK_DATA_FORMAT
231 }
233 void
234 print_generic_summary_data(ParallelCompactData& summary_data,
235 HeapWord* const beg_addr,
236 HeapWord* const end_addr)
237 {
238 size_t total_words = 0;
239 size_t i = summary_data.addr_to_chunk_idx(beg_addr);
240 const size_t last = summary_data.addr_to_chunk_idx(end_addr);
241 HeapWord* pdest = 0;
243 while (i <= last) {
244 ParallelCompactData::ChunkData* c = summary_data.chunk(i);
245 if (c->data_size() != 0 || c->destination() != pdest) {
246 print_generic_summary_chunk(i, c);
247 total_words += c->data_size();
248 pdest = c->destination();
249 }
250 ++i;
251 }
253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
254 }
256 void
257 print_generic_summary_data(ParallelCompactData& summary_data,
258 SpaceInfo* space_info)
259 {
260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
261 const MutableSpace* space = space_info[id].space();
262 print_generic_summary_data(summary_data, space->bottom(),
263 MAX2(space->top(), space_info[id].new_top()));
264 }
265 }
267 void
268 print_initial_summary_chunk(size_t i,
269 const ParallelCompactData::ChunkData* c,
270 bool newline = true)
271 {
272 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
273 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
275 i, c->destination(),
276 c->partial_obj_size(), c->live_obj_size(),
277 c->data_size(), c->source_chunk(), c->destination_count());
278 if (newline) tty->cr();
279 }
281 void
282 print_initial_summary_data(ParallelCompactData& summary_data,
283 const MutableSpace* space) {
284 if (space->top() == space->bottom()) {
285 return;
286 }
288 const size_t chunk_size = ParallelCompactData::ChunkSize;
289 HeapWord* const top_aligned_up = summary_data.chunk_align_up(space->top());
290 const size_t end_chunk = summary_data.addr_to_chunk_idx(top_aligned_up);
291 const ParallelCompactData::ChunkData* c = summary_data.chunk(end_chunk - 1);
292 HeapWord* end_addr = c->destination() + c->data_size();
293 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
295 // Print (and count) the full chunks at the beginning of the space.
296 size_t full_chunk_count = 0;
297 size_t i = summary_data.addr_to_chunk_idx(space->bottom());
298 while (i < end_chunk && summary_data.chunk(i)->data_size() == chunk_size) {
299 print_initial_summary_chunk(i, summary_data.chunk(i));
300 ++full_chunk_count;
301 ++i;
302 }
304 size_t live_to_right = live_in_space - full_chunk_count * chunk_size;
306 double max_reclaimed_ratio = 0.0;
307 size_t max_reclaimed_ratio_chunk = 0;
308 size_t max_dead_to_right = 0;
309 size_t max_live_to_right = 0;
311 // Print the 'reclaimed ratio' for chunks while there is something live in the
312 // chunk or to the right of it. The remaining chunks are empty (and
313 // uninteresting), and computing the ratio will result in division by 0.
314 while (i < end_chunk && live_to_right > 0) {
315 c = summary_data.chunk(i);
316 HeapWord* const chunk_addr = summary_data.chunk_to_addr(i);
317 const size_t used_to_right = pointer_delta(space->top(), chunk_addr);
318 const size_t dead_to_right = used_to_right - live_to_right;
319 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
321 if (reclaimed_ratio > max_reclaimed_ratio) {
322 max_reclaimed_ratio = reclaimed_ratio;
323 max_reclaimed_ratio_chunk = i;
324 max_dead_to_right = dead_to_right;
325 max_live_to_right = live_to_right;
326 }
328 print_initial_summary_chunk(i, c, false);
329 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
330 reclaimed_ratio, dead_to_right, live_to_right);
332 live_to_right -= c->data_size();
333 ++i;
334 }
336 // Any remaining chunks are empty. Print one more if there is one.
337 if (i < end_chunk) {
338 print_initial_summary_chunk(i, summary_data.chunk(i));
339 }
341 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
342 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
343 max_reclaimed_ratio_chunk, max_dead_to_right,
344 max_live_to_right, max_reclaimed_ratio);
345 }
347 void
348 print_initial_summary_data(ParallelCompactData& summary_data,
349 SpaceInfo* space_info) {
350 unsigned int id = PSParallelCompact::perm_space_id;
351 const MutableSpace* space;
352 do {
353 space = space_info[id].space();
354 print_initial_summary_data(summary_data, space);
355 } while (++id < PSParallelCompact::eden_space_id);
357 do {
358 space = space_info[id].space();
359 print_generic_summary_data(summary_data, space->bottom(), space->top());
360 } while (++id < PSParallelCompact::last_space_id);
361 }
362 #endif // #ifndef PRODUCT
364 #ifdef ASSERT
365 size_t add_obj_count;
366 size_t add_obj_size;
367 size_t mark_bitmap_count;
368 size_t mark_bitmap_size;
369 #endif // #ifdef ASSERT
371 ParallelCompactData::ParallelCompactData()
372 {
373 _region_start = 0;
375 _chunk_vspace = 0;
376 _chunk_data = 0;
377 _chunk_count = 0;
379 _block_vspace = 0;
380 _block_data = 0;
381 _block_count = 0;
382 }
384 bool ParallelCompactData::initialize(MemRegion covered_region)
385 {
386 _region_start = covered_region.start();
387 const size_t region_size = covered_region.word_size();
388 DEBUG_ONLY(_region_end = _region_start + region_size;)
390 assert(chunk_align_down(_region_start) == _region_start,
391 "region start not aligned");
392 assert((region_size & ChunkSizeOffsetMask) == 0,
393 "region size not a multiple of ChunkSize");
395 bool result = initialize_chunk_data(region_size);
397 // Initialize the block data if it will be used for updating pointers, or if
398 // this is a debug build.
399 if (!UseParallelOldGCChunkPointerCalc || trueInDebug) {
400 result = result && initialize_block_data(region_size);
401 }
403 return result;
404 }
406 PSVirtualSpace*
407 ParallelCompactData::create_vspace(size_t count, size_t element_size)
408 {
409 const size_t raw_bytes = count * element_size;
410 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
411 const size_t granularity = os::vm_allocation_granularity();
412 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
414 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
415 MAX2(page_sz, granularity);
416 ReservedSpace rs(bytes, rs_align, rs_align > 0);
417 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
418 rs.size());
419 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
420 if (vspace != 0) {
421 if (vspace->expand_by(bytes)) {
422 return vspace;
423 }
424 delete vspace;
425 // Release memory reserved in the space.
426 rs.release();
427 }
429 return 0;
430 }
432 bool ParallelCompactData::initialize_chunk_data(size_t region_size)
433 {
434 const size_t count = (region_size + ChunkSizeOffsetMask) >> Log2ChunkSize;
435 _chunk_vspace = create_vspace(count, sizeof(ChunkData));
436 if (_chunk_vspace != 0) {
437 _chunk_data = (ChunkData*)_chunk_vspace->reserved_low_addr();
438 _chunk_count = count;
439 return true;
440 }
441 return false;
442 }
444 bool ParallelCompactData::initialize_block_data(size_t region_size)
445 {
446 const size_t count = (region_size + BlockOffsetMask) >> Log2BlockSize;
447 _block_vspace = create_vspace(count, sizeof(BlockData));
448 if (_block_vspace != 0) {
449 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
450 _block_count = count;
451 return true;
452 }
453 return false;
454 }
456 void ParallelCompactData::clear()
457 {
458 if (_block_data) {
459 memset(_block_data, 0, _block_vspace->committed_size());
460 }
461 memset(_chunk_data, 0, _chunk_vspace->committed_size());
462 }
464 void ParallelCompactData::clear_range(size_t beg_chunk, size_t end_chunk) {
465 assert(beg_chunk <= _chunk_count, "beg_chunk out of range");
466 assert(end_chunk <= _chunk_count, "end_chunk out of range");
467 assert(ChunkSize % BlockSize == 0, "ChunkSize not a multiple of BlockSize");
469 const size_t chunk_cnt = end_chunk - beg_chunk;
471 if (_block_data) {
472 const size_t blocks_per_chunk = ChunkSize / BlockSize;
473 const size_t beg_block = beg_chunk * blocks_per_chunk;
474 const size_t block_cnt = chunk_cnt * blocks_per_chunk;
475 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
476 }
477 memset(_chunk_data + beg_chunk, 0, chunk_cnt * sizeof(ChunkData));
478 }
480 HeapWord* ParallelCompactData::partial_obj_end(size_t chunk_idx) const
481 {
482 const ChunkData* cur_cp = chunk(chunk_idx);
483 const ChunkData* const end_cp = chunk(chunk_count() - 1);
485 HeapWord* result = chunk_to_addr(chunk_idx);
486 if (cur_cp < end_cp) {
487 do {
488 result += cur_cp->partial_obj_size();
489 } while (cur_cp->partial_obj_size() == ChunkSize && ++cur_cp < end_cp);
490 }
491 return result;
492 }
494 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
495 {
496 const size_t obj_ofs = pointer_delta(addr, _region_start);
497 const size_t beg_chunk = obj_ofs >> Log2ChunkSize;
498 const size_t end_chunk = (obj_ofs + len - 1) >> Log2ChunkSize;
500 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
501 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
503 if (beg_chunk == end_chunk) {
504 // All in one chunk.
505 _chunk_data[beg_chunk].add_live_obj(len);
506 return;
507 }
509 // First chunk.
510 const size_t beg_ofs = chunk_offset(addr);
511 _chunk_data[beg_chunk].add_live_obj(ChunkSize - beg_ofs);
513 klassOop klass = ((oop)addr)->klass();
514 // Middle chunks--completely spanned by this object.
515 for (size_t chunk = beg_chunk + 1; chunk < end_chunk; ++chunk) {
516 _chunk_data[chunk].set_partial_obj_size(ChunkSize);
517 _chunk_data[chunk].set_partial_obj_addr(addr);
518 }
520 // Last chunk.
521 const size_t end_ofs = chunk_offset(addr + len - 1);
522 _chunk_data[end_chunk].set_partial_obj_size(end_ofs + 1);
523 _chunk_data[end_chunk].set_partial_obj_addr(addr);
524 }
526 void
527 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
528 {
529 assert(chunk_offset(beg) == 0, "not ChunkSize aligned");
530 assert(chunk_offset(end) == 0, "not ChunkSize aligned");
532 size_t cur_chunk = addr_to_chunk_idx(beg);
533 const size_t end_chunk = addr_to_chunk_idx(end);
534 HeapWord* addr = beg;
535 while (cur_chunk < end_chunk) {
536 _chunk_data[cur_chunk].set_destination(addr);
537 _chunk_data[cur_chunk].set_destination_count(0);
538 _chunk_data[cur_chunk].set_source_chunk(cur_chunk);
539 _chunk_data[cur_chunk].set_data_location(addr);
541 // Update live_obj_size so the chunk appears completely full.
542 size_t live_size = ChunkSize - _chunk_data[cur_chunk].partial_obj_size();
543 _chunk_data[cur_chunk].set_live_obj_size(live_size);
545 ++cur_chunk;
546 addr += ChunkSize;
547 }
548 }
550 bool ParallelCompactData::summarize(HeapWord* target_beg, HeapWord* target_end,
551 HeapWord* source_beg, HeapWord* source_end,
552 HeapWord** target_next,
553 HeapWord** source_next) {
554 // This is too strict.
555 // assert(chunk_offset(source_beg) == 0, "not ChunkSize aligned");
557 if (TraceParallelOldGCSummaryPhase) {
558 tty->print_cr("tb=" PTR_FORMAT " te=" PTR_FORMAT " "
559 "sb=" PTR_FORMAT " se=" PTR_FORMAT " "
560 "tn=" PTR_FORMAT " sn=" PTR_FORMAT,
561 target_beg, target_end,
562 source_beg, source_end,
563 target_next != 0 ? *target_next : (HeapWord*) 0,
564 source_next != 0 ? *source_next : (HeapWord*) 0);
565 }
567 size_t cur_chunk = addr_to_chunk_idx(source_beg);
568 const size_t end_chunk = addr_to_chunk_idx(chunk_align_up(source_end));
570 HeapWord *dest_addr = target_beg;
571 while (cur_chunk < end_chunk) {
572 size_t words = _chunk_data[cur_chunk].data_size();
574 #if 1
575 assert(pointer_delta(target_end, dest_addr) >= words,
576 "source region does not fit into target region");
577 #else
578 // XXX - need some work on the corner cases here. If the chunk does not
579 // fit, then must either make sure any partial_obj from the chunk fits, or
580 // 'undo' the initial part of the partial_obj that is in the previous chunk.
581 if (dest_addr + words >= target_end) {
582 // Let the caller know where to continue.
583 *target_next = dest_addr;
584 *source_next = chunk_to_addr(cur_chunk);
585 return false;
586 }
587 #endif // #if 1
589 _chunk_data[cur_chunk].set_destination(dest_addr);
591 // Set the destination_count for cur_chunk, and if necessary, update
592 // source_chunk for a destination chunk. The source_chunk field is updated
593 // if cur_chunk is the first (left-most) chunk to be copied to a destination
594 // chunk.
595 //
596 // The destination_count calculation is a bit subtle. A chunk that has data
597 // that compacts into itself does not count itself as a destination. This
598 // maintains the invariant that a zero count means the chunk is available
599 // and can be claimed and then filled.
600 if (words > 0) {
601 HeapWord* const last_addr = dest_addr + words - 1;
602 const size_t dest_chunk_1 = addr_to_chunk_idx(dest_addr);
603 const size_t dest_chunk_2 = addr_to_chunk_idx(last_addr);
604 #if 0
605 // Initially assume that the destination chunks will be the same and
606 // adjust the value below if necessary. Under this assumption, if
607 // cur_chunk == dest_chunk_2, then cur_chunk will be compacted completely
608 // into itself.
609 uint destination_count = cur_chunk == dest_chunk_2 ? 0 : 1;
610 if (dest_chunk_1 != dest_chunk_2) {
611 // Destination chunks differ; adjust destination_count.
612 destination_count += 1;
613 // Data from cur_chunk will be copied to the start of dest_chunk_2.
614 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk);
615 } else if (chunk_offset(dest_addr) == 0) {
616 // Data from cur_chunk will be copied to the start of the destination
617 // chunk.
618 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk);
619 }
620 #else
621 // Initially assume that the destination chunks will be different and
622 // adjust the value below if necessary. Under this assumption, if
623 // cur_chunk == dest_chunk2, then cur_chunk will be compacted partially
624 // into dest_chunk_1 and partially into itself.
625 uint destination_count = cur_chunk == dest_chunk_2 ? 1 : 2;
626 if (dest_chunk_1 != dest_chunk_2) {
627 // Data from cur_chunk will be copied to the start of dest_chunk_2.
628 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk);
629 } else {
630 // Destination chunks are the same; adjust destination_count.
631 destination_count -= 1;
632 if (chunk_offset(dest_addr) == 0) {
633 // Data from cur_chunk will be copied to the start of the destination
634 // chunk.
635 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk);
636 }
637 }
638 #endif // #if 0
640 _chunk_data[cur_chunk].set_destination_count(destination_count);
641 _chunk_data[cur_chunk].set_data_location(chunk_to_addr(cur_chunk));
642 dest_addr += words;
643 }
645 ++cur_chunk;
646 }
648 *target_next = dest_addr;
649 return true;
650 }
652 bool ParallelCompactData::partial_obj_ends_in_block(size_t block_index) {
653 HeapWord* block_addr = block_to_addr(block_index);
654 HeapWord* block_end_addr = block_addr + BlockSize;
655 size_t chunk_index = addr_to_chunk_idx(block_addr);
656 HeapWord* partial_obj_end_addr = partial_obj_end(chunk_index);
658 // An object that ends at the end of the block, ends
659 // in the block (the last word of the object is to
660 // the left of the end).
661 if ((block_addr < partial_obj_end_addr) &&
662 (partial_obj_end_addr <= block_end_addr)) {
663 return true;
664 }
666 return false;
667 }
669 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
670 HeapWord* result = NULL;
671 if (UseParallelOldGCChunkPointerCalc) {
672 result = chunk_calc_new_pointer(addr);
673 } else {
674 result = block_calc_new_pointer(addr);
675 }
676 return result;
677 }
679 // This method is overly complicated (expensive) to be called
680 // for every reference.
681 // Try to restructure this so that a NULL is returned if
682 // the object is dead. But don't wast the cycles to explicitly check
683 // that it is dead since only live objects should be passed in.
685 HeapWord* ParallelCompactData::chunk_calc_new_pointer(HeapWord* addr) {
686 assert(addr != NULL, "Should detect NULL oop earlier");
687 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
688 #ifdef ASSERT
689 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
690 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
691 }
692 #endif
693 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
695 // Chunk covering the object.
696 size_t chunk_index = addr_to_chunk_idx(addr);
697 const ChunkData* const chunk_ptr = chunk(chunk_index);
698 HeapWord* const chunk_addr = chunk_align_down(addr);
700 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object");
701 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check");
703 HeapWord* result = chunk_ptr->destination();
705 // If all the data in the chunk is live, then the new location of the object
706 // can be calculated from the destination of the chunk plus the offset of the
707 // object in the chunk.
708 if (chunk_ptr->data_size() == ChunkSize) {
709 result += pointer_delta(addr, chunk_addr);
710 return result;
711 }
713 // The new location of the object is
714 // chunk destination +
715 // size of the partial object extending onto the chunk +
716 // sizes of the live objects in the Chunk that are to the left of addr
717 const size_t partial_obj_size = chunk_ptr->partial_obj_size();
718 HeapWord* const search_start = chunk_addr + partial_obj_size;
720 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
721 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
723 result += partial_obj_size + live_to_left;
724 assert(result <= addr, "object cannot move to the right");
725 return result;
726 }
728 HeapWord* ParallelCompactData::block_calc_new_pointer(HeapWord* addr) {
729 assert(addr != NULL, "Should detect NULL oop earlier");
730 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
731 #ifdef ASSERT
732 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
733 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
734 }
735 #endif
736 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
738 // Chunk covering the object.
739 size_t chunk_index = addr_to_chunk_idx(addr);
740 const ChunkData* const chunk_ptr = chunk(chunk_index);
741 HeapWord* const chunk_addr = chunk_align_down(addr);
743 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object");
744 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check");
746 HeapWord* result = chunk_ptr->destination();
748 // If all the data in the chunk is live, then the new location of the object
749 // can be calculated from the destination of the chunk plus the offset of the
750 // object in the chunk.
751 if (chunk_ptr->data_size() == ChunkSize) {
752 result += pointer_delta(addr, chunk_addr);
753 return result;
754 }
756 // The new location of the object is
757 // chunk destination +
758 // block offset +
759 // sizes of the live objects in the Block that are to the left of addr
760 const size_t block_offset = addr_to_block_ptr(addr)->offset();
761 HeapWord* const search_start = chunk_addr + block_offset;
763 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
764 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
766 result += block_offset + live_to_left;
767 assert(result <= addr, "object cannot move to the right");
768 assert(result == chunk_calc_new_pointer(addr), "Should match");
769 return result;
770 }
772 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
773 klassOop updated_klass;
774 if (PSParallelCompact::should_update_klass(old_klass)) {
775 updated_klass = (klassOop) calc_new_pointer(old_klass);
776 } else {
777 updated_klass = old_klass;
778 }
780 return updated_klass;
781 }
783 #ifdef ASSERT
784 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
785 {
786 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
787 const size_t* const end = (const size_t*)vspace->committed_high_addr();
788 for (const size_t* p = beg; p < end; ++p) {
789 assert(*p == 0, "not zero");
790 }
791 }
793 void ParallelCompactData::verify_clear()
794 {
795 verify_clear(_chunk_vspace);
796 verify_clear(_block_vspace);
797 }
798 #endif // #ifdef ASSERT
800 #ifdef NOT_PRODUCT
801 ParallelCompactData::ChunkData* debug_chunk(size_t chunk_index) {
802 ParallelCompactData& sd = PSParallelCompact::summary_data();
803 return sd.chunk(chunk_index);
804 }
805 #endif
807 elapsedTimer PSParallelCompact::_accumulated_time;
808 unsigned int PSParallelCompact::_total_invocations = 0;
809 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
810 jlong PSParallelCompact::_time_of_last_gc = 0;
811 CollectorCounters* PSParallelCompact::_counters = NULL;
812 ParMarkBitMap PSParallelCompact::_mark_bitmap;
813 ParallelCompactData PSParallelCompact::_summary_data;
815 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
817 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
818 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
820 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
821 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
823 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
824 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
826 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); }
827 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
829 void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); }
831 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); }
832 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
834 void PSParallelCompact::post_initialize() {
835 ParallelScavengeHeap* heap = gc_heap();
836 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
838 MemRegion mr = heap->reserved_region();
839 _ref_processor = ReferenceProcessor::create_ref_processor(
840 mr, // span
841 true, // atomic_discovery
842 true, // mt_discovery
843 &_is_alive_closure,
844 ParallelGCThreads,
845 ParallelRefProcEnabled);
846 _counters = new CollectorCounters("PSParallelCompact", 1);
848 // Initialize static fields in ParCompactionManager.
849 ParCompactionManager::initialize(mark_bitmap());
850 }
852 bool PSParallelCompact::initialize() {
853 ParallelScavengeHeap* heap = gc_heap();
854 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
855 MemRegion mr = heap->reserved_region();
857 // Was the old gen get allocated successfully?
858 if (!heap->old_gen()->is_allocated()) {
859 return false;
860 }
862 initialize_space_info();
863 initialize_dead_wood_limiter();
865 if (!_mark_bitmap.initialize(mr)) {
866 vm_shutdown_during_initialization("Unable to allocate bit map for "
867 "parallel garbage collection for the requested heap size.");
868 return false;
869 }
871 if (!_summary_data.initialize(mr)) {
872 vm_shutdown_during_initialization("Unable to allocate tables for "
873 "parallel garbage collection for the requested heap size.");
874 return false;
875 }
877 return true;
878 }
880 void PSParallelCompact::initialize_space_info()
881 {
882 memset(&_space_info, 0, sizeof(_space_info));
884 ParallelScavengeHeap* heap = gc_heap();
885 PSYoungGen* young_gen = heap->young_gen();
886 MutableSpace* perm_space = heap->perm_gen()->object_space();
888 _space_info[perm_space_id].set_space(perm_space);
889 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
890 _space_info[eden_space_id].set_space(young_gen->eden_space());
891 _space_info[from_space_id].set_space(young_gen->from_space());
892 _space_info[to_space_id].set_space(young_gen->to_space());
894 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
895 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
897 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
898 if (TraceParallelOldGCDensePrefix) {
899 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
900 _space_info[perm_space_id].min_dense_prefix());
901 }
902 }
904 void PSParallelCompact::initialize_dead_wood_limiter()
905 {
906 const size_t max = 100;
907 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
908 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
909 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
910 DEBUG_ONLY(_dwl_initialized = true;)
911 _dwl_adjustment = normal_distribution(1.0);
912 }
914 // Simple class for storing info about the heap at the start of GC, to be used
915 // after GC for comparison/printing.
916 class PreGCValues {
917 public:
918 PreGCValues() { }
919 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
921 void fill(ParallelScavengeHeap* heap) {
922 _heap_used = heap->used();
923 _young_gen_used = heap->young_gen()->used_in_bytes();
924 _old_gen_used = heap->old_gen()->used_in_bytes();
925 _perm_gen_used = heap->perm_gen()->used_in_bytes();
926 };
928 size_t heap_used() const { return _heap_used; }
929 size_t young_gen_used() const { return _young_gen_used; }
930 size_t old_gen_used() const { return _old_gen_used; }
931 size_t perm_gen_used() const { return _perm_gen_used; }
933 private:
934 size_t _heap_used;
935 size_t _young_gen_used;
936 size_t _old_gen_used;
937 size_t _perm_gen_used;
938 };
940 void
941 PSParallelCompact::clear_data_covering_space(SpaceId id)
942 {
943 // At this point, top is the value before GC, new_top() is the value that will
944 // be set at the end of GC. The marking bitmap is cleared to top; nothing
945 // should be marked above top. The summary data is cleared to the larger of
946 // top & new_top.
947 MutableSpace* const space = _space_info[id].space();
948 HeapWord* const bot = space->bottom();
949 HeapWord* const top = space->top();
950 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
952 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
953 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
954 _mark_bitmap.clear_range(beg_bit, end_bit);
956 const size_t beg_chunk = _summary_data.addr_to_chunk_idx(bot);
957 const size_t end_chunk =
958 _summary_data.addr_to_chunk_idx(_summary_data.chunk_align_up(max_top));
959 _summary_data.clear_range(beg_chunk, end_chunk);
960 }
962 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
963 {
964 // Update the from & to space pointers in space_info, since they are swapped
965 // at each young gen gc. Do the update unconditionally (even though a
966 // promotion failure does not swap spaces) because an unknown number of minor
967 // collections will have swapped the spaces an unknown number of times.
968 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
969 ParallelScavengeHeap* heap = gc_heap();
970 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
971 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
973 pre_gc_values->fill(heap);
975 ParCompactionManager::reset();
976 NOT_PRODUCT(_mark_bitmap.reset_counters());
977 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
978 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
980 // Increment the invocation count
981 heap->increment_total_collections(true);
983 // We need to track unique mark sweep invocations as well.
984 _total_invocations++;
986 if (PrintHeapAtGC) {
987 Universe::print_heap_before_gc();
988 }
990 // Fill in TLABs
991 heap->accumulate_statistics_all_tlabs();
992 heap->ensure_parsability(true); // retire TLABs
994 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
995 HandleMark hm; // Discard invalid handles created during verification
996 gclog_or_tty->print(" VerifyBeforeGC:");
997 Universe::verify(true);
998 }
1000 // Verify object start arrays
1001 if (VerifyObjectStartArray &&
1002 VerifyBeforeGC) {
1003 heap->old_gen()->verify_object_start_array();
1004 heap->perm_gen()->verify_object_start_array();
1005 }
1007 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1008 DEBUG_ONLY(summary_data().verify_clear();)
1010 // Have worker threads release resources the next time they run a task.
1011 gc_task_manager()->release_all_resources();
1012 }
1014 void PSParallelCompact::post_compact()
1015 {
1016 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
1018 // Clear the marking bitmap and summary data and update top() in each space.
1019 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1020 clear_data_covering_space(SpaceId(id));
1021 _space_info[id].space()->set_top(_space_info[id].new_top());
1022 }
1024 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1025 MutableSpace* const from_space = _space_info[from_space_id].space();
1026 MutableSpace* const to_space = _space_info[to_space_id].space();
1028 ParallelScavengeHeap* heap = gc_heap();
1029 bool eden_empty = eden_space->is_empty();
1030 if (!eden_empty) {
1031 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1032 heap->young_gen(), heap->old_gen());
1033 }
1035 // Update heap occupancy information which is used as input to the soft ref
1036 // clearing policy at the next gc.
1037 Universe::update_heap_info_at_gc();
1039 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1040 to_space->is_empty();
1042 BarrierSet* bs = heap->barrier_set();
1043 if (bs->is_a(BarrierSet::ModRef)) {
1044 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1045 MemRegion old_mr = heap->old_gen()->reserved();
1046 MemRegion perm_mr = heap->perm_gen()->reserved();
1047 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1049 if (young_gen_empty) {
1050 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1051 } else {
1052 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1053 }
1054 }
1056 Threads::gc_epilogue();
1057 CodeCache::gc_epilogue();
1059 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1061 ref_processor()->enqueue_discovered_references(NULL);
1063 if (ZapUnusedHeapArea) {
1064 heap->gen_mangle_unused_area();
1065 }
1067 // Update time of last GC
1068 reset_millis_since_last_gc();
1069 }
1071 HeapWord*
1072 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1073 bool maximum_compaction)
1074 {
1075 const size_t chunk_size = ParallelCompactData::ChunkSize;
1076 const ParallelCompactData& sd = summary_data();
1078 const MutableSpace* const space = _space_info[id].space();
1079 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top());
1080 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(space->bottom());
1081 const ChunkData* const end_cp = sd.addr_to_chunk_ptr(top_aligned_up);
1083 // Skip full chunks at the beginning of the space--they are necessarily part
1084 // of the dense prefix.
1085 size_t full_count = 0;
1086 const ChunkData* cp;
1087 for (cp = beg_cp; cp < end_cp && cp->data_size() == chunk_size; ++cp) {
1088 ++full_count;
1089 }
1091 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1092 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1093 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1094 if (maximum_compaction || cp == end_cp || interval_ended) {
1095 _maximum_compaction_gc_num = total_invocations();
1096 return sd.chunk_to_addr(cp);
1097 }
1099 HeapWord* const new_top = _space_info[id].new_top();
1100 const size_t space_live = pointer_delta(new_top, space->bottom());
1101 const size_t space_used = space->used_in_words();
1102 const size_t space_capacity = space->capacity_in_words();
1104 const double cur_density = double(space_live) / space_capacity;
1105 const double deadwood_density =
1106 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1107 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1109 if (TraceParallelOldGCDensePrefix) {
1110 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1111 cur_density, deadwood_density, deadwood_goal);
1112 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1113 "space_cap=" SIZE_FORMAT,
1114 space_live, space_used,
1115 space_capacity);
1116 }
1118 // XXX - Use binary search?
1119 HeapWord* dense_prefix = sd.chunk_to_addr(cp);
1120 const ChunkData* full_cp = cp;
1121 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(space->top() - 1);
1122 while (cp < end_cp) {
1123 HeapWord* chunk_destination = cp->destination();
1124 const size_t cur_deadwood = pointer_delta(dense_prefix, chunk_destination);
1125 if (TraceParallelOldGCDensePrefix && Verbose) {
1126 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1127 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1128 sd.chunk(cp), chunk_destination,
1129 dense_prefix, cur_deadwood);
1130 }
1132 if (cur_deadwood >= deadwood_goal) {
1133 // Found the chunk that has the correct amount of deadwood to the left.
1134 // This typically occurs after crossing a fairly sparse set of chunks, so
1135 // iterate backwards over those sparse chunks, looking for the chunk that
1136 // has the lowest density of live objects 'to the right.'
1137 size_t space_to_left = sd.chunk(cp) * chunk_size;
1138 size_t live_to_left = space_to_left - cur_deadwood;
1139 size_t space_to_right = space_capacity - space_to_left;
1140 size_t live_to_right = space_live - live_to_left;
1141 double density_to_right = double(live_to_right) / space_to_right;
1142 while (cp > full_cp) {
1143 --cp;
1144 const size_t prev_chunk_live_to_right = live_to_right - cp->data_size();
1145 const size_t prev_chunk_space_to_right = space_to_right + chunk_size;
1146 double prev_chunk_density_to_right =
1147 double(prev_chunk_live_to_right) / prev_chunk_space_to_right;
1148 if (density_to_right <= prev_chunk_density_to_right) {
1149 return dense_prefix;
1150 }
1151 if (TraceParallelOldGCDensePrefix && Verbose) {
1152 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1153 "pc_d2r=%10.8f", sd.chunk(cp), density_to_right,
1154 prev_chunk_density_to_right);
1155 }
1156 dense_prefix -= chunk_size;
1157 live_to_right = prev_chunk_live_to_right;
1158 space_to_right = prev_chunk_space_to_right;
1159 density_to_right = prev_chunk_density_to_right;
1160 }
1161 return dense_prefix;
1162 }
1164 dense_prefix += chunk_size;
1165 ++cp;
1166 }
1168 return dense_prefix;
1169 }
1171 #ifndef PRODUCT
1172 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1173 const SpaceId id,
1174 const bool maximum_compaction,
1175 HeapWord* const addr)
1176 {
1177 const size_t chunk_idx = summary_data().addr_to_chunk_idx(addr);
1178 ChunkData* const cp = summary_data().chunk(chunk_idx);
1179 const MutableSpace* const space = _space_info[id].space();
1180 HeapWord* const new_top = _space_info[id].new_top();
1182 const size_t space_live = pointer_delta(new_top, space->bottom());
1183 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1184 const size_t space_cap = space->capacity_in_words();
1185 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1186 const size_t live_to_right = new_top - cp->destination();
1187 const size_t dead_to_right = space->top() - addr - live_to_right;
1189 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1190 "spl=" SIZE_FORMAT " "
1191 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1192 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1193 " ratio=%10.8f",
1194 algorithm, addr, chunk_idx,
1195 space_live,
1196 dead_to_left, dead_to_left_pct,
1197 dead_to_right, live_to_right,
1198 double(dead_to_right) / live_to_right);
1199 }
1200 #endif // #ifndef PRODUCT
1202 // Return a fraction indicating how much of the generation can be treated as
1203 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1204 // based on the density of live objects in the generation to determine a limit,
1205 // which is then adjusted so the return value is min_percent when the density is
1206 // 1.
1207 //
1208 // The following table shows some return values for a different values of the
1209 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1210 // min_percent is 1.
1211 //
1212 // fraction allowed as dead wood
1213 // -----------------------------------------------------------------
1214 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1215 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1216 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1217 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1218 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1219 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1220 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1221 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1222 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1223 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1224 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1225 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1226 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1227 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1228 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1229 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1230 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1231 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1232 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1233 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1234 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1235 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1236 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1238 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1239 {
1240 assert(_dwl_initialized, "uninitialized");
1242 // The raw limit is the value of the normal distribution at x = density.
1243 const double raw_limit = normal_distribution(density);
1245 // Adjust the raw limit so it becomes the minimum when the density is 1.
1246 //
1247 // First subtract the adjustment value (which is simply the precomputed value
1248 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1249 // Then add the minimum value, so the minimum is returned when the density is
1250 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1251 const double min = double(min_percent) / 100.0;
1252 const double limit = raw_limit - _dwl_adjustment + min;
1253 return MAX2(limit, 0.0);
1254 }
1256 ParallelCompactData::ChunkData*
1257 PSParallelCompact::first_dead_space_chunk(const ChunkData* beg,
1258 const ChunkData* end)
1259 {
1260 const size_t chunk_size = ParallelCompactData::ChunkSize;
1261 ParallelCompactData& sd = summary_data();
1262 size_t left = sd.chunk(beg);
1263 size_t right = end > beg ? sd.chunk(end) - 1 : left;
1265 // Binary search.
1266 while (left < right) {
1267 // Equivalent to (left + right) / 2, but does not overflow.
1268 const size_t middle = left + (right - left) / 2;
1269 ChunkData* const middle_ptr = sd.chunk(middle);
1270 HeapWord* const dest = middle_ptr->destination();
1271 HeapWord* const addr = sd.chunk_to_addr(middle);
1272 assert(dest != NULL, "sanity");
1273 assert(dest <= addr, "must move left");
1275 if (middle > left && dest < addr) {
1276 right = middle - 1;
1277 } else if (middle < right && middle_ptr->data_size() == chunk_size) {
1278 left = middle + 1;
1279 } else {
1280 return middle_ptr;
1281 }
1282 }
1283 return sd.chunk(left);
1284 }
1286 ParallelCompactData::ChunkData*
1287 PSParallelCompact::dead_wood_limit_chunk(const ChunkData* beg,
1288 const ChunkData* end,
1289 size_t dead_words)
1290 {
1291 ParallelCompactData& sd = summary_data();
1292 size_t left = sd.chunk(beg);
1293 size_t right = end > beg ? sd.chunk(end) - 1 : left;
1295 // Binary search.
1296 while (left < right) {
1297 // Equivalent to (left + right) / 2, but does not overflow.
1298 const size_t middle = left + (right - left) / 2;
1299 ChunkData* const middle_ptr = sd.chunk(middle);
1300 HeapWord* const dest = middle_ptr->destination();
1301 HeapWord* const addr = sd.chunk_to_addr(middle);
1302 assert(dest != NULL, "sanity");
1303 assert(dest <= addr, "must move left");
1305 const size_t dead_to_left = pointer_delta(addr, dest);
1306 if (middle > left && dead_to_left > dead_words) {
1307 right = middle - 1;
1308 } else if (middle < right && dead_to_left < dead_words) {
1309 left = middle + 1;
1310 } else {
1311 return middle_ptr;
1312 }
1313 }
1314 return sd.chunk(left);
1315 }
1317 // The result is valid during the summary phase, after the initial summarization
1318 // of each space into itself, and before final summarization.
1319 inline double
1320 PSParallelCompact::reclaimed_ratio(const ChunkData* const cp,
1321 HeapWord* const bottom,
1322 HeapWord* const top,
1323 HeapWord* const new_top)
1324 {
1325 ParallelCompactData& sd = summary_data();
1327 assert(cp != NULL, "sanity");
1328 assert(bottom != NULL, "sanity");
1329 assert(top != NULL, "sanity");
1330 assert(new_top != NULL, "sanity");
1331 assert(top >= new_top, "summary data problem?");
1332 assert(new_top > bottom, "space is empty; should not be here");
1333 assert(new_top >= cp->destination(), "sanity");
1334 assert(top >= sd.chunk_to_addr(cp), "sanity");
1336 HeapWord* const destination = cp->destination();
1337 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1338 const size_t compacted_region_live = pointer_delta(new_top, destination);
1339 const size_t compacted_region_used = pointer_delta(top, sd.chunk_to_addr(cp));
1340 const size_t reclaimable = compacted_region_used - compacted_region_live;
1342 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1343 return double(reclaimable) / divisor;
1344 }
1346 // Return the address of the end of the dense prefix, a.k.a. the start of the
1347 // compacted region. The address is always on a chunk boundary.
1348 //
1349 // Completely full chunks at the left are skipped, since no compaction can occur
1350 // in those chunks. Then the maximum amount of dead wood to allow is computed,
1351 // based on the density (amount live / capacity) of the generation; the chunk
1352 // with approximately that amount of dead space to the left is identified as the
1353 // limit chunk. Chunks between the last completely full chunk and the limit
1354 // chunk are scanned and the one that has the best (maximum) reclaimed_ratio()
1355 // is selected.
1356 HeapWord*
1357 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1358 bool maximum_compaction)
1359 {
1360 const size_t chunk_size = ParallelCompactData::ChunkSize;
1361 const ParallelCompactData& sd = summary_data();
1363 const MutableSpace* const space = _space_info[id].space();
1364 HeapWord* const top = space->top();
1365 HeapWord* const top_aligned_up = sd.chunk_align_up(top);
1366 HeapWord* const new_top = _space_info[id].new_top();
1367 HeapWord* const new_top_aligned_up = sd.chunk_align_up(new_top);
1368 HeapWord* const bottom = space->bottom();
1369 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(bottom);
1370 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up);
1371 const ChunkData* const new_top_cp = sd.addr_to_chunk_ptr(new_top_aligned_up);
1373 // Skip full chunks at the beginning of the space--they are necessarily part
1374 // of the dense prefix.
1375 const ChunkData* const full_cp = first_dead_space_chunk(beg_cp, new_top_cp);
1376 assert(full_cp->destination() == sd.chunk_to_addr(full_cp) ||
1377 space->is_empty(), "no dead space allowed to the left");
1378 assert(full_cp->data_size() < chunk_size || full_cp == new_top_cp - 1,
1379 "chunk must have dead space");
1381 // The gc number is saved whenever a maximum compaction is done, and used to
1382 // determine when the maximum compaction interval has expired. This avoids
1383 // successive max compactions for different reasons.
1384 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1385 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1386 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1387 total_invocations() == HeapFirstMaximumCompactionCount;
1388 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1389 _maximum_compaction_gc_num = total_invocations();
1390 return sd.chunk_to_addr(full_cp);
1391 }
1393 const size_t space_live = pointer_delta(new_top, bottom);
1394 const size_t space_used = space->used_in_words();
1395 const size_t space_capacity = space->capacity_in_words();
1397 const double density = double(space_live) / double(space_capacity);
1398 const size_t min_percent_free =
1399 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1400 const double limiter = dead_wood_limiter(density, min_percent_free);
1401 const size_t dead_wood_max = space_used - space_live;
1402 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1403 dead_wood_max);
1405 if (TraceParallelOldGCDensePrefix) {
1406 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1407 "space_cap=" SIZE_FORMAT,
1408 space_live, space_used,
1409 space_capacity);
1410 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1411 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1412 density, min_percent_free, limiter,
1413 dead_wood_max, dead_wood_limit);
1414 }
1416 // Locate the chunk with the desired amount of dead space to the left.
1417 const ChunkData* const limit_cp =
1418 dead_wood_limit_chunk(full_cp, top_cp, dead_wood_limit);
1420 // Scan from the first chunk with dead space to the limit chunk and find the
1421 // one with the best (largest) reclaimed ratio.
1422 double best_ratio = 0.0;
1423 const ChunkData* best_cp = full_cp;
1424 for (const ChunkData* cp = full_cp; cp < limit_cp; ++cp) {
1425 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1426 if (tmp_ratio > best_ratio) {
1427 best_cp = cp;
1428 best_ratio = tmp_ratio;
1429 }
1430 }
1432 #if 0
1433 // Something to consider: if the chunk with the best ratio is 'close to' the
1434 // first chunk w/free space, choose the first chunk with free space
1435 // ("first-free"). The first-free chunk is usually near the start of the
1436 // heap, which means we are copying most of the heap already, so copy a bit
1437 // more to get complete compaction.
1438 if (pointer_delta(best_cp, full_cp, sizeof(ChunkData)) < 4) {
1439 _maximum_compaction_gc_num = total_invocations();
1440 best_cp = full_cp;
1441 }
1442 #endif // #if 0
1444 return sd.chunk_to_addr(best_cp);
1445 }
1447 void PSParallelCompact::summarize_spaces_quick()
1448 {
1449 for (unsigned int i = 0; i < last_space_id; ++i) {
1450 const MutableSpace* space = _space_info[i].space();
1451 bool result = _summary_data.summarize(space->bottom(), space->end(),
1452 space->bottom(), space->top(),
1453 _space_info[i].new_top_addr());
1454 assert(result, "should never fail");
1455 _space_info[i].set_dense_prefix(space->bottom());
1456 }
1457 }
1459 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1460 {
1461 HeapWord* const dense_prefix_end = dense_prefix(id);
1462 const ChunkData* chunk = _summary_data.addr_to_chunk_ptr(dense_prefix_end);
1463 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1464 if (dead_space_crosses_boundary(chunk, dense_prefix_bit)) {
1465 // Only enough dead space is filled so that any remaining dead space to the
1466 // left is larger than the minimum filler object. (The remainder is filled
1467 // during the copy/update phase.)
1468 //
1469 // The size of the dead space to the right of the boundary is not a
1470 // concern, since compaction will be able to use whatever space is
1471 // available.
1472 //
1473 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1474 // surrounds the space to be filled with an object.
1475 //
1476 // In the 32-bit VM, each bit represents two 32-bit words:
1477 // +---+
1478 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1479 // end_bits: ... x x x | 0 | || 0 x x ...
1480 // +---+
1481 //
1482 // In the 64-bit VM, each bit represents one 64-bit word:
1483 // +------------+
1484 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1485 // end_bits: ... x x 1 | 0 || 0 | x x ...
1486 // +------------+
1487 // +-------+
1488 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1489 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1490 // +-------+
1491 // +-----------+
1492 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1493 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1494 // +-----------+
1495 // +-------+
1496 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1497 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1498 // +-------+
1500 // Initially assume case a, c or e will apply.
1501 size_t obj_len = (size_t)oopDesc::header_size();
1502 HeapWord* obj_beg = dense_prefix_end - obj_len;
1504 #ifdef _LP64
1505 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1506 // Case b above.
1507 obj_beg = dense_prefix_end - 1;
1508 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1509 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1510 // Case d above.
1511 obj_beg = dense_prefix_end - 3;
1512 obj_len = 3;
1513 }
1514 #endif // #ifdef _LP64
1516 MemRegion region(obj_beg, obj_len);
1517 SharedHeap::fill_region_with_object(region);
1518 _mark_bitmap.mark_obj(obj_beg, obj_len);
1519 _summary_data.add_obj(obj_beg, obj_len);
1520 assert(start_array(id) != NULL, "sanity");
1521 start_array(id)->allocate_block(obj_beg);
1522 }
1523 }
1525 void
1526 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1527 {
1528 assert(id < last_space_id, "id out of range");
1529 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1530 "should have been set in summarize_spaces_quick()");
1532 const MutableSpace* space = _space_info[id].space();
1533 if (_space_info[id].new_top() != space->bottom()) {
1534 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1535 _space_info[id].set_dense_prefix(dense_prefix_end);
1537 #ifndef PRODUCT
1538 if (TraceParallelOldGCDensePrefix) {
1539 print_dense_prefix_stats("ratio", id, maximum_compaction,
1540 dense_prefix_end);
1541 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1542 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1543 }
1544 #endif // #ifndef PRODUCT
1546 // If dead space crosses the dense prefix boundary, it is (at least
1547 // partially) filled with a dummy object, marked live and added to the
1548 // summary data. This simplifies the copy/update phase and must be done
1549 // before the final locations of objects are determined, to prevent leaving
1550 // a fragment of dead space that is too small to fill with an object.
1551 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1552 fill_dense_prefix_end(id);
1553 }
1555 // Compute the destination of each Chunk, and thus each object.
1556 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1557 _summary_data.summarize(dense_prefix_end, space->end(),
1558 dense_prefix_end, space->top(),
1559 _space_info[id].new_top_addr());
1560 }
1562 if (TraceParallelOldGCSummaryPhase) {
1563 const size_t chunk_size = ParallelCompactData::ChunkSize;
1564 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1565 const size_t dp_chunk = _summary_data.addr_to_chunk_idx(dense_prefix_end);
1566 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1567 HeapWord* const new_top = _space_info[id].new_top();
1568 const HeapWord* nt_aligned_up = _summary_data.chunk_align_up(new_top);
1569 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1570 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1571 "dp_chunk=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1572 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1573 id, space->capacity_in_words(), dense_prefix_end,
1574 dp_chunk, dp_words / chunk_size,
1575 cr_words / chunk_size, new_top);
1576 }
1577 }
1579 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1580 bool maximum_compaction)
1581 {
1582 EventMark m("2 summarize");
1583 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1584 // trace("2");
1586 #ifdef ASSERT
1587 if (VerifyParallelOldWithMarkSweep &&
1588 (PSParallelCompact::total_invocations() %
1589 VerifyParallelOldWithMarkSweepInterval) == 0) {
1590 verify_mark_bitmap(_mark_bitmap);
1591 }
1592 if (TraceParallelOldGCMarkingPhase) {
1593 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1594 "add_obj_bytes=" SIZE_FORMAT,
1595 add_obj_count, add_obj_size * HeapWordSize);
1596 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1597 "mark_bitmap_bytes=" SIZE_FORMAT,
1598 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1599 }
1600 #endif // #ifdef ASSERT
1602 // Quick summarization of each space into itself, to see how much is live.
1603 summarize_spaces_quick();
1605 if (TraceParallelOldGCSummaryPhase) {
1606 tty->print_cr("summary_phase: after summarizing each space to self");
1607 Universe::print();
1608 NOT_PRODUCT(print_chunk_ranges());
1609 if (Verbose) {
1610 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1611 }
1612 }
1614 // The amount of live data that will end up in old space (assuming it fits).
1615 size_t old_space_total_live = 0;
1616 unsigned int id;
1617 for (id = old_space_id; id < last_space_id; ++id) {
1618 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1619 _space_info[id].space()->bottom());
1620 }
1622 const MutableSpace* old_space = _space_info[old_space_id].space();
1623 if (old_space_total_live > old_space->capacity_in_words()) {
1624 // XXX - should also try to expand
1625 maximum_compaction = true;
1626 } else if (!UseParallelOldGCDensePrefix) {
1627 maximum_compaction = true;
1628 }
1630 // Permanent and Old generations.
1631 summarize_space(perm_space_id, maximum_compaction);
1632 summarize_space(old_space_id, maximum_compaction);
1634 // Summarize the remaining spaces (those in the young gen) into old space. If
1635 // the live data from a space doesn't fit, the existing summarization is left
1636 // intact, so the data is compacted down within the space itself.
1637 HeapWord** new_top_addr = _space_info[old_space_id].new_top_addr();
1638 HeapWord* const target_space_end = old_space->end();
1639 for (id = eden_space_id; id < last_space_id; ++id) {
1640 const MutableSpace* space = _space_info[id].space();
1641 const size_t live = pointer_delta(_space_info[id].new_top(),
1642 space->bottom());
1643 const size_t available = pointer_delta(target_space_end, *new_top_addr);
1644 if (live > 0 && live <= available) {
1645 // All the live data will fit.
1646 if (TraceParallelOldGCSummaryPhase) {
1647 tty->print_cr("summarizing %d into old_space @ " PTR_FORMAT,
1648 id, *new_top_addr);
1649 }
1650 _summary_data.summarize(*new_top_addr, target_space_end,
1651 space->bottom(), space->top(),
1652 new_top_addr);
1654 // Clear the source_chunk field for each chunk in the space.
1655 HeapWord* const new_top = _space_info[id].new_top();
1656 HeapWord* const clear_end = _summary_data.chunk_align_up(new_top);
1657 ChunkData* beg_chunk = _summary_data.addr_to_chunk_ptr(space->bottom());
1658 ChunkData* end_chunk = _summary_data.addr_to_chunk_ptr(clear_end);
1659 while (beg_chunk < end_chunk) {
1660 beg_chunk->set_source_chunk(0);
1661 ++beg_chunk;
1662 }
1664 // Reset the new_top value for the space.
1665 _space_info[id].set_new_top(space->bottom());
1666 }
1667 }
1669 // Fill in the block data after any changes to the chunks have
1670 // been made.
1671 #ifdef ASSERT
1672 summarize_blocks(cm, perm_space_id);
1673 summarize_blocks(cm, old_space_id);
1674 #else
1675 if (!UseParallelOldGCChunkPointerCalc) {
1676 summarize_blocks(cm, perm_space_id);
1677 summarize_blocks(cm, old_space_id);
1678 }
1679 #endif
1681 if (TraceParallelOldGCSummaryPhase) {
1682 tty->print_cr("summary_phase: after final summarization");
1683 Universe::print();
1684 NOT_PRODUCT(print_chunk_ranges());
1685 if (Verbose) {
1686 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1687 }
1688 }
1689 }
1691 // Fill in the BlockData.
1692 // Iterate over the spaces and within each space iterate over
1693 // the chunks and fill in the BlockData for each chunk.
1695 void PSParallelCompact::summarize_blocks(ParCompactionManager* cm,
1696 SpaceId first_compaction_space_id) {
1697 #if 0
1698 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(1);)
1699 for (SpaceId cur_space_id = first_compaction_space_id;
1700 cur_space_id != last_space_id;
1701 cur_space_id = next_compaction_space_id(cur_space_id)) {
1702 // Iterate over the chunks in the space
1703 size_t start_chunk_index =
1704 _summary_data.addr_to_chunk_idx(space(cur_space_id)->bottom());
1705 BitBlockUpdateClosure bbu(mark_bitmap(),
1706 cm,
1707 start_chunk_index);
1708 // Iterate over blocks.
1709 for (size_t chunk_index = start_chunk_index;
1710 chunk_index < _summary_data.chunk_count() &&
1711 _summary_data.chunk_to_addr(chunk_index) < space(cur_space_id)->top();
1712 chunk_index++) {
1714 // Reset the closure for the new chunk. Note that the closure
1715 // maintains some data that does not get reset for each chunk
1716 // so a new instance of the closure is no appropriate.
1717 bbu.reset_chunk(chunk_index);
1719 // Start the iteration with the first live object. This
1720 // may return the end of the chunk. That is acceptable since
1721 // it will properly limit the iterations.
1722 ParMarkBitMap::idx_t left_offset = mark_bitmap()->addr_to_bit(
1723 _summary_data.first_live_or_end_in_chunk(chunk_index));
1725 // End the iteration at the end of the chunk.
1726 HeapWord* chunk_addr = _summary_data.chunk_to_addr(chunk_index);
1727 HeapWord* chunk_end = chunk_addr + ParallelCompactData::ChunkSize;
1728 ParMarkBitMap::idx_t right_offset =
1729 mark_bitmap()->addr_to_bit(chunk_end);
1731 // Blocks that have not objects starting in them can be
1732 // skipped because their data will never be used.
1733 if (left_offset < right_offset) {
1735 // Iterate through the objects in the chunk.
1736 ParMarkBitMap::idx_t last_offset =
1737 mark_bitmap()->pair_iterate(&bbu, left_offset, right_offset);
1739 // If last_offset is less than right_offset, then the iterations
1740 // terminated while it was looking for an end bit. "last_offset"
1741 // is then the offset for the last start bit. In this situation
1742 // the "offset" field for the next block to the right (_cur_block + 1)
1743 // will not have been update although there may be live data
1744 // to the left of the chunk.
1746 size_t cur_block_plus_1 = bbu.cur_block() + 1;
1747 HeapWord* cur_block_plus_1_addr =
1748 _summary_data.block_to_addr(bbu.cur_block()) +
1749 ParallelCompactData::BlockSize;
1750 HeapWord* last_offset_addr = mark_bitmap()->bit_to_addr(last_offset);
1751 #if 1 // This code works. The else doesn't but should. Why does it?
1752 // The current block (cur_block()) has already been updated.
1753 // The last block that may need to be updated is either the
1754 // next block (current block + 1) or the block where the
1755 // last object starts (which can be greater than the
1756 // next block if there were no objects found in intervening
1757 // blocks).
1758 size_t last_block =
1759 MAX2(bbu.cur_block() + 1,
1760 _summary_data.addr_to_block_idx(last_offset_addr));
1761 #else
1762 // The current block has already been updated. The only block
1763 // that remains to be updated is the block where the last
1764 // object in the chunk starts.
1765 size_t last_block = _summary_data.addr_to_block_idx(last_offset_addr);
1766 #endif
1767 assert_bit_is_start(last_offset);
1768 assert((last_block == _summary_data.block_count()) ||
1769 (_summary_data.block(last_block)->raw_offset() == 0),
1770 "Should not have been set");
1771 // Is the last block still in the current chunk? If still
1772 // in this chunk, update the last block (the counting that
1773 // included the current block is meant for the offset of the last
1774 // block). If not in this chunk, do nothing. Should not
1775 // update a block in the next chunk.
1776 if (ParallelCompactData::chunk_contains_block(bbu.chunk_index(),
1777 last_block)) {
1778 if (last_offset < right_offset) {
1779 // The last object started in this chunk but ends beyond
1780 // this chunk. Update the block for this last object.
1781 assert(mark_bitmap()->is_marked(last_offset), "Should be marked");
1782 // No end bit was found. The closure takes care of
1783 // the cases where
1784 // an objects crosses over into the next block
1785 // an objects starts and ends in the next block
1786 // It does not handle the case where an object is
1787 // the first object in a later block and extends
1788 // past the end of the chunk (i.e., the closure
1789 // only handles complete objects that are in the range
1790 // it is given). That object is handed back here
1791 // for any special consideration necessary.
1792 //
1793 // Is the first bit in the last block a start or end bit?
1794 //
1795 // If the partial object ends in the last block L,
1796 // then the 1st bit in L may be an end bit.
1797 //
1798 // Else does the last object start in a block after the current
1799 // block? A block AA will already have been updated if an
1800 // object ends in the next block AA+1. An object found to end in
1801 // the AA+1 is the trigger that updates AA. Objects are being
1802 // counted in the current block for updaing a following
1803 // block. An object may start in later block
1804 // block but may extend beyond the last block in the chunk.
1805 // Updates are only done when the end of an object has been
1806 // found. If the last object (covered by block L) starts
1807 // beyond the current block, then no object ends in L (otherwise
1808 // L would be the current block). So the first bit in L is
1809 // a start bit.
1810 //
1811 // Else the last objects start in the current block and ends
1812 // beyond the chunk. The current block has already been
1813 // updated and there is no later block (with an object
1814 // starting in it) that needs to be updated.
1815 //
1816 if (_summary_data.partial_obj_ends_in_block(last_block)) {
1817 _summary_data.block(last_block)->set_end_bit_offset(
1818 bbu.live_data_left());
1819 } else if (last_offset_addr >= cur_block_plus_1_addr) {
1820 // The start of the object is on a later block
1821 // (to the right of the current block and there are no
1822 // complete live objects to the left of this last object
1823 // within the chunk.
1824 // The first bit in the block is for the start of the
1825 // last object.
1826 _summary_data.block(last_block)->set_start_bit_offset(
1827 bbu.live_data_left());
1828 } else {
1829 // The start of the last object was found in
1830 // the current chunk (which has already
1831 // been updated).
1832 assert(bbu.cur_block() ==
1833 _summary_data.addr_to_block_idx(last_offset_addr),
1834 "Should be a block already processed");
1835 }
1836 #ifdef ASSERT
1837 // Is there enough block information to find this object?
1838 // The destination of the chunk has not been set so the
1839 // values returned by calc_new_pointer() and
1840 // block_calc_new_pointer() will only be
1841 // offsets. But they should agree.
1842 HeapWord* moved_obj_with_chunks =
1843 _summary_data.chunk_calc_new_pointer(last_offset_addr);
1844 HeapWord* moved_obj_with_blocks =
1845 _summary_data.calc_new_pointer(last_offset_addr);
1846 assert(moved_obj_with_chunks == moved_obj_with_blocks,
1847 "Block calculation is wrong");
1848 #endif
1849 } else if (last_block < _summary_data.block_count()) {
1850 // Iterations ended looking for a start bit (but
1851 // did not run off the end of the block table).
1852 _summary_data.block(last_block)->set_start_bit_offset(
1853 bbu.live_data_left());
1854 }
1855 }
1856 #ifdef ASSERT
1857 // Is there enough block information to find this object?
1858 HeapWord* left_offset_addr = mark_bitmap()->bit_to_addr(left_offset);
1859 HeapWord* moved_obj_with_chunks =
1860 _summary_data.calc_new_pointer(left_offset_addr);
1861 HeapWord* moved_obj_with_blocks =
1862 _summary_data.calc_new_pointer(left_offset_addr);
1863 assert(moved_obj_with_chunks == moved_obj_with_blocks,
1864 "Block calculation is wrong");
1865 #endif
1867 // Is there another block after the end of this chunk?
1868 #ifdef ASSERT
1869 if (last_block < _summary_data.block_count()) {
1870 // No object may have been found in a block. If that
1871 // block is at the end of the chunk, the iteration will
1872 // terminate without incrementing the current block so
1873 // that the current block is not the last block in the
1874 // chunk. That situation precludes asserting that the
1875 // current block is the last block in the chunk. Assert
1876 // the lesser condition that the current block does not
1877 // exceed the chunk.
1878 assert(_summary_data.block_to_addr(last_block) <=
1879 (_summary_data.chunk_to_addr(chunk_index) +
1880 ParallelCompactData::ChunkSize),
1881 "Chunk and block inconsistency");
1882 assert(last_offset <= right_offset, "Iteration over ran end");
1883 }
1884 #endif
1885 }
1886 #ifdef ASSERT
1887 if (PrintGCDetails && Verbose) {
1888 if (_summary_data.chunk(chunk_index)->partial_obj_size() == 1) {
1889 size_t first_block =
1890 chunk_index / ParallelCompactData::BlocksPerChunk;
1891 gclog_or_tty->print_cr("first_block " PTR_FORMAT
1892 " _offset " PTR_FORMAT
1893 "_first_is_start_bit %d",
1894 first_block,
1895 _summary_data.block(first_block)->raw_offset(),
1896 _summary_data.block(first_block)->first_is_start_bit());
1897 }
1898 }
1899 #endif
1900 }
1901 }
1902 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(16);)
1903 #endif // #if 0
1904 }
1906 // This method should contain all heap-specific policy for invoking a full
1907 // collection. invoke_no_policy() will only attempt to compact the heap; it
1908 // will do nothing further. If we need to bail out for policy reasons, scavenge
1909 // before full gc, or any other specialized behavior, it needs to be added here.
1910 //
1911 // Note that this method should only be called from the vm_thread while at a
1912 // safepoint.
1913 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1914 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1915 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1916 "should be in vm thread");
1917 ParallelScavengeHeap* heap = gc_heap();
1918 GCCause::Cause gc_cause = heap->gc_cause();
1919 assert(!heap->is_gc_active(), "not reentrant");
1921 PSAdaptiveSizePolicy* policy = heap->size_policy();
1923 // Before each allocation/collection attempt, find out from the
1924 // policy object if GCs are, on the whole, taking too long. If so,
1925 // bail out without attempting a collection. The exceptions are
1926 // for explicitly requested GC's.
1927 if (!policy->gc_time_limit_exceeded() ||
1928 GCCause::is_user_requested_gc(gc_cause) ||
1929 GCCause::is_serviceability_requested_gc(gc_cause)) {
1930 IsGCActiveMark mark;
1932 if (ScavengeBeforeFullGC) {
1933 PSScavenge::invoke_no_policy();
1934 }
1936 PSParallelCompact::invoke_no_policy(maximum_heap_compaction);
1937 }
1938 }
1940 bool ParallelCompactData::chunk_contains(size_t chunk_index, HeapWord* addr) {
1941 size_t addr_chunk_index = addr_to_chunk_idx(addr);
1942 return chunk_index == addr_chunk_index;
1943 }
1945 bool ParallelCompactData::chunk_contains_block(size_t chunk_index,
1946 size_t block_index) {
1947 size_t first_block_in_chunk = chunk_index * BlocksPerChunk;
1948 size_t last_block_in_chunk = (chunk_index + 1) * BlocksPerChunk - 1;
1950 return (first_block_in_chunk <= block_index) &&
1951 (block_index <= last_block_in_chunk);
1952 }
1954 // This method contains no policy. You should probably
1955 // be calling invoke() instead.
1956 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1957 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1958 assert(ref_processor() != NULL, "Sanity");
1960 if (GC_locker::check_active_before_gc()) {
1961 return;
1962 }
1964 TimeStamp marking_start;
1965 TimeStamp compaction_start;
1966 TimeStamp collection_exit;
1968 ParallelScavengeHeap* heap = gc_heap();
1969 GCCause::Cause gc_cause = heap->gc_cause();
1970 PSYoungGen* young_gen = heap->young_gen();
1971 PSOldGen* old_gen = heap->old_gen();
1972 PSPermGen* perm_gen = heap->perm_gen();
1973 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1975 if (ZapUnusedHeapArea) {
1976 // Save information needed to minimize mangling
1977 heap->record_gen_tops_before_GC();
1978 }
1980 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
1982 // Make sure data structures are sane, make the heap parsable, and do other
1983 // miscellaneous bookkeeping.
1984 PreGCValues pre_gc_values;
1985 pre_compact(&pre_gc_values);
1987 // Get the compaction manager reserved for the VM thread.
1988 ParCompactionManager* const vmthread_cm =
1989 ParCompactionManager::manager_array(gc_task_manager()->workers());
1991 // Place after pre_compact() where the number of invocations is incremented.
1992 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
1994 {
1995 ResourceMark rm;
1996 HandleMark hm;
1998 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
2000 // This is useful for debugging but don't change the output the
2001 // the customer sees.
2002 const char* gc_cause_str = "Full GC";
2003 if (is_system_gc && PrintGCDetails) {
2004 gc_cause_str = "Full GC (System)";
2005 }
2006 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2007 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2008 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
2009 TraceCollectorStats tcs(counters());
2010 TraceMemoryManagerStats tms(true /* Full GC */);
2012 if (TraceGen1Time) accumulated_time()->start();
2014 // Let the size policy know we're starting
2015 size_policy->major_collection_begin();
2017 // When collecting the permanent generation methodOops may be moving,
2018 // so we either have to flush all bcp data or convert it into bci.
2019 CodeCache::gc_prologue();
2020 Threads::gc_prologue();
2022 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2023 COMPILER2_PRESENT(DerivedPointerTable::clear());
2025 ref_processor()->enable_discovery();
2027 bool marked_for_unloading = false;
2029 marking_start.update();
2030 marking_phase(vmthread_cm, maximum_heap_compaction);
2032 #ifndef PRODUCT
2033 if (TraceParallelOldGCMarkingPhase) {
2034 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2035 "cas_by_another %d",
2036 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2037 mark_bitmap()->cas_by_another());
2038 }
2039 #endif // #ifndef PRODUCT
2041 #ifdef ASSERT
2042 if (VerifyParallelOldWithMarkSweep &&
2043 (PSParallelCompact::total_invocations() %
2044 VerifyParallelOldWithMarkSweepInterval) == 0) {
2045 gclog_or_tty->print_cr("Verify marking with mark_sweep_phase1()");
2046 if (PrintGCDetails && Verbose) {
2047 gclog_or_tty->print_cr("mark_sweep_phase1:");
2048 }
2049 // Clear the discovered lists so that discovered objects
2050 // don't look like they have been discovered twice.
2051 ref_processor()->clear_discovered_references();
2053 PSMarkSweep::allocate_stacks();
2054 MemRegion mr = Universe::heap()->reserved_region();
2055 PSMarkSweep::ref_processor()->enable_discovery();
2056 PSMarkSweep::mark_sweep_phase1(maximum_heap_compaction);
2057 }
2058 #endif
2060 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
2061 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2063 #ifdef ASSERT
2064 if (VerifyParallelOldWithMarkSweep &&
2065 (PSParallelCompact::total_invocations() %
2066 VerifyParallelOldWithMarkSweepInterval) == 0) {
2067 if (PrintGCDetails && Verbose) {
2068 gclog_or_tty->print_cr("mark_sweep_phase2:");
2069 }
2070 PSMarkSweep::mark_sweep_phase2();
2071 }
2072 #endif
2074 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2075 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2077 // adjust_roots() updates Universe::_intArrayKlassObj which is
2078 // needed by the compaction for filling holes in the dense prefix.
2079 adjust_roots();
2081 #ifdef ASSERT
2082 if (VerifyParallelOldWithMarkSweep &&
2083 (PSParallelCompact::total_invocations() %
2084 VerifyParallelOldWithMarkSweepInterval) == 0) {
2085 // Do a separate verify phase so that the verify
2086 // code can use the the forwarding pointers to
2087 // check the new pointer calculation. The restore_marks()
2088 // has to be done before the real compact.
2089 vmthread_cm->set_action(ParCompactionManager::VerifyUpdate);
2090 compact_perm(vmthread_cm);
2091 compact_serial(vmthread_cm);
2092 vmthread_cm->set_action(ParCompactionManager::ResetObjects);
2093 compact_perm(vmthread_cm);
2094 compact_serial(vmthread_cm);
2095 vmthread_cm->set_action(ParCompactionManager::UpdateAndCopy);
2097 // For debugging only
2098 PSMarkSweep::restore_marks();
2099 PSMarkSweep::deallocate_stacks();
2100 }
2101 #endif
2103 compaction_start.update();
2104 // Does the perm gen always have to be done serially because
2105 // klasses are used in the update of an object?
2106 compact_perm(vmthread_cm);
2108 if (UseParallelOldGCCompacting) {
2109 compact();
2110 } else {
2111 compact_serial(vmthread_cm);
2112 }
2114 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2115 // done before resizing.
2116 post_compact();
2118 // Let the size policy know we're done
2119 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2121 if (UseAdaptiveSizePolicy) {
2122 if (PrintAdaptiveSizePolicy) {
2123 gclog_or_tty->print("AdaptiveSizeStart: ");
2124 gclog_or_tty->stamp();
2125 gclog_or_tty->print_cr(" collection: %d ",
2126 heap->total_collections());
2127 if (Verbose) {
2128 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2129 " perm_gen_capacity: %d ",
2130 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2131 perm_gen->capacity_in_bytes());
2132 }
2133 }
2135 // Don't check if the size_policy is ready here. Let
2136 // the size_policy check that internally.
2137 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2138 ((gc_cause != GCCause::_java_lang_system_gc) ||
2139 UseAdaptiveSizePolicyWithSystemGC)) {
2140 // Calculate optimal free space amounts
2141 assert(young_gen->max_size() >
2142 young_gen->from_space()->capacity_in_bytes() +
2143 young_gen->to_space()->capacity_in_bytes(),
2144 "Sizes of space in young gen are out-of-bounds");
2145 size_t max_eden_size = young_gen->max_size() -
2146 young_gen->from_space()->capacity_in_bytes() -
2147 young_gen->to_space()->capacity_in_bytes();
2148 size_policy->compute_generation_free_space(
2149 young_gen->used_in_bytes(),
2150 young_gen->eden_space()->used_in_bytes(),
2151 old_gen->used_in_bytes(),
2152 perm_gen->used_in_bytes(),
2153 young_gen->eden_space()->capacity_in_bytes(),
2154 old_gen->max_gen_size(),
2155 max_eden_size,
2156 true /* full gc*/,
2157 gc_cause);
2159 heap->resize_old_gen(
2160 size_policy->calculated_old_free_size_in_bytes());
2162 // Don't resize the young generation at an major collection. A
2163 // desired young generation size may have been calculated but
2164 // resizing the young generation complicates the code because the
2165 // resizing of the old generation may have moved the boundary
2166 // between the young generation and the old generation. Let the
2167 // young generation resizing happen at the minor collections.
2168 }
2169 if (PrintAdaptiveSizePolicy) {
2170 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2171 heap->total_collections());
2172 }
2173 }
2175 if (UsePerfData) {
2176 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2177 counters->update_counters();
2178 counters->update_old_capacity(old_gen->capacity_in_bytes());
2179 counters->update_young_capacity(young_gen->capacity_in_bytes());
2180 }
2182 heap->resize_all_tlabs();
2184 // We collected the perm gen, so we'll resize it here.
2185 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2187 if (TraceGen1Time) accumulated_time()->stop();
2189 if (PrintGC) {
2190 if (PrintGCDetails) {
2191 // No GC timestamp here. This is after GC so it would be confusing.
2192 young_gen->print_used_change(pre_gc_values.young_gen_used());
2193 old_gen->print_used_change(pre_gc_values.old_gen_used());
2194 heap->print_heap_change(pre_gc_values.heap_used());
2195 // Print perm gen last (print_heap_change() excludes the perm gen).
2196 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2197 } else {
2198 heap->print_heap_change(pre_gc_values.heap_used());
2199 }
2200 }
2202 // Track memory usage and detect low memory
2203 MemoryService::track_memory_usage();
2204 heap->update_counters();
2206 if (PrintGCDetails) {
2207 if (size_policy->print_gc_time_limit_would_be_exceeded()) {
2208 if (size_policy->gc_time_limit_exceeded()) {
2209 gclog_or_tty->print_cr(" GC time is exceeding GCTimeLimit "
2210 "of %d%%", GCTimeLimit);
2211 } else {
2212 gclog_or_tty->print_cr(" GC time would exceed GCTimeLimit "
2213 "of %d%%", GCTimeLimit);
2214 }
2215 }
2216 size_policy->set_print_gc_time_limit_would_be_exceeded(false);
2217 }
2218 }
2220 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2221 HandleMark hm; // Discard invalid handles created during verification
2222 gclog_or_tty->print(" VerifyAfterGC:");
2223 Universe::verify(false);
2224 }
2226 // Re-verify object start arrays
2227 if (VerifyObjectStartArray &&
2228 VerifyAfterGC) {
2229 old_gen->verify_object_start_array();
2230 perm_gen->verify_object_start_array();
2231 }
2233 if (ZapUnusedHeapArea) {
2234 old_gen->object_space()->check_mangled_unused_area_complete();
2235 perm_gen->object_space()->check_mangled_unused_area_complete();
2236 }
2238 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2240 collection_exit.update();
2242 if (PrintHeapAtGC) {
2243 Universe::print_heap_after_gc();
2244 }
2245 if (PrintGCTaskTimeStamps) {
2246 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2247 INT64_FORMAT,
2248 marking_start.ticks(), compaction_start.ticks(),
2249 collection_exit.ticks());
2250 gc_task_manager()->print_task_time_stamps();
2251 }
2252 }
2254 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2255 PSYoungGen* young_gen,
2256 PSOldGen* old_gen) {
2257 MutableSpace* const eden_space = young_gen->eden_space();
2258 assert(!eden_space->is_empty(), "eden must be non-empty");
2259 assert(young_gen->virtual_space()->alignment() ==
2260 old_gen->virtual_space()->alignment(), "alignments do not match");
2262 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2263 return false;
2264 }
2266 // Both generations must be completely committed.
2267 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2268 return false;
2269 }
2270 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2271 return false;
2272 }
2274 // Figure out how much to take from eden. Include the average amount promoted
2275 // in the total; otherwise the next young gen GC will simply bail out to a
2276 // full GC.
2277 const size_t alignment = old_gen->virtual_space()->alignment();
2278 const size_t eden_used = eden_space->used_in_bytes();
2279 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2280 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2281 const size_t eden_capacity = eden_space->capacity_in_bytes();
2283 if (absorb_size >= eden_capacity) {
2284 return false; // Must leave some space in eden.
2285 }
2287 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2288 if (new_young_size < young_gen->min_gen_size()) {
2289 return false; // Respect young gen minimum size.
2290 }
2292 if (TraceAdaptiveGCBoundary && Verbose) {
2293 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2294 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2295 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2296 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2297 absorb_size / K,
2298 eden_capacity / K, (eden_capacity - absorb_size) / K,
2299 young_gen->from_space()->used_in_bytes() / K,
2300 young_gen->to_space()->used_in_bytes() / K,
2301 young_gen->capacity_in_bytes() / K, new_young_size / K);
2302 }
2304 // Fill the unused part of the old gen.
2305 MutableSpace* const old_space = old_gen->object_space();
2306 MemRegion old_gen_unused(old_space->top(), old_space->end());
2307 if (!old_gen_unused.is_empty()) {
2308 SharedHeap::fill_region_with_object(old_gen_unused);
2309 }
2311 // Take the live data from eden and set both top and end in the old gen to
2312 // eden top. (Need to set end because reset_after_change() mangles the region
2313 // from end to virtual_space->high() in debug builds).
2314 HeapWord* const new_top = eden_space->top();
2315 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2316 absorb_size);
2317 young_gen->reset_after_change();
2318 old_space->set_top(new_top);
2319 old_space->set_end(new_top);
2320 old_gen->reset_after_change();
2322 // Update the object start array for the filler object and the data from eden.
2323 ObjectStartArray* const start_array = old_gen->start_array();
2324 HeapWord* const start = old_gen_unused.start();
2325 for (HeapWord* addr = start; addr < new_top; addr += oop(addr)->size()) {
2326 start_array->allocate_block(addr);
2327 }
2329 // Could update the promoted average here, but it is not typically updated at
2330 // full GCs and the value to use is unclear. Something like
2331 //
2332 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2334 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2335 return true;
2336 }
2338 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2339 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2340 "shouldn't return NULL");
2341 return ParallelScavengeHeap::gc_task_manager();
2342 }
2344 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2345 bool maximum_heap_compaction) {
2346 // Recursively traverse all live objects and mark them
2347 EventMark m("1 mark object");
2348 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2350 ParallelScavengeHeap* heap = gc_heap();
2351 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2352 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array();
2353 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2355 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2356 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2358 {
2359 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2361 GCTaskQueue* q = GCTaskQueue::create();
2363 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2364 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2365 // We scan the thread roots in parallel
2366 Threads::create_thread_roots_marking_tasks(q);
2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2368 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2369 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2371 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2372 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols));
2374 if (parallel_gc_threads > 1) {
2375 for (uint j = 0; j < parallel_gc_threads; j++) {
2376 q->enqueue(new StealMarkingTask(&terminator));
2377 }
2378 }
2380 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2381 q->enqueue(fin);
2383 gc_task_manager()->add_list(q);
2385 fin->wait_for();
2387 // We have to release the barrier tasks!
2388 WaitForBarrierGCTask::destroy(fin);
2389 }
2391 // Process reference objects found during marking
2392 {
2393 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2394 ReferencePolicy *soft_ref_policy;
2395 if (maximum_heap_compaction) {
2396 soft_ref_policy = new AlwaysClearPolicy();
2397 } else {
2398 #ifdef COMPILER2
2399 soft_ref_policy = new LRUMaxHeapPolicy();
2400 #else
2401 soft_ref_policy = new LRUCurrentHeapPolicy();
2402 #endif // COMPILER2
2403 }
2404 assert(soft_ref_policy != NULL, "No soft reference policy");
2405 if (ref_processor()->processing_is_mt()) {
2406 RefProcTaskExecutor task_executor;
2407 ref_processor()->process_discovered_references(
2408 soft_ref_policy, is_alive_closure(), &mark_and_push_closure,
2409 &follow_stack_closure, &task_executor);
2410 } else {
2411 ref_processor()->process_discovered_references(
2412 soft_ref_policy, is_alive_closure(), &mark_and_push_closure,
2413 &follow_stack_closure, NULL);
2414 }
2415 }
2417 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2418 // Follow system dictionary roots and unload classes.
2419 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2421 // Follow code cache roots.
2422 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2423 purged_class);
2424 follow_stack(cm); // Flush marking stack.
2426 // Update subklass/sibling/implementor links of live klasses
2427 // revisit_klass_stack is used in follow_weak_klass_links().
2428 follow_weak_klass_links(cm);
2430 // Visit symbol and interned string tables and delete unmarked oops
2431 SymbolTable::unlink(is_alive_closure());
2432 StringTable::unlink(is_alive_closure());
2434 assert(cm->marking_stack()->size() == 0, "stack should be empty by now");
2435 assert(cm->overflow_stack()->is_empty(), "stack should be empty by now");
2436 }
2438 // This should be moved to the shared markSweep code!
2439 class PSAlwaysTrueClosure: public BoolObjectClosure {
2440 public:
2441 void do_object(oop p) { ShouldNotReachHere(); }
2442 bool do_object_b(oop p) { return true; }
2443 };
2444 static PSAlwaysTrueClosure always_true;
2446 void PSParallelCompact::adjust_roots() {
2447 // Adjust the pointers to reflect the new locations
2448 EventMark m("3 adjust roots");
2449 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2451 // General strong roots.
2452 Universe::oops_do(adjust_root_pointer_closure());
2453 ReferenceProcessor::oops_do(adjust_root_pointer_closure());
2454 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
2455 Threads::oops_do(adjust_root_pointer_closure());
2456 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2457 FlatProfiler::oops_do(adjust_root_pointer_closure());
2458 Management::oops_do(adjust_root_pointer_closure());
2459 JvmtiExport::oops_do(adjust_root_pointer_closure());
2460 // SO_AllClasses
2461 SystemDictionary::oops_do(adjust_root_pointer_closure());
2462 vmSymbols::oops_do(adjust_root_pointer_closure());
2464 // Now adjust pointers in remaining weak roots. (All of which should
2465 // have been cleared if they pointed to non-surviving objects.)
2466 // Global (weak) JNI handles
2467 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2469 CodeCache::oops_do(adjust_pointer_closure());
2470 SymbolTable::oops_do(adjust_root_pointer_closure());
2471 StringTable::oops_do(adjust_root_pointer_closure());
2472 ref_processor()->weak_oops_do(adjust_root_pointer_closure());
2473 // Roots were visited so references into the young gen in roots
2474 // may have been scanned. Process them also.
2475 // Should the reference processor have a span that excludes
2476 // young gen objects?
2477 PSScavenge::reference_processor()->weak_oops_do(
2478 adjust_root_pointer_closure());
2479 }
2481 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2482 EventMark m("4 compact perm");
2483 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
2484 // trace("4");
2486 gc_heap()->perm_gen()->start_array()->reset();
2487 move_and_update(cm, perm_space_id);
2488 }
2490 void PSParallelCompact::enqueue_chunk_draining_tasks(GCTaskQueue* q,
2491 uint parallel_gc_threads) {
2492 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2494 const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
2495 for (unsigned int j = 0; j < task_count; j++) {
2496 q->enqueue(new DrainStacksCompactionTask());
2497 }
2499 // Find all chunks that are available (can be filled immediately) and
2500 // distribute them to the thread stacks. The iteration is done in reverse
2501 // order (high to low) so the chunks will be removed in ascending order.
2503 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2505 size_t fillable_chunks = 0; // A count for diagnostic purposes.
2506 unsigned int which = 0; // The worker thread number.
2508 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2509 SpaceInfo* const space_info = _space_info + id;
2510 MutableSpace* const space = space_info->space();
2511 HeapWord* const new_top = space_info->new_top();
2513 const size_t beg_chunk = sd.addr_to_chunk_idx(space_info->dense_prefix());
2514 const size_t end_chunk = sd.addr_to_chunk_idx(sd.chunk_align_up(new_top));
2515 assert(end_chunk > 0, "perm gen cannot be empty");
2517 for (size_t cur = end_chunk - 1; cur >= beg_chunk; --cur) {
2518 if (sd.chunk(cur)->claim_unsafe()) {
2519 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2520 cm->save_for_processing(cur);
2522 if (TraceParallelOldGCCompactionPhase && Verbose) {
2523 const size_t count_mod_8 = fillable_chunks & 7;
2524 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2525 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2526 if (count_mod_8 == 7) gclog_or_tty->cr();
2527 }
2529 NOT_PRODUCT(++fillable_chunks;)
2531 // Assign chunks to threads in round-robin fashion.
2532 if (++which == task_count) {
2533 which = 0;
2534 }
2535 }
2536 }
2537 }
2539 if (TraceParallelOldGCCompactionPhase) {
2540 if (Verbose && (fillable_chunks & 7) != 0) gclog_or_tty->cr();
2541 gclog_or_tty->print_cr("%u initially fillable chunks", fillable_chunks);
2542 }
2543 }
2545 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2547 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2548 uint parallel_gc_threads) {
2549 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2551 ParallelCompactData& sd = PSParallelCompact::summary_data();
2553 // Iterate over all the spaces adding tasks for updating
2554 // chunks in the dense prefix. Assume that 1 gc thread
2555 // will work on opening the gaps and the remaining gc threads
2556 // will work on the dense prefix.
2557 SpaceId space_id = old_space_id;
2558 while (space_id != last_space_id) {
2559 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2560 const MutableSpace* const space = _space_info[space_id].space();
2562 if (dense_prefix_end == space->bottom()) {
2563 // There is no dense prefix for this space.
2564 space_id = next_compaction_space_id(space_id);
2565 continue;
2566 }
2568 // The dense prefix is before this chunk.
2569 size_t chunk_index_end_dense_prefix =
2570 sd.addr_to_chunk_idx(dense_prefix_end);
2571 ChunkData* const dense_prefix_cp = sd.chunk(chunk_index_end_dense_prefix);
2572 assert(dense_prefix_end == space->end() ||
2573 dense_prefix_cp->available() ||
2574 dense_prefix_cp->claimed(),
2575 "The chunk after the dense prefix should always be ready to fill");
2577 size_t chunk_index_start = sd.addr_to_chunk_idx(space->bottom());
2579 // Is there dense prefix work?
2580 size_t total_dense_prefix_chunks =
2581 chunk_index_end_dense_prefix - chunk_index_start;
2582 // How many chunks of the dense prefix should be given to
2583 // each thread?
2584 if (total_dense_prefix_chunks > 0) {
2585 uint tasks_for_dense_prefix = 1;
2586 if (UseParallelDensePrefixUpdate) {
2587 if (total_dense_prefix_chunks <=
2588 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2589 // Don't over partition. This assumes that
2590 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2591 // so there are not many chunks to process.
2592 tasks_for_dense_prefix = parallel_gc_threads;
2593 } else {
2594 // Over partition
2595 tasks_for_dense_prefix = parallel_gc_threads *
2596 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2597 }
2598 }
2599 size_t chunks_per_thread = total_dense_prefix_chunks /
2600 tasks_for_dense_prefix;
2601 // Give each thread at least 1 chunk.
2602 if (chunks_per_thread == 0) {
2603 chunks_per_thread = 1;
2604 }
2606 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2607 if (chunk_index_start >= chunk_index_end_dense_prefix) {
2608 break;
2609 }
2610 // chunk_index_end is not processed
2611 size_t chunk_index_end = MIN2(chunk_index_start + chunks_per_thread,
2612 chunk_index_end_dense_prefix);
2613 q->enqueue(new UpdateDensePrefixTask(
2614 space_id,
2615 chunk_index_start,
2616 chunk_index_end));
2617 chunk_index_start = chunk_index_end;
2618 }
2619 }
2620 // This gets any part of the dense prefix that did not
2621 // fit evenly.
2622 if (chunk_index_start < chunk_index_end_dense_prefix) {
2623 q->enqueue(new UpdateDensePrefixTask(
2624 space_id,
2625 chunk_index_start,
2626 chunk_index_end_dense_prefix));
2627 }
2628 space_id = next_compaction_space_id(space_id);
2629 } // End tasks for dense prefix
2630 }
2632 void PSParallelCompact::enqueue_chunk_stealing_tasks(
2633 GCTaskQueue* q,
2634 ParallelTaskTerminator* terminator_ptr,
2635 uint parallel_gc_threads) {
2636 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2638 // Once a thread has drained it's stack, it should try to steal chunks from
2639 // other threads.
2640 if (parallel_gc_threads > 1) {
2641 for (uint j = 0; j < parallel_gc_threads; j++) {
2642 q->enqueue(new StealChunkCompactionTask(terminator_ptr));
2643 }
2644 }
2645 }
2647 void PSParallelCompact::compact() {
2648 EventMark m("5 compact");
2649 // trace("5");
2650 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2652 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2653 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2654 PSOldGen* old_gen = heap->old_gen();
2655 old_gen->start_array()->reset();
2656 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2657 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array();
2658 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2660 GCTaskQueue* q = GCTaskQueue::create();
2661 enqueue_chunk_draining_tasks(q, parallel_gc_threads);
2662 enqueue_dense_prefix_tasks(q, parallel_gc_threads);
2663 enqueue_chunk_stealing_tasks(q, &terminator, parallel_gc_threads);
2665 {
2666 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2668 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2669 q->enqueue(fin);
2671 gc_task_manager()->add_list(q);
2673 fin->wait_for();
2675 // We have to release the barrier tasks!
2676 WaitForBarrierGCTask::destroy(fin);
2678 #ifdef ASSERT
2679 // Verify that all chunks have been processed before the deferred updates.
2680 // Note that perm_space_id is skipped; this type of verification is not
2681 // valid until the perm gen is compacted by chunks.
2682 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2683 verify_complete(SpaceId(id));
2684 }
2685 #endif
2686 }
2688 {
2689 // Update the deferred objects, if any. Any compaction manager can be used.
2690 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2691 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2692 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2693 update_deferred_objects(cm, SpaceId(id));
2694 }
2695 }
2696 }
2698 #ifdef ASSERT
2699 void PSParallelCompact::verify_complete(SpaceId space_id) {
2700 // All Chunks between space bottom() to new_top() should be marked as filled
2701 // and all Chunks between new_top() and top() should be available (i.e.,
2702 // should have been emptied).
2703 ParallelCompactData& sd = summary_data();
2704 SpaceInfo si = _space_info[space_id];
2705 HeapWord* new_top_addr = sd.chunk_align_up(si.new_top());
2706 HeapWord* old_top_addr = sd.chunk_align_up(si.space()->top());
2707 const size_t beg_chunk = sd.addr_to_chunk_idx(si.space()->bottom());
2708 const size_t new_top_chunk = sd.addr_to_chunk_idx(new_top_addr);
2709 const size_t old_top_chunk = sd.addr_to_chunk_idx(old_top_addr);
2711 bool issued_a_warning = false;
2713 size_t cur_chunk;
2714 for (cur_chunk = beg_chunk; cur_chunk < new_top_chunk; ++cur_chunk) {
2715 const ChunkData* const c = sd.chunk(cur_chunk);
2716 if (!c->completed()) {
2717 warning("chunk " SIZE_FORMAT " not filled: "
2718 "destination_count=" SIZE_FORMAT,
2719 cur_chunk, c->destination_count());
2720 issued_a_warning = true;
2721 }
2722 }
2724 for (cur_chunk = new_top_chunk; cur_chunk < old_top_chunk; ++cur_chunk) {
2725 const ChunkData* const c = sd.chunk(cur_chunk);
2726 if (!c->available()) {
2727 warning("chunk " SIZE_FORMAT " not empty: "
2728 "destination_count=" SIZE_FORMAT,
2729 cur_chunk, c->destination_count());
2730 issued_a_warning = true;
2731 }
2732 }
2734 if (issued_a_warning) {
2735 print_chunk_ranges();
2736 }
2737 }
2738 #endif // #ifdef ASSERT
2740 void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
2741 EventMark m("5 compact serial");
2742 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty);
2744 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2745 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2747 PSYoungGen* young_gen = heap->young_gen();
2748 PSOldGen* old_gen = heap->old_gen();
2750 old_gen->start_array()->reset();
2751 old_gen->move_and_update(cm);
2752 young_gen->move_and_update(cm);
2753 }
2756 void PSParallelCompact::follow_stack(ParCompactionManager* cm) {
2757 while(!cm->overflow_stack()->is_empty()) {
2758 oop obj = cm->overflow_stack()->pop();
2759 obj->follow_contents(cm);
2760 }
2762 oop obj;
2763 // obj is a reference!!!
2764 while (cm->marking_stack()->pop_local(obj)) {
2765 // It would be nice to assert about the type of objects we might
2766 // pop, but they can come from anywhere, unfortunately.
2767 obj->follow_contents(cm);
2768 }
2769 }
2771 void
2772 PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) {
2773 // All klasses on the revisit stack are marked at this point.
2774 // Update and follow all subklass, sibling and implementor links.
2775 for (uint i = 0; i < ParallelGCThreads+1; i++) {
2776 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2777 KeepAliveClosure keep_alive_closure(cm);
2778 for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) {
2779 cm->revisit_klass_stack()->at(i)->follow_weak_klass_links(
2780 is_alive_closure(),
2781 &keep_alive_closure);
2782 }
2783 follow_stack(cm);
2784 }
2785 }
2787 void
2788 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2789 cm->revisit_klass_stack()->push(k);
2790 }
2792 #ifdef VALIDATE_MARK_SWEEP
2794 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2795 if (!ValidateMarkSweep)
2796 return;
2798 if (!isroot) {
2799 if (_pointer_tracking) {
2800 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2801 _adjusted_pointers->remove(p);
2802 }
2803 } else {
2804 ptrdiff_t index = _root_refs_stack->find(p);
2805 if (index != -1) {
2806 int l = _root_refs_stack->length();
2807 if (l > 0 && l - 1 != index) {
2808 void* last = _root_refs_stack->pop();
2809 assert(last != p, "should be different");
2810 _root_refs_stack->at_put(index, last);
2811 } else {
2812 _root_refs_stack->remove(p);
2813 }
2814 }
2815 }
2816 }
2819 void PSParallelCompact::check_adjust_pointer(void* p) {
2820 _adjusted_pointers->push(p);
2821 }
2824 class AdjusterTracker: public OopClosure {
2825 public:
2826 AdjusterTracker() {};
2827 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
2828 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); }
2829 };
2832 void PSParallelCompact::track_interior_pointers(oop obj) {
2833 if (ValidateMarkSweep) {
2834 _adjusted_pointers->clear();
2835 _pointer_tracking = true;
2837 AdjusterTracker checker;
2838 obj->oop_iterate(&checker);
2839 }
2840 }
2843 void PSParallelCompact::check_interior_pointers() {
2844 if (ValidateMarkSweep) {
2845 _pointer_tracking = false;
2846 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2847 }
2848 }
2851 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2852 if (ValidateMarkSweep) {
2853 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2854 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2855 }
2856 }
2859 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2860 if (ValidateMarkSweep) {
2861 _live_oops->push(p);
2862 _live_oops_size->push(size);
2863 _live_oops_index++;
2864 }
2865 }
2867 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2868 if (ValidateMarkSweep) {
2869 oop obj = _live_oops->at((int)_live_oops_index);
2870 guarantee(obj == p, "should be the same object");
2871 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2872 _live_oops_index++;
2873 }
2874 }
2876 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2877 HeapWord* compaction_top) {
2878 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2879 "should be moved to forwarded location");
2880 if (ValidateMarkSweep) {
2881 PSParallelCompact::validate_live_oop(oop(q), size);
2882 _live_oops_moved_to->push(oop(compaction_top));
2883 }
2884 if (RecordMarkSweepCompaction) {
2885 _cur_gc_live_oops->push(q);
2886 _cur_gc_live_oops_moved_to->push(compaction_top);
2887 _cur_gc_live_oops_size->push(size);
2888 }
2889 }
2892 void PSParallelCompact::compaction_complete() {
2893 if (RecordMarkSweepCompaction) {
2894 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
2895 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2896 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
2898 _cur_gc_live_oops = _last_gc_live_oops;
2899 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
2900 _cur_gc_live_oops_size = _last_gc_live_oops_size;
2901 _last_gc_live_oops = _tmp_live_oops;
2902 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2903 _last_gc_live_oops_size = _tmp_live_oops_size;
2904 }
2905 }
2908 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2909 if (!RecordMarkSweepCompaction) {
2910 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2911 return;
2912 }
2914 if (_last_gc_live_oops == NULL) {
2915 tty->print_cr("No compaction information gathered yet");
2916 return;
2917 }
2919 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2920 HeapWord* old_oop = _last_gc_live_oops->at(i);
2921 size_t sz = _last_gc_live_oops_size->at(i);
2922 if (old_oop <= q && q < (old_oop + sz)) {
2923 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2924 size_t offset = (q - old_oop);
2925 tty->print_cr("Address " PTR_FORMAT, q);
2926 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2927 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2928 return;
2929 }
2930 }
2932 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2933 }
2934 #endif //VALIDATE_MARK_SWEEP
2936 // Update interior oops in the ranges of chunks [beg_chunk, end_chunk).
2937 void
2938 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2939 SpaceId space_id,
2940 size_t beg_chunk,
2941 size_t end_chunk) {
2942 ParallelCompactData& sd = summary_data();
2943 ParMarkBitMap* const mbm = mark_bitmap();
2945 HeapWord* beg_addr = sd.chunk_to_addr(beg_chunk);
2946 HeapWord* const end_addr = sd.chunk_to_addr(end_chunk);
2947 assert(beg_chunk <= end_chunk, "bad chunk range");
2948 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2950 #ifdef ASSERT
2951 // Claim the chunks to avoid triggering an assert when they are marked as
2952 // filled.
2953 for (size_t claim_chunk = beg_chunk; claim_chunk < end_chunk; ++claim_chunk) {
2954 assert(sd.chunk(claim_chunk)->claim_unsafe(), "claim() failed");
2955 }
2956 #endif // #ifdef ASSERT
2958 if (beg_addr != space(space_id)->bottom()) {
2959 // Find the first live object or block of dead space that *starts* in this
2960 // range of chunks. If a partial object crosses onto the chunk, skip it; it
2961 // will be marked for 'deferred update' when the object head is processed.
2962 // If dead space crosses onto the chunk, it is also skipped; it will be
2963 // filled when the prior chunk is processed. If neither of those apply, the
2964 // first word in the chunk is the start of a live object or dead space.
2965 assert(beg_addr > space(space_id)->bottom(), "sanity");
2966 const ChunkData* const cp = sd.chunk(beg_chunk);
2967 if (cp->partial_obj_size() != 0) {
2968 beg_addr = sd.partial_obj_end(beg_chunk);
2969 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2970 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2971 }
2972 }
2974 if (beg_addr < end_addr) {
2975 // A live object or block of dead space starts in this range of Chunks.
2976 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2978 // Create closures and iterate.
2979 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2980 FillClosure fill_closure(cm, space_id);
2981 ParMarkBitMap::IterationStatus status;
2982 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2983 dense_prefix_end);
2984 if (status == ParMarkBitMap::incomplete) {
2985 update_closure.do_addr(update_closure.source());
2986 }
2987 }
2989 // Mark the chunks as filled.
2990 ChunkData* const beg_cp = sd.chunk(beg_chunk);
2991 ChunkData* const end_cp = sd.chunk(end_chunk);
2992 for (ChunkData* cp = beg_cp; cp < end_cp; ++cp) {
2993 cp->set_completed();
2994 }
2995 }
2997 // Return the SpaceId for the space containing addr. If addr is not in the
2998 // heap, last_space_id is returned. In debug mode it expects the address to be
2999 // in the heap and asserts such.
3000 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
3001 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
3003 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
3004 if (_space_info[id].space()->contains(addr)) {
3005 return SpaceId(id);
3006 }
3007 }
3009 assert(false, "no space contains the addr");
3010 return last_space_id;
3011 }
3013 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
3014 SpaceId id) {
3015 assert(id < last_space_id, "bad space id");
3017 ParallelCompactData& sd = summary_data();
3018 const SpaceInfo* const space_info = _space_info + id;
3019 ObjectStartArray* const start_array = space_info->start_array();
3021 const MutableSpace* const space = space_info->space();
3022 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
3023 HeapWord* const beg_addr = space_info->dense_prefix();
3024 HeapWord* const end_addr = sd.chunk_align_up(space_info->new_top());
3026 const ChunkData* const beg_chunk = sd.addr_to_chunk_ptr(beg_addr);
3027 const ChunkData* const end_chunk = sd.addr_to_chunk_ptr(end_addr);
3028 const ChunkData* cur_chunk;
3029 for (cur_chunk = beg_chunk; cur_chunk < end_chunk; ++cur_chunk) {
3030 HeapWord* const addr = cur_chunk->deferred_obj_addr();
3031 if (addr != NULL) {
3032 if (start_array != NULL) {
3033 start_array->allocate_block(addr);
3034 }
3035 oop(addr)->update_contents(cm);
3036 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3037 }
3038 }
3039 }
3041 // Skip over count live words starting from beg, and return the address of the
3042 // next live word. Unless marked, the word corresponding to beg is assumed to
3043 // be dead. Callers must either ensure beg does not correspond to the middle of
3044 // an object, or account for those live words in some other way. Callers must
3045 // also ensure that there are enough live words in the range [beg, end) to skip.
3046 HeapWord*
3047 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3048 {
3049 assert(count > 0, "sanity");
3051 ParMarkBitMap* m = mark_bitmap();
3052 idx_t bits_to_skip = m->words_to_bits(count);
3053 idx_t cur_beg = m->addr_to_bit(beg);
3054 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3056 do {
3057 cur_beg = m->find_obj_beg(cur_beg, search_end);
3058 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3059 const size_t obj_bits = cur_end - cur_beg + 1;
3060 if (obj_bits > bits_to_skip) {
3061 return m->bit_to_addr(cur_beg + bits_to_skip);
3062 }
3063 bits_to_skip -= obj_bits;
3064 cur_beg = cur_end + 1;
3065 } while (bits_to_skip > 0);
3067 // Skipping the desired number of words landed just past the end of an object.
3068 // Find the start of the next object.
3069 cur_beg = m->find_obj_beg(cur_beg, search_end);
3070 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3071 return m->bit_to_addr(cur_beg);
3072 }
3074 HeapWord*
3075 PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3076 size_t src_chunk_idx)
3077 {
3078 ParMarkBitMap* const bitmap = mark_bitmap();
3079 const ParallelCompactData& sd = summary_data();
3080 const size_t ChunkSize = ParallelCompactData::ChunkSize;
3082 assert(sd.is_chunk_aligned(dest_addr), "not aligned");
3084 const ChunkData* const src_chunk_ptr = sd.chunk(src_chunk_idx);
3085 const size_t partial_obj_size = src_chunk_ptr->partial_obj_size();
3086 HeapWord* const src_chunk_destination = src_chunk_ptr->destination();
3088 assert(dest_addr >= src_chunk_destination, "wrong src chunk");
3089 assert(src_chunk_ptr->data_size() > 0, "src chunk cannot be empty");
3091 HeapWord* const src_chunk_beg = sd.chunk_to_addr(src_chunk_idx);
3092 HeapWord* const src_chunk_end = src_chunk_beg + ChunkSize;
3094 HeapWord* addr = src_chunk_beg;
3095 if (dest_addr == src_chunk_destination) {
3096 // Return the first live word in the source chunk.
3097 if (partial_obj_size == 0) {
3098 addr = bitmap->find_obj_beg(addr, src_chunk_end);
3099 assert(addr < src_chunk_end, "no objects start in src chunk");
3100 }
3101 return addr;
3102 }
3104 // Must skip some live data.
3105 size_t words_to_skip = dest_addr - src_chunk_destination;
3106 assert(src_chunk_ptr->data_size() > words_to_skip, "wrong src chunk");
3108 if (partial_obj_size >= words_to_skip) {
3109 // All the live words to skip are part of the partial object.
3110 addr += words_to_skip;
3111 if (partial_obj_size == words_to_skip) {
3112 // Find the first live word past the partial object.
3113 addr = bitmap->find_obj_beg(addr, src_chunk_end);
3114 assert(addr < src_chunk_end, "wrong src chunk");
3115 }
3116 return addr;
3117 }
3119 // Skip over the partial object (if any).
3120 if (partial_obj_size != 0) {
3121 words_to_skip -= partial_obj_size;
3122 addr += partial_obj_size;
3123 }
3125 // Skip over live words due to objects that start in the chunk.
3126 addr = skip_live_words(addr, src_chunk_end, words_to_skip);
3127 assert(addr < src_chunk_end, "wrong src chunk");
3128 return addr;
3129 }
3131 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3132 size_t beg_chunk,
3133 HeapWord* end_addr)
3134 {
3135 ParallelCompactData& sd = summary_data();
3136 ChunkData* const beg = sd.chunk(beg_chunk);
3137 HeapWord* const end_addr_aligned_up = sd.chunk_align_up(end_addr);
3138 ChunkData* const end = sd.addr_to_chunk_ptr(end_addr_aligned_up);
3139 size_t cur_idx = beg_chunk;
3140 for (ChunkData* cur = beg; cur < end; ++cur, ++cur_idx) {
3141 assert(cur->data_size() > 0, "chunk must have live data");
3142 cur->decrement_destination_count();
3143 if (cur_idx <= cur->source_chunk() && cur->available() && cur->claim()) {
3144 cm->save_for_processing(cur_idx);
3145 }
3146 }
3147 }
3149 size_t PSParallelCompact::next_src_chunk(MoveAndUpdateClosure& closure,
3150 SpaceId& src_space_id,
3151 HeapWord*& src_space_top,
3152 HeapWord* end_addr)
3153 {
3154 typedef ParallelCompactData::ChunkData ChunkData;
3156 ParallelCompactData& sd = PSParallelCompact::summary_data();
3157 const size_t chunk_size = ParallelCompactData::ChunkSize;
3159 size_t src_chunk_idx = 0;
3161 // Skip empty chunks (if any) up to the top of the space.
3162 HeapWord* const src_aligned_up = sd.chunk_align_up(end_addr);
3163 ChunkData* src_chunk_ptr = sd.addr_to_chunk_ptr(src_aligned_up);
3164 HeapWord* const top_aligned_up = sd.chunk_align_up(src_space_top);
3165 const ChunkData* const top_chunk_ptr = sd.addr_to_chunk_ptr(top_aligned_up);
3166 while (src_chunk_ptr < top_chunk_ptr && src_chunk_ptr->data_size() == 0) {
3167 ++src_chunk_ptr;
3168 }
3170 if (src_chunk_ptr < top_chunk_ptr) {
3171 // The next source chunk is in the current space. Update src_chunk_idx and
3172 // the source address to match src_chunk_ptr.
3173 src_chunk_idx = sd.chunk(src_chunk_ptr);
3174 HeapWord* const src_chunk_addr = sd.chunk_to_addr(src_chunk_idx);
3175 if (src_chunk_addr > closure.source()) {
3176 closure.set_source(src_chunk_addr);
3177 }
3178 return src_chunk_idx;
3179 }
3181 // Switch to a new source space and find the first non-empty chunk.
3182 unsigned int space_id = src_space_id + 1;
3183 assert(space_id < last_space_id, "not enough spaces");
3185 HeapWord* const destination = closure.destination();
3187 do {
3188 MutableSpace* space = _space_info[space_id].space();
3189 HeapWord* const bottom = space->bottom();
3190 const ChunkData* const bottom_cp = sd.addr_to_chunk_ptr(bottom);
3192 // Iterate over the spaces that do not compact into themselves.
3193 if (bottom_cp->destination() != bottom) {
3194 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top());
3195 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up);
3197 for (const ChunkData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3198 if (src_cp->live_obj_size() > 0) {
3199 // Found it.
3200 assert(src_cp->destination() == destination,
3201 "first live obj in the space must match the destination");
3202 assert(src_cp->partial_obj_size() == 0,
3203 "a space cannot begin with a partial obj");
3205 src_space_id = SpaceId(space_id);
3206 src_space_top = space->top();
3207 const size_t src_chunk_idx = sd.chunk(src_cp);
3208 closure.set_source(sd.chunk_to_addr(src_chunk_idx));
3209 return src_chunk_idx;
3210 } else {
3211 assert(src_cp->data_size() == 0, "sanity");
3212 }
3213 }
3214 }
3215 } while (++space_id < last_space_id);
3217 assert(false, "no source chunk was found");
3218 return 0;
3219 }
3221 void PSParallelCompact::fill_chunk(ParCompactionManager* cm, size_t chunk_idx)
3222 {
3223 typedef ParMarkBitMap::IterationStatus IterationStatus;
3224 const size_t ChunkSize = ParallelCompactData::ChunkSize;
3225 ParMarkBitMap* const bitmap = mark_bitmap();
3226 ParallelCompactData& sd = summary_data();
3227 ChunkData* const chunk_ptr = sd.chunk(chunk_idx);
3229 // Get the items needed to construct the closure.
3230 HeapWord* dest_addr = sd.chunk_to_addr(chunk_idx);
3231 SpaceId dest_space_id = space_id(dest_addr);
3232 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3233 HeapWord* new_top = _space_info[dest_space_id].new_top();
3234 assert(dest_addr < new_top, "sanity");
3235 const size_t words = MIN2(pointer_delta(new_top, dest_addr), ChunkSize);
3237 // Get the source chunk and related info.
3238 size_t src_chunk_idx = chunk_ptr->source_chunk();
3239 SpaceId src_space_id = space_id(sd.chunk_to_addr(src_chunk_idx));
3240 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3242 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3243 closure.set_source(first_src_addr(dest_addr, src_chunk_idx));
3245 // Adjust src_chunk_idx to prepare for decrementing destination counts (the
3246 // destination count is not decremented when a chunk is copied to itself).
3247 if (src_chunk_idx == chunk_idx) {
3248 src_chunk_idx += 1;
3249 }
3251 if (bitmap->is_unmarked(closure.source())) {
3252 // The first source word is in the middle of an object; copy the remainder
3253 // of the object or as much as will fit. The fact that pointer updates were
3254 // deferred will be noted when the object header is processed.
3255 HeapWord* const old_src_addr = closure.source();
3256 closure.copy_partial_obj();
3257 if (closure.is_full()) {
3258 decrement_destination_counts(cm, src_chunk_idx, closure.source());
3259 chunk_ptr->set_deferred_obj_addr(NULL);
3260 chunk_ptr->set_completed();
3261 return;
3262 }
3264 HeapWord* const end_addr = sd.chunk_align_down(closure.source());
3265 if (sd.chunk_align_down(old_src_addr) != end_addr) {
3266 // The partial object was copied from more than one source chunk.
3267 decrement_destination_counts(cm, src_chunk_idx, end_addr);
3269 // Move to the next source chunk, possibly switching spaces as well. All
3270 // args except end_addr may be modified.
3271 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top,
3272 end_addr);
3273 }
3274 }
3276 do {
3277 HeapWord* const cur_addr = closure.source();
3278 HeapWord* const end_addr = MIN2(sd.chunk_align_up(cur_addr + 1),
3279 src_space_top);
3280 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3282 if (status == ParMarkBitMap::incomplete) {
3283 // The last obj that starts in the source chunk does not end in the chunk.
3284 assert(closure.source() < end_addr, "sanity")
3285 HeapWord* const obj_beg = closure.source();
3286 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3287 src_space_top);
3288 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3289 if (obj_end < range_end) {
3290 // The end was found; the entire object will fit.
3291 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3292 assert(status != ParMarkBitMap::would_overflow, "sanity");
3293 } else {
3294 // The end was not found; the object will not fit.
3295 assert(range_end < src_space_top, "obj cannot cross space boundary");
3296 status = ParMarkBitMap::would_overflow;
3297 }
3298 }
3300 if (status == ParMarkBitMap::would_overflow) {
3301 // The last object did not fit. Note that interior oop updates were
3302 // deferred, then copy enough of the object to fill the chunk.
3303 chunk_ptr->set_deferred_obj_addr(closure.destination());
3304 status = closure.copy_until_full(); // copies from closure.source()
3306 decrement_destination_counts(cm, src_chunk_idx, closure.source());
3307 chunk_ptr->set_completed();
3308 return;
3309 }
3311 if (status == ParMarkBitMap::full) {
3312 decrement_destination_counts(cm, src_chunk_idx, closure.source());
3313 chunk_ptr->set_deferred_obj_addr(NULL);
3314 chunk_ptr->set_completed();
3315 return;
3316 }
3318 decrement_destination_counts(cm, src_chunk_idx, end_addr);
3320 // Move to the next source chunk, possibly switching spaces as well. All
3321 // args except end_addr may be modified.
3322 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top,
3323 end_addr);
3324 } while (true);
3325 }
3327 void
3328 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3329 const MutableSpace* sp = space(space_id);
3330 if (sp->is_empty()) {
3331 return;
3332 }
3334 ParallelCompactData& sd = PSParallelCompact::summary_data();
3335 ParMarkBitMap* const bitmap = mark_bitmap();
3336 HeapWord* const dp_addr = dense_prefix(space_id);
3337 HeapWord* beg_addr = sp->bottom();
3338 HeapWord* end_addr = sp->top();
3340 #ifdef ASSERT
3341 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3342 if (cm->should_verify_only()) {
3343 VerifyUpdateClosure verify_update(cm, sp);
3344 bitmap->iterate(&verify_update, beg_addr, end_addr);
3345 return;
3346 }
3348 if (cm->should_reset_only()) {
3349 ResetObjectsClosure reset_objects(cm);
3350 bitmap->iterate(&reset_objects, beg_addr, end_addr);
3351 return;
3352 }
3353 #endif
3355 const size_t beg_chunk = sd.addr_to_chunk_idx(beg_addr);
3356 const size_t dp_chunk = sd.addr_to_chunk_idx(dp_addr);
3357 if (beg_chunk < dp_chunk) {
3358 update_and_deadwood_in_dense_prefix(cm, space_id, beg_chunk, dp_chunk);
3359 }
3361 // The destination of the first live object that starts in the chunk is one
3362 // past the end of the partial object entering the chunk (if any).
3363 HeapWord* const dest_addr = sd.partial_obj_end(dp_chunk);
3364 HeapWord* const new_top = _space_info[space_id].new_top();
3365 assert(new_top >= dest_addr, "bad new_top value");
3366 const size_t words = pointer_delta(new_top, dest_addr);
3368 if (words > 0) {
3369 ObjectStartArray* start_array = _space_info[space_id].start_array();
3370 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3372 ParMarkBitMap::IterationStatus status;
3373 status = bitmap->iterate(&closure, dest_addr, end_addr);
3374 assert(status == ParMarkBitMap::full, "iteration not complete");
3375 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3376 "live objects skipped because closure is full");
3377 }
3378 }
3380 jlong PSParallelCompact::millis_since_last_gc() {
3381 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
3382 // XXX See note in genCollectedHeap::millis_since_last_gc().
3383 if (ret_val < 0) {
3384 NOT_PRODUCT(warning("time warp: %d", ret_val);)
3385 return 0;
3386 }
3387 return ret_val;
3388 }
3390 void PSParallelCompact::reset_millis_since_last_gc() {
3391 _time_of_last_gc = os::javaTimeMillis();
3392 }
3394 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3395 {
3396 if (source() != destination()) {
3397 assert(source() > destination(), "must copy to the left");
3398 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3399 }
3400 update_state(words_remaining());
3401 assert(is_full(), "sanity");
3402 return ParMarkBitMap::full;
3403 }
3405 void MoveAndUpdateClosure::copy_partial_obj()
3406 {
3407 size_t words = words_remaining();
3409 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3410 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3411 if (end_addr < range_end) {
3412 words = bitmap()->obj_size(source(), end_addr);
3413 }
3415 // This test is necessary; if omitted, the pointer updates to a partial object
3416 // that crosses the dense prefix boundary could be overwritten.
3417 if (source() != destination()) {
3418 assert(source() > destination(), "must copy to the left");
3419 Copy::aligned_conjoint_words(source(), destination(), words);
3420 }
3421 update_state(words);
3422 }
3424 ParMarkBitMapClosure::IterationStatus
3425 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3426 assert(destination() != NULL, "sanity");
3427 assert(bitmap()->obj_size(addr) == words, "bad size");
3429 _source = addr;
3430 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3431 destination(), "wrong destination");
3433 if (words > words_remaining()) {
3434 return ParMarkBitMap::would_overflow;
3435 }
3437 // The start_array must be updated even if the object is not moving.
3438 if (_start_array != NULL) {
3439 _start_array->allocate_block(destination());
3440 }
3442 if (destination() != source()) {
3443 assert(destination() < source(), "must copy to the left");
3444 Copy::aligned_conjoint_words(source(), destination(), words);
3445 }
3447 oop moved_oop = (oop) destination();
3448 moved_oop->update_contents(compaction_manager());
3449 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3451 update_state(words);
3452 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3453 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3454 }
3456 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3457 ParCompactionManager* cm,
3458 PSParallelCompact::SpaceId space_id) :
3459 ParMarkBitMapClosure(mbm, cm),
3460 _space_id(space_id),
3461 _start_array(PSParallelCompact::start_array(space_id))
3462 {
3463 }
3465 // Updates the references in the object to their new values.
3466 ParMarkBitMapClosure::IterationStatus
3467 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3468 do_addr(addr);
3469 return ParMarkBitMap::incomplete;
3470 }
3472 BitBlockUpdateClosure::BitBlockUpdateClosure(ParMarkBitMap* mbm,
3473 ParCompactionManager* cm,
3474 size_t chunk_index) :
3475 ParMarkBitMapClosure(mbm, cm),
3476 _live_data_left(0),
3477 _cur_block(0) {
3478 _chunk_start =
3479 PSParallelCompact::summary_data().chunk_to_addr(chunk_index);
3480 _chunk_end =
3481 PSParallelCompact::summary_data().chunk_to_addr(chunk_index) +
3482 ParallelCompactData::ChunkSize;
3483 _chunk_index = chunk_index;
3484 _cur_block =
3485 PSParallelCompact::summary_data().addr_to_block_idx(_chunk_start);
3486 }
3488 bool BitBlockUpdateClosure::chunk_contains_cur_block() {
3489 return ParallelCompactData::chunk_contains_block(_chunk_index, _cur_block);
3490 }
3492 void BitBlockUpdateClosure::reset_chunk(size_t chunk_index) {
3493 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(7);)
3494 ParallelCompactData& sd = PSParallelCompact::summary_data();
3495 _chunk_index = chunk_index;
3496 _live_data_left = 0;
3497 _chunk_start = sd.chunk_to_addr(chunk_index);
3498 _chunk_end = sd.chunk_to_addr(chunk_index) + ParallelCompactData::ChunkSize;
3500 // The first block in this chunk
3501 size_t first_block = sd.addr_to_block_idx(_chunk_start);
3502 size_t partial_live_size = sd.chunk(chunk_index)->partial_obj_size();
3504 // Set the offset to 0. By definition it should have that value
3505 // but it may have been written while processing an earlier chunk.
3506 if (partial_live_size == 0) {
3507 // No live object extends onto the chunk. The first bit
3508 // in the bit map for the first chunk must be a start bit.
3509 // Although there may not be any marked bits, it is safe
3510 // to set it as a start bit.
3511 sd.block(first_block)->set_start_bit_offset(0);
3512 sd.block(first_block)->set_first_is_start_bit(true);
3513 } else if (sd.partial_obj_ends_in_block(first_block)) {
3514 sd.block(first_block)->set_end_bit_offset(0);
3515 sd.block(first_block)->set_first_is_start_bit(false);
3516 } else {
3517 // The partial object extends beyond the first block.
3518 // There is no object starting in the first block
3519 // so the offset and bit parity are not needed.
3520 // Set the the bit parity to start bit so assertions
3521 // work when not bit is found.
3522 sd.block(first_block)->set_end_bit_offset(0);
3523 sd.block(first_block)->set_first_is_start_bit(false);
3524 }
3525 _cur_block = first_block;
3526 #ifdef ASSERT
3527 if (sd.block(first_block)->first_is_start_bit()) {
3528 assert(!sd.partial_obj_ends_in_block(first_block),
3529 "Partial object cannot end in first block");
3530 }
3532 if (PrintGCDetails && Verbose) {
3533 if (partial_live_size == 1) {
3534 gclog_or_tty->print_cr("first_block " PTR_FORMAT
3535 " _offset " PTR_FORMAT
3536 " _first_is_start_bit %d",
3537 first_block,
3538 sd.block(first_block)->raw_offset(),
3539 sd.block(first_block)->first_is_start_bit());
3540 }
3541 }
3542 #endif
3543 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(17);)
3544 }
3546 // This method is called when a object has been found (both beginning
3547 // and end of the object) in the range of iteration. This method is
3548 // calculating the words of live data to the left of a block. That live
3549 // data includes any object starting to the left of the block (i.e.,
3550 // the live-data-to-the-left of block AAA will include the full size
3551 // of any object entering AAA).
3553 ParMarkBitMapClosure::IterationStatus
3554 BitBlockUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3555 // add the size to the block data.
3556 HeapWord* obj = addr;
3557 ParallelCompactData& sd = PSParallelCompact::summary_data();
3559 assert(bitmap()->obj_size(obj) == words, "bad size");
3560 assert(_chunk_start <= obj, "object is not in chunk");
3561 assert(obj + words <= _chunk_end, "object is not in chunk");
3563 // Update the live data to the left
3564 size_t prev_live_data_left = _live_data_left;
3565 _live_data_left = _live_data_left + words;
3567 // Is this object in the current block.
3568 size_t block_of_obj = sd.addr_to_block_idx(obj);
3569 size_t block_of_obj_last = sd.addr_to_block_idx(obj + words - 1);
3570 HeapWord* block_of_obj_last_addr = sd.block_to_addr(block_of_obj_last);
3571 if (_cur_block < block_of_obj) {
3573 //
3574 // No object crossed the block boundary and this object was found
3575 // on the other side of the block boundary. Update the offset for
3576 // the new block with the data size that does not include this object.
3577 //
3578 // The first bit in block_of_obj is a start bit except in the
3579 // case where the partial object for the chunk extends into
3580 // this block.
3581 if (sd.partial_obj_ends_in_block(block_of_obj)) {
3582 sd.block(block_of_obj)->set_end_bit_offset(prev_live_data_left);
3583 } else {
3584 sd.block(block_of_obj)->set_start_bit_offset(prev_live_data_left);
3585 }
3587 // Does this object pass beyond the its block?
3588 if (block_of_obj < block_of_obj_last) {
3589 // Object crosses block boundary. Two blocks need to be udpated:
3590 // the current block where the object started
3591 // the block where the object ends
3592 //
3593 // The offset for blocks with no objects starting in them
3594 // (e.g., blocks between _cur_block and block_of_obj_last)
3595 // should not be needed.
3596 // Note that block_of_obj_last may be in another chunk. If so,
3597 // it should be overwritten later. This is a problem (writting
3598 // into a block in a later chunk) for parallel execution.
3599 assert(obj < block_of_obj_last_addr,
3600 "Object should start in previous block");
3602 // obj is crossing into block_of_obj_last so the first bit
3603 // is and end bit.
3604 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left);
3606 _cur_block = block_of_obj_last;
3607 } else {
3608 // _first_is_start_bit has already been set correctly
3609 // in the if-then-else above so don't reset it here.
3610 _cur_block = block_of_obj;
3611 }
3612 } else {
3613 // The current block only changes if the object extends beyound
3614 // the block it starts in.
3615 //
3616 // The object starts in the current block.
3617 // Does this object pass beyond the end of it?
3618 if (block_of_obj < block_of_obj_last) {
3619 // Object crosses block boundary.
3620 // See note above on possible blocks between block_of_obj and
3621 // block_of_obj_last
3622 assert(obj < block_of_obj_last_addr,
3623 "Object should start in previous block");
3625 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left);
3627 _cur_block = block_of_obj_last;
3628 }
3629 }
3631 // Return incomplete if there are more blocks to be done.
3632 if (chunk_contains_cur_block()) {
3633 return ParMarkBitMap::incomplete;
3634 }
3635 return ParMarkBitMap::complete;
3636 }
3638 // Verify the new location using the forwarding pointer
3639 // from MarkSweep::mark_sweep_phase2(). Set the mark_word
3640 // to the initial value.
3641 ParMarkBitMapClosure::IterationStatus
3642 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3643 // The second arg (words) is not used.
3644 oop obj = (oop) addr;
3645 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
3646 HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
3647 if (forwarding_ptr == NULL) {
3648 // The object is dead or not moving.
3649 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
3650 "Object liveness is wrong.");
3651 return ParMarkBitMap::incomplete;
3652 }
3653 assert(UseParallelOldGCDensePrefix ||
3654 (HeapMaximumCompactionInterval > 1) ||
3655 (MarkSweepAlwaysCompactCount > 1) ||
3656 (forwarding_ptr == new_pointer),
3657 "Calculation of new location is incorrect");
3658 return ParMarkBitMap::incomplete;
3659 }
3661 // Reset objects modified for debug checking.
3662 ParMarkBitMapClosure::IterationStatus
3663 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
3664 // The second arg (words) is not used.
3665 oop obj = (oop) addr;
3666 obj->init_mark();
3667 return ParMarkBitMap::incomplete;
3668 }
3670 // Prepare for compaction. This method is executed once
3671 // (i.e., by a single thread) before compaction.
3672 // Save the updated location of the intArrayKlassObj for
3673 // filling holes in the dense prefix.
3674 void PSParallelCompact::compact_prologue() {
3675 _updated_int_array_klass_obj = (klassOop)
3676 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3677 }
3679 // The initial implementation of this method created a field
3680 // _next_compaction_space_id in SpaceInfo and initialized
3681 // that field in SpaceInfo::initialize_space_info(). That
3682 // required that _next_compaction_space_id be declared a
3683 // SpaceId in SpaceInfo and that would have required that
3684 // either SpaceId be declared in a separate class or that
3685 // it be declared in SpaceInfo. It didn't seem consistent
3686 // to declare it in SpaceInfo (didn't really fit logically).
3687 // Alternatively, defining a separate class to define SpaceId
3688 // seem excessive. This implementation is simple and localizes
3689 // the knowledge.
3691 PSParallelCompact::SpaceId
3692 PSParallelCompact::next_compaction_space_id(SpaceId id) {
3693 assert(id < last_space_id, "id out of range");
3694 switch (id) {
3695 case perm_space_id :
3696 return last_space_id;
3697 case old_space_id :
3698 return eden_space_id;
3699 case eden_space_id :
3700 return from_space_id;
3701 case from_space_id :
3702 return to_space_id;
3703 case to_space_id :
3704 return last_space_id;
3705 default:
3706 assert(false, "Bad space id");
3707 return last_space_id;
3708 }
3709 }
3711 // Here temporarily for debugging
3712 #ifdef ASSERT
3713 size_t ParallelCompactData::block_idx(BlockData* block) {
3714 size_t index = pointer_delta(block,
3715 PSParallelCompact::summary_data()._block_data, sizeof(BlockData));
3716 return index;
3717 }
3718 #endif