Sun, 04 May 2008 03:29:31 -0700
Merge
1 /*
2 * Copyright 2005-2007 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 }
427 return 0;
428 }
430 bool ParallelCompactData::initialize_chunk_data(size_t region_size)
431 {
432 const size_t count = (region_size + ChunkSizeOffsetMask) >> Log2ChunkSize;
433 _chunk_vspace = create_vspace(count, sizeof(ChunkData));
434 if (_chunk_vspace != 0) {
435 _chunk_data = (ChunkData*)_chunk_vspace->reserved_low_addr();
436 _chunk_count = count;
437 return true;
438 }
439 return false;
440 }
442 bool ParallelCompactData::initialize_block_data(size_t region_size)
443 {
444 const size_t count = (region_size + BlockOffsetMask) >> Log2BlockSize;
445 _block_vspace = create_vspace(count, sizeof(BlockData));
446 if (_block_vspace != 0) {
447 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
448 _block_count = count;
449 return true;
450 }
451 return false;
452 }
454 void ParallelCompactData::clear()
455 {
456 if (_block_data) {
457 memset(_block_data, 0, _block_vspace->committed_size());
458 }
459 memset(_chunk_data, 0, _chunk_vspace->committed_size());
460 }
462 void ParallelCompactData::clear_range(size_t beg_chunk, size_t end_chunk) {
463 assert(beg_chunk <= _chunk_count, "beg_chunk out of range");
464 assert(end_chunk <= _chunk_count, "end_chunk out of range");
465 assert(ChunkSize % BlockSize == 0, "ChunkSize not a multiple of BlockSize");
467 const size_t chunk_cnt = end_chunk - beg_chunk;
469 if (_block_data) {
470 const size_t blocks_per_chunk = ChunkSize / BlockSize;
471 const size_t beg_block = beg_chunk * blocks_per_chunk;
472 const size_t block_cnt = chunk_cnt * blocks_per_chunk;
473 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
474 }
475 memset(_chunk_data + beg_chunk, 0, chunk_cnt * sizeof(ChunkData));
476 }
478 HeapWord* ParallelCompactData::partial_obj_end(size_t chunk_idx) const
479 {
480 const ChunkData* cur_cp = chunk(chunk_idx);
481 const ChunkData* const end_cp = chunk(chunk_count() - 1);
483 HeapWord* result = chunk_to_addr(chunk_idx);
484 if (cur_cp < end_cp) {
485 do {
486 result += cur_cp->partial_obj_size();
487 } while (cur_cp->partial_obj_size() == ChunkSize && ++cur_cp < end_cp);
488 }
489 return result;
490 }
492 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
493 {
494 const size_t obj_ofs = pointer_delta(addr, _region_start);
495 const size_t beg_chunk = obj_ofs >> Log2ChunkSize;
496 const size_t end_chunk = (obj_ofs + len - 1) >> Log2ChunkSize;
498 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
499 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
501 if (beg_chunk == end_chunk) {
502 // All in one chunk.
503 _chunk_data[beg_chunk].add_live_obj(len);
504 return;
505 }
507 // First chunk.
508 const size_t beg_ofs = chunk_offset(addr);
509 _chunk_data[beg_chunk].add_live_obj(ChunkSize - beg_ofs);
511 klassOop klass = ((oop)addr)->klass();
512 // Middle chunks--completely spanned by this object.
513 for (size_t chunk = beg_chunk + 1; chunk < end_chunk; ++chunk) {
514 _chunk_data[chunk].set_partial_obj_size(ChunkSize);
515 _chunk_data[chunk].set_partial_obj_addr(addr);
516 }
518 // Last chunk.
519 const size_t end_ofs = chunk_offset(addr + len - 1);
520 _chunk_data[end_chunk].set_partial_obj_size(end_ofs + 1);
521 _chunk_data[end_chunk].set_partial_obj_addr(addr);
522 }
524 void
525 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
526 {
527 assert(chunk_offset(beg) == 0, "not ChunkSize aligned");
528 assert(chunk_offset(end) == 0, "not ChunkSize aligned");
530 size_t cur_chunk = addr_to_chunk_idx(beg);
531 const size_t end_chunk = addr_to_chunk_idx(end);
532 HeapWord* addr = beg;
533 while (cur_chunk < end_chunk) {
534 _chunk_data[cur_chunk].set_destination(addr);
535 _chunk_data[cur_chunk].set_destination_count(0);
536 _chunk_data[cur_chunk].set_source_chunk(cur_chunk);
537 _chunk_data[cur_chunk].set_data_location(addr);
539 // Update live_obj_size so the chunk appears completely full.
540 size_t live_size = ChunkSize - _chunk_data[cur_chunk].partial_obj_size();
541 _chunk_data[cur_chunk].set_live_obj_size(live_size);
543 ++cur_chunk;
544 addr += ChunkSize;
545 }
546 }
548 bool ParallelCompactData::summarize(HeapWord* target_beg, HeapWord* target_end,
549 HeapWord* source_beg, HeapWord* source_end,
550 HeapWord** target_next,
551 HeapWord** source_next) {
552 // This is too strict.
553 // assert(chunk_offset(source_beg) == 0, "not ChunkSize aligned");
555 if (TraceParallelOldGCSummaryPhase) {
556 tty->print_cr("tb=" PTR_FORMAT " te=" PTR_FORMAT " "
557 "sb=" PTR_FORMAT " se=" PTR_FORMAT " "
558 "tn=" PTR_FORMAT " sn=" PTR_FORMAT,
559 target_beg, target_end,
560 source_beg, source_end,
561 target_next != 0 ? *target_next : (HeapWord*) 0,
562 source_next != 0 ? *source_next : (HeapWord*) 0);
563 }
565 size_t cur_chunk = addr_to_chunk_idx(source_beg);
566 const size_t end_chunk = addr_to_chunk_idx(chunk_align_up(source_end));
568 HeapWord *dest_addr = target_beg;
569 while (cur_chunk < end_chunk) {
570 size_t words = _chunk_data[cur_chunk].data_size();
572 #if 1
573 assert(pointer_delta(target_end, dest_addr) >= words,
574 "source region does not fit into target region");
575 #else
576 // XXX - need some work on the corner cases here. If the chunk does not
577 // fit, then must either make sure any partial_obj from the chunk fits, or
578 // 'undo' the initial part of the partial_obj that is in the previous chunk.
579 if (dest_addr + words >= target_end) {
580 // Let the caller know where to continue.
581 *target_next = dest_addr;
582 *source_next = chunk_to_addr(cur_chunk);
583 return false;
584 }
585 #endif // #if 1
587 _chunk_data[cur_chunk].set_destination(dest_addr);
589 // Set the destination_count for cur_chunk, and if necessary, update
590 // source_chunk for a destination chunk. The source_chunk field is updated
591 // if cur_chunk is the first (left-most) chunk to be copied to a destination
592 // chunk.
593 //
594 // The destination_count calculation is a bit subtle. A chunk that has data
595 // that compacts into itself does not count itself as a destination. This
596 // maintains the invariant that a zero count means the chunk is available
597 // and can be claimed and then filled.
598 if (words > 0) {
599 HeapWord* const last_addr = dest_addr + words - 1;
600 const size_t dest_chunk_1 = addr_to_chunk_idx(dest_addr);
601 const size_t dest_chunk_2 = addr_to_chunk_idx(last_addr);
602 #if 0
603 // Initially assume that the destination chunks will be the same and
604 // adjust the value below if necessary. Under this assumption, if
605 // cur_chunk == dest_chunk_2, then cur_chunk will be compacted completely
606 // into itself.
607 uint destination_count = cur_chunk == dest_chunk_2 ? 0 : 1;
608 if (dest_chunk_1 != dest_chunk_2) {
609 // Destination chunks differ; adjust destination_count.
610 destination_count += 1;
611 // Data from cur_chunk will be copied to the start of dest_chunk_2.
612 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk);
613 } else if (chunk_offset(dest_addr) == 0) {
614 // Data from cur_chunk will be copied to the start of the destination
615 // chunk.
616 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk);
617 }
618 #else
619 // Initially assume that the destination chunks will be different and
620 // adjust the value below if necessary. Under this assumption, if
621 // cur_chunk == dest_chunk2, then cur_chunk will be compacted partially
622 // into dest_chunk_1 and partially into itself.
623 uint destination_count = cur_chunk == dest_chunk_2 ? 1 : 2;
624 if (dest_chunk_1 != dest_chunk_2) {
625 // Data from cur_chunk will be copied to the start of dest_chunk_2.
626 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk);
627 } else {
628 // Destination chunks are the same; adjust destination_count.
629 destination_count -= 1;
630 if (chunk_offset(dest_addr) == 0) {
631 // Data from cur_chunk will be copied to the start of the destination
632 // chunk.
633 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk);
634 }
635 }
636 #endif // #if 0
638 _chunk_data[cur_chunk].set_destination_count(destination_count);
639 _chunk_data[cur_chunk].set_data_location(chunk_to_addr(cur_chunk));
640 dest_addr += words;
641 }
643 ++cur_chunk;
644 }
646 *target_next = dest_addr;
647 return true;
648 }
650 bool ParallelCompactData::partial_obj_ends_in_block(size_t block_index) {
651 HeapWord* block_addr = block_to_addr(block_index);
652 HeapWord* block_end_addr = block_addr + BlockSize;
653 size_t chunk_index = addr_to_chunk_idx(block_addr);
654 HeapWord* partial_obj_end_addr = partial_obj_end(chunk_index);
656 // An object that ends at the end of the block, ends
657 // in the block (the last word of the object is to
658 // the left of the end).
659 if ((block_addr < partial_obj_end_addr) &&
660 (partial_obj_end_addr <= block_end_addr)) {
661 return true;
662 }
664 return false;
665 }
667 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
668 HeapWord* result = NULL;
669 if (UseParallelOldGCChunkPointerCalc) {
670 result = chunk_calc_new_pointer(addr);
671 } else {
672 result = block_calc_new_pointer(addr);
673 }
674 return result;
675 }
677 // This method is overly complicated (expensive) to be called
678 // for every reference.
679 // Try to restructure this so that a NULL is returned if
680 // the object is dead. But don't wast the cycles to explicitly check
681 // that it is dead since only live objects should be passed in.
683 HeapWord* ParallelCompactData::chunk_calc_new_pointer(HeapWord* addr) {
684 assert(addr != NULL, "Should detect NULL oop earlier");
685 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
686 #ifdef ASSERT
687 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
688 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
689 }
690 #endif
691 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
693 // Chunk covering the object.
694 size_t chunk_index = addr_to_chunk_idx(addr);
695 const ChunkData* const chunk_ptr = chunk(chunk_index);
696 HeapWord* const chunk_addr = chunk_align_down(addr);
698 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object");
699 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check");
701 HeapWord* result = chunk_ptr->destination();
703 // If all the data in the chunk is live, then the new location of the object
704 // can be calculated from the destination of the chunk plus the offset of the
705 // object in the chunk.
706 if (chunk_ptr->data_size() == ChunkSize) {
707 result += pointer_delta(addr, chunk_addr);
708 return result;
709 }
711 // The new location of the object is
712 // chunk destination +
713 // size of the partial object extending onto the chunk +
714 // sizes of the live objects in the Chunk that are to the left of addr
715 const size_t partial_obj_size = chunk_ptr->partial_obj_size();
716 HeapWord* const search_start = chunk_addr + partial_obj_size;
718 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
719 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
721 result += partial_obj_size + live_to_left;
722 assert(result <= addr, "object cannot move to the right");
723 return result;
724 }
726 HeapWord* ParallelCompactData::block_calc_new_pointer(HeapWord* addr) {
727 assert(addr != NULL, "Should detect NULL oop earlier");
728 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
729 #ifdef ASSERT
730 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
731 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
732 }
733 #endif
734 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
736 // Chunk covering the object.
737 size_t chunk_index = addr_to_chunk_idx(addr);
738 const ChunkData* const chunk_ptr = chunk(chunk_index);
739 HeapWord* const chunk_addr = chunk_align_down(addr);
741 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object");
742 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check");
744 HeapWord* result = chunk_ptr->destination();
746 // If all the data in the chunk is live, then the new location of the object
747 // can be calculated from the destination of the chunk plus the offset of the
748 // object in the chunk.
749 if (chunk_ptr->data_size() == ChunkSize) {
750 result += pointer_delta(addr, chunk_addr);
751 return result;
752 }
754 // The new location of the object is
755 // chunk destination +
756 // block offset +
757 // sizes of the live objects in the Block that are to the left of addr
758 const size_t block_offset = addr_to_block_ptr(addr)->offset();
759 HeapWord* const search_start = chunk_addr + block_offset;
761 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
762 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
764 result += block_offset + live_to_left;
765 assert(result <= addr, "object cannot move to the right");
766 assert(result == chunk_calc_new_pointer(addr), "Should match");
767 return result;
768 }
770 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
771 klassOop updated_klass;
772 if (PSParallelCompact::should_update_klass(old_klass)) {
773 updated_klass = (klassOop) calc_new_pointer(old_klass);
774 } else {
775 updated_klass = old_klass;
776 }
778 return updated_klass;
779 }
781 #ifdef ASSERT
782 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
783 {
784 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
785 const size_t* const end = (const size_t*)vspace->committed_high_addr();
786 for (const size_t* p = beg; p < end; ++p) {
787 assert(*p == 0, "not zero");
788 }
789 }
791 void ParallelCompactData::verify_clear()
792 {
793 verify_clear(_chunk_vspace);
794 verify_clear(_block_vspace);
795 }
796 #endif // #ifdef ASSERT
798 #ifdef NOT_PRODUCT
799 ParallelCompactData::ChunkData* debug_chunk(size_t chunk_index) {
800 ParallelCompactData& sd = PSParallelCompact::summary_data();
801 return sd.chunk(chunk_index);
802 }
803 #endif
805 elapsedTimer PSParallelCompact::_accumulated_time;
806 unsigned int PSParallelCompact::_total_invocations = 0;
807 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
808 jlong PSParallelCompact::_time_of_last_gc = 0;
809 CollectorCounters* PSParallelCompact::_counters = NULL;
810 ParMarkBitMap PSParallelCompact::_mark_bitmap;
811 ParallelCompactData PSParallelCompact::_summary_data;
813 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
815 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
816 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
818 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
819 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
821 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
822 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
824 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); }
825 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
827 void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); }
829 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); }
830 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
832 void PSParallelCompact::post_initialize() {
833 ParallelScavengeHeap* heap = gc_heap();
834 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
836 MemRegion mr = heap->reserved_region();
837 _ref_processor = ReferenceProcessor::create_ref_processor(
838 mr, // span
839 true, // atomic_discovery
840 true, // mt_discovery
841 &_is_alive_closure,
842 ParallelGCThreads,
843 ParallelRefProcEnabled);
844 _counters = new CollectorCounters("PSParallelCompact", 1);
846 // Initialize static fields in ParCompactionManager.
847 ParCompactionManager::initialize(mark_bitmap());
848 }
850 bool PSParallelCompact::initialize() {
851 ParallelScavengeHeap* heap = gc_heap();
852 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
853 MemRegion mr = heap->reserved_region();
855 // Was the old gen get allocated successfully?
856 if (!heap->old_gen()->is_allocated()) {
857 return false;
858 }
860 initialize_space_info();
861 initialize_dead_wood_limiter();
863 if (!_mark_bitmap.initialize(mr)) {
864 vm_shutdown_during_initialization("Unable to allocate bit map for "
865 "parallel garbage collection for the requested heap size.");
866 return false;
867 }
869 if (!_summary_data.initialize(mr)) {
870 vm_shutdown_during_initialization("Unable to allocate tables for "
871 "parallel garbage collection for the requested heap size.");
872 return false;
873 }
875 return true;
876 }
878 void PSParallelCompact::initialize_space_info()
879 {
880 memset(&_space_info, 0, sizeof(_space_info));
882 ParallelScavengeHeap* heap = gc_heap();
883 PSYoungGen* young_gen = heap->young_gen();
884 MutableSpace* perm_space = heap->perm_gen()->object_space();
886 _space_info[perm_space_id].set_space(perm_space);
887 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
888 _space_info[eden_space_id].set_space(young_gen->eden_space());
889 _space_info[from_space_id].set_space(young_gen->from_space());
890 _space_info[to_space_id].set_space(young_gen->to_space());
892 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
893 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
895 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
896 if (TraceParallelOldGCDensePrefix) {
897 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
898 _space_info[perm_space_id].min_dense_prefix());
899 }
900 }
902 void PSParallelCompact::initialize_dead_wood_limiter()
903 {
904 const size_t max = 100;
905 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
906 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
907 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
908 DEBUG_ONLY(_dwl_initialized = true;)
909 _dwl_adjustment = normal_distribution(1.0);
910 }
912 // Simple class for storing info about the heap at the start of GC, to be used
913 // after GC for comparison/printing.
914 class PreGCValues {
915 public:
916 PreGCValues() { }
917 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
919 void fill(ParallelScavengeHeap* heap) {
920 _heap_used = heap->used();
921 _young_gen_used = heap->young_gen()->used_in_bytes();
922 _old_gen_used = heap->old_gen()->used_in_bytes();
923 _perm_gen_used = heap->perm_gen()->used_in_bytes();
924 };
926 size_t heap_used() const { return _heap_used; }
927 size_t young_gen_used() const { return _young_gen_used; }
928 size_t old_gen_used() const { return _old_gen_used; }
929 size_t perm_gen_used() const { return _perm_gen_used; }
931 private:
932 size_t _heap_used;
933 size_t _young_gen_used;
934 size_t _old_gen_used;
935 size_t _perm_gen_used;
936 };
938 void
939 PSParallelCompact::clear_data_covering_space(SpaceId id)
940 {
941 // At this point, top is the value before GC, new_top() is the value that will
942 // be set at the end of GC. The marking bitmap is cleared to top; nothing
943 // should be marked above top. The summary data is cleared to the larger of
944 // top & new_top.
945 MutableSpace* const space = _space_info[id].space();
946 HeapWord* const bot = space->bottom();
947 HeapWord* const top = space->top();
948 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
950 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
951 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
952 _mark_bitmap.clear_range(beg_bit, end_bit);
954 const size_t beg_chunk = _summary_data.addr_to_chunk_idx(bot);
955 const size_t end_chunk =
956 _summary_data.addr_to_chunk_idx(_summary_data.chunk_align_up(max_top));
957 _summary_data.clear_range(beg_chunk, end_chunk);
958 }
960 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
961 {
962 // Update the from & to space pointers in space_info, since they are swapped
963 // at each young gen gc. Do the update unconditionally (even though a
964 // promotion failure does not swap spaces) because an unknown number of minor
965 // collections will have swapped the spaces an unknown number of times.
966 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
967 ParallelScavengeHeap* heap = gc_heap();
968 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
969 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
971 pre_gc_values->fill(heap);
973 ParCompactionManager::reset();
974 NOT_PRODUCT(_mark_bitmap.reset_counters());
975 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
976 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
978 // Increment the invocation count
979 heap->increment_total_collections(true);
981 // We need to track unique mark sweep invocations as well.
982 _total_invocations++;
984 if (PrintHeapAtGC) {
985 Universe::print_heap_before_gc();
986 }
988 // Fill in TLABs
989 heap->accumulate_statistics_all_tlabs();
990 heap->ensure_parsability(true); // retire TLABs
992 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
993 HandleMark hm; // Discard invalid handles created during verification
994 gclog_or_tty->print(" VerifyBeforeGC:");
995 Universe::verify(true);
996 }
998 // Verify object start arrays
999 if (VerifyObjectStartArray &&
1000 VerifyBeforeGC) {
1001 heap->old_gen()->verify_object_start_array();
1002 heap->perm_gen()->verify_object_start_array();
1003 }
1005 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1006 DEBUG_ONLY(summary_data().verify_clear();)
1007 }
1009 void PSParallelCompact::post_compact()
1010 {
1011 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
1013 // Clear the marking bitmap and summary data and update top() in each space.
1014 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1015 clear_data_covering_space(SpaceId(id));
1016 _space_info[id].space()->set_top(_space_info[id].new_top());
1017 }
1019 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1020 MutableSpace* const from_space = _space_info[from_space_id].space();
1021 MutableSpace* const to_space = _space_info[to_space_id].space();
1023 ParallelScavengeHeap* heap = gc_heap();
1024 bool eden_empty = eden_space->is_empty();
1025 if (!eden_empty) {
1026 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1027 heap->young_gen(), heap->old_gen());
1028 }
1030 // Update heap occupancy information which is used as input to the soft ref
1031 // clearing policy at the next gc.
1032 Universe::update_heap_info_at_gc();
1034 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1035 to_space->is_empty();
1037 BarrierSet* bs = heap->barrier_set();
1038 if (bs->is_a(BarrierSet::ModRef)) {
1039 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1040 MemRegion old_mr = heap->old_gen()->reserved();
1041 MemRegion perm_mr = heap->perm_gen()->reserved();
1042 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1044 if (young_gen_empty) {
1045 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1046 } else {
1047 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1048 }
1049 }
1051 Threads::gc_epilogue();
1052 CodeCache::gc_epilogue();
1054 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1056 ref_processor()->enqueue_discovered_references(NULL);
1058 // Update time of last GC
1059 reset_millis_since_last_gc();
1060 }
1062 HeapWord*
1063 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1064 bool maximum_compaction)
1065 {
1066 const size_t chunk_size = ParallelCompactData::ChunkSize;
1067 const ParallelCompactData& sd = summary_data();
1069 const MutableSpace* const space = _space_info[id].space();
1070 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top());
1071 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(space->bottom());
1072 const ChunkData* const end_cp = sd.addr_to_chunk_ptr(top_aligned_up);
1074 // Skip full chunks at the beginning of the space--they are necessarily part
1075 // of the dense prefix.
1076 size_t full_count = 0;
1077 const ChunkData* cp;
1078 for (cp = beg_cp; cp < end_cp && cp->data_size() == chunk_size; ++cp) {
1079 ++full_count;
1080 }
1082 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1083 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1084 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1085 if (maximum_compaction || cp == end_cp || interval_ended) {
1086 _maximum_compaction_gc_num = total_invocations();
1087 return sd.chunk_to_addr(cp);
1088 }
1090 HeapWord* const new_top = _space_info[id].new_top();
1091 const size_t space_live = pointer_delta(new_top, space->bottom());
1092 const size_t space_used = space->used_in_words();
1093 const size_t space_capacity = space->capacity_in_words();
1095 const double cur_density = double(space_live) / space_capacity;
1096 const double deadwood_density =
1097 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1098 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1100 if (TraceParallelOldGCDensePrefix) {
1101 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1102 cur_density, deadwood_density, deadwood_goal);
1103 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1104 "space_cap=" SIZE_FORMAT,
1105 space_live, space_used,
1106 space_capacity);
1107 }
1109 // XXX - Use binary search?
1110 HeapWord* dense_prefix = sd.chunk_to_addr(cp);
1111 const ChunkData* full_cp = cp;
1112 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(space->top() - 1);
1113 while (cp < end_cp) {
1114 HeapWord* chunk_destination = cp->destination();
1115 const size_t cur_deadwood = pointer_delta(dense_prefix, chunk_destination);
1116 if (TraceParallelOldGCDensePrefix && Verbose) {
1117 tty->print_cr("c#=" SIZE_FORMAT_W("04") " dst=" PTR_FORMAT " "
1118 "dp=" SIZE_FORMAT_W("08") " " "cdw=" SIZE_FORMAT_W("08"),
1119 sd.chunk(cp), chunk_destination,
1120 dense_prefix, cur_deadwood);
1121 }
1123 if (cur_deadwood >= deadwood_goal) {
1124 // Found the chunk that has the correct amount of deadwood to the left.
1125 // This typically occurs after crossing a fairly sparse set of chunks, so
1126 // iterate backwards over those sparse chunks, looking for the chunk that
1127 // has the lowest density of live objects 'to the right.'
1128 size_t space_to_left = sd.chunk(cp) * chunk_size;
1129 size_t live_to_left = space_to_left - cur_deadwood;
1130 size_t space_to_right = space_capacity - space_to_left;
1131 size_t live_to_right = space_live - live_to_left;
1132 double density_to_right = double(live_to_right) / space_to_right;
1133 while (cp > full_cp) {
1134 --cp;
1135 const size_t prev_chunk_live_to_right = live_to_right - cp->data_size();
1136 const size_t prev_chunk_space_to_right = space_to_right + chunk_size;
1137 double prev_chunk_density_to_right =
1138 double(prev_chunk_live_to_right) / prev_chunk_space_to_right;
1139 if (density_to_right <= prev_chunk_density_to_right) {
1140 return dense_prefix;
1141 }
1142 if (TraceParallelOldGCDensePrefix && Verbose) {
1143 tty->print_cr("backing up from c=" SIZE_FORMAT_W("4") " d2r=%10.8f "
1144 "pc_d2r=%10.8f", sd.chunk(cp), density_to_right,
1145 prev_chunk_density_to_right);
1146 }
1147 dense_prefix -= chunk_size;
1148 live_to_right = prev_chunk_live_to_right;
1149 space_to_right = prev_chunk_space_to_right;
1150 density_to_right = prev_chunk_density_to_right;
1151 }
1152 return dense_prefix;
1153 }
1155 dense_prefix += chunk_size;
1156 ++cp;
1157 }
1159 return dense_prefix;
1160 }
1162 #ifndef PRODUCT
1163 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1164 const SpaceId id,
1165 const bool maximum_compaction,
1166 HeapWord* const addr)
1167 {
1168 const size_t chunk_idx = summary_data().addr_to_chunk_idx(addr);
1169 ChunkData* const cp = summary_data().chunk(chunk_idx);
1170 const MutableSpace* const space = _space_info[id].space();
1171 HeapWord* const new_top = _space_info[id].new_top();
1173 const size_t space_live = pointer_delta(new_top, space->bottom());
1174 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1175 const size_t space_cap = space->capacity_in_words();
1176 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1177 const size_t live_to_right = new_top - cp->destination();
1178 const size_t dead_to_right = space->top() - addr - live_to_right;
1180 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W("05") " "
1181 "spl=" SIZE_FORMAT " "
1182 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1183 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1184 " ratio=%10.8f",
1185 algorithm, addr, chunk_idx,
1186 space_live,
1187 dead_to_left, dead_to_left_pct,
1188 dead_to_right, live_to_right,
1189 double(dead_to_right) / live_to_right);
1190 }
1191 #endif // #ifndef PRODUCT
1193 // Return a fraction indicating how much of the generation can be treated as
1194 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1195 // based on the density of live objects in the generation to determine a limit,
1196 // which is then adjusted so the return value is min_percent when the density is
1197 // 1.
1198 //
1199 // The following table shows some return values for a different values of the
1200 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1201 // min_percent is 1.
1202 //
1203 // fraction allowed as dead wood
1204 // -----------------------------------------------------------------
1205 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1206 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1207 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1208 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1209 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1210 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1211 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1212 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1213 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1214 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1215 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1216 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1217 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1218 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1219 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1220 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1221 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1222 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1223 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1224 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1225 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1226 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1227 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1229 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1230 {
1231 assert(_dwl_initialized, "uninitialized");
1233 // The raw limit is the value of the normal distribution at x = density.
1234 const double raw_limit = normal_distribution(density);
1236 // Adjust the raw limit so it becomes the minimum when the density is 1.
1237 //
1238 // First subtract the adjustment value (which is simply the precomputed value
1239 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1240 // Then add the minimum value, so the minimum is returned when the density is
1241 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1242 const double min = double(min_percent) / 100.0;
1243 const double limit = raw_limit - _dwl_adjustment + min;
1244 return MAX2(limit, 0.0);
1245 }
1247 ParallelCompactData::ChunkData*
1248 PSParallelCompact::first_dead_space_chunk(const ChunkData* beg,
1249 const ChunkData* end)
1250 {
1251 const size_t chunk_size = ParallelCompactData::ChunkSize;
1252 ParallelCompactData& sd = summary_data();
1253 size_t left = sd.chunk(beg);
1254 size_t right = end > beg ? sd.chunk(end) - 1 : left;
1256 // Binary search.
1257 while (left < right) {
1258 // Equivalent to (left + right) / 2, but does not overflow.
1259 const size_t middle = left + (right - left) / 2;
1260 ChunkData* const middle_ptr = sd.chunk(middle);
1261 HeapWord* const dest = middle_ptr->destination();
1262 HeapWord* const addr = sd.chunk_to_addr(middle);
1263 assert(dest != NULL, "sanity");
1264 assert(dest <= addr, "must move left");
1266 if (middle > left && dest < addr) {
1267 right = middle - 1;
1268 } else if (middle < right && middle_ptr->data_size() == chunk_size) {
1269 left = middle + 1;
1270 } else {
1271 return middle_ptr;
1272 }
1273 }
1274 return sd.chunk(left);
1275 }
1277 ParallelCompactData::ChunkData*
1278 PSParallelCompact::dead_wood_limit_chunk(const ChunkData* beg,
1279 const ChunkData* end,
1280 size_t dead_words)
1281 {
1282 ParallelCompactData& sd = summary_data();
1283 size_t left = sd.chunk(beg);
1284 size_t right = end > beg ? sd.chunk(end) - 1 : left;
1286 // Binary search.
1287 while (left < right) {
1288 // Equivalent to (left + right) / 2, but does not overflow.
1289 const size_t middle = left + (right - left) / 2;
1290 ChunkData* const middle_ptr = sd.chunk(middle);
1291 HeapWord* const dest = middle_ptr->destination();
1292 HeapWord* const addr = sd.chunk_to_addr(middle);
1293 assert(dest != NULL, "sanity");
1294 assert(dest <= addr, "must move left");
1296 const size_t dead_to_left = pointer_delta(addr, dest);
1297 if (middle > left && dead_to_left > dead_words) {
1298 right = middle - 1;
1299 } else if (middle < right && dead_to_left < dead_words) {
1300 left = middle + 1;
1301 } else {
1302 return middle_ptr;
1303 }
1304 }
1305 return sd.chunk(left);
1306 }
1308 // The result is valid during the summary phase, after the initial summarization
1309 // of each space into itself, and before final summarization.
1310 inline double
1311 PSParallelCompact::reclaimed_ratio(const ChunkData* const cp,
1312 HeapWord* const bottom,
1313 HeapWord* const top,
1314 HeapWord* const new_top)
1315 {
1316 ParallelCompactData& sd = summary_data();
1318 assert(cp != NULL, "sanity");
1319 assert(bottom != NULL, "sanity");
1320 assert(top != NULL, "sanity");
1321 assert(new_top != NULL, "sanity");
1322 assert(top >= new_top, "summary data problem?");
1323 assert(new_top > bottom, "space is empty; should not be here");
1324 assert(new_top >= cp->destination(), "sanity");
1325 assert(top >= sd.chunk_to_addr(cp), "sanity");
1327 HeapWord* const destination = cp->destination();
1328 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1329 const size_t compacted_region_live = pointer_delta(new_top, destination);
1330 const size_t compacted_region_used = pointer_delta(top, sd.chunk_to_addr(cp));
1331 const size_t reclaimable = compacted_region_used - compacted_region_live;
1333 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1334 return double(reclaimable) / divisor;
1335 }
1337 // Return the address of the end of the dense prefix, a.k.a. the start of the
1338 // compacted region. The address is always on a chunk boundary.
1339 //
1340 // Completely full chunks at the left are skipped, since no compaction can occur
1341 // in those chunks. Then the maximum amount of dead wood to allow is computed,
1342 // based on the density (amount live / capacity) of the generation; the chunk
1343 // with approximately that amount of dead space to the left is identified as the
1344 // limit chunk. Chunks between the last completely full chunk and the limit
1345 // chunk are scanned and the one that has the best (maximum) reclaimed_ratio()
1346 // is selected.
1347 HeapWord*
1348 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1349 bool maximum_compaction)
1350 {
1351 const size_t chunk_size = ParallelCompactData::ChunkSize;
1352 const ParallelCompactData& sd = summary_data();
1354 const MutableSpace* const space = _space_info[id].space();
1355 HeapWord* const top = space->top();
1356 HeapWord* const top_aligned_up = sd.chunk_align_up(top);
1357 HeapWord* const new_top = _space_info[id].new_top();
1358 HeapWord* const new_top_aligned_up = sd.chunk_align_up(new_top);
1359 HeapWord* const bottom = space->bottom();
1360 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(bottom);
1361 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up);
1362 const ChunkData* const new_top_cp = sd.addr_to_chunk_ptr(new_top_aligned_up);
1364 // Skip full chunks at the beginning of the space--they are necessarily part
1365 // of the dense prefix.
1366 const ChunkData* const full_cp = first_dead_space_chunk(beg_cp, new_top_cp);
1367 assert(full_cp->destination() == sd.chunk_to_addr(full_cp) ||
1368 space->is_empty(), "no dead space allowed to the left");
1369 assert(full_cp->data_size() < chunk_size || full_cp == new_top_cp - 1,
1370 "chunk must have dead space");
1372 // The gc number is saved whenever a maximum compaction is done, and used to
1373 // determine when the maximum compaction interval has expired. This avoids
1374 // successive max compactions for different reasons.
1375 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1376 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1377 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1378 total_invocations() == HeapFirstMaximumCompactionCount;
1379 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1380 _maximum_compaction_gc_num = total_invocations();
1381 return sd.chunk_to_addr(full_cp);
1382 }
1384 const size_t space_live = pointer_delta(new_top, bottom);
1385 const size_t space_used = space->used_in_words();
1386 const size_t space_capacity = space->capacity_in_words();
1388 const double density = double(space_live) / double(space_capacity);
1389 const size_t min_percent_free =
1390 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1391 const double limiter = dead_wood_limiter(density, min_percent_free);
1392 const size_t dead_wood_max = space_used - space_live;
1393 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1394 dead_wood_max);
1396 if (TraceParallelOldGCDensePrefix) {
1397 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1398 "space_cap=" SIZE_FORMAT,
1399 space_live, space_used,
1400 space_capacity);
1401 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1402 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1403 density, min_percent_free, limiter,
1404 dead_wood_max, dead_wood_limit);
1405 }
1407 // Locate the chunk with the desired amount of dead space to the left.
1408 const ChunkData* const limit_cp =
1409 dead_wood_limit_chunk(full_cp, top_cp, dead_wood_limit);
1411 // Scan from the first chunk with dead space to the limit chunk and find the
1412 // one with the best (largest) reclaimed ratio.
1413 double best_ratio = 0.0;
1414 const ChunkData* best_cp = full_cp;
1415 for (const ChunkData* cp = full_cp; cp < limit_cp; ++cp) {
1416 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1417 if (tmp_ratio > best_ratio) {
1418 best_cp = cp;
1419 best_ratio = tmp_ratio;
1420 }
1421 }
1423 #if 0
1424 // Something to consider: if the chunk with the best ratio is 'close to' the
1425 // first chunk w/free space, choose the first chunk with free space
1426 // ("first-free"). The first-free chunk is usually near the start of the
1427 // heap, which means we are copying most of the heap already, so copy a bit
1428 // more to get complete compaction.
1429 if (pointer_delta(best_cp, full_cp, sizeof(ChunkData)) < 4) {
1430 _maximum_compaction_gc_num = total_invocations();
1431 best_cp = full_cp;
1432 }
1433 #endif // #if 0
1435 return sd.chunk_to_addr(best_cp);
1436 }
1438 void PSParallelCompact::summarize_spaces_quick()
1439 {
1440 for (unsigned int i = 0; i < last_space_id; ++i) {
1441 const MutableSpace* space = _space_info[i].space();
1442 bool result = _summary_data.summarize(space->bottom(), space->end(),
1443 space->bottom(), space->top(),
1444 _space_info[i].new_top_addr());
1445 assert(result, "should never fail");
1446 _space_info[i].set_dense_prefix(space->bottom());
1447 }
1448 }
1450 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1451 {
1452 HeapWord* const dense_prefix_end = dense_prefix(id);
1453 const ChunkData* chunk = _summary_data.addr_to_chunk_ptr(dense_prefix_end);
1454 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1455 if (dead_space_crosses_boundary(chunk, dense_prefix_bit)) {
1456 // Only enough dead space is filled so that any remaining dead space to the
1457 // left is larger than the minimum filler object. (The remainder is filled
1458 // during the copy/update phase.)
1459 //
1460 // The size of the dead space to the right of the boundary is not a
1461 // concern, since compaction will be able to use whatever space is
1462 // available.
1463 //
1464 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1465 // surrounds the space to be filled with an object.
1466 //
1467 // In the 32-bit VM, each bit represents two 32-bit words:
1468 // +---+
1469 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1470 // end_bits: ... x x x | 0 | || 0 x x ...
1471 // +---+
1472 //
1473 // In the 64-bit VM, each bit represents one 64-bit word:
1474 // +------------+
1475 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1476 // end_bits: ... x x 1 | 0 || 0 | x x ...
1477 // +------------+
1478 // +-------+
1479 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1480 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1481 // +-------+
1482 // +-----------+
1483 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1484 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1485 // +-----------+
1486 // +-------+
1487 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1488 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1489 // +-------+
1491 // Initially assume case a, c or e will apply.
1492 size_t obj_len = (size_t)oopDesc::header_size();
1493 HeapWord* obj_beg = dense_prefix_end - obj_len;
1495 #ifdef _LP64
1496 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1497 // Case b above.
1498 obj_beg = dense_prefix_end - 1;
1499 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1500 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1501 // Case d above.
1502 obj_beg = dense_prefix_end - 3;
1503 obj_len = 3;
1504 }
1505 #endif // #ifdef _LP64
1507 MemRegion region(obj_beg, obj_len);
1508 SharedHeap::fill_region_with_object(region);
1509 _mark_bitmap.mark_obj(obj_beg, obj_len);
1510 _summary_data.add_obj(obj_beg, obj_len);
1511 assert(start_array(id) != NULL, "sanity");
1512 start_array(id)->allocate_block(obj_beg);
1513 }
1514 }
1516 void
1517 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1518 {
1519 assert(id < last_space_id, "id out of range");
1521 const MutableSpace* space = _space_info[id].space();
1522 HeapWord** new_top_addr = _space_info[id].new_top_addr();
1524 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1525 _space_info[id].set_dense_prefix(dense_prefix_end);
1527 #ifndef PRODUCT
1528 if (TraceParallelOldGCDensePrefix) {
1529 print_dense_prefix_stats("ratio", id, maximum_compaction, dense_prefix_end);
1530 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1531 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1532 }
1533 #endif // #ifndef PRODUCT
1535 // If dead space crosses the dense prefix boundary, it is (at least partially)
1536 // filled with a dummy object, marked live and added to the summary data.
1537 // This simplifies the copy/update phase and must be done before the final
1538 // locations of objects are determined, to prevent leaving a fragment of dead
1539 // space that is too small to fill with an object.
1540 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1541 fill_dense_prefix_end(id);
1542 }
1544 // Compute the destination of each Chunk, and thus each object.
1545 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1546 _summary_data.summarize(dense_prefix_end, space->end(),
1547 dense_prefix_end, space->top(),
1548 new_top_addr);
1550 if (TraceParallelOldGCSummaryPhase) {
1551 const size_t chunk_size = ParallelCompactData::ChunkSize;
1552 const size_t dp_chunk = _summary_data.addr_to_chunk_idx(dense_prefix_end);
1553 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1554 const HeapWord* nt_aligned_up = _summary_data.chunk_align_up(*new_top_addr);
1555 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1556 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1557 "dp_chunk=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1558 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1559 id, space->capacity_in_words(), dense_prefix_end,
1560 dp_chunk, dp_words / chunk_size,
1561 cr_words / chunk_size, *new_top_addr);
1562 }
1563 }
1565 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1566 bool maximum_compaction)
1567 {
1568 EventMark m("2 summarize");
1569 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1570 // trace("2");
1572 #ifdef ASSERT
1573 if (VerifyParallelOldWithMarkSweep &&
1574 (PSParallelCompact::total_invocations() %
1575 VerifyParallelOldWithMarkSweepInterval) == 0) {
1576 verify_mark_bitmap(_mark_bitmap);
1577 }
1578 if (TraceParallelOldGCMarkingPhase) {
1579 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1580 "add_obj_bytes=" SIZE_FORMAT,
1581 add_obj_count, add_obj_size * HeapWordSize);
1582 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1583 "mark_bitmap_bytes=" SIZE_FORMAT,
1584 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1585 }
1586 #endif // #ifdef ASSERT
1588 // Quick summarization of each space into itself, to see how much is live.
1589 summarize_spaces_quick();
1591 if (TraceParallelOldGCSummaryPhase) {
1592 tty->print_cr("summary_phase: after summarizing each space to self");
1593 Universe::print();
1594 NOT_PRODUCT(print_chunk_ranges());
1595 if (Verbose) {
1596 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1597 }
1598 }
1600 // The amount of live data that will end up in old space (assuming it fits).
1601 size_t old_space_total_live = 0;
1602 unsigned int id;
1603 for (id = old_space_id; id < last_space_id; ++id) {
1604 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1605 _space_info[id].space()->bottom());
1606 }
1608 const MutableSpace* old_space = _space_info[old_space_id].space();
1609 if (old_space_total_live > old_space->capacity_in_words()) {
1610 // XXX - should also try to expand
1611 maximum_compaction = true;
1612 } else if (!UseParallelOldGCDensePrefix) {
1613 maximum_compaction = true;
1614 }
1616 // Permanent and Old generations.
1617 summarize_space(perm_space_id, maximum_compaction);
1618 summarize_space(old_space_id, maximum_compaction);
1620 // Summarize the remaining spaces (those in the young gen) into old space. If
1621 // the live data from a space doesn't fit, the existing summarization is left
1622 // intact, so the data is compacted down within the space itself.
1623 HeapWord** new_top_addr = _space_info[old_space_id].new_top_addr();
1624 HeapWord* const target_space_end = old_space->end();
1625 for (id = eden_space_id; id < last_space_id; ++id) {
1626 const MutableSpace* space = _space_info[id].space();
1627 const size_t live = pointer_delta(_space_info[id].new_top(),
1628 space->bottom());
1629 const size_t available = pointer_delta(target_space_end, *new_top_addr);
1630 if (live <= available) {
1631 // All the live data will fit.
1632 if (TraceParallelOldGCSummaryPhase) {
1633 tty->print_cr("summarizing %d into old_space @ " PTR_FORMAT,
1634 id, *new_top_addr);
1635 }
1636 _summary_data.summarize(*new_top_addr, target_space_end,
1637 space->bottom(), space->top(),
1638 new_top_addr);
1640 // Reset the new_top value for the space.
1641 _space_info[id].set_new_top(space->bottom());
1643 // Clear the source_chunk field for each chunk in the space.
1644 ChunkData* beg_chunk = _summary_data.addr_to_chunk_ptr(space->bottom());
1645 ChunkData* end_chunk = _summary_data.addr_to_chunk_ptr(space->top() - 1);
1646 while (beg_chunk <= end_chunk) {
1647 beg_chunk->set_source_chunk(0);
1648 ++beg_chunk;
1649 }
1650 }
1651 }
1653 // Fill in the block data after any changes to the chunks have
1654 // been made.
1655 #ifdef ASSERT
1656 summarize_blocks(cm, perm_space_id);
1657 summarize_blocks(cm, old_space_id);
1658 #else
1659 if (!UseParallelOldGCChunkPointerCalc) {
1660 summarize_blocks(cm, perm_space_id);
1661 summarize_blocks(cm, old_space_id);
1662 }
1663 #endif
1665 if (TraceParallelOldGCSummaryPhase) {
1666 tty->print_cr("summary_phase: after final summarization");
1667 Universe::print();
1668 NOT_PRODUCT(print_chunk_ranges());
1669 if (Verbose) {
1670 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1671 }
1672 }
1673 }
1675 // Fill in the BlockData.
1676 // Iterate over the spaces and within each space iterate over
1677 // the chunks and fill in the BlockData for each chunk.
1679 void PSParallelCompact::summarize_blocks(ParCompactionManager* cm,
1680 SpaceId first_compaction_space_id) {
1681 #if 0
1682 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(1);)
1683 for (SpaceId cur_space_id = first_compaction_space_id;
1684 cur_space_id != last_space_id;
1685 cur_space_id = next_compaction_space_id(cur_space_id)) {
1686 // Iterate over the chunks in the space
1687 size_t start_chunk_index =
1688 _summary_data.addr_to_chunk_idx(space(cur_space_id)->bottom());
1689 BitBlockUpdateClosure bbu(mark_bitmap(),
1690 cm,
1691 start_chunk_index);
1692 // Iterate over blocks.
1693 for (size_t chunk_index = start_chunk_index;
1694 chunk_index < _summary_data.chunk_count() &&
1695 _summary_data.chunk_to_addr(chunk_index) < space(cur_space_id)->top();
1696 chunk_index++) {
1698 // Reset the closure for the new chunk. Note that the closure
1699 // maintains some data that does not get reset for each chunk
1700 // so a new instance of the closure is no appropriate.
1701 bbu.reset_chunk(chunk_index);
1703 // Start the iteration with the first live object. This
1704 // may return the end of the chunk. That is acceptable since
1705 // it will properly limit the iterations.
1706 ParMarkBitMap::idx_t left_offset = mark_bitmap()->addr_to_bit(
1707 _summary_data.first_live_or_end_in_chunk(chunk_index));
1709 // End the iteration at the end of the chunk.
1710 HeapWord* chunk_addr = _summary_data.chunk_to_addr(chunk_index);
1711 HeapWord* chunk_end = chunk_addr + ParallelCompactData::ChunkSize;
1712 ParMarkBitMap::idx_t right_offset =
1713 mark_bitmap()->addr_to_bit(chunk_end);
1715 // Blocks that have not objects starting in them can be
1716 // skipped because their data will never be used.
1717 if (left_offset < right_offset) {
1719 // Iterate through the objects in the chunk.
1720 ParMarkBitMap::idx_t last_offset =
1721 mark_bitmap()->pair_iterate(&bbu, left_offset, right_offset);
1723 // If last_offset is less than right_offset, then the iterations
1724 // terminated while it was looking for an end bit. "last_offset"
1725 // is then the offset for the last start bit. In this situation
1726 // the "offset" field for the next block to the right (_cur_block + 1)
1727 // will not have been update although there may be live data
1728 // to the left of the chunk.
1730 size_t cur_block_plus_1 = bbu.cur_block() + 1;
1731 HeapWord* cur_block_plus_1_addr =
1732 _summary_data.block_to_addr(bbu.cur_block()) +
1733 ParallelCompactData::BlockSize;
1734 HeapWord* last_offset_addr = mark_bitmap()->bit_to_addr(last_offset);
1735 #if 1 // This code works. The else doesn't but should. Why does it?
1736 // The current block (cur_block()) has already been updated.
1737 // The last block that may need to be updated is either the
1738 // next block (current block + 1) or the block where the
1739 // last object starts (which can be greater than the
1740 // next block if there were no objects found in intervening
1741 // blocks).
1742 size_t last_block =
1743 MAX2(bbu.cur_block() + 1,
1744 _summary_data.addr_to_block_idx(last_offset_addr));
1745 #else
1746 // The current block has already been updated. The only block
1747 // that remains to be updated is the block where the last
1748 // object in the chunk starts.
1749 size_t last_block = _summary_data.addr_to_block_idx(last_offset_addr);
1750 #endif
1751 assert_bit_is_start(last_offset);
1752 assert((last_block == _summary_data.block_count()) ||
1753 (_summary_data.block(last_block)->raw_offset() == 0),
1754 "Should not have been set");
1755 // Is the last block still in the current chunk? If still
1756 // in this chunk, update the last block (the counting that
1757 // included the current block is meant for the offset of the last
1758 // block). If not in this chunk, do nothing. Should not
1759 // update a block in the next chunk.
1760 if (ParallelCompactData::chunk_contains_block(bbu.chunk_index(),
1761 last_block)) {
1762 if (last_offset < right_offset) {
1763 // The last object started in this chunk but ends beyond
1764 // this chunk. Update the block for this last object.
1765 assert(mark_bitmap()->is_marked(last_offset), "Should be marked");
1766 // No end bit was found. The closure takes care of
1767 // the cases where
1768 // an objects crosses over into the next block
1769 // an objects starts and ends in the next block
1770 // It does not handle the case where an object is
1771 // the first object in a later block and extends
1772 // past the end of the chunk (i.e., the closure
1773 // only handles complete objects that are in the range
1774 // it is given). That object is handed back here
1775 // for any special consideration necessary.
1776 //
1777 // Is the first bit in the last block a start or end bit?
1778 //
1779 // If the partial object ends in the last block L,
1780 // then the 1st bit in L may be an end bit.
1781 //
1782 // Else does the last object start in a block after the current
1783 // block? A block AA will already have been updated if an
1784 // object ends in the next block AA+1. An object found to end in
1785 // the AA+1 is the trigger that updates AA. Objects are being
1786 // counted in the current block for updaing a following
1787 // block. An object may start in later block
1788 // block but may extend beyond the last block in the chunk.
1789 // Updates are only done when the end of an object has been
1790 // found. If the last object (covered by block L) starts
1791 // beyond the current block, then no object ends in L (otherwise
1792 // L would be the current block). So the first bit in L is
1793 // a start bit.
1794 //
1795 // Else the last objects start in the current block and ends
1796 // beyond the chunk. The current block has already been
1797 // updated and there is no later block (with an object
1798 // starting in it) that needs to be updated.
1799 //
1800 if (_summary_data.partial_obj_ends_in_block(last_block)) {
1801 _summary_data.block(last_block)->set_end_bit_offset(
1802 bbu.live_data_left());
1803 } else if (last_offset_addr >= cur_block_plus_1_addr) {
1804 // The start of the object is on a later block
1805 // (to the right of the current block and there are no
1806 // complete live objects to the left of this last object
1807 // within the chunk.
1808 // The first bit in the block is for the start of the
1809 // last object.
1810 _summary_data.block(last_block)->set_start_bit_offset(
1811 bbu.live_data_left());
1812 } else {
1813 // The start of the last object was found in
1814 // the current chunk (which has already
1815 // been updated).
1816 assert(bbu.cur_block() ==
1817 _summary_data.addr_to_block_idx(last_offset_addr),
1818 "Should be a block already processed");
1819 }
1820 #ifdef ASSERT
1821 // Is there enough block information to find this object?
1822 // The destination of the chunk has not been set so the
1823 // values returned by calc_new_pointer() and
1824 // block_calc_new_pointer() will only be
1825 // offsets. But they should agree.
1826 HeapWord* moved_obj_with_chunks =
1827 _summary_data.chunk_calc_new_pointer(last_offset_addr);
1828 HeapWord* moved_obj_with_blocks =
1829 _summary_data.calc_new_pointer(last_offset_addr);
1830 assert(moved_obj_with_chunks == moved_obj_with_blocks,
1831 "Block calculation is wrong");
1832 #endif
1833 } else if (last_block < _summary_data.block_count()) {
1834 // Iterations ended looking for a start bit (but
1835 // did not run off the end of the block table).
1836 _summary_data.block(last_block)->set_start_bit_offset(
1837 bbu.live_data_left());
1838 }
1839 }
1840 #ifdef ASSERT
1841 // Is there enough block information to find this object?
1842 HeapWord* left_offset_addr = mark_bitmap()->bit_to_addr(left_offset);
1843 HeapWord* moved_obj_with_chunks =
1844 _summary_data.calc_new_pointer(left_offset_addr);
1845 HeapWord* moved_obj_with_blocks =
1846 _summary_data.calc_new_pointer(left_offset_addr);
1847 assert(moved_obj_with_chunks == moved_obj_with_blocks,
1848 "Block calculation is wrong");
1849 #endif
1851 // Is there another block after the end of this chunk?
1852 #ifdef ASSERT
1853 if (last_block < _summary_data.block_count()) {
1854 // No object may have been found in a block. If that
1855 // block is at the end of the chunk, the iteration will
1856 // terminate without incrementing the current block so
1857 // that the current block is not the last block in the
1858 // chunk. That situation precludes asserting that the
1859 // current block is the last block in the chunk. Assert
1860 // the lesser condition that the current block does not
1861 // exceed the chunk.
1862 assert(_summary_data.block_to_addr(last_block) <=
1863 (_summary_data.chunk_to_addr(chunk_index) +
1864 ParallelCompactData::ChunkSize),
1865 "Chunk and block inconsistency");
1866 assert(last_offset <= right_offset, "Iteration over ran end");
1867 }
1868 #endif
1869 }
1870 #ifdef ASSERT
1871 if (PrintGCDetails && Verbose) {
1872 if (_summary_data.chunk(chunk_index)->partial_obj_size() == 1) {
1873 size_t first_block =
1874 chunk_index / ParallelCompactData::BlocksPerChunk;
1875 gclog_or_tty->print_cr("first_block " PTR_FORMAT
1876 " _offset " PTR_FORMAT
1877 "_first_is_start_bit %d",
1878 first_block,
1879 _summary_data.block(first_block)->raw_offset(),
1880 _summary_data.block(first_block)->first_is_start_bit());
1881 }
1882 }
1883 #endif
1884 }
1885 }
1886 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(16);)
1887 #endif // #if 0
1888 }
1890 // This method should contain all heap-specific policy for invoking a full
1891 // collection. invoke_no_policy() will only attempt to compact the heap; it
1892 // will do nothing further. If we need to bail out for policy reasons, scavenge
1893 // before full gc, or any other specialized behavior, it needs to be added here.
1894 //
1895 // Note that this method should only be called from the vm_thread while at a
1896 // safepoint.
1897 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1898 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1899 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1900 "should be in vm thread");
1901 ParallelScavengeHeap* heap = gc_heap();
1902 GCCause::Cause gc_cause = heap->gc_cause();
1903 assert(!heap->is_gc_active(), "not reentrant");
1905 PSAdaptiveSizePolicy* policy = heap->size_policy();
1907 // Before each allocation/collection attempt, find out from the
1908 // policy object if GCs are, on the whole, taking too long. If so,
1909 // bail out without attempting a collection. The exceptions are
1910 // for explicitly requested GC's.
1911 if (!policy->gc_time_limit_exceeded() ||
1912 GCCause::is_user_requested_gc(gc_cause) ||
1913 GCCause::is_serviceability_requested_gc(gc_cause)) {
1914 IsGCActiveMark mark;
1916 if (ScavengeBeforeFullGC) {
1917 PSScavenge::invoke_no_policy();
1918 }
1920 PSParallelCompact::invoke_no_policy(maximum_heap_compaction);
1921 }
1922 }
1924 bool ParallelCompactData::chunk_contains(size_t chunk_index, HeapWord* addr) {
1925 size_t addr_chunk_index = addr_to_chunk_idx(addr);
1926 return chunk_index == addr_chunk_index;
1927 }
1929 bool ParallelCompactData::chunk_contains_block(size_t chunk_index,
1930 size_t block_index) {
1931 size_t first_block_in_chunk = chunk_index * BlocksPerChunk;
1932 size_t last_block_in_chunk = (chunk_index + 1) * BlocksPerChunk - 1;
1934 return (first_block_in_chunk <= block_index) &&
1935 (block_index <= last_block_in_chunk);
1936 }
1938 // This method contains no policy. You should probably
1939 // be calling invoke() instead.
1940 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1941 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1942 assert(ref_processor() != NULL, "Sanity");
1944 if (GC_locker::check_active_before_gc()) {
1945 return;
1946 }
1948 TimeStamp marking_start;
1949 TimeStamp compaction_start;
1950 TimeStamp collection_exit;
1952 // "serial_CM" is needed until the parallel implementation
1953 // of the move and update is done.
1954 ParCompactionManager* serial_CM = new ParCompactionManager();
1955 // Don't initialize more than once.
1956 // serial_CM->initialize(&summary_data(), mark_bitmap());
1958 ParallelScavengeHeap* heap = gc_heap();
1959 GCCause::Cause gc_cause = heap->gc_cause();
1960 PSYoungGen* young_gen = heap->young_gen();
1961 PSOldGen* old_gen = heap->old_gen();
1962 PSPermGen* perm_gen = heap->perm_gen();
1963 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1965 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
1967 // Make sure data structures are sane, make the heap parsable, and do other
1968 // miscellaneous bookkeeping.
1969 PreGCValues pre_gc_values;
1970 pre_compact(&pre_gc_values);
1972 // Place after pre_compact() where the number of invocations is incremented.
1973 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
1975 {
1976 ResourceMark rm;
1977 HandleMark hm;
1979 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
1981 // This is useful for debugging but don't change the output the
1982 // the customer sees.
1983 const char* gc_cause_str = "Full GC";
1984 if (is_system_gc && PrintGCDetails) {
1985 gc_cause_str = "Full GC (System)";
1986 }
1987 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1988 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1989 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
1990 TraceCollectorStats tcs(counters());
1991 TraceMemoryManagerStats tms(true /* Full GC */);
1993 if (TraceGen1Time) accumulated_time()->start();
1995 // Let the size policy know we're starting
1996 size_policy->major_collection_begin();
1998 // When collecting the permanent generation methodOops may be moving,
1999 // so we either have to flush all bcp data or convert it into bci.
2000 CodeCache::gc_prologue();
2001 Threads::gc_prologue();
2003 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2004 COMPILER2_PRESENT(DerivedPointerTable::clear());
2006 ref_processor()->enable_discovery();
2008 bool marked_for_unloading = false;
2010 marking_start.update();
2011 marking_phase(serial_CM, maximum_heap_compaction);
2013 #ifndef PRODUCT
2014 if (TraceParallelOldGCMarkingPhase) {
2015 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2016 "cas_by_another %d",
2017 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2018 mark_bitmap()->cas_by_another());
2019 }
2020 #endif // #ifndef PRODUCT
2022 #ifdef ASSERT
2023 if (VerifyParallelOldWithMarkSweep &&
2024 (PSParallelCompact::total_invocations() %
2025 VerifyParallelOldWithMarkSweepInterval) == 0) {
2026 gclog_or_tty->print_cr("Verify marking with mark_sweep_phase1()");
2027 if (PrintGCDetails && Verbose) {
2028 gclog_or_tty->print_cr("mark_sweep_phase1:");
2029 }
2030 // Clear the discovered lists so that discovered objects
2031 // don't look like they have been discovered twice.
2032 ref_processor()->clear_discovered_references();
2034 PSMarkSweep::allocate_stacks();
2035 MemRegion mr = Universe::heap()->reserved_region();
2036 PSMarkSweep::ref_processor()->enable_discovery();
2037 PSMarkSweep::mark_sweep_phase1(maximum_heap_compaction);
2038 }
2039 #endif
2041 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
2042 summary_phase(serial_CM, maximum_heap_compaction || max_on_system_gc);
2044 #ifdef ASSERT
2045 if (VerifyParallelOldWithMarkSweep &&
2046 (PSParallelCompact::total_invocations() %
2047 VerifyParallelOldWithMarkSweepInterval) == 0) {
2048 if (PrintGCDetails && Verbose) {
2049 gclog_or_tty->print_cr("mark_sweep_phase2:");
2050 }
2051 PSMarkSweep::mark_sweep_phase2();
2052 }
2053 #endif
2055 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2056 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2058 // adjust_roots() updates Universe::_intArrayKlassObj which is
2059 // needed by the compaction for filling holes in the dense prefix.
2060 adjust_roots();
2062 #ifdef ASSERT
2063 if (VerifyParallelOldWithMarkSweep &&
2064 (PSParallelCompact::total_invocations() %
2065 VerifyParallelOldWithMarkSweepInterval) == 0) {
2066 // Do a separate verify phase so that the verify
2067 // code can use the the forwarding pointers to
2068 // check the new pointer calculation. The restore_marks()
2069 // has to be done before the real compact.
2070 serial_CM->set_action(ParCompactionManager::VerifyUpdate);
2071 compact_perm(serial_CM);
2072 compact_serial(serial_CM);
2073 serial_CM->set_action(ParCompactionManager::ResetObjects);
2074 compact_perm(serial_CM);
2075 compact_serial(serial_CM);
2076 serial_CM->set_action(ParCompactionManager::UpdateAndCopy);
2078 // For debugging only
2079 PSMarkSweep::restore_marks();
2080 PSMarkSweep::deallocate_stacks();
2081 }
2082 #endif
2084 compaction_start.update();
2085 // Does the perm gen always have to be done serially because
2086 // klasses are used in the update of an object?
2087 compact_perm(serial_CM);
2089 if (UseParallelOldGCCompacting) {
2090 compact();
2091 } else {
2092 compact_serial(serial_CM);
2093 }
2095 delete serial_CM;
2097 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2098 // done before resizing.
2099 post_compact();
2101 // Let the size policy know we're done
2102 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2104 if (UseAdaptiveSizePolicy) {
2105 if (PrintAdaptiveSizePolicy) {
2106 gclog_or_tty->print("AdaptiveSizeStart: ");
2107 gclog_or_tty->stamp();
2108 gclog_or_tty->print_cr(" collection: %d ",
2109 heap->total_collections());
2110 if (Verbose) {
2111 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2112 " perm_gen_capacity: %d ",
2113 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2114 perm_gen->capacity_in_bytes());
2115 }
2116 }
2118 // Don't check if the size_policy is ready here. Let
2119 // the size_policy check that internally.
2120 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2121 ((gc_cause != GCCause::_java_lang_system_gc) ||
2122 UseAdaptiveSizePolicyWithSystemGC)) {
2123 // Calculate optimal free space amounts
2124 assert(young_gen->max_size() >
2125 young_gen->from_space()->capacity_in_bytes() +
2126 young_gen->to_space()->capacity_in_bytes(),
2127 "Sizes of space in young gen are out-of-bounds");
2128 size_t max_eden_size = young_gen->max_size() -
2129 young_gen->from_space()->capacity_in_bytes() -
2130 young_gen->to_space()->capacity_in_bytes();
2131 size_policy->compute_generation_free_space(young_gen->used_in_bytes(),
2132 young_gen->eden_space()->used_in_bytes(),
2133 old_gen->used_in_bytes(),
2134 perm_gen->used_in_bytes(),
2135 young_gen->eden_space()->capacity_in_bytes(),
2136 old_gen->max_gen_size(),
2137 max_eden_size,
2138 true /* full gc*/,
2139 gc_cause);
2141 heap->resize_old_gen(size_policy->calculated_old_free_size_in_bytes());
2143 // Don't resize the young generation at an major collection. A
2144 // desired young generation size may have been calculated but
2145 // resizing the young generation complicates the code because the
2146 // resizing of the old generation may have moved the boundary
2147 // between the young generation and the old generation. Let the
2148 // young generation resizing happen at the minor collections.
2149 }
2150 if (PrintAdaptiveSizePolicy) {
2151 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2152 heap->total_collections());
2153 }
2154 }
2156 if (UsePerfData) {
2157 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2158 counters->update_counters();
2159 counters->update_old_capacity(old_gen->capacity_in_bytes());
2160 counters->update_young_capacity(young_gen->capacity_in_bytes());
2161 }
2163 heap->resize_all_tlabs();
2165 // We collected the perm gen, so we'll resize it here.
2166 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2168 if (TraceGen1Time) accumulated_time()->stop();
2170 if (PrintGC) {
2171 if (PrintGCDetails) {
2172 // No GC timestamp here. This is after GC so it would be confusing.
2173 young_gen->print_used_change(pre_gc_values.young_gen_used());
2174 old_gen->print_used_change(pre_gc_values.old_gen_used());
2175 heap->print_heap_change(pre_gc_values.heap_used());
2176 // Print perm gen last (print_heap_change() excludes the perm gen).
2177 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2178 } else {
2179 heap->print_heap_change(pre_gc_values.heap_used());
2180 }
2181 }
2183 // Track memory usage and detect low memory
2184 MemoryService::track_memory_usage();
2185 heap->update_counters();
2187 if (PrintGCDetails) {
2188 if (size_policy->print_gc_time_limit_would_be_exceeded()) {
2189 if (size_policy->gc_time_limit_exceeded()) {
2190 gclog_or_tty->print_cr(" GC time is exceeding GCTimeLimit "
2191 "of %d%%", GCTimeLimit);
2192 } else {
2193 gclog_or_tty->print_cr(" GC time would exceed GCTimeLimit "
2194 "of %d%%", GCTimeLimit);
2195 }
2196 }
2197 size_policy->set_print_gc_time_limit_would_be_exceeded(false);
2198 }
2199 }
2201 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2202 HandleMark hm; // Discard invalid handles created during verification
2203 gclog_or_tty->print(" VerifyAfterGC:");
2204 Universe::verify(false);
2205 }
2207 // Re-verify object start arrays
2208 if (VerifyObjectStartArray &&
2209 VerifyAfterGC) {
2210 old_gen->verify_object_start_array();
2211 perm_gen->verify_object_start_array();
2212 }
2214 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2216 collection_exit.update();
2218 if (PrintHeapAtGC) {
2219 Universe::print_heap_after_gc();
2220 }
2221 if (PrintGCTaskTimeStamps) {
2222 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2223 INT64_FORMAT,
2224 marking_start.ticks(), compaction_start.ticks(),
2225 collection_exit.ticks());
2226 gc_task_manager()->print_task_time_stamps();
2227 }
2228 }
2230 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2231 PSYoungGen* young_gen,
2232 PSOldGen* old_gen) {
2233 MutableSpace* const eden_space = young_gen->eden_space();
2234 assert(!eden_space->is_empty(), "eden must be non-empty");
2235 assert(young_gen->virtual_space()->alignment() ==
2236 old_gen->virtual_space()->alignment(), "alignments do not match");
2238 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2239 return false;
2240 }
2242 // Both generations must be completely committed.
2243 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2244 return false;
2245 }
2246 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2247 return false;
2248 }
2250 // Figure out how much to take from eden. Include the average amount promoted
2251 // in the total; otherwise the next young gen GC will simply bail out to a
2252 // full GC.
2253 const size_t alignment = old_gen->virtual_space()->alignment();
2254 const size_t eden_used = eden_space->used_in_bytes();
2255 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2256 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2257 const size_t eden_capacity = eden_space->capacity_in_bytes();
2259 if (absorb_size >= eden_capacity) {
2260 return false; // Must leave some space in eden.
2261 }
2263 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2264 if (new_young_size < young_gen->min_gen_size()) {
2265 return false; // Respect young gen minimum size.
2266 }
2268 if (TraceAdaptiveGCBoundary && Verbose) {
2269 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2270 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2271 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2272 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2273 absorb_size / K,
2274 eden_capacity / K, (eden_capacity - absorb_size) / K,
2275 young_gen->from_space()->used_in_bytes() / K,
2276 young_gen->to_space()->used_in_bytes() / K,
2277 young_gen->capacity_in_bytes() / K, new_young_size / K);
2278 }
2280 // Fill the unused part of the old gen.
2281 MutableSpace* const old_space = old_gen->object_space();
2282 MemRegion old_gen_unused(old_space->top(), old_space->end());
2283 if (!old_gen_unused.is_empty()) {
2284 SharedHeap::fill_region_with_object(old_gen_unused);
2285 }
2287 // Take the live data from eden and set both top and end in the old gen to
2288 // eden top. (Need to set end because reset_after_change() mangles the region
2289 // from end to virtual_space->high() in debug builds).
2290 HeapWord* const new_top = eden_space->top();
2291 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2292 absorb_size);
2293 young_gen->reset_after_change();
2294 old_space->set_top(new_top);
2295 old_space->set_end(new_top);
2296 old_gen->reset_after_change();
2298 // Update the object start array for the filler object and the data from eden.
2299 ObjectStartArray* const start_array = old_gen->start_array();
2300 HeapWord* const start = old_gen_unused.start();
2301 for (HeapWord* addr = start; addr < new_top; addr += oop(addr)->size()) {
2302 start_array->allocate_block(addr);
2303 }
2305 // Could update the promoted average here, but it is not typically updated at
2306 // full GCs and the value to use is unclear. Something like
2307 //
2308 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2310 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2311 return true;
2312 }
2314 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2315 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2316 "shouldn't return NULL");
2317 return ParallelScavengeHeap::gc_task_manager();
2318 }
2320 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2321 bool maximum_heap_compaction) {
2322 // Recursively traverse all live objects and mark them
2323 EventMark m("1 mark object");
2324 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2326 ParallelScavengeHeap* heap = gc_heap();
2327 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2328 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array();
2329 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2331 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2332 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2334 {
2335 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2337 GCTaskQueue* q = GCTaskQueue::create();
2339 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2340 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2341 // We scan the thread roots in parallel
2342 Threads::create_thread_roots_marking_tasks(q);
2343 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2344 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2345 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2346 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2347 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2348 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols));
2350 if (parallel_gc_threads > 1) {
2351 for (uint j = 0; j < parallel_gc_threads; j++) {
2352 q->enqueue(new StealMarkingTask(&terminator));
2353 }
2354 }
2356 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2357 q->enqueue(fin);
2359 gc_task_manager()->add_list(q);
2361 fin->wait_for();
2363 // We have to release the barrier tasks!
2364 WaitForBarrierGCTask::destroy(fin);
2365 }
2367 // Process reference objects found during marking
2368 {
2369 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2370 ReferencePolicy *soft_ref_policy;
2371 if (maximum_heap_compaction) {
2372 soft_ref_policy = new AlwaysClearPolicy();
2373 } else {
2374 #ifdef COMPILER2
2375 soft_ref_policy = new LRUMaxHeapPolicy();
2376 #else
2377 soft_ref_policy = new LRUCurrentHeapPolicy();
2378 #endif // COMPILER2
2379 }
2380 assert(soft_ref_policy != NULL, "No soft reference policy");
2381 if (ref_processor()->processing_is_mt()) {
2382 RefProcTaskExecutor task_executor;
2383 ref_processor()->process_discovered_references(
2384 soft_ref_policy, is_alive_closure(), &mark_and_push_closure,
2385 &follow_stack_closure, &task_executor);
2386 } else {
2387 ref_processor()->process_discovered_references(
2388 soft_ref_policy, is_alive_closure(), &mark_and_push_closure,
2389 &follow_stack_closure, NULL);
2390 }
2391 }
2393 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2394 // Follow system dictionary roots and unload classes.
2395 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2397 // Follow code cache roots.
2398 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2399 purged_class);
2400 follow_stack(cm); // Flush marking stack.
2402 // Update subklass/sibling/implementor links of live klasses
2403 // revisit_klass_stack is used in follow_weak_klass_links().
2404 follow_weak_klass_links(cm);
2406 // Visit symbol and interned string tables and delete unmarked oops
2407 SymbolTable::unlink(is_alive_closure());
2408 StringTable::unlink(is_alive_closure());
2410 assert(cm->marking_stack()->size() == 0, "stack should be empty by now");
2411 assert(cm->overflow_stack()->is_empty(), "stack should be empty by now");
2412 }
2414 // This should be moved to the shared markSweep code!
2415 class PSAlwaysTrueClosure: public BoolObjectClosure {
2416 public:
2417 void do_object(oop p) { ShouldNotReachHere(); }
2418 bool do_object_b(oop p) { return true; }
2419 };
2420 static PSAlwaysTrueClosure always_true;
2422 void PSParallelCompact::adjust_roots() {
2423 // Adjust the pointers to reflect the new locations
2424 EventMark m("3 adjust roots");
2425 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2427 // General strong roots.
2428 Universe::oops_do(adjust_root_pointer_closure());
2429 ReferenceProcessor::oops_do(adjust_root_pointer_closure());
2430 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
2431 Threads::oops_do(adjust_root_pointer_closure());
2432 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2433 FlatProfiler::oops_do(adjust_root_pointer_closure());
2434 Management::oops_do(adjust_root_pointer_closure());
2435 JvmtiExport::oops_do(adjust_root_pointer_closure());
2436 // SO_AllClasses
2437 SystemDictionary::oops_do(adjust_root_pointer_closure());
2438 vmSymbols::oops_do(adjust_root_pointer_closure());
2440 // Now adjust pointers in remaining weak roots. (All of which should
2441 // have been cleared if they pointed to non-surviving objects.)
2442 // Global (weak) JNI handles
2443 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2445 CodeCache::oops_do(adjust_pointer_closure());
2446 SymbolTable::oops_do(adjust_root_pointer_closure());
2447 StringTable::oops_do(adjust_root_pointer_closure());
2448 ref_processor()->weak_oops_do(adjust_root_pointer_closure());
2449 // Roots were visited so references into the young gen in roots
2450 // may have been scanned. Process them also.
2451 // Should the reference processor have a span that excludes
2452 // young gen objects?
2453 PSScavenge::reference_processor()->weak_oops_do(
2454 adjust_root_pointer_closure());
2455 }
2457 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2458 EventMark m("4 compact perm");
2459 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
2460 // trace("4");
2462 gc_heap()->perm_gen()->start_array()->reset();
2463 move_and_update(cm, perm_space_id);
2464 }
2466 void PSParallelCompact::enqueue_chunk_draining_tasks(GCTaskQueue* q,
2467 uint parallel_gc_threads) {
2468 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2470 const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
2471 for (unsigned int j = 0; j < task_count; j++) {
2472 q->enqueue(new DrainStacksCompactionTask());
2473 }
2475 // Find all chunks that are available (can be filled immediately) and
2476 // distribute them to the thread stacks. The iteration is done in reverse
2477 // order (high to low) so the chunks will be removed in ascending order.
2479 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2481 size_t fillable_chunks = 0; // A count for diagnostic purposes.
2482 unsigned int which = 0; // The worker thread number.
2484 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2485 SpaceInfo* const space_info = _space_info + id;
2486 MutableSpace* const space = space_info->space();
2487 HeapWord* const new_top = space_info->new_top();
2489 const size_t beg_chunk = sd.addr_to_chunk_idx(space_info->dense_prefix());
2490 const size_t end_chunk = sd.addr_to_chunk_idx(sd.chunk_align_up(new_top));
2491 assert(end_chunk > 0, "perm gen cannot be empty");
2493 for (size_t cur = end_chunk - 1; cur >= beg_chunk; --cur) {
2494 if (sd.chunk(cur)->claim_unsafe()) {
2495 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2496 cm->save_for_processing(cur);
2498 if (TraceParallelOldGCCompactionPhase && Verbose) {
2499 const size_t count_mod_8 = fillable_chunks & 7;
2500 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2501 gclog_or_tty->print(" " SIZE_FORMAT_W("7"), cur);
2502 if (count_mod_8 == 7) gclog_or_tty->cr();
2503 }
2505 NOT_PRODUCT(++fillable_chunks;)
2507 // Assign chunks to threads in round-robin fashion.
2508 if (++which == task_count) {
2509 which = 0;
2510 }
2511 }
2512 }
2513 }
2515 if (TraceParallelOldGCCompactionPhase) {
2516 if (Verbose && (fillable_chunks & 7) != 0) gclog_or_tty->cr();
2517 gclog_or_tty->print_cr("%u initially fillable chunks", fillable_chunks);
2518 }
2519 }
2521 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2523 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2524 uint parallel_gc_threads) {
2525 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2527 ParallelCompactData& sd = PSParallelCompact::summary_data();
2529 // Iterate over all the spaces adding tasks for updating
2530 // chunks in the dense prefix. Assume that 1 gc thread
2531 // will work on opening the gaps and the remaining gc threads
2532 // will work on the dense prefix.
2533 SpaceId space_id = old_space_id;
2534 while (space_id != last_space_id) {
2535 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2536 const MutableSpace* const space = _space_info[space_id].space();
2538 if (dense_prefix_end == space->bottom()) {
2539 // There is no dense prefix for this space.
2540 space_id = next_compaction_space_id(space_id);
2541 continue;
2542 }
2544 // The dense prefix is before this chunk.
2545 size_t chunk_index_end_dense_prefix =
2546 sd.addr_to_chunk_idx(dense_prefix_end);
2547 ChunkData* const dense_prefix_cp = sd.chunk(chunk_index_end_dense_prefix);
2548 assert(dense_prefix_end == space->end() ||
2549 dense_prefix_cp->available() ||
2550 dense_prefix_cp->claimed(),
2551 "The chunk after the dense prefix should always be ready to fill");
2553 size_t chunk_index_start = sd.addr_to_chunk_idx(space->bottom());
2555 // Is there dense prefix work?
2556 size_t total_dense_prefix_chunks =
2557 chunk_index_end_dense_prefix - chunk_index_start;
2558 // How many chunks of the dense prefix should be given to
2559 // each thread?
2560 if (total_dense_prefix_chunks > 0) {
2561 uint tasks_for_dense_prefix = 1;
2562 if (UseParallelDensePrefixUpdate) {
2563 if (total_dense_prefix_chunks <=
2564 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2565 // Don't over partition. This assumes that
2566 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2567 // so there are not many chunks to process.
2568 tasks_for_dense_prefix = parallel_gc_threads;
2569 } else {
2570 // Over partition
2571 tasks_for_dense_prefix = parallel_gc_threads *
2572 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2573 }
2574 }
2575 size_t chunks_per_thread = total_dense_prefix_chunks /
2576 tasks_for_dense_prefix;
2577 // Give each thread at least 1 chunk.
2578 if (chunks_per_thread == 0) {
2579 chunks_per_thread = 1;
2580 }
2582 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2583 if (chunk_index_start >= chunk_index_end_dense_prefix) {
2584 break;
2585 }
2586 // chunk_index_end is not processed
2587 size_t chunk_index_end = MIN2(chunk_index_start + chunks_per_thread,
2588 chunk_index_end_dense_prefix);
2589 q->enqueue(new UpdateDensePrefixTask(
2590 space_id,
2591 chunk_index_start,
2592 chunk_index_end));
2593 chunk_index_start = chunk_index_end;
2594 }
2595 }
2596 // This gets any part of the dense prefix that did not
2597 // fit evenly.
2598 if (chunk_index_start < chunk_index_end_dense_prefix) {
2599 q->enqueue(new UpdateDensePrefixTask(
2600 space_id,
2601 chunk_index_start,
2602 chunk_index_end_dense_prefix));
2603 }
2604 space_id = next_compaction_space_id(space_id);
2605 } // End tasks for dense prefix
2606 }
2608 void PSParallelCompact::enqueue_chunk_stealing_tasks(
2609 GCTaskQueue* q,
2610 ParallelTaskTerminator* terminator_ptr,
2611 uint parallel_gc_threads) {
2612 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2614 // Once a thread has drained it's stack, it should try to steal chunks from
2615 // other threads.
2616 if (parallel_gc_threads > 1) {
2617 for (uint j = 0; j < parallel_gc_threads; j++) {
2618 q->enqueue(new StealChunkCompactionTask(terminator_ptr));
2619 }
2620 }
2621 }
2623 void PSParallelCompact::compact() {
2624 EventMark m("5 compact");
2625 // trace("5");
2626 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2628 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2629 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2630 PSOldGen* old_gen = heap->old_gen();
2631 old_gen->start_array()->reset();
2632 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2633 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array();
2634 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2636 GCTaskQueue* q = GCTaskQueue::create();
2637 enqueue_chunk_draining_tasks(q, parallel_gc_threads);
2638 enqueue_dense_prefix_tasks(q, parallel_gc_threads);
2639 enqueue_chunk_stealing_tasks(q, &terminator, parallel_gc_threads);
2641 {
2642 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2644 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2645 q->enqueue(fin);
2647 gc_task_manager()->add_list(q);
2649 fin->wait_for();
2651 // We have to release the barrier tasks!
2652 WaitForBarrierGCTask::destroy(fin);
2654 #ifdef ASSERT
2655 // Verify that all chunks have been processed before the deferred updates.
2656 // Note that perm_space_id is skipped; this type of verification is not
2657 // valid until the perm gen is compacted by chunks.
2658 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2659 verify_complete(SpaceId(id));
2660 }
2661 #endif
2662 }
2664 {
2665 // Update the deferred objects, if any. Any compaction manager can be used.
2666 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2667 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2668 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2669 update_deferred_objects(cm, SpaceId(id));
2670 }
2671 }
2672 }
2674 #ifdef ASSERT
2675 void PSParallelCompact::verify_complete(SpaceId space_id) {
2676 // All Chunks between space bottom() to new_top() should be marked as filled
2677 // and all Chunks between new_top() and top() should be available (i.e.,
2678 // should have been emptied).
2679 ParallelCompactData& sd = summary_data();
2680 SpaceInfo si = _space_info[space_id];
2681 HeapWord* new_top_addr = sd.chunk_align_up(si.new_top());
2682 HeapWord* old_top_addr = sd.chunk_align_up(si.space()->top());
2683 const size_t beg_chunk = sd.addr_to_chunk_idx(si.space()->bottom());
2684 const size_t new_top_chunk = sd.addr_to_chunk_idx(new_top_addr);
2685 const size_t old_top_chunk = sd.addr_to_chunk_idx(old_top_addr);
2687 bool issued_a_warning = false;
2689 size_t cur_chunk;
2690 for (cur_chunk = beg_chunk; cur_chunk < new_top_chunk; ++cur_chunk) {
2691 const ChunkData* const c = sd.chunk(cur_chunk);
2692 if (!c->completed()) {
2693 warning("chunk " SIZE_FORMAT " not filled: "
2694 "destination_count=" SIZE_FORMAT,
2695 cur_chunk, c->destination_count());
2696 issued_a_warning = true;
2697 }
2698 }
2700 for (cur_chunk = new_top_chunk; cur_chunk < old_top_chunk; ++cur_chunk) {
2701 const ChunkData* const c = sd.chunk(cur_chunk);
2702 if (!c->available()) {
2703 warning("chunk " SIZE_FORMAT " not empty: "
2704 "destination_count=" SIZE_FORMAT,
2705 cur_chunk, c->destination_count());
2706 issued_a_warning = true;
2707 }
2708 }
2710 if (issued_a_warning) {
2711 print_chunk_ranges();
2712 }
2713 }
2714 #endif // #ifdef ASSERT
2716 void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
2717 EventMark m("5 compact serial");
2718 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty);
2720 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2721 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2723 PSYoungGen* young_gen = heap->young_gen();
2724 PSOldGen* old_gen = heap->old_gen();
2726 old_gen->start_array()->reset();
2727 old_gen->move_and_update(cm);
2728 young_gen->move_and_update(cm);
2729 }
2732 void PSParallelCompact::follow_stack(ParCompactionManager* cm) {
2733 while(!cm->overflow_stack()->is_empty()) {
2734 oop obj = cm->overflow_stack()->pop();
2735 obj->follow_contents(cm);
2736 }
2738 oop obj;
2739 // obj is a reference!!!
2740 while (cm->marking_stack()->pop_local(obj)) {
2741 // It would be nice to assert about the type of objects we might
2742 // pop, but they can come from anywhere, unfortunately.
2743 obj->follow_contents(cm);
2744 }
2745 }
2747 void
2748 PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) {
2749 // All klasses on the revisit stack are marked at this point.
2750 // Update and follow all subklass, sibling and implementor links.
2751 for (uint i = 0; i < ParallelGCThreads+1; i++) {
2752 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2753 KeepAliveClosure keep_alive_closure(cm);
2754 for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) {
2755 cm->revisit_klass_stack()->at(i)->follow_weak_klass_links(
2756 is_alive_closure(),
2757 &keep_alive_closure);
2758 }
2759 follow_stack(cm);
2760 }
2761 }
2763 void
2764 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2765 cm->revisit_klass_stack()->push(k);
2766 }
2768 #ifdef VALIDATE_MARK_SWEEP
2770 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2771 if (!ValidateMarkSweep)
2772 return;
2774 if (!isroot) {
2775 if (_pointer_tracking) {
2776 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2777 _adjusted_pointers->remove(p);
2778 }
2779 } else {
2780 ptrdiff_t index = _root_refs_stack->find(p);
2781 if (index != -1) {
2782 int l = _root_refs_stack->length();
2783 if (l > 0 && l - 1 != index) {
2784 void* last = _root_refs_stack->pop();
2785 assert(last != p, "should be different");
2786 _root_refs_stack->at_put(index, last);
2787 } else {
2788 _root_refs_stack->remove(p);
2789 }
2790 }
2791 }
2792 }
2795 void PSParallelCompact::check_adjust_pointer(void* p) {
2796 _adjusted_pointers->push(p);
2797 }
2800 class AdjusterTracker: public OopClosure {
2801 public:
2802 AdjusterTracker() {};
2803 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
2804 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); }
2805 };
2808 void PSParallelCompact::track_interior_pointers(oop obj) {
2809 if (ValidateMarkSweep) {
2810 _adjusted_pointers->clear();
2811 _pointer_tracking = true;
2813 AdjusterTracker checker;
2814 obj->oop_iterate(&checker);
2815 }
2816 }
2819 void PSParallelCompact::check_interior_pointers() {
2820 if (ValidateMarkSweep) {
2821 _pointer_tracking = false;
2822 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2823 }
2824 }
2827 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2828 if (ValidateMarkSweep) {
2829 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2830 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2831 }
2832 }
2835 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2836 if (ValidateMarkSweep) {
2837 _live_oops->push(p);
2838 _live_oops_size->push(size);
2839 _live_oops_index++;
2840 }
2841 }
2843 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2844 if (ValidateMarkSweep) {
2845 oop obj = _live_oops->at((int)_live_oops_index);
2846 guarantee(obj == p, "should be the same object");
2847 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2848 _live_oops_index++;
2849 }
2850 }
2852 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2853 HeapWord* compaction_top) {
2854 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2855 "should be moved to forwarded location");
2856 if (ValidateMarkSweep) {
2857 PSParallelCompact::validate_live_oop(oop(q), size);
2858 _live_oops_moved_to->push(oop(compaction_top));
2859 }
2860 if (RecordMarkSweepCompaction) {
2861 _cur_gc_live_oops->push(q);
2862 _cur_gc_live_oops_moved_to->push(compaction_top);
2863 _cur_gc_live_oops_size->push(size);
2864 }
2865 }
2868 void PSParallelCompact::compaction_complete() {
2869 if (RecordMarkSweepCompaction) {
2870 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
2871 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2872 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
2874 _cur_gc_live_oops = _last_gc_live_oops;
2875 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
2876 _cur_gc_live_oops_size = _last_gc_live_oops_size;
2877 _last_gc_live_oops = _tmp_live_oops;
2878 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2879 _last_gc_live_oops_size = _tmp_live_oops_size;
2880 }
2881 }
2884 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2885 if (!RecordMarkSweepCompaction) {
2886 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2887 return;
2888 }
2890 if (_last_gc_live_oops == NULL) {
2891 tty->print_cr("No compaction information gathered yet");
2892 return;
2893 }
2895 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2896 HeapWord* old_oop = _last_gc_live_oops->at(i);
2897 size_t sz = _last_gc_live_oops_size->at(i);
2898 if (old_oop <= q && q < (old_oop + sz)) {
2899 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2900 size_t offset = (q - old_oop);
2901 tty->print_cr("Address " PTR_FORMAT, q);
2902 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2903 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2904 return;
2905 }
2906 }
2908 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2909 }
2910 #endif //VALIDATE_MARK_SWEEP
2912 // Update interior oops in the ranges of chunks [beg_chunk, end_chunk).
2913 void
2914 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2915 SpaceId space_id,
2916 size_t beg_chunk,
2917 size_t end_chunk) {
2918 ParallelCompactData& sd = summary_data();
2919 ParMarkBitMap* const mbm = mark_bitmap();
2921 HeapWord* beg_addr = sd.chunk_to_addr(beg_chunk);
2922 HeapWord* const end_addr = sd.chunk_to_addr(end_chunk);
2923 assert(beg_chunk <= end_chunk, "bad chunk range");
2924 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2926 #ifdef ASSERT
2927 // Claim the chunks to avoid triggering an assert when they are marked as
2928 // filled.
2929 for (size_t claim_chunk = beg_chunk; claim_chunk < end_chunk; ++claim_chunk) {
2930 assert(sd.chunk(claim_chunk)->claim_unsafe(), "claim() failed");
2931 }
2932 #endif // #ifdef ASSERT
2934 if (beg_addr != space(space_id)->bottom()) {
2935 // Find the first live object or block of dead space that *starts* in this
2936 // range of chunks. If a partial object crosses onto the chunk, skip it; it
2937 // will be marked for 'deferred update' when the object head is processed.
2938 // If dead space crosses onto the chunk, it is also skipped; it will be
2939 // filled when the prior chunk is processed. If neither of those apply, the
2940 // first word in the chunk is the start of a live object or dead space.
2941 assert(beg_addr > space(space_id)->bottom(), "sanity");
2942 const ChunkData* const cp = sd.chunk(beg_chunk);
2943 if (cp->partial_obj_size() != 0) {
2944 beg_addr = sd.partial_obj_end(beg_chunk);
2945 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2946 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2947 }
2948 }
2950 if (beg_addr < end_addr) {
2951 // A live object or block of dead space starts in this range of Chunks.
2952 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2954 // Create closures and iterate.
2955 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2956 FillClosure fill_closure(cm, space_id);
2957 ParMarkBitMap::IterationStatus status;
2958 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2959 dense_prefix_end);
2960 if (status == ParMarkBitMap::incomplete) {
2961 update_closure.do_addr(update_closure.source());
2962 }
2963 }
2965 // Mark the chunks as filled.
2966 ChunkData* const beg_cp = sd.chunk(beg_chunk);
2967 ChunkData* const end_cp = sd.chunk(end_chunk);
2968 for (ChunkData* cp = beg_cp; cp < end_cp; ++cp) {
2969 cp->set_completed();
2970 }
2971 }
2973 // Return the SpaceId for the space containing addr. If addr is not in the
2974 // heap, last_space_id is returned. In debug mode it expects the address to be
2975 // in the heap and asserts such.
2976 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2977 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2979 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
2980 if (_space_info[id].space()->contains(addr)) {
2981 return SpaceId(id);
2982 }
2983 }
2985 assert(false, "no space contains the addr");
2986 return last_space_id;
2987 }
2989 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2990 SpaceId id) {
2991 assert(id < last_space_id, "bad space id");
2993 ParallelCompactData& sd = summary_data();
2994 const SpaceInfo* const space_info = _space_info + id;
2995 ObjectStartArray* const start_array = space_info->start_array();
2997 const MutableSpace* const space = space_info->space();
2998 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2999 HeapWord* const beg_addr = space_info->dense_prefix();
3000 HeapWord* const end_addr = sd.chunk_align_up(space_info->new_top());
3002 const ChunkData* const beg_chunk = sd.addr_to_chunk_ptr(beg_addr);
3003 const ChunkData* const end_chunk = sd.addr_to_chunk_ptr(end_addr);
3004 const ChunkData* cur_chunk;
3005 for (cur_chunk = beg_chunk; cur_chunk < end_chunk; ++cur_chunk) {
3006 HeapWord* const addr = cur_chunk->deferred_obj_addr();
3007 if (addr != NULL) {
3008 if (start_array != NULL) {
3009 start_array->allocate_block(addr);
3010 }
3011 oop(addr)->update_contents(cm);
3012 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3013 }
3014 }
3015 }
3017 // Skip over count live words starting from beg, and return the address of the
3018 // next live word. Unless marked, the word corresponding to beg is assumed to
3019 // be dead. Callers must either ensure beg does not correspond to the middle of
3020 // an object, or account for those live words in some other way. Callers must
3021 // also ensure that there are enough live words in the range [beg, end) to skip.
3022 HeapWord*
3023 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3024 {
3025 assert(count > 0, "sanity");
3027 ParMarkBitMap* m = mark_bitmap();
3028 idx_t bits_to_skip = m->words_to_bits(count);
3029 idx_t cur_beg = m->addr_to_bit(beg);
3030 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3032 do {
3033 cur_beg = m->find_obj_beg(cur_beg, search_end);
3034 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3035 const size_t obj_bits = cur_end - cur_beg + 1;
3036 if (obj_bits > bits_to_skip) {
3037 return m->bit_to_addr(cur_beg + bits_to_skip);
3038 }
3039 bits_to_skip -= obj_bits;
3040 cur_beg = cur_end + 1;
3041 } while (bits_to_skip > 0);
3043 // Skipping the desired number of words landed just past the end of an object.
3044 // Find the start of the next object.
3045 cur_beg = m->find_obj_beg(cur_beg, search_end);
3046 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3047 return m->bit_to_addr(cur_beg);
3048 }
3050 HeapWord*
3051 PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3052 size_t src_chunk_idx)
3053 {
3054 ParMarkBitMap* const bitmap = mark_bitmap();
3055 const ParallelCompactData& sd = summary_data();
3056 const size_t ChunkSize = ParallelCompactData::ChunkSize;
3058 assert(sd.is_chunk_aligned(dest_addr), "not aligned");
3060 const ChunkData* const src_chunk_ptr = sd.chunk(src_chunk_idx);
3061 const size_t partial_obj_size = src_chunk_ptr->partial_obj_size();
3062 HeapWord* const src_chunk_destination = src_chunk_ptr->destination();
3064 assert(dest_addr >= src_chunk_destination, "wrong src chunk");
3065 assert(src_chunk_ptr->data_size() > 0, "src chunk cannot be empty");
3067 HeapWord* const src_chunk_beg = sd.chunk_to_addr(src_chunk_idx);
3068 HeapWord* const src_chunk_end = src_chunk_beg + ChunkSize;
3070 HeapWord* addr = src_chunk_beg;
3071 if (dest_addr == src_chunk_destination) {
3072 // Return the first live word in the source chunk.
3073 if (partial_obj_size == 0) {
3074 addr = bitmap->find_obj_beg(addr, src_chunk_end);
3075 assert(addr < src_chunk_end, "no objects start in src chunk");
3076 }
3077 return addr;
3078 }
3080 // Must skip some live data.
3081 size_t words_to_skip = dest_addr - src_chunk_destination;
3082 assert(src_chunk_ptr->data_size() > words_to_skip, "wrong src chunk");
3084 if (partial_obj_size >= words_to_skip) {
3085 // All the live words to skip are part of the partial object.
3086 addr += words_to_skip;
3087 if (partial_obj_size == words_to_skip) {
3088 // Find the first live word past the partial object.
3089 addr = bitmap->find_obj_beg(addr, src_chunk_end);
3090 assert(addr < src_chunk_end, "wrong src chunk");
3091 }
3092 return addr;
3093 }
3095 // Skip over the partial object (if any).
3096 if (partial_obj_size != 0) {
3097 words_to_skip -= partial_obj_size;
3098 addr += partial_obj_size;
3099 }
3101 // Skip over live words due to objects that start in the chunk.
3102 addr = skip_live_words(addr, src_chunk_end, words_to_skip);
3103 assert(addr < src_chunk_end, "wrong src chunk");
3104 return addr;
3105 }
3107 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3108 size_t beg_chunk,
3109 HeapWord* end_addr)
3110 {
3111 ParallelCompactData& sd = summary_data();
3112 ChunkData* const beg = sd.chunk(beg_chunk);
3113 HeapWord* const end_addr_aligned_up = sd.chunk_align_up(end_addr);
3114 ChunkData* const end = sd.addr_to_chunk_ptr(end_addr_aligned_up);
3115 size_t cur_idx = beg_chunk;
3116 for (ChunkData* cur = beg; cur < end; ++cur, ++cur_idx) {
3117 assert(cur->data_size() > 0, "chunk must have live data");
3118 cur->decrement_destination_count();
3119 if (cur_idx <= cur->source_chunk() && cur->available() && cur->claim()) {
3120 cm->save_for_processing(cur_idx);
3121 }
3122 }
3123 }
3125 size_t PSParallelCompact::next_src_chunk(MoveAndUpdateClosure& closure,
3126 SpaceId& src_space_id,
3127 HeapWord*& src_space_top,
3128 HeapWord* end_addr)
3129 {
3130 typedef ParallelCompactData::ChunkData ChunkData;
3132 ParallelCompactData& sd = PSParallelCompact::summary_data();
3133 const size_t chunk_size = ParallelCompactData::ChunkSize;
3135 size_t src_chunk_idx = 0;
3137 // Skip empty chunks (if any) up to the top of the space.
3138 HeapWord* const src_aligned_up = sd.chunk_align_up(end_addr);
3139 ChunkData* src_chunk_ptr = sd.addr_to_chunk_ptr(src_aligned_up);
3140 HeapWord* const top_aligned_up = sd.chunk_align_up(src_space_top);
3141 const ChunkData* const top_chunk_ptr = sd.addr_to_chunk_ptr(top_aligned_up);
3142 while (src_chunk_ptr < top_chunk_ptr && src_chunk_ptr->data_size() == 0) {
3143 ++src_chunk_ptr;
3144 }
3146 if (src_chunk_ptr < top_chunk_ptr) {
3147 // The next source chunk is in the current space. Update src_chunk_idx and
3148 // the source address to match src_chunk_ptr.
3149 src_chunk_idx = sd.chunk(src_chunk_ptr);
3150 HeapWord* const src_chunk_addr = sd.chunk_to_addr(src_chunk_idx);
3151 if (src_chunk_addr > closure.source()) {
3152 closure.set_source(src_chunk_addr);
3153 }
3154 return src_chunk_idx;
3155 }
3157 // Switch to a new source space and find the first non-empty chunk.
3158 unsigned int space_id = src_space_id + 1;
3159 assert(space_id < last_space_id, "not enough spaces");
3161 HeapWord* const destination = closure.destination();
3163 do {
3164 MutableSpace* space = _space_info[space_id].space();
3165 HeapWord* const bottom = space->bottom();
3166 const ChunkData* const bottom_cp = sd.addr_to_chunk_ptr(bottom);
3168 // Iterate over the spaces that do not compact into themselves.
3169 if (bottom_cp->destination() != bottom) {
3170 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top());
3171 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up);
3173 for (const ChunkData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3174 if (src_cp->live_obj_size() > 0) {
3175 // Found it.
3176 assert(src_cp->destination() == destination,
3177 "first live obj in the space must match the destination");
3178 assert(src_cp->partial_obj_size() == 0,
3179 "a space cannot begin with a partial obj");
3181 src_space_id = SpaceId(space_id);
3182 src_space_top = space->top();
3183 const size_t src_chunk_idx = sd.chunk(src_cp);
3184 closure.set_source(sd.chunk_to_addr(src_chunk_idx));
3185 return src_chunk_idx;
3186 } else {
3187 assert(src_cp->data_size() == 0, "sanity");
3188 }
3189 }
3190 }
3191 } while (++space_id < last_space_id);
3193 assert(false, "no source chunk was found");
3194 return 0;
3195 }
3197 void PSParallelCompact::fill_chunk(ParCompactionManager* cm, size_t chunk_idx)
3198 {
3199 typedef ParMarkBitMap::IterationStatus IterationStatus;
3200 const size_t ChunkSize = ParallelCompactData::ChunkSize;
3201 ParMarkBitMap* const bitmap = mark_bitmap();
3202 ParallelCompactData& sd = summary_data();
3203 ChunkData* const chunk_ptr = sd.chunk(chunk_idx);
3205 // Get the items needed to construct the closure.
3206 HeapWord* dest_addr = sd.chunk_to_addr(chunk_idx);
3207 SpaceId dest_space_id = space_id(dest_addr);
3208 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3209 HeapWord* new_top = _space_info[dest_space_id].new_top();
3210 assert(dest_addr < new_top, "sanity");
3211 const size_t words = MIN2(pointer_delta(new_top, dest_addr), ChunkSize);
3213 // Get the source chunk and related info.
3214 size_t src_chunk_idx = chunk_ptr->source_chunk();
3215 SpaceId src_space_id = space_id(sd.chunk_to_addr(src_chunk_idx));
3216 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3218 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3219 closure.set_source(first_src_addr(dest_addr, src_chunk_idx));
3221 // Adjust src_chunk_idx to prepare for decrementing destination counts (the
3222 // destination count is not decremented when a chunk is copied to itself).
3223 if (src_chunk_idx == chunk_idx) {
3224 src_chunk_idx += 1;
3225 }
3227 if (bitmap->is_unmarked(closure.source())) {
3228 // The first source word is in the middle of an object; copy the remainder
3229 // of the object or as much as will fit. The fact that pointer updates were
3230 // deferred will be noted when the object header is processed.
3231 HeapWord* const old_src_addr = closure.source();
3232 closure.copy_partial_obj();
3233 if (closure.is_full()) {
3234 decrement_destination_counts(cm, src_chunk_idx, closure.source());
3235 chunk_ptr->set_deferred_obj_addr(NULL);
3236 chunk_ptr->set_completed();
3237 return;
3238 }
3240 HeapWord* const end_addr = sd.chunk_align_down(closure.source());
3241 if (sd.chunk_align_down(old_src_addr) != end_addr) {
3242 // The partial object was copied from more than one source chunk.
3243 decrement_destination_counts(cm, src_chunk_idx, end_addr);
3245 // Move to the next source chunk, possibly switching spaces as well. All
3246 // args except end_addr may be modified.
3247 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top,
3248 end_addr);
3249 }
3250 }
3252 do {
3253 HeapWord* const cur_addr = closure.source();
3254 HeapWord* const end_addr = MIN2(sd.chunk_align_up(cur_addr + 1),
3255 src_space_top);
3256 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3258 if (status == ParMarkBitMap::incomplete) {
3259 // The last obj that starts in the source chunk does not end in the chunk.
3260 assert(closure.source() < end_addr, "sanity")
3261 HeapWord* const obj_beg = closure.source();
3262 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3263 src_space_top);
3264 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3265 if (obj_end < range_end) {
3266 // The end was found; the entire object will fit.
3267 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3268 assert(status != ParMarkBitMap::would_overflow, "sanity");
3269 } else {
3270 // The end was not found; the object will not fit.
3271 assert(range_end < src_space_top, "obj cannot cross space boundary");
3272 status = ParMarkBitMap::would_overflow;
3273 }
3274 }
3276 if (status == ParMarkBitMap::would_overflow) {
3277 // The last object did not fit. Note that interior oop updates were
3278 // deferred, then copy enough of the object to fill the chunk.
3279 chunk_ptr->set_deferred_obj_addr(closure.destination());
3280 status = closure.copy_until_full(); // copies from closure.source()
3282 decrement_destination_counts(cm, src_chunk_idx, closure.source());
3283 chunk_ptr->set_completed();
3284 return;
3285 }
3287 if (status == ParMarkBitMap::full) {
3288 decrement_destination_counts(cm, src_chunk_idx, closure.source());
3289 chunk_ptr->set_deferred_obj_addr(NULL);
3290 chunk_ptr->set_completed();
3291 return;
3292 }
3294 decrement_destination_counts(cm, src_chunk_idx, end_addr);
3296 // Move to the next source chunk, possibly switching spaces as well. All
3297 // args except end_addr may be modified.
3298 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top,
3299 end_addr);
3300 } while (true);
3301 }
3303 void
3304 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3305 const MutableSpace* sp = space(space_id);
3306 if (sp->is_empty()) {
3307 return;
3308 }
3310 ParallelCompactData& sd = PSParallelCompact::summary_data();
3311 ParMarkBitMap* const bitmap = mark_bitmap();
3312 HeapWord* const dp_addr = dense_prefix(space_id);
3313 HeapWord* beg_addr = sp->bottom();
3314 HeapWord* end_addr = sp->top();
3316 #ifdef ASSERT
3317 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3318 if (cm->should_verify_only()) {
3319 VerifyUpdateClosure verify_update(cm, sp);
3320 bitmap->iterate(&verify_update, beg_addr, end_addr);
3321 return;
3322 }
3324 if (cm->should_reset_only()) {
3325 ResetObjectsClosure reset_objects(cm);
3326 bitmap->iterate(&reset_objects, beg_addr, end_addr);
3327 return;
3328 }
3329 #endif
3331 const size_t beg_chunk = sd.addr_to_chunk_idx(beg_addr);
3332 const size_t dp_chunk = sd.addr_to_chunk_idx(dp_addr);
3333 if (beg_chunk < dp_chunk) {
3334 update_and_deadwood_in_dense_prefix(cm, space_id, beg_chunk, dp_chunk);
3335 }
3337 // The destination of the first live object that starts in the chunk is one
3338 // past the end of the partial object entering the chunk (if any).
3339 HeapWord* const dest_addr = sd.partial_obj_end(dp_chunk);
3340 HeapWord* const new_top = _space_info[space_id].new_top();
3341 assert(new_top >= dest_addr, "bad new_top value");
3342 const size_t words = pointer_delta(new_top, dest_addr);
3344 if (words > 0) {
3345 ObjectStartArray* start_array = _space_info[space_id].start_array();
3346 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3348 ParMarkBitMap::IterationStatus status;
3349 status = bitmap->iterate(&closure, dest_addr, end_addr);
3350 assert(status == ParMarkBitMap::full, "iteration not complete");
3351 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3352 "live objects skipped because closure is full");
3353 }
3354 }
3356 jlong PSParallelCompact::millis_since_last_gc() {
3357 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
3358 // XXX See note in genCollectedHeap::millis_since_last_gc().
3359 if (ret_val < 0) {
3360 NOT_PRODUCT(warning("time warp: %d", ret_val);)
3361 return 0;
3362 }
3363 return ret_val;
3364 }
3366 void PSParallelCompact::reset_millis_since_last_gc() {
3367 _time_of_last_gc = os::javaTimeMillis();
3368 }
3370 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3371 {
3372 if (source() != destination()) {
3373 assert(source() > destination(), "must copy to the left");
3374 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3375 }
3376 update_state(words_remaining());
3377 assert(is_full(), "sanity");
3378 return ParMarkBitMap::full;
3379 }
3381 void MoveAndUpdateClosure::copy_partial_obj()
3382 {
3383 size_t words = words_remaining();
3385 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3386 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3387 if (end_addr < range_end) {
3388 words = bitmap()->obj_size(source(), end_addr);
3389 }
3391 // This test is necessary; if omitted, the pointer updates to a partial object
3392 // that crosses the dense prefix boundary could be overwritten.
3393 if (source() != destination()) {
3394 assert(source() > destination(), "must copy to the left");
3395 Copy::aligned_conjoint_words(source(), destination(), words);
3396 }
3397 update_state(words);
3398 }
3400 ParMarkBitMapClosure::IterationStatus
3401 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3402 assert(destination() != NULL, "sanity");
3403 assert(bitmap()->obj_size(addr) == words, "bad size");
3405 _source = addr;
3406 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3407 destination(), "wrong destination");
3409 if (words > words_remaining()) {
3410 return ParMarkBitMap::would_overflow;
3411 }
3413 // The start_array must be updated even if the object is not moving.
3414 if (_start_array != NULL) {
3415 _start_array->allocate_block(destination());
3416 }
3418 if (destination() != source()) {
3419 assert(destination() < source(), "must copy to the left");
3420 Copy::aligned_conjoint_words(source(), destination(), words);
3421 }
3423 oop moved_oop = (oop) destination();
3424 moved_oop->update_contents(compaction_manager());
3425 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3427 update_state(words);
3428 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3429 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3430 }
3432 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3433 ParCompactionManager* cm,
3434 PSParallelCompact::SpaceId space_id) :
3435 ParMarkBitMapClosure(mbm, cm),
3436 _space_id(space_id),
3437 _start_array(PSParallelCompact::start_array(space_id))
3438 {
3439 }
3441 // Updates the references in the object to their new values.
3442 ParMarkBitMapClosure::IterationStatus
3443 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3444 do_addr(addr);
3445 return ParMarkBitMap::incomplete;
3446 }
3448 BitBlockUpdateClosure::BitBlockUpdateClosure(ParMarkBitMap* mbm,
3449 ParCompactionManager* cm,
3450 size_t chunk_index) :
3451 ParMarkBitMapClosure(mbm, cm),
3452 _live_data_left(0),
3453 _cur_block(0) {
3454 _chunk_start =
3455 PSParallelCompact::summary_data().chunk_to_addr(chunk_index);
3456 _chunk_end =
3457 PSParallelCompact::summary_data().chunk_to_addr(chunk_index) +
3458 ParallelCompactData::ChunkSize;
3459 _chunk_index = chunk_index;
3460 _cur_block =
3461 PSParallelCompact::summary_data().addr_to_block_idx(_chunk_start);
3462 }
3464 bool BitBlockUpdateClosure::chunk_contains_cur_block() {
3465 return ParallelCompactData::chunk_contains_block(_chunk_index, _cur_block);
3466 }
3468 void BitBlockUpdateClosure::reset_chunk(size_t chunk_index) {
3469 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(7);)
3470 ParallelCompactData& sd = PSParallelCompact::summary_data();
3471 _chunk_index = chunk_index;
3472 _live_data_left = 0;
3473 _chunk_start = sd.chunk_to_addr(chunk_index);
3474 _chunk_end = sd.chunk_to_addr(chunk_index) + ParallelCompactData::ChunkSize;
3476 // The first block in this chunk
3477 size_t first_block = sd.addr_to_block_idx(_chunk_start);
3478 size_t partial_live_size = sd.chunk(chunk_index)->partial_obj_size();
3480 // Set the offset to 0. By definition it should have that value
3481 // but it may have been written while processing an earlier chunk.
3482 if (partial_live_size == 0) {
3483 // No live object extends onto the chunk. The first bit
3484 // in the bit map for the first chunk must be a start bit.
3485 // Although there may not be any marked bits, it is safe
3486 // to set it as a start bit.
3487 sd.block(first_block)->set_start_bit_offset(0);
3488 sd.block(first_block)->set_first_is_start_bit(true);
3489 } else if (sd.partial_obj_ends_in_block(first_block)) {
3490 sd.block(first_block)->set_end_bit_offset(0);
3491 sd.block(first_block)->set_first_is_start_bit(false);
3492 } else {
3493 // The partial object extends beyond the first block.
3494 // There is no object starting in the first block
3495 // so the offset and bit parity are not needed.
3496 // Set the the bit parity to start bit so assertions
3497 // work when not bit is found.
3498 sd.block(first_block)->set_end_bit_offset(0);
3499 sd.block(first_block)->set_first_is_start_bit(false);
3500 }
3501 _cur_block = first_block;
3502 #ifdef ASSERT
3503 if (sd.block(first_block)->first_is_start_bit()) {
3504 assert(!sd.partial_obj_ends_in_block(first_block),
3505 "Partial object cannot end in first block");
3506 }
3508 if (PrintGCDetails && Verbose) {
3509 if (partial_live_size == 1) {
3510 gclog_or_tty->print_cr("first_block " PTR_FORMAT
3511 " _offset " PTR_FORMAT
3512 " _first_is_start_bit %d",
3513 first_block,
3514 sd.block(first_block)->raw_offset(),
3515 sd.block(first_block)->first_is_start_bit());
3516 }
3517 }
3518 #endif
3519 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(17);)
3520 }
3522 // This method is called when a object has been found (both beginning
3523 // and end of the object) in the range of iteration. This method is
3524 // calculating the words of live data to the left of a block. That live
3525 // data includes any object starting to the left of the block (i.e.,
3526 // the live-data-to-the-left of block AAA will include the full size
3527 // of any object entering AAA).
3529 ParMarkBitMapClosure::IterationStatus
3530 BitBlockUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3531 // add the size to the block data.
3532 HeapWord* obj = addr;
3533 ParallelCompactData& sd = PSParallelCompact::summary_data();
3535 assert(bitmap()->obj_size(obj) == words, "bad size");
3536 assert(_chunk_start <= obj, "object is not in chunk");
3537 assert(obj + words <= _chunk_end, "object is not in chunk");
3539 // Update the live data to the left
3540 size_t prev_live_data_left = _live_data_left;
3541 _live_data_left = _live_data_left + words;
3543 // Is this object in the current block.
3544 size_t block_of_obj = sd.addr_to_block_idx(obj);
3545 size_t block_of_obj_last = sd.addr_to_block_idx(obj + words - 1);
3546 HeapWord* block_of_obj_last_addr = sd.block_to_addr(block_of_obj_last);
3547 if (_cur_block < block_of_obj) {
3549 //
3550 // No object crossed the block boundary and this object was found
3551 // on the other side of the block boundary. Update the offset for
3552 // the new block with the data size that does not include this object.
3553 //
3554 // The first bit in block_of_obj is a start bit except in the
3555 // case where the partial object for the chunk extends into
3556 // this block.
3557 if (sd.partial_obj_ends_in_block(block_of_obj)) {
3558 sd.block(block_of_obj)->set_end_bit_offset(prev_live_data_left);
3559 } else {
3560 sd.block(block_of_obj)->set_start_bit_offset(prev_live_data_left);
3561 }
3563 // Does this object pass beyond the its block?
3564 if (block_of_obj < block_of_obj_last) {
3565 // Object crosses block boundary. Two blocks need to be udpated:
3566 // the current block where the object started
3567 // the block where the object ends
3568 //
3569 // The offset for blocks with no objects starting in them
3570 // (e.g., blocks between _cur_block and block_of_obj_last)
3571 // should not be needed.
3572 // Note that block_of_obj_last may be in another chunk. If so,
3573 // it should be overwritten later. This is a problem (writting
3574 // into a block in a later chunk) for parallel execution.
3575 assert(obj < block_of_obj_last_addr,
3576 "Object should start in previous block");
3578 // obj is crossing into block_of_obj_last so the first bit
3579 // is and end bit.
3580 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left);
3582 _cur_block = block_of_obj_last;
3583 } else {
3584 // _first_is_start_bit has already been set correctly
3585 // in the if-then-else above so don't reset it here.
3586 _cur_block = block_of_obj;
3587 }
3588 } else {
3589 // The current block only changes if the object extends beyound
3590 // the block it starts in.
3591 //
3592 // The object starts in the current block.
3593 // Does this object pass beyond the end of it?
3594 if (block_of_obj < block_of_obj_last) {
3595 // Object crosses block boundary.
3596 // See note above on possible blocks between block_of_obj and
3597 // block_of_obj_last
3598 assert(obj < block_of_obj_last_addr,
3599 "Object should start in previous block");
3601 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left);
3603 _cur_block = block_of_obj_last;
3604 }
3605 }
3607 // Return incomplete if there are more blocks to be done.
3608 if (chunk_contains_cur_block()) {
3609 return ParMarkBitMap::incomplete;
3610 }
3611 return ParMarkBitMap::complete;
3612 }
3614 // Verify the new location using the forwarding pointer
3615 // from MarkSweep::mark_sweep_phase2(). Set the mark_word
3616 // to the initial value.
3617 ParMarkBitMapClosure::IterationStatus
3618 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3619 // The second arg (words) is not used.
3620 oop obj = (oop) addr;
3621 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
3622 HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
3623 if (forwarding_ptr == NULL) {
3624 // The object is dead or not moving.
3625 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
3626 "Object liveness is wrong.");
3627 return ParMarkBitMap::incomplete;
3628 }
3629 assert(UseParallelOldGCDensePrefix ||
3630 (HeapMaximumCompactionInterval > 1) ||
3631 (MarkSweepAlwaysCompactCount > 1) ||
3632 (forwarding_ptr == new_pointer),
3633 "Calculation of new location is incorrect");
3634 return ParMarkBitMap::incomplete;
3635 }
3637 // Reset objects modified for debug checking.
3638 ParMarkBitMapClosure::IterationStatus
3639 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
3640 // The second arg (words) is not used.
3641 oop obj = (oop) addr;
3642 obj->init_mark();
3643 return ParMarkBitMap::incomplete;
3644 }
3646 // Prepare for compaction. This method is executed once
3647 // (i.e., by a single thread) before compaction.
3648 // Save the updated location of the intArrayKlassObj for
3649 // filling holes in the dense prefix.
3650 void PSParallelCompact::compact_prologue() {
3651 _updated_int_array_klass_obj = (klassOop)
3652 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3653 }
3655 // The initial implementation of this method created a field
3656 // _next_compaction_space_id in SpaceInfo and initialized
3657 // that field in SpaceInfo::initialize_space_info(). That
3658 // required that _next_compaction_space_id be declared a
3659 // SpaceId in SpaceInfo and that would have required that
3660 // either SpaceId be declared in a separate class or that
3661 // it be declared in SpaceInfo. It didn't seem consistent
3662 // to declare it in SpaceInfo (didn't really fit logically).
3663 // Alternatively, defining a separate class to define SpaceId
3664 // seem excessive. This implementation is simple and localizes
3665 // the knowledge.
3667 PSParallelCompact::SpaceId
3668 PSParallelCompact::next_compaction_space_id(SpaceId id) {
3669 assert(id < last_space_id, "id out of range");
3670 switch (id) {
3671 case perm_space_id :
3672 return last_space_id;
3673 case old_space_id :
3674 return eden_space_id;
3675 case eden_space_id :
3676 return from_space_id;
3677 case from_space_id :
3678 return to_space_id;
3679 case to_space_id :
3680 return last_space_id;
3681 default:
3682 assert(false, "Bad space id");
3683 return last_space_id;
3684 }
3685 }
3687 // Here temporarily for debugging
3688 #ifdef ASSERT
3689 size_t ParallelCompactData::block_idx(BlockData* block) {
3690 size_t index = pointer_delta(block,
3691 PSParallelCompact::summary_data()._block_data, sizeof(BlockData));
3692 return index;
3693 }
3694 #endif