Print this page
rev 4515 : 7176220: 'Full GC' events miss date stamp information occasionally
Summary: Move date stamp logic into GCTraceTime
Reviewed-by: johnc, brutisso, jmasa
Split |
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
+++ new/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
1 1 /*
2 2 * Copyright (c) 2005, 2012, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/symbolTable.hpp"
27 27 #include "classfile/systemDictionary.hpp"
28 28 #include "code/codeCache.hpp"
29 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
30 30 #include "gc_implementation/parallelScavenge/generationSizer.hpp"
31 31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
32 32 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
33 33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
34 34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
35 35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
36 36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
37 37 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
38 38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
39 39 #include "gc_implementation/parallelScavenge/psPermGen.hpp"
40 40 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
41 41 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
42 42 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
43 43 #include "gc_implementation/shared/gcHeapSummary.hpp"
44 44 #include "gc_implementation/shared/gcTimer.hpp"
45 45 #include "gc_implementation/shared/gcTrace.hpp"
46 46 #include "gc_implementation/shared/gcTraceTime.hpp"
47 47 #include "gc_implementation/shared/isGCActiveMark.hpp"
48 48 #include "gc_interface/gcCause.hpp"
49 49 #include "memory/gcLocker.inline.hpp"
50 50 #include "memory/referencePolicy.hpp"
51 51 #include "memory/referenceProcessor.hpp"
52 52 #include "oops/methodDataOop.hpp"
53 53 #include "oops/oop.inline.hpp"
54 54 #include "oops/oop.pcgc.inline.hpp"
55 55 #include "runtime/fprofiler.hpp"
56 56 #include "runtime/safepoint.hpp"
57 57 #include "runtime/vmThread.hpp"
58 58 #include "services/management.hpp"
59 59 #include "services/memoryService.hpp"
60 60 #include "services/memTracker.hpp"
61 61 #include "utilities/events.hpp"
62 62 #include "utilities/stack.inline.hpp"
63 63
64 64 #include <math.h>
65 65
66 66 // All sizes are in HeapWords.
67 67 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words
68 68 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
69 69 const size_t ParallelCompactData::RegionSizeBytes =
70 70 RegionSize << LogHeapWordSize;
71 71 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
72 72 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
73 73 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
74 74
75 75 const ParallelCompactData::RegionData::region_sz_t
76 76 ParallelCompactData::RegionData::dc_shift = 27;
77 77
78 78 const ParallelCompactData::RegionData::region_sz_t
79 79 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
80 80
81 81 const ParallelCompactData::RegionData::region_sz_t
82 82 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
83 83
84 84 const ParallelCompactData::RegionData::region_sz_t
85 85 ParallelCompactData::RegionData::los_mask = ~dc_mask;
86 86
87 87 const ParallelCompactData::RegionData::region_sz_t
88 88 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
89 89
90 90 const ParallelCompactData::RegionData::region_sz_t
91 91 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
92 92
93 93 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
94 94 bool PSParallelCompact::_print_phases = false;
95 95
96 96 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
97 97 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL;
98 98
99 99 double PSParallelCompact::_dwl_mean;
100 100 double PSParallelCompact::_dwl_std_dev;
101 101 double PSParallelCompact::_dwl_first_term;
102 102 double PSParallelCompact::_dwl_adjustment;
103 103 #ifdef ASSERT
104 104 bool PSParallelCompact::_dwl_initialized = false;
105 105 #endif // #ifdef ASSERT
106 106
107 107 #ifdef VALIDATE_MARK_SWEEP
108 108 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL;
109 109 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL;
110 110 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL;
111 111 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL;
112 112 size_t PSParallelCompact::_live_oops_index = 0;
113 113 size_t PSParallelCompact::_live_oops_index_at_perm = 0;
114 114 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL;
115 115 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL;
116 116 bool PSParallelCompact::_pointer_tracking = false;
117 117 bool PSParallelCompact::_root_tracking = true;
118 118
119 119 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
120 120 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
121 121 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL;
122 122 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
123 123 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
124 124 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL;
125 125 #endif
126 126
127 127 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
128 128 HeapWord* destination)
129 129 {
130 130 assert(src_region_idx != 0, "invalid src_region_idx");
131 131 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
132 132 assert(destination != NULL, "invalid destination argument");
133 133
134 134 _src_region_idx = src_region_idx;
135 135 _partial_obj_size = partial_obj_size;
136 136 _destination = destination;
137 137
138 138 // These fields may not be updated below, so make sure they're clear.
139 139 assert(_dest_region_addr == NULL, "should have been cleared");
140 140 assert(_first_src_addr == NULL, "should have been cleared");
141 141
142 142 // Determine the number of destination regions for the partial object.
143 143 HeapWord* const last_word = destination + partial_obj_size - 1;
144 144 const ParallelCompactData& sd = PSParallelCompact::summary_data();
145 145 HeapWord* const beg_region_addr = sd.region_align_down(destination);
146 146 HeapWord* const end_region_addr = sd.region_align_down(last_word);
147 147
148 148 if (beg_region_addr == end_region_addr) {
149 149 // One destination region.
150 150 _destination_count = 1;
151 151 if (end_region_addr == destination) {
152 152 // The destination falls on a region boundary, thus the first word of the
153 153 // partial object will be the first word copied to the destination region.
154 154 _dest_region_addr = end_region_addr;
155 155 _first_src_addr = sd.region_to_addr(src_region_idx);
156 156 }
157 157 } else {
158 158 // Two destination regions. When copied, the partial object will cross a
159 159 // destination region boundary, so a word somewhere within the partial
160 160 // object will be the first word copied to the second destination region.
161 161 _destination_count = 2;
162 162 _dest_region_addr = end_region_addr;
163 163 const size_t ofs = pointer_delta(end_region_addr, destination);
164 164 assert(ofs < _partial_obj_size, "sanity");
165 165 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
166 166 }
167 167 }
168 168
169 169 void SplitInfo::clear()
170 170 {
171 171 _src_region_idx = 0;
172 172 _partial_obj_size = 0;
173 173 _destination = NULL;
174 174 _destination_count = 0;
175 175 _dest_region_addr = NULL;
176 176 _first_src_addr = NULL;
177 177 assert(!is_valid(), "sanity");
178 178 }
179 179
180 180 #ifdef ASSERT
181 181 void SplitInfo::verify_clear()
182 182 {
183 183 assert(_src_region_idx == 0, "not clear");
184 184 assert(_partial_obj_size == 0, "not clear");
185 185 assert(_destination == NULL, "not clear");
186 186 assert(_destination_count == 0, "not clear");
187 187 assert(_dest_region_addr == NULL, "not clear");
188 188 assert(_first_src_addr == NULL, "not clear");
189 189 }
190 190 #endif // #ifdef ASSERT
191 191
192 192
193 193 #ifndef PRODUCT
194 194 const char* PSParallelCompact::space_names[] = {
195 195 "perm", "old ", "eden", "from", "to "
196 196 };
197 197
198 198 void PSParallelCompact::print_region_ranges()
199 199 {
200 200 tty->print_cr("space bottom top end new_top");
201 201 tty->print_cr("------ ---------- ---------- ---------- ----------");
202 202
203 203 for (unsigned int id = 0; id < last_space_id; ++id) {
204 204 const MutableSpace* space = _space_info[id].space();
205 205 tty->print_cr("%u %s "
206 206 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
207 207 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
208 208 id, space_names[id],
209 209 summary_data().addr_to_region_idx(space->bottom()),
210 210 summary_data().addr_to_region_idx(space->top()),
211 211 summary_data().addr_to_region_idx(space->end()),
212 212 summary_data().addr_to_region_idx(_space_info[id].new_top()));
213 213 }
214 214 }
215 215
216 216 void
217 217 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
218 218 {
219 219 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
220 220 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
221 221
222 222 ParallelCompactData& sd = PSParallelCompact::summary_data();
223 223 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
224 224 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
225 225 REGION_IDX_FORMAT " " PTR_FORMAT " "
226 226 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
227 227 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
228 228 i, c->data_location(), dci, c->destination(),
229 229 c->partial_obj_size(), c->live_obj_size(),
230 230 c->data_size(), c->source_region(), c->destination_count());
231 231
232 232 #undef REGION_IDX_FORMAT
233 233 #undef REGION_DATA_FORMAT
234 234 }
235 235
236 236 void
237 237 print_generic_summary_data(ParallelCompactData& summary_data,
238 238 HeapWord* const beg_addr,
239 239 HeapWord* const end_addr)
240 240 {
241 241 size_t total_words = 0;
242 242 size_t i = summary_data.addr_to_region_idx(beg_addr);
243 243 const size_t last = summary_data.addr_to_region_idx(end_addr);
244 244 HeapWord* pdest = 0;
245 245
246 246 while (i <= last) {
247 247 ParallelCompactData::RegionData* c = summary_data.region(i);
248 248 if (c->data_size() != 0 || c->destination() != pdest) {
249 249 print_generic_summary_region(i, c);
250 250 total_words += c->data_size();
251 251 pdest = c->destination();
252 252 }
253 253 ++i;
254 254 }
255 255
256 256 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
257 257 }
258 258
259 259 void
260 260 print_generic_summary_data(ParallelCompactData& summary_data,
261 261 SpaceInfo* space_info)
262 262 {
263 263 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
264 264 const MutableSpace* space = space_info[id].space();
265 265 print_generic_summary_data(summary_data, space->bottom(),
266 266 MAX2(space->top(), space_info[id].new_top()));
267 267 }
268 268 }
269 269
270 270 void
271 271 print_initial_summary_region(size_t i,
272 272 const ParallelCompactData::RegionData* c,
273 273 bool newline = true)
274 274 {
275 275 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
276 276 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
277 277 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
278 278 i, c->destination(),
279 279 c->partial_obj_size(), c->live_obj_size(),
280 280 c->data_size(), c->source_region(), c->destination_count());
281 281 if (newline) tty->cr();
282 282 }
283 283
284 284 void
285 285 print_initial_summary_data(ParallelCompactData& summary_data,
286 286 const MutableSpace* space) {
287 287 if (space->top() == space->bottom()) {
288 288 return;
289 289 }
290 290
291 291 const size_t region_size = ParallelCompactData::RegionSize;
292 292 typedef ParallelCompactData::RegionData RegionData;
293 293 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
294 294 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
295 295 const RegionData* c = summary_data.region(end_region - 1);
296 296 HeapWord* end_addr = c->destination() + c->data_size();
297 297 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
298 298
299 299 // Print (and count) the full regions at the beginning of the space.
300 300 size_t full_region_count = 0;
301 301 size_t i = summary_data.addr_to_region_idx(space->bottom());
302 302 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
303 303 print_initial_summary_region(i, summary_data.region(i));
304 304 ++full_region_count;
305 305 ++i;
306 306 }
307 307
308 308 size_t live_to_right = live_in_space - full_region_count * region_size;
309 309
310 310 double max_reclaimed_ratio = 0.0;
311 311 size_t max_reclaimed_ratio_region = 0;
312 312 size_t max_dead_to_right = 0;
313 313 size_t max_live_to_right = 0;
314 314
315 315 // Print the 'reclaimed ratio' for regions while there is something live in
316 316 // the region or to the right of it. The remaining regions are empty (and
317 317 // uninteresting), and computing the ratio will result in division by 0.
318 318 while (i < end_region && live_to_right > 0) {
319 319 c = summary_data.region(i);
320 320 HeapWord* const region_addr = summary_data.region_to_addr(i);
321 321 const size_t used_to_right = pointer_delta(space->top(), region_addr);
322 322 const size_t dead_to_right = used_to_right - live_to_right;
323 323 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
324 324
325 325 if (reclaimed_ratio > max_reclaimed_ratio) {
326 326 max_reclaimed_ratio = reclaimed_ratio;
327 327 max_reclaimed_ratio_region = i;
328 328 max_dead_to_right = dead_to_right;
329 329 max_live_to_right = live_to_right;
330 330 }
331 331
332 332 print_initial_summary_region(i, c, false);
333 333 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
334 334 reclaimed_ratio, dead_to_right, live_to_right);
335 335
336 336 live_to_right -= c->data_size();
337 337 ++i;
338 338 }
339 339
340 340 // Any remaining regions are empty. Print one more if there is one.
341 341 if (i < end_region) {
342 342 print_initial_summary_region(i, summary_data.region(i));
343 343 }
344 344
345 345 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
346 346 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
347 347 max_reclaimed_ratio_region, max_dead_to_right,
348 348 max_live_to_right, max_reclaimed_ratio);
349 349 }
350 350
351 351 void
352 352 print_initial_summary_data(ParallelCompactData& summary_data,
353 353 SpaceInfo* space_info) {
354 354 unsigned int id = PSParallelCompact::perm_space_id;
355 355 const MutableSpace* space;
356 356 do {
357 357 space = space_info[id].space();
358 358 print_initial_summary_data(summary_data, space);
359 359 } while (++id < PSParallelCompact::eden_space_id);
360 360
361 361 do {
362 362 space = space_info[id].space();
363 363 print_generic_summary_data(summary_data, space->bottom(), space->top());
364 364 } while (++id < PSParallelCompact::last_space_id);
365 365 }
366 366 #endif // #ifndef PRODUCT
367 367
368 368 #ifdef ASSERT
369 369 size_t add_obj_count;
370 370 size_t add_obj_size;
371 371 size_t mark_bitmap_count;
372 372 size_t mark_bitmap_size;
373 373 #endif // #ifdef ASSERT
374 374
375 375 ParallelCompactData::ParallelCompactData()
376 376 {
377 377 _region_start = 0;
378 378
379 379 _region_vspace = 0;
380 380 _region_data = 0;
381 381 _region_count = 0;
382 382 }
383 383
384 384 bool ParallelCompactData::initialize(MemRegion covered_region)
385 385 {
386 386 _region_start = covered_region.start();
387 387 const size_t region_size = covered_region.word_size();
388 388 DEBUG_ONLY(_region_end = _region_start + region_size;)
389 389
390 390 assert(region_align_down(_region_start) == _region_start,
391 391 "region start not aligned");
392 392 assert((region_size & RegionSizeOffsetMask) == 0,
393 393 "region size not a multiple of RegionSize");
394 394
395 395 bool result = initialize_region_data(region_size);
396 396
397 397 return result;
398 398 }
399 399
400 400 PSVirtualSpace*
401 401 ParallelCompactData::create_vspace(size_t count, size_t element_size)
402 402 {
403 403 const size_t raw_bytes = count * element_size;
404 404 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
405 405 const size_t granularity = os::vm_allocation_granularity();
406 406 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
407 407
408 408 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
409 409 MAX2(page_sz, granularity);
410 410 ReservedSpace rs(bytes, rs_align, rs_align > 0);
411 411 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
412 412 rs.size());
413 413
414 414 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
415 415
416 416 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
417 417 if (vspace != 0) {
418 418 if (vspace->expand_by(bytes)) {
419 419 return vspace;
420 420 }
421 421 delete vspace;
422 422 // Release memory reserved in the space.
423 423 rs.release();
424 424 }
425 425
426 426 return 0;
427 427 }
428 428
429 429 bool ParallelCompactData::initialize_region_data(size_t region_size)
430 430 {
431 431 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
432 432 _region_vspace = create_vspace(count, sizeof(RegionData));
433 433 if (_region_vspace != 0) {
434 434 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
435 435 _region_count = count;
436 436 return true;
437 437 }
438 438 return false;
439 439 }
440 440
441 441 void ParallelCompactData::clear()
442 442 {
443 443 memset(_region_data, 0, _region_vspace->committed_size());
444 444 }
445 445
446 446 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
447 447 assert(beg_region <= _region_count, "beg_region out of range");
448 448 assert(end_region <= _region_count, "end_region out of range");
449 449
450 450 const size_t region_cnt = end_region - beg_region;
451 451 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
452 452 }
453 453
454 454 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
455 455 {
456 456 const RegionData* cur_cp = region(region_idx);
457 457 const RegionData* const end_cp = region(region_count() - 1);
458 458
459 459 HeapWord* result = region_to_addr(region_idx);
460 460 if (cur_cp < end_cp) {
461 461 do {
462 462 result += cur_cp->partial_obj_size();
463 463 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
464 464 }
465 465 return result;
466 466 }
467 467
468 468 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
469 469 {
470 470 const size_t obj_ofs = pointer_delta(addr, _region_start);
471 471 const size_t beg_region = obj_ofs >> Log2RegionSize;
472 472 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
473 473
474 474 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
475 475 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
476 476
477 477 if (beg_region == end_region) {
478 478 // All in one region.
479 479 _region_data[beg_region].add_live_obj(len);
480 480 return;
481 481 }
482 482
483 483 // First region.
484 484 const size_t beg_ofs = region_offset(addr);
485 485 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
486 486
487 487 klassOop klass = ((oop)addr)->klass();
488 488 // Middle regions--completely spanned by this object.
489 489 for (size_t region = beg_region + 1; region < end_region; ++region) {
490 490 _region_data[region].set_partial_obj_size(RegionSize);
491 491 _region_data[region].set_partial_obj_addr(addr);
492 492 }
493 493
494 494 // Last region.
495 495 const size_t end_ofs = region_offset(addr + len - 1);
496 496 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
497 497 _region_data[end_region].set_partial_obj_addr(addr);
498 498 }
499 499
500 500 void
501 501 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
502 502 {
503 503 assert(region_offset(beg) == 0, "not RegionSize aligned");
504 504 assert(region_offset(end) == 0, "not RegionSize aligned");
505 505
506 506 size_t cur_region = addr_to_region_idx(beg);
507 507 const size_t end_region = addr_to_region_idx(end);
508 508 HeapWord* addr = beg;
509 509 while (cur_region < end_region) {
510 510 _region_data[cur_region].set_destination(addr);
511 511 _region_data[cur_region].set_destination_count(0);
512 512 _region_data[cur_region].set_source_region(cur_region);
513 513 _region_data[cur_region].set_data_location(addr);
514 514
515 515 // Update live_obj_size so the region appears completely full.
516 516 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
517 517 _region_data[cur_region].set_live_obj_size(live_size);
518 518
519 519 ++cur_region;
520 520 addr += RegionSize;
521 521 }
522 522 }
523 523
524 524 // Find the point at which a space can be split and, if necessary, record the
525 525 // split point.
526 526 //
527 527 // If the current src region (which overflowed the destination space) doesn't
528 528 // have a partial object, the split point is at the beginning of the current src
529 529 // region (an "easy" split, no extra bookkeeping required).
530 530 //
531 531 // If the current src region has a partial object, the split point is in the
532 532 // region where that partial object starts (call it the split_region). If
533 533 // split_region has a partial object, then the split point is just after that
534 534 // partial object (a "hard" split where we have to record the split data and
535 535 // zero the partial_obj_size field). With a "hard" split, we know that the
536 536 // partial_obj ends within split_region because the partial object that caused
537 537 // the overflow starts in split_region. If split_region doesn't have a partial
538 538 // obj, then the split is at the beginning of split_region (another "easy"
539 539 // split).
540 540 HeapWord*
541 541 ParallelCompactData::summarize_split_space(size_t src_region,
542 542 SplitInfo& split_info,
543 543 HeapWord* destination,
544 544 HeapWord* target_end,
545 545 HeapWord** target_next)
546 546 {
547 547 assert(destination <= target_end, "sanity");
548 548 assert(destination + _region_data[src_region].data_size() > target_end,
549 549 "region should not fit into target space");
550 550 assert(is_region_aligned(target_end), "sanity");
551 551
552 552 size_t split_region = src_region;
553 553 HeapWord* split_destination = destination;
554 554 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
555 555
556 556 if (destination + partial_obj_size > target_end) {
557 557 // The split point is just after the partial object (if any) in the
558 558 // src_region that contains the start of the object that overflowed the
559 559 // destination space.
560 560 //
561 561 // Find the start of the "overflow" object and set split_region to the
562 562 // region containing it.
563 563 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
564 564 split_region = addr_to_region_idx(overflow_obj);
565 565
566 566 // Clear the source_region field of all destination regions whose first word
567 567 // came from data after the split point (a non-null source_region field
568 568 // implies a region must be filled).
569 569 //
570 570 // An alternative to the simple loop below: clear during post_compact(),
571 571 // which uses memcpy instead of individual stores, and is easy to
572 572 // parallelize. (The downside is that it clears the entire RegionData
573 573 // object as opposed to just one field.)
574 574 //
575 575 // post_compact() would have to clear the summary data up to the highest
576 576 // address that was written during the summary phase, which would be
577 577 //
578 578 // max(top, max(new_top, clear_top))
579 579 //
580 580 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
581 581 // to target_end.
582 582 const RegionData* const sr = region(split_region);
583 583 const size_t beg_idx =
584 584 addr_to_region_idx(region_align_up(sr->destination() +
585 585 sr->partial_obj_size()));
586 586 const size_t end_idx = addr_to_region_idx(target_end);
587 587
588 588 if (TraceParallelOldGCSummaryPhase) {
589 589 gclog_or_tty->print_cr("split: clearing source_region field in ["
590 590 SIZE_FORMAT ", " SIZE_FORMAT ")",
591 591 beg_idx, end_idx);
592 592 }
593 593 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
594 594 _region_data[idx].set_source_region(0);
595 595 }
596 596
597 597 // Set split_destination and partial_obj_size to reflect the split region.
598 598 split_destination = sr->destination();
599 599 partial_obj_size = sr->partial_obj_size();
600 600 }
601 601
602 602 // The split is recorded only if a partial object extends onto the region.
603 603 if (partial_obj_size != 0) {
604 604 _region_data[split_region].set_partial_obj_size(0);
605 605 split_info.record(split_region, partial_obj_size, split_destination);
606 606 }
607 607
608 608 // Setup the continuation addresses.
609 609 *target_next = split_destination + partial_obj_size;
610 610 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
611 611
612 612 if (TraceParallelOldGCSummaryPhase) {
613 613 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
614 614 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
615 615 " pos=" SIZE_FORMAT,
616 616 split_type, source_next, split_region,
617 617 partial_obj_size);
618 618 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
619 619 " tn=" PTR_FORMAT,
620 620 split_type, split_destination,
621 621 addr_to_region_idx(split_destination),
622 622 *target_next);
623 623
624 624 if (partial_obj_size != 0) {
625 625 HeapWord* const po_beg = split_info.destination();
626 626 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
627 627 gclog_or_tty->print_cr("%s split: "
628 628 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
629 629 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
630 630 split_type,
631 631 po_beg, addr_to_region_idx(po_beg),
632 632 po_end, addr_to_region_idx(po_end));
633 633 }
634 634 }
635 635
636 636 return source_next;
637 637 }
638 638
639 639 bool ParallelCompactData::summarize(SplitInfo& split_info,
640 640 HeapWord* source_beg, HeapWord* source_end,
641 641 HeapWord** source_next,
642 642 HeapWord* target_beg, HeapWord* target_end,
643 643 HeapWord** target_next)
644 644 {
645 645 if (TraceParallelOldGCSummaryPhase) {
646 646 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
647 647 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
648 648 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
649 649 source_beg, source_end, source_next_val,
650 650 target_beg, target_end, *target_next);
651 651 }
652 652
653 653 size_t cur_region = addr_to_region_idx(source_beg);
654 654 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
655 655
656 656 HeapWord *dest_addr = target_beg;
657 657 while (cur_region < end_region) {
658 658 // The destination must be set even if the region has no data.
659 659 _region_data[cur_region].set_destination(dest_addr);
660 660
661 661 size_t words = _region_data[cur_region].data_size();
662 662 if (words > 0) {
663 663 // If cur_region does not fit entirely into the target space, find a point
664 664 // at which the source space can be 'split' so that part is copied to the
665 665 // target space and the rest is copied elsewhere.
666 666 if (dest_addr + words > target_end) {
667 667 assert(source_next != NULL, "source_next is NULL when splitting");
668 668 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
669 669 target_end, target_next);
670 670 return false;
671 671 }
672 672
673 673 // Compute the destination_count for cur_region, and if necessary, update
674 674 // source_region for a destination region. The source_region field is
675 675 // updated if cur_region is the first (left-most) region to be copied to a
676 676 // destination region.
677 677 //
678 678 // The destination_count calculation is a bit subtle. A region that has
679 679 // data that compacts into itself does not count itself as a destination.
680 680 // This maintains the invariant that a zero count means the region is
681 681 // available and can be claimed and then filled.
682 682 uint destination_count = 0;
683 683 if (split_info.is_split(cur_region)) {
684 684 // The current region has been split: the partial object will be copied
685 685 // to one destination space and the remaining data will be copied to
686 686 // another destination space. Adjust the initial destination_count and,
687 687 // if necessary, set the source_region field if the partial object will
688 688 // cross a destination region boundary.
689 689 destination_count = split_info.destination_count();
690 690 if (destination_count == 2) {
691 691 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
692 692 _region_data[dest_idx].set_source_region(cur_region);
693 693 }
694 694 }
695 695
696 696 HeapWord* const last_addr = dest_addr + words - 1;
697 697 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
698 698 const size_t dest_region_2 = addr_to_region_idx(last_addr);
699 699
700 700 // Initially assume that the destination regions will be the same and
701 701 // adjust the value below if necessary. Under this assumption, if
702 702 // cur_region == dest_region_2, then cur_region will be compacted
703 703 // completely into itself.
704 704 destination_count += cur_region == dest_region_2 ? 0 : 1;
705 705 if (dest_region_1 != dest_region_2) {
706 706 // Destination regions differ; adjust destination_count.
707 707 destination_count += 1;
708 708 // Data from cur_region will be copied to the start of dest_region_2.
709 709 _region_data[dest_region_2].set_source_region(cur_region);
710 710 } else if (region_offset(dest_addr) == 0) {
711 711 // Data from cur_region will be copied to the start of the destination
712 712 // region.
713 713 _region_data[dest_region_1].set_source_region(cur_region);
714 714 }
715 715
716 716 _region_data[cur_region].set_destination_count(destination_count);
717 717 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
718 718 dest_addr += words;
719 719 }
720 720
721 721 ++cur_region;
722 722 }
723 723
724 724 *target_next = dest_addr;
725 725 return true;
726 726 }
727 727
728 728 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
729 729 assert(addr != NULL, "Should detect NULL oop earlier");
730 730 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
731 731 #ifdef ASSERT
732 732 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
733 733 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
734 734 }
735 735 #endif
736 736 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
737 737
738 738 // Region covering the object.
739 739 size_t region_index = addr_to_region_idx(addr);
740 740 const RegionData* const region_ptr = region(region_index);
741 741 HeapWord* const region_addr = region_align_down(addr);
742 742
743 743 assert(addr < region_addr + RegionSize, "Region does not cover object");
744 744 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
745 745
746 746 HeapWord* result = region_ptr->destination();
747 747
748 748 // If all the data in the region is live, then the new location of the object
749 749 // can be calculated from the destination of the region plus the offset of the
750 750 // object in the region.
751 751 if (region_ptr->data_size() == RegionSize) {
752 752 result += pointer_delta(addr, region_addr);
753 753 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
754 754 return result;
755 755 }
756 756
757 757 // The new location of the object is
758 758 // region destination +
759 759 // size of the partial object extending onto the region +
760 760 // sizes of the live objects in the Region that are to the left of addr
761 761 const size_t partial_obj_size = region_ptr->partial_obj_size();
762 762 HeapWord* const search_start = region_addr + partial_obj_size;
763 763
764 764 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
765 765 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
766 766
767 767 result += partial_obj_size + live_to_left;
768 768 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
769 769 return result;
770 770 }
771 771
772 772 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
773 773 klassOop updated_klass;
774 774 if (PSParallelCompact::should_update_klass(old_klass)) {
775 775 updated_klass = (klassOop) calc_new_pointer(old_klass);
776 776 } else {
777 777 updated_klass = old_klass;
778 778 }
779 779
780 780 return updated_klass;
781 781 }
782 782
783 783 #ifdef ASSERT
784 784 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
785 785 {
786 786 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
787 787 const size_t* const end = (const size_t*)vspace->committed_high_addr();
788 788 for (const size_t* p = beg; p < end; ++p) {
789 789 assert(*p == 0, "not zero");
790 790 }
791 791 }
792 792
793 793 void ParallelCompactData::verify_clear()
794 794 {
795 795 verify_clear(_region_vspace);
796 796 }
797 797 #endif // #ifdef ASSERT
798 798
799 799 #ifdef NOT_PRODUCT
800 800 ParallelCompactData::RegionData* debug_region(size_t region_index) {
801 801 ParallelCompactData& sd = PSParallelCompact::summary_data();
802 802 return sd.region(region_index);
803 803 }
804 804 #endif
805 805
806 806 STWGCTimer PSParallelCompact::_gc_timer;
807 807 ParallelOldTracer PSParallelCompact::_gc_tracer;
808 808 elapsedTimer PSParallelCompact::_accumulated_time;
809 809 unsigned int PSParallelCompact::_total_invocations = 0;
810 810 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
811 811 jlong PSParallelCompact::_time_of_last_gc = 0;
812 812 CollectorCounters* PSParallelCompact::_counters = NULL;
813 813 ParMarkBitMap PSParallelCompact::_mark_bitmap;
814 814 ParallelCompactData PSParallelCompact::_summary_data;
815 815
816 816 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
817 817
818 818 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
819 819 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
820 820
821 821 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
822 822 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
823 823
824 824 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
825 825 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
826 826
827 827 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); }
828 828 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
829 829
830 830 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
831 831
832 832 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); }
833 833 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
834 834
835 835 void PSParallelCompact::post_initialize() {
836 836 ParallelScavengeHeap* heap = gc_heap();
837 837 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
838 838
839 839 MemRegion mr = heap->reserved_region();
840 840 _ref_processor =
841 841 new ReferenceProcessor(mr, // span
842 842 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
843 843 (int) ParallelGCThreads, // mt processing degree
844 844 true, // mt discovery
845 845 (int) ParallelGCThreads, // mt discovery degree
846 846 true, // atomic_discovery
847 847 &_is_alive_closure, // non-header is alive closure
848 848 false); // write barrier for next field updates
849 849 _counters = new CollectorCounters("PSParallelCompact", 1);
850 850
851 851 // Initialize static fields in ParCompactionManager.
852 852 ParCompactionManager::initialize(mark_bitmap());
853 853 }
854 854
855 855 bool PSParallelCompact::initialize() {
856 856 ParallelScavengeHeap* heap = gc_heap();
857 857 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
858 858 MemRegion mr = heap->reserved_region();
859 859
860 860 // Was the old gen get allocated successfully?
861 861 if (!heap->old_gen()->is_allocated()) {
862 862 return false;
863 863 }
864 864
865 865 initialize_space_info();
866 866 initialize_dead_wood_limiter();
867 867
868 868 if (!_mark_bitmap.initialize(mr)) {
869 869 vm_shutdown_during_initialization("Unable to allocate bit map for "
870 870 "parallel garbage collection for the requested heap size.");
871 871 return false;
872 872 }
873 873
874 874 if (!_summary_data.initialize(mr)) {
875 875 vm_shutdown_during_initialization("Unable to allocate tables for "
876 876 "parallel garbage collection for the requested heap size.");
877 877 return false;
878 878 }
879 879
880 880 return true;
881 881 }
882 882
883 883 void PSParallelCompact::initialize_space_info()
884 884 {
885 885 memset(&_space_info, 0, sizeof(_space_info));
886 886
887 887 ParallelScavengeHeap* heap = gc_heap();
888 888 PSYoungGen* young_gen = heap->young_gen();
889 889 MutableSpace* perm_space = heap->perm_gen()->object_space();
890 890
891 891 _space_info[perm_space_id].set_space(perm_space);
892 892 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
893 893 _space_info[eden_space_id].set_space(young_gen->eden_space());
894 894 _space_info[from_space_id].set_space(young_gen->from_space());
895 895 _space_info[to_space_id].set_space(young_gen->to_space());
896 896
897 897 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
898 898 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
899 899
900 900 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
901 901 if (TraceParallelOldGCDensePrefix) {
902 902 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
903 903 _space_info[perm_space_id].min_dense_prefix());
904 904 }
905 905 }
906 906
907 907 void PSParallelCompact::initialize_dead_wood_limiter()
908 908 {
909 909 const size_t max = 100;
910 910 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
911 911 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
912 912 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
913 913 DEBUG_ONLY(_dwl_initialized = true;)
914 914 _dwl_adjustment = normal_distribution(1.0);
915 915 }
916 916
917 917 // Simple class for storing info about the heap at the start of GC, to be used
918 918 // after GC for comparison/printing.
919 919 class PreGCValues {
920 920 public:
921 921 PreGCValues() { }
922 922 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
923 923
924 924 void fill(ParallelScavengeHeap* heap) {
925 925 _heap_used = heap->used();
926 926 _young_gen_used = heap->young_gen()->used_in_bytes();
927 927 _old_gen_used = heap->old_gen()->used_in_bytes();
928 928 _perm_gen_used = heap->perm_gen()->used_in_bytes();
929 929 };
930 930
931 931 size_t heap_used() const { return _heap_used; }
932 932 size_t young_gen_used() const { return _young_gen_used; }
933 933 size_t old_gen_used() const { return _old_gen_used; }
934 934 size_t perm_gen_used() const { return _perm_gen_used; }
935 935
936 936 private:
937 937 size_t _heap_used;
938 938 size_t _young_gen_used;
939 939 size_t _old_gen_used;
940 940 size_t _perm_gen_used;
941 941 };
942 942
943 943 void
944 944 PSParallelCompact::clear_data_covering_space(SpaceId id)
945 945 {
946 946 // At this point, top is the value before GC, new_top() is the value that will
947 947 // be set at the end of GC. The marking bitmap is cleared to top; nothing
948 948 // should be marked above top. The summary data is cleared to the larger of
949 949 // top & new_top.
950 950 MutableSpace* const space = _space_info[id].space();
951 951 HeapWord* const bot = space->bottom();
952 952 HeapWord* const top = space->top();
953 953 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
954 954
955 955 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
956 956 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
957 957 _mark_bitmap.clear_range(beg_bit, end_bit);
958 958
959 959 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
960 960 const size_t end_region =
961 961 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
962 962 _summary_data.clear_range(beg_region, end_region);
963 963
964 964 // Clear the data used to 'split' regions.
965 965 SplitInfo& split_info = _space_info[id].split_info();
966 966 if (split_info.is_valid()) {
967 967 split_info.clear();
968 968 }
969 969 DEBUG_ONLY(split_info.verify_clear();)
970 970 }
971 971
972 972 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
973 973 {
974 974 // Update the from & to space pointers in space_info, since they are swapped
975 975 // at each young gen gc. Do the update unconditionally (even though a
976 976 // promotion failure does not swap spaces) because an unknown number of minor
977 977 // collections will have swapped the spaces an unknown number of times.
978 978 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer);
979 979 ParallelScavengeHeap* heap = gc_heap();
980 980 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
981 981 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
982 982
983 983 pre_gc_values->fill(heap);
984 984
985 985 ParCompactionManager::reset();
986 986 NOT_PRODUCT(_mark_bitmap.reset_counters());
987 987 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
988 988 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
989 989
990 990 // Increment the invocation count
991 991 heap->increment_total_collections(true);
992 992
993 993 // We need to track unique mark sweep invocations as well.
994 994 _total_invocations++;
995 995
996 996 heap->print_heap_before_gc();
997 997 heap->trace_heap_before_gc(&_gc_tracer);
998 998
999 999 // Fill in TLABs
1000 1000 heap->accumulate_statistics_all_tlabs();
1001 1001 heap->ensure_parsability(true); // retire TLABs
1002 1002
1003 1003 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1004 1004 HandleMark hm; // Discard invalid handles created during verification
1005 1005 gclog_or_tty->print(" VerifyBeforeGC:");
1006 1006 Universe::verify();
1007 1007 }
1008 1008
1009 1009 // Verify object start arrays
1010 1010 if (VerifyObjectStartArray &&
1011 1011 VerifyBeforeGC) {
1012 1012 heap->old_gen()->verify_object_start_array();
1013 1013 heap->perm_gen()->verify_object_start_array();
1014 1014 }
1015 1015
1016 1016 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1017 1017 DEBUG_ONLY(summary_data().verify_clear();)
1018 1018
1019 1019 // Have worker threads release resources the next time they run a task.
1020 1020 gc_task_manager()->release_all_resources();
1021 1021 }
1022 1022
1023 1023 void PSParallelCompact::post_compact()
1024 1024 {
1025 1025 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer);
1026 1026
1027 1027 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1028 1028 // Clear the marking bitmap, summary data and split info.
1029 1029 clear_data_covering_space(SpaceId(id));
1030 1030 // Update top(). Must be done after clearing the bitmap and summary data.
1031 1031 _space_info[id].publish_new_top();
1032 1032 }
1033 1033
1034 1034 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1035 1035 MutableSpace* const from_space = _space_info[from_space_id].space();
1036 1036 MutableSpace* const to_space = _space_info[to_space_id].space();
1037 1037
1038 1038 ParallelScavengeHeap* heap = gc_heap();
1039 1039 bool eden_empty = eden_space->is_empty();
1040 1040 if (!eden_empty) {
1041 1041 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1042 1042 heap->young_gen(), heap->old_gen());
1043 1043 }
1044 1044
1045 1045 // Update heap occupancy information which is used as input to the soft ref
1046 1046 // clearing policy at the next gc.
1047 1047 Universe::update_heap_info_at_gc();
1048 1048
1049 1049 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1050 1050 to_space->is_empty();
1051 1051
1052 1052 BarrierSet* bs = heap->barrier_set();
1053 1053 if (bs->is_a(BarrierSet::ModRef)) {
1054 1054 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1055 1055 MemRegion old_mr = heap->old_gen()->reserved();
1056 1056 MemRegion perm_mr = heap->perm_gen()->reserved();
1057 1057 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1058 1058
1059 1059 if (young_gen_empty) {
1060 1060 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1061 1061 } else {
1062 1062 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1063 1063 }
1064 1064 }
1065 1065
1066 1066 Threads::gc_epilogue();
1067 1067 CodeCache::gc_epilogue();
1068 1068 JvmtiExport::gc_epilogue();
1069 1069
1070 1070 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1071 1071
1072 1072 ref_processor()->enqueue_discovered_references(NULL);
1073 1073
1074 1074 if (ZapUnusedHeapArea) {
1075 1075 heap->gen_mangle_unused_area();
1076 1076 }
1077 1077
1078 1078 // Update time of last GC
1079 1079 reset_millis_since_last_gc();
1080 1080 }
1081 1081
1082 1082 HeapWord*
1083 1083 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1084 1084 bool maximum_compaction)
1085 1085 {
1086 1086 const size_t region_size = ParallelCompactData::RegionSize;
1087 1087 const ParallelCompactData& sd = summary_data();
1088 1088
1089 1089 const MutableSpace* const space = _space_info[id].space();
1090 1090 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1091 1091 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1092 1092 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1093 1093
1094 1094 // Skip full regions at the beginning of the space--they are necessarily part
1095 1095 // of the dense prefix.
1096 1096 size_t full_count = 0;
1097 1097 const RegionData* cp;
1098 1098 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1099 1099 ++full_count;
1100 1100 }
1101 1101
1102 1102 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1103 1103 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1104 1104 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1105 1105 if (maximum_compaction || cp == end_cp || interval_ended) {
1106 1106 _maximum_compaction_gc_num = total_invocations();
1107 1107 return sd.region_to_addr(cp);
1108 1108 }
1109 1109
1110 1110 HeapWord* const new_top = _space_info[id].new_top();
1111 1111 const size_t space_live = pointer_delta(new_top, space->bottom());
1112 1112 const size_t space_used = space->used_in_words();
1113 1113 const size_t space_capacity = space->capacity_in_words();
1114 1114
1115 1115 const double cur_density = double(space_live) / space_capacity;
1116 1116 const double deadwood_density =
1117 1117 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1118 1118 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1119 1119
1120 1120 if (TraceParallelOldGCDensePrefix) {
1121 1121 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1122 1122 cur_density, deadwood_density, deadwood_goal);
1123 1123 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1124 1124 "space_cap=" SIZE_FORMAT,
1125 1125 space_live, space_used,
1126 1126 space_capacity);
1127 1127 }
1128 1128
1129 1129 // XXX - Use binary search?
1130 1130 HeapWord* dense_prefix = sd.region_to_addr(cp);
1131 1131 const RegionData* full_cp = cp;
1132 1132 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1133 1133 while (cp < end_cp) {
1134 1134 HeapWord* region_destination = cp->destination();
1135 1135 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1136 1136 if (TraceParallelOldGCDensePrefix && Verbose) {
1137 1137 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1138 1138 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1139 1139 sd.region(cp), region_destination,
1140 1140 dense_prefix, cur_deadwood);
1141 1141 }
1142 1142
1143 1143 if (cur_deadwood >= deadwood_goal) {
1144 1144 // Found the region that has the correct amount of deadwood to the left.
1145 1145 // This typically occurs after crossing a fairly sparse set of regions, so
1146 1146 // iterate backwards over those sparse regions, looking for the region
1147 1147 // that has the lowest density of live objects 'to the right.'
1148 1148 size_t space_to_left = sd.region(cp) * region_size;
1149 1149 size_t live_to_left = space_to_left - cur_deadwood;
1150 1150 size_t space_to_right = space_capacity - space_to_left;
1151 1151 size_t live_to_right = space_live - live_to_left;
1152 1152 double density_to_right = double(live_to_right) / space_to_right;
1153 1153 while (cp > full_cp) {
1154 1154 --cp;
1155 1155 const size_t prev_region_live_to_right = live_to_right -
1156 1156 cp->data_size();
1157 1157 const size_t prev_region_space_to_right = space_to_right + region_size;
1158 1158 double prev_region_density_to_right =
1159 1159 double(prev_region_live_to_right) / prev_region_space_to_right;
1160 1160 if (density_to_right <= prev_region_density_to_right) {
1161 1161 return dense_prefix;
1162 1162 }
1163 1163 if (TraceParallelOldGCDensePrefix && Verbose) {
1164 1164 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1165 1165 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1166 1166 prev_region_density_to_right);
1167 1167 }
1168 1168 dense_prefix -= region_size;
1169 1169 live_to_right = prev_region_live_to_right;
1170 1170 space_to_right = prev_region_space_to_right;
1171 1171 density_to_right = prev_region_density_to_right;
1172 1172 }
1173 1173 return dense_prefix;
1174 1174 }
1175 1175
1176 1176 dense_prefix += region_size;
1177 1177 ++cp;
1178 1178 }
1179 1179
1180 1180 return dense_prefix;
1181 1181 }
1182 1182
1183 1183 #ifndef PRODUCT
1184 1184 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1185 1185 const SpaceId id,
1186 1186 const bool maximum_compaction,
1187 1187 HeapWord* const addr)
1188 1188 {
1189 1189 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1190 1190 RegionData* const cp = summary_data().region(region_idx);
1191 1191 const MutableSpace* const space = _space_info[id].space();
1192 1192 HeapWord* const new_top = _space_info[id].new_top();
1193 1193
1194 1194 const size_t space_live = pointer_delta(new_top, space->bottom());
1195 1195 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1196 1196 const size_t space_cap = space->capacity_in_words();
1197 1197 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1198 1198 const size_t live_to_right = new_top - cp->destination();
1199 1199 const size_t dead_to_right = space->top() - addr - live_to_right;
1200 1200
1201 1201 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1202 1202 "spl=" SIZE_FORMAT " "
1203 1203 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1204 1204 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1205 1205 " ratio=%10.8f",
1206 1206 algorithm, addr, region_idx,
1207 1207 space_live,
1208 1208 dead_to_left, dead_to_left_pct,
1209 1209 dead_to_right, live_to_right,
1210 1210 double(dead_to_right) / live_to_right);
1211 1211 }
1212 1212 #endif // #ifndef PRODUCT
1213 1213
1214 1214 // Return a fraction indicating how much of the generation can be treated as
1215 1215 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1216 1216 // based on the density of live objects in the generation to determine a limit,
1217 1217 // which is then adjusted so the return value is min_percent when the density is
1218 1218 // 1.
1219 1219 //
1220 1220 // The following table shows some return values for a different values of the
1221 1221 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1222 1222 // min_percent is 1.
1223 1223 //
1224 1224 // fraction allowed as dead wood
1225 1225 // -----------------------------------------------------------------
1226 1226 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1227 1227 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1228 1228 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1229 1229 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1230 1230 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1231 1231 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1232 1232 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1233 1233 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1234 1234 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1235 1235 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1236 1236 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1237 1237 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1238 1238 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1239 1239 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1240 1240 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1241 1241 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1242 1242 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1243 1243 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1244 1244 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1245 1245 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1246 1246 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1247 1247 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1248 1248 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1249 1249
1250 1250 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1251 1251 {
1252 1252 assert(_dwl_initialized, "uninitialized");
1253 1253
1254 1254 // The raw limit is the value of the normal distribution at x = density.
1255 1255 const double raw_limit = normal_distribution(density);
1256 1256
1257 1257 // Adjust the raw limit so it becomes the minimum when the density is 1.
1258 1258 //
1259 1259 // First subtract the adjustment value (which is simply the precomputed value
1260 1260 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1261 1261 // Then add the minimum value, so the minimum is returned when the density is
1262 1262 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1263 1263 const double min = double(min_percent) / 100.0;
1264 1264 const double limit = raw_limit - _dwl_adjustment + min;
1265 1265 return MAX2(limit, 0.0);
1266 1266 }
1267 1267
1268 1268 ParallelCompactData::RegionData*
1269 1269 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1270 1270 const RegionData* end)
1271 1271 {
1272 1272 const size_t region_size = ParallelCompactData::RegionSize;
1273 1273 ParallelCompactData& sd = summary_data();
1274 1274 size_t left = sd.region(beg);
1275 1275 size_t right = end > beg ? sd.region(end) - 1 : left;
1276 1276
1277 1277 // Binary search.
1278 1278 while (left < right) {
1279 1279 // Equivalent to (left + right) / 2, but does not overflow.
1280 1280 const size_t middle = left + (right - left) / 2;
1281 1281 RegionData* const middle_ptr = sd.region(middle);
1282 1282 HeapWord* const dest = middle_ptr->destination();
1283 1283 HeapWord* const addr = sd.region_to_addr(middle);
1284 1284 assert(dest != NULL, "sanity");
1285 1285 assert(dest <= addr, "must move left");
1286 1286
1287 1287 if (middle > left && dest < addr) {
1288 1288 right = middle - 1;
1289 1289 } else if (middle < right && middle_ptr->data_size() == region_size) {
1290 1290 left = middle + 1;
1291 1291 } else {
1292 1292 return middle_ptr;
1293 1293 }
1294 1294 }
1295 1295 return sd.region(left);
1296 1296 }
1297 1297
1298 1298 ParallelCompactData::RegionData*
1299 1299 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1300 1300 const RegionData* end,
1301 1301 size_t dead_words)
1302 1302 {
1303 1303 ParallelCompactData& sd = summary_data();
1304 1304 size_t left = sd.region(beg);
1305 1305 size_t right = end > beg ? sd.region(end) - 1 : left;
1306 1306
1307 1307 // Binary search.
1308 1308 while (left < right) {
1309 1309 // Equivalent to (left + right) / 2, but does not overflow.
1310 1310 const size_t middle = left + (right - left) / 2;
1311 1311 RegionData* const middle_ptr = sd.region(middle);
1312 1312 HeapWord* const dest = middle_ptr->destination();
1313 1313 HeapWord* const addr = sd.region_to_addr(middle);
1314 1314 assert(dest != NULL, "sanity");
1315 1315 assert(dest <= addr, "must move left");
1316 1316
1317 1317 const size_t dead_to_left = pointer_delta(addr, dest);
1318 1318 if (middle > left && dead_to_left > dead_words) {
1319 1319 right = middle - 1;
1320 1320 } else if (middle < right && dead_to_left < dead_words) {
1321 1321 left = middle + 1;
1322 1322 } else {
1323 1323 return middle_ptr;
1324 1324 }
1325 1325 }
1326 1326 return sd.region(left);
1327 1327 }
1328 1328
1329 1329 // The result is valid during the summary phase, after the initial summarization
1330 1330 // of each space into itself, and before final summarization.
1331 1331 inline double
1332 1332 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1333 1333 HeapWord* const bottom,
1334 1334 HeapWord* const top,
1335 1335 HeapWord* const new_top)
1336 1336 {
1337 1337 ParallelCompactData& sd = summary_data();
1338 1338
1339 1339 assert(cp != NULL, "sanity");
1340 1340 assert(bottom != NULL, "sanity");
1341 1341 assert(top != NULL, "sanity");
1342 1342 assert(new_top != NULL, "sanity");
1343 1343 assert(top >= new_top, "summary data problem?");
1344 1344 assert(new_top > bottom, "space is empty; should not be here");
1345 1345 assert(new_top >= cp->destination(), "sanity");
1346 1346 assert(top >= sd.region_to_addr(cp), "sanity");
1347 1347
1348 1348 HeapWord* const destination = cp->destination();
1349 1349 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1350 1350 const size_t compacted_region_live = pointer_delta(new_top, destination);
1351 1351 const size_t compacted_region_used = pointer_delta(top,
1352 1352 sd.region_to_addr(cp));
1353 1353 const size_t reclaimable = compacted_region_used - compacted_region_live;
1354 1354
1355 1355 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1356 1356 return double(reclaimable) / divisor;
1357 1357 }
1358 1358
1359 1359 // Return the address of the end of the dense prefix, a.k.a. the start of the
1360 1360 // compacted region. The address is always on a region boundary.
1361 1361 //
1362 1362 // Completely full regions at the left are skipped, since no compaction can
1363 1363 // occur in those regions. Then the maximum amount of dead wood to allow is
1364 1364 // computed, based on the density (amount live / capacity) of the generation;
1365 1365 // the region with approximately that amount of dead space to the left is
1366 1366 // identified as the limit region. Regions between the last completely full
1367 1367 // region and the limit region are scanned and the one that has the best
1368 1368 // (maximum) reclaimed_ratio() is selected.
1369 1369 HeapWord*
1370 1370 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1371 1371 bool maximum_compaction)
1372 1372 {
1373 1373 if (ParallelOldGCSplitALot) {
1374 1374 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1375 1375 // The value was chosen to provoke splitting a young gen space; use it.
1376 1376 return _space_info[id].dense_prefix();
1377 1377 }
1378 1378 }
1379 1379
1380 1380 const size_t region_size = ParallelCompactData::RegionSize;
1381 1381 const ParallelCompactData& sd = summary_data();
1382 1382
1383 1383 const MutableSpace* const space = _space_info[id].space();
1384 1384 HeapWord* const top = space->top();
1385 1385 HeapWord* const top_aligned_up = sd.region_align_up(top);
1386 1386 HeapWord* const new_top = _space_info[id].new_top();
1387 1387 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1388 1388 HeapWord* const bottom = space->bottom();
1389 1389 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1390 1390 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1391 1391 const RegionData* const new_top_cp =
1392 1392 sd.addr_to_region_ptr(new_top_aligned_up);
1393 1393
1394 1394 // Skip full regions at the beginning of the space--they are necessarily part
1395 1395 // of the dense prefix.
1396 1396 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1397 1397 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1398 1398 space->is_empty(), "no dead space allowed to the left");
1399 1399 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1400 1400 "region must have dead space");
1401 1401
1402 1402 // The gc number is saved whenever a maximum compaction is done, and used to
1403 1403 // determine when the maximum compaction interval has expired. This avoids
1404 1404 // successive max compactions for different reasons.
1405 1405 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1406 1406 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1407 1407 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1408 1408 total_invocations() == HeapFirstMaximumCompactionCount;
1409 1409 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1410 1410 _maximum_compaction_gc_num = total_invocations();
1411 1411 return sd.region_to_addr(full_cp);
1412 1412 }
1413 1413
1414 1414 const size_t space_live = pointer_delta(new_top, bottom);
1415 1415 const size_t space_used = space->used_in_words();
1416 1416 const size_t space_capacity = space->capacity_in_words();
1417 1417
1418 1418 const double density = double(space_live) / double(space_capacity);
1419 1419 const size_t min_percent_free =
1420 1420 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1421 1421 const double limiter = dead_wood_limiter(density, min_percent_free);
1422 1422 const size_t dead_wood_max = space_used - space_live;
1423 1423 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1424 1424 dead_wood_max);
1425 1425
1426 1426 if (TraceParallelOldGCDensePrefix) {
1427 1427 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1428 1428 "space_cap=" SIZE_FORMAT,
1429 1429 space_live, space_used,
1430 1430 space_capacity);
1431 1431 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1432 1432 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1433 1433 density, min_percent_free, limiter,
1434 1434 dead_wood_max, dead_wood_limit);
1435 1435 }
1436 1436
1437 1437 // Locate the region with the desired amount of dead space to the left.
1438 1438 const RegionData* const limit_cp =
1439 1439 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1440 1440
1441 1441 // Scan from the first region with dead space to the limit region and find the
1442 1442 // one with the best (largest) reclaimed ratio.
1443 1443 double best_ratio = 0.0;
1444 1444 const RegionData* best_cp = full_cp;
1445 1445 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1446 1446 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1447 1447 if (tmp_ratio > best_ratio) {
1448 1448 best_cp = cp;
1449 1449 best_ratio = tmp_ratio;
1450 1450 }
1451 1451 }
1452 1452
1453 1453 #if 0
1454 1454 // Something to consider: if the region with the best ratio is 'close to' the
1455 1455 // first region w/free space, choose the first region with free space
1456 1456 // ("first-free"). The first-free region is usually near the start of the
1457 1457 // heap, which means we are copying most of the heap already, so copy a bit
1458 1458 // more to get complete compaction.
1459 1459 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1460 1460 _maximum_compaction_gc_num = total_invocations();
1461 1461 best_cp = full_cp;
1462 1462 }
1463 1463 #endif // #if 0
1464 1464
1465 1465 return sd.region_to_addr(best_cp);
1466 1466 }
1467 1467
1468 1468 #ifndef PRODUCT
1469 1469 void
1470 1470 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1471 1471 size_t words)
1472 1472 {
1473 1473 if (TraceParallelOldGCSummaryPhase) {
1474 1474 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1475 1475 SIZE_FORMAT, start, start + words, words);
1476 1476 }
1477 1477
1478 1478 ObjectStartArray* const start_array = _space_info[id].start_array();
1479 1479 CollectedHeap::fill_with_objects(start, words);
1480 1480 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1481 1481 _mark_bitmap.mark_obj(p, words);
1482 1482 _summary_data.add_obj(p, words);
1483 1483 start_array->allocate_block(p);
1484 1484 }
1485 1485 }
1486 1486
1487 1487 void
1488 1488 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1489 1489 {
1490 1490 ParallelCompactData& sd = summary_data();
1491 1491 MutableSpace* space = _space_info[id].space();
1492 1492
1493 1493 // Find the source and destination start addresses.
1494 1494 HeapWord* const src_addr = sd.region_align_down(start);
1495 1495 HeapWord* dst_addr;
1496 1496 if (src_addr < start) {
1497 1497 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1498 1498 } else if (src_addr > space->bottom()) {
1499 1499 // The start (the original top() value) is aligned to a region boundary so
1500 1500 // the associated region does not have a destination. Compute the
1501 1501 // destination from the previous region.
1502 1502 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1503 1503 dst_addr = cp->destination() + cp->data_size();
1504 1504 } else {
1505 1505 // Filling the entire space.
1506 1506 dst_addr = space->bottom();
1507 1507 }
1508 1508 assert(dst_addr != NULL, "sanity");
1509 1509
1510 1510 // Update the summary data.
1511 1511 bool result = _summary_data.summarize(_space_info[id].split_info(),
1512 1512 src_addr, space->top(), NULL,
1513 1513 dst_addr, space->end(),
1514 1514 _space_info[id].new_top_addr());
1515 1515 assert(result, "should not fail: bad filler object size");
1516 1516 }
1517 1517
1518 1518 void
1519 1519 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1520 1520 {
1521 1521 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1522 1522 return;
1523 1523 }
1524 1524
1525 1525 MutableSpace* const space = _space_info[id].space();
1526 1526 if (space->is_empty()) {
1527 1527 HeapWord* b = space->bottom();
1528 1528 HeapWord* t = b + space->capacity_in_words() / 2;
1529 1529 space->set_top(t);
1530 1530 if (ZapUnusedHeapArea) {
1531 1531 space->set_top_for_allocations();
1532 1532 }
1533 1533
1534 1534 size_t min_size = CollectedHeap::min_fill_size();
1535 1535 size_t obj_len = min_size;
1536 1536 while (b + obj_len <= t) {
1537 1537 CollectedHeap::fill_with_object(b, obj_len);
1538 1538 mark_bitmap()->mark_obj(b, obj_len);
1539 1539 summary_data().add_obj(b, obj_len);
1540 1540 b += obj_len;
1541 1541 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1542 1542 }
1543 1543 if (b < t) {
1544 1544 // The loop didn't completely fill to t (top); adjust top downward.
1545 1545 space->set_top(b);
1546 1546 if (ZapUnusedHeapArea) {
1547 1547 space->set_top_for_allocations();
1548 1548 }
1549 1549 }
1550 1550
1551 1551 HeapWord** nta = _space_info[id].new_top_addr();
1552 1552 bool result = summary_data().summarize(_space_info[id].split_info(),
1553 1553 space->bottom(), space->top(), NULL,
1554 1554 space->bottom(), space->end(), nta);
1555 1555 assert(result, "space must fit into itself");
1556 1556 }
1557 1557 }
1558 1558
1559 1559 void
1560 1560 PSParallelCompact::provoke_split(bool & max_compaction)
1561 1561 {
1562 1562 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1563 1563 return;
1564 1564 }
1565 1565
1566 1566 const size_t region_size = ParallelCompactData::RegionSize;
1567 1567 ParallelCompactData& sd = summary_data();
1568 1568
1569 1569 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1570 1570 MutableSpace* const from_space = _space_info[from_space_id].space();
1571 1571 const size_t eden_live = pointer_delta(eden_space->top(),
1572 1572 _space_info[eden_space_id].new_top());
1573 1573 const size_t from_live = pointer_delta(from_space->top(),
1574 1574 _space_info[from_space_id].new_top());
1575 1575
1576 1576 const size_t min_fill_size = CollectedHeap::min_fill_size();
1577 1577 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1578 1578 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1579 1579 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1580 1580 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1581 1581
1582 1582 // Choose the space to split; need at least 2 regions live (or fillable).
1583 1583 SpaceId id;
1584 1584 MutableSpace* space;
1585 1585 size_t live_words;
1586 1586 size_t fill_words;
1587 1587 if (eden_live + eden_fillable >= region_size * 2) {
1588 1588 id = eden_space_id;
1589 1589 space = eden_space;
1590 1590 live_words = eden_live;
1591 1591 fill_words = eden_fillable;
1592 1592 } else if (from_live + from_fillable >= region_size * 2) {
1593 1593 id = from_space_id;
1594 1594 space = from_space;
1595 1595 live_words = from_live;
1596 1596 fill_words = from_fillable;
1597 1597 } else {
1598 1598 return; // Give up.
1599 1599 }
1600 1600 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1601 1601
1602 1602 if (live_words < region_size * 2) {
1603 1603 // Fill from top() to end() w/live objects of mixed sizes.
1604 1604 HeapWord* const fill_start = space->top();
1605 1605 live_words += fill_words;
1606 1606
1607 1607 space->set_top(fill_start + fill_words);
1608 1608 if (ZapUnusedHeapArea) {
1609 1609 space->set_top_for_allocations();
1610 1610 }
1611 1611
1612 1612 HeapWord* cur_addr = fill_start;
1613 1613 while (fill_words > 0) {
1614 1614 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1615 1615 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1616 1616 if (fill_words - cur_size < min_fill_size) {
1617 1617 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1618 1618 }
1619 1619
1620 1620 CollectedHeap::fill_with_object(cur_addr, cur_size);
1621 1621 mark_bitmap()->mark_obj(cur_addr, cur_size);
1622 1622 sd.add_obj(cur_addr, cur_size);
1623 1623
1624 1624 cur_addr += cur_size;
1625 1625 fill_words -= cur_size;
1626 1626 }
1627 1627
1628 1628 summarize_new_objects(id, fill_start);
1629 1629 }
1630 1630
1631 1631 max_compaction = false;
1632 1632
1633 1633 // Manipulate the old gen so that it has room for about half of the live data
1634 1634 // in the target young gen space (live_words / 2).
1635 1635 id = old_space_id;
1636 1636 space = _space_info[id].space();
1637 1637 const size_t free_at_end = space->free_in_words();
1638 1638 const size_t free_target = align_object_size(live_words / 2);
1639 1639 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1640 1640
1641 1641 if (free_at_end >= free_target + min_fill_size) {
1642 1642 // Fill space above top() and set the dense prefix so everything survives.
1643 1643 HeapWord* const fill_start = space->top();
1644 1644 const size_t fill_size = free_at_end - free_target;
1645 1645 space->set_top(space->top() + fill_size);
1646 1646 if (ZapUnusedHeapArea) {
1647 1647 space->set_top_for_allocations();
1648 1648 }
1649 1649 fill_with_live_objects(id, fill_start, fill_size);
1650 1650 summarize_new_objects(id, fill_start);
1651 1651 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1652 1652 } else if (dead + free_at_end > free_target) {
1653 1653 // Find a dense prefix that makes the right amount of space available.
1654 1654 HeapWord* cur = sd.region_align_down(space->top());
1655 1655 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1656 1656 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1657 1657 while (dead_to_right < free_target) {
1658 1658 cur -= region_size;
1659 1659 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1660 1660 dead_to_right = pointer_delta(space->end(), cur_destination);
1661 1661 }
1662 1662 _space_info[id].set_dense_prefix(cur);
1663 1663 }
1664 1664 }
1665 1665 #endif // #ifndef PRODUCT
1666 1666
1667 1667 void PSParallelCompact::summarize_spaces_quick()
1668 1668 {
1669 1669 for (unsigned int i = 0; i < last_space_id; ++i) {
1670 1670 const MutableSpace* space = _space_info[i].space();
1671 1671 HeapWord** nta = _space_info[i].new_top_addr();
1672 1672 bool result = _summary_data.summarize(_space_info[i].split_info(),
1673 1673 space->bottom(), space->top(), NULL,
1674 1674 space->bottom(), space->end(), nta);
1675 1675 assert(result, "space must fit into itself");
1676 1676 _space_info[i].set_dense_prefix(space->bottom());
1677 1677 }
1678 1678
1679 1679 #ifndef PRODUCT
1680 1680 if (ParallelOldGCSplitALot) {
1681 1681 provoke_split_fill_survivor(to_space_id);
1682 1682 }
1683 1683 #endif // #ifndef PRODUCT
1684 1684 }
1685 1685
1686 1686 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1687 1687 {
1688 1688 HeapWord* const dense_prefix_end = dense_prefix(id);
1689 1689 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1690 1690 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1691 1691 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1692 1692 // Only enough dead space is filled so that any remaining dead space to the
1693 1693 // left is larger than the minimum filler object. (The remainder is filled
1694 1694 // during the copy/update phase.)
1695 1695 //
1696 1696 // The size of the dead space to the right of the boundary is not a
1697 1697 // concern, since compaction will be able to use whatever space is
1698 1698 // available.
1699 1699 //
1700 1700 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1701 1701 // surrounds the space to be filled with an object.
1702 1702 //
1703 1703 // In the 32-bit VM, each bit represents two 32-bit words:
1704 1704 // +---+
1705 1705 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1706 1706 // end_bits: ... x x x | 0 | || 0 x x ...
1707 1707 // +---+
1708 1708 //
1709 1709 // In the 64-bit VM, each bit represents one 64-bit word:
1710 1710 // +------------+
1711 1711 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1712 1712 // end_bits: ... x x 1 | 0 || 0 | x x ...
1713 1713 // +------------+
1714 1714 // +-------+
1715 1715 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1716 1716 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1717 1717 // +-------+
1718 1718 // +-----------+
1719 1719 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1720 1720 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1721 1721 // +-----------+
1722 1722 // +-------+
1723 1723 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1724 1724 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1725 1725 // +-------+
1726 1726
1727 1727 // Initially assume case a, c or e will apply.
1728 1728 size_t obj_len = CollectedHeap::min_fill_size();
1729 1729 HeapWord* obj_beg = dense_prefix_end - obj_len;
1730 1730
1731 1731 #ifdef _LP64
1732 1732 if (MinObjAlignment > 1) { // object alignment > heap word size
1733 1733 // Cases a, c or e.
1734 1734 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1735 1735 // Case b above.
1736 1736 obj_beg = dense_prefix_end - 1;
1737 1737 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1738 1738 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1739 1739 // Case d above.
1740 1740 obj_beg = dense_prefix_end - 3;
1741 1741 obj_len = 3;
1742 1742 }
1743 1743 #endif // #ifdef _LP64
1744 1744
1745 1745 CollectedHeap::fill_with_object(obj_beg, obj_len);
1746 1746 _mark_bitmap.mark_obj(obj_beg, obj_len);
1747 1747 _summary_data.add_obj(obj_beg, obj_len);
1748 1748 assert(start_array(id) != NULL, "sanity");
1749 1749 start_array(id)->allocate_block(obj_beg);
1750 1750 }
1751 1751 }
1752 1752
1753 1753 void
1754 1754 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1755 1755 {
1756 1756 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1757 1757 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1758 1758 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1759 1759 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1760 1760 cur->set_source_region(0);
1761 1761 }
1762 1762 }
1763 1763
1764 1764 void
1765 1765 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1766 1766 {
1767 1767 assert(id < last_space_id, "id out of range");
1768 1768 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1769 1769 ParallelOldGCSplitALot && id == old_space_id,
1770 1770 "should have been reset in summarize_spaces_quick()");
1771 1771
1772 1772 const MutableSpace* space = _space_info[id].space();
1773 1773 if (_space_info[id].new_top() != space->bottom()) {
1774 1774 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1775 1775 _space_info[id].set_dense_prefix(dense_prefix_end);
1776 1776
1777 1777 #ifndef PRODUCT
1778 1778 if (TraceParallelOldGCDensePrefix) {
1779 1779 print_dense_prefix_stats("ratio", id, maximum_compaction,
1780 1780 dense_prefix_end);
1781 1781 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1782 1782 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1783 1783 }
1784 1784 #endif // #ifndef PRODUCT
1785 1785
1786 1786 // Recompute the summary data, taking into account the dense prefix. If
1787 1787 // every last byte will be reclaimed, then the existing summary data which
1788 1788 // compacts everything can be left in place.
1789 1789 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1790 1790 // If dead space crosses the dense prefix boundary, it is (at least
1791 1791 // partially) filled with a dummy object, marked live and added to the
1792 1792 // summary data. This simplifies the copy/update phase and must be done
1793 1793 // before the final locations of objects are determined, to prevent
1794 1794 // leaving a fragment of dead space that is too small to fill.
1795 1795 fill_dense_prefix_end(id);
1796 1796
1797 1797 // Compute the destination of each Region, and thus each object.
1798 1798 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1799 1799 _summary_data.summarize(_space_info[id].split_info(),
1800 1800 dense_prefix_end, space->top(), NULL,
1801 1801 dense_prefix_end, space->end(),
1802 1802 _space_info[id].new_top_addr());
1803 1803 }
1804 1804 }
1805 1805
1806 1806 if (TraceParallelOldGCSummaryPhase) {
1807 1807 const size_t region_size = ParallelCompactData::RegionSize;
1808 1808 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1809 1809 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1810 1810 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1811 1811 HeapWord* const new_top = _space_info[id].new_top();
1812 1812 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1813 1813 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1814 1814 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1815 1815 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1816 1816 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1817 1817 id, space->capacity_in_words(), dense_prefix_end,
1818 1818 dp_region, dp_words / region_size,
1819 1819 cr_words / region_size, new_top);
1820 1820 }
1821 1821 }
1822 1822
1823 1823 #ifndef PRODUCT
1824 1824 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1825 1825 HeapWord* dst_beg, HeapWord* dst_end,
1826 1826 SpaceId src_space_id,
1827 1827 HeapWord* src_beg, HeapWord* src_end)
1828 1828 {
1829 1829 if (TraceParallelOldGCSummaryPhase) {
1830 1830 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1831 1831 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1832 1832 SIZE_FORMAT "-" SIZE_FORMAT " "
1833 1833 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1834 1834 SIZE_FORMAT "-" SIZE_FORMAT,
1835 1835 src_space_id, space_names[src_space_id],
1836 1836 dst_space_id, space_names[dst_space_id],
1837 1837 src_beg, src_end,
1838 1838 _summary_data.addr_to_region_idx(src_beg),
1839 1839 _summary_data.addr_to_region_idx(src_end),
1840 1840 dst_beg, dst_end,
1841 1841 _summary_data.addr_to_region_idx(dst_beg),
1842 1842 _summary_data.addr_to_region_idx(dst_end));
1843 1843 }
1844 1844 }
1845 1845 #endif // #ifndef PRODUCT
1846 1846
1847 1847 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1848 1848 bool maximum_compaction)
1849 1849 {
1850 1850 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer);
1851 1851 // trace("2");
1852 1852
1853 1853 #ifdef ASSERT
1854 1854 if (TraceParallelOldGCMarkingPhase) {
1855 1855 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1856 1856 "add_obj_bytes=" SIZE_FORMAT,
1857 1857 add_obj_count, add_obj_size * HeapWordSize);
1858 1858 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1859 1859 "mark_bitmap_bytes=" SIZE_FORMAT,
1860 1860 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1861 1861 }
1862 1862 #endif // #ifdef ASSERT
1863 1863
1864 1864 // Quick summarization of each space into itself, to see how much is live.
1865 1865 summarize_spaces_quick();
1866 1866
1867 1867 if (TraceParallelOldGCSummaryPhase) {
1868 1868 tty->print_cr("summary_phase: after summarizing each space to self");
1869 1869 Universe::print();
1870 1870 NOT_PRODUCT(print_region_ranges());
1871 1871 if (Verbose) {
1872 1872 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1873 1873 }
1874 1874 }
1875 1875
1876 1876 // The amount of live data that will end up in old space (assuming it fits).
1877 1877 size_t old_space_total_live = 0;
1878 1878 assert(perm_space_id < old_space_id, "should not count perm data here");
1879 1879 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1880 1880 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1881 1881 _space_info[id].space()->bottom());
1882 1882 }
1883 1883
1884 1884 MutableSpace* const old_space = _space_info[old_space_id].space();
1885 1885 const size_t old_capacity = old_space->capacity_in_words();
1886 1886 if (old_space_total_live > old_capacity) {
1887 1887 // XXX - should also try to expand
1888 1888 maximum_compaction = true;
1889 1889 }
1890 1890 #ifndef PRODUCT
1891 1891 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1892 1892 provoke_split(maximum_compaction);
1893 1893 }
1894 1894 #endif // #ifndef PRODUCT
1895 1895
1896 1896 // Permanent and Old generations.
1897 1897 summarize_space(perm_space_id, maximum_compaction);
1898 1898 summarize_space(old_space_id, maximum_compaction);
1899 1899
1900 1900 // Summarize the remaining spaces in the young gen. The initial target space
1901 1901 // is the old gen. If a space does not fit entirely into the target, then the
1902 1902 // remainder is compacted into the space itself and that space becomes the new
1903 1903 // target.
1904 1904 SpaceId dst_space_id = old_space_id;
1905 1905 HeapWord* dst_space_end = old_space->end();
1906 1906 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1907 1907 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1908 1908 const MutableSpace* space = _space_info[id].space();
1909 1909 const size_t live = pointer_delta(_space_info[id].new_top(),
1910 1910 space->bottom());
1911 1911 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1912 1912
1913 1913 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1914 1914 SpaceId(id), space->bottom(), space->top());)
1915 1915 if (live > 0 && live <= available) {
1916 1916 // All the live data will fit.
1917 1917 bool done = _summary_data.summarize(_space_info[id].split_info(),
1918 1918 space->bottom(), space->top(),
1919 1919 NULL,
1920 1920 *new_top_addr, dst_space_end,
1921 1921 new_top_addr);
1922 1922 assert(done, "space must fit into old gen");
1923 1923
1924 1924 // Reset the new_top value for the space.
1925 1925 _space_info[id].set_new_top(space->bottom());
1926 1926 } else if (live > 0) {
1927 1927 // Attempt to fit part of the source space into the target space.
1928 1928 HeapWord* next_src_addr = NULL;
1929 1929 bool done = _summary_data.summarize(_space_info[id].split_info(),
1930 1930 space->bottom(), space->top(),
1931 1931 &next_src_addr,
1932 1932 *new_top_addr, dst_space_end,
1933 1933 new_top_addr);
1934 1934 assert(!done, "space should not fit into old gen");
1935 1935 assert(next_src_addr != NULL, "sanity");
1936 1936
1937 1937 // The source space becomes the new target, so the remainder is compacted
1938 1938 // within the space itself.
1939 1939 dst_space_id = SpaceId(id);
1940 1940 dst_space_end = space->end();
1941 1941 new_top_addr = _space_info[id].new_top_addr();
1942 1942 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1943 1943 space->bottom(), dst_space_end,
1944 1944 SpaceId(id), next_src_addr, space->top());)
1945 1945 done = _summary_data.summarize(_space_info[id].split_info(),
1946 1946 next_src_addr, space->top(),
1947 1947 NULL,
1948 1948 space->bottom(), dst_space_end,
1949 1949 new_top_addr);
1950 1950 assert(done, "space must fit when compacted into itself");
1951 1951 assert(*new_top_addr <= space->top(), "usage should not grow");
1952 1952 }
1953 1953 }
1954 1954
1955 1955 if (TraceParallelOldGCSummaryPhase) {
1956 1956 tty->print_cr("summary_phase: after final summarization");
1957 1957 Universe::print();
1958 1958 NOT_PRODUCT(print_region_ranges());
1959 1959 if (Verbose) {
1960 1960 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1961 1961 }
1962 1962 }
1963 1963 }
1964 1964
1965 1965 // This method should contain all heap-specific policy for invoking a full
1966 1966 // collection. invoke_no_policy() will only attempt to compact the heap; it
1967 1967 // will do nothing further. If we need to bail out for policy reasons, scavenge
1968 1968 // before full gc, or any other specialized behavior, it needs to be added here.
1969 1969 //
1970 1970 // Note that this method should only be called from the vm_thread while at a
1971 1971 // safepoint.
1972 1972 //
1973 1973 // Note that the all_soft_refs_clear flag in the collector policy
1974 1974 // may be true because this method can be called without intervening
1975 1975 // activity. For example when the heap space is tight and full measure
1976 1976 // are being taken to free space.
1977 1977 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1978 1978 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1979 1979 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1980 1980 "should be in vm thread");
1981 1981
1982 1982 ParallelScavengeHeap* heap = gc_heap();
1983 1983 GCCause::Cause gc_cause = heap->gc_cause();
1984 1984 assert(!heap->is_gc_active(), "not reentrant");
1985 1985
1986 1986 PSAdaptiveSizePolicy* policy = heap->size_policy();
1987 1987 IsGCActiveMark mark;
1988 1988
1989 1989 if (ScavengeBeforeFullGC) {
1990 1990 PSScavenge::invoke_no_policy();
1991 1991 }
1992 1992
1993 1993 const bool clear_all_soft_refs =
1994 1994 heap->collector_policy()->should_clear_all_soft_refs();
1995 1995
1996 1996 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1997 1997 maximum_heap_compaction);
1998 1998 }
1999 1999
2000 2000 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
2001 2001 size_t addr_region_index = addr_to_region_idx(addr);
2002 2002 return region_index == addr_region_index;
2003 2003 }
2004 2004
2005 2005 // This method contains no policy. You should probably
2006 2006 // be calling invoke() instead.
2007 2007 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
2008 2008 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
2009 2009 assert(ref_processor() != NULL, "Sanity");
2010 2010
2011 2011 if (GC_locker::check_active_before_gc()) {
2012 2012 return false;
2013 2013 }
2014 2014
2015 2015 ParallelScavengeHeap* heap = gc_heap();
2016 2016
2017 2017 _gc_timer.register_gc_start(os::elapsed_counter());
2018 2018 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
2019 2019
2020 2020 TimeStamp marking_start;
2021 2021 TimeStamp compaction_start;
2022 2022 TimeStamp collection_exit;
2023 2023
2024 2024 GCCause::Cause gc_cause = heap->gc_cause();
2025 2025 PSYoungGen* young_gen = heap->young_gen();
2026 2026 PSOldGen* old_gen = heap->old_gen();
2027 2027 PSPermGen* perm_gen = heap->perm_gen();
2028 2028 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2029 2029
2030 2030 // The scope of casr should end after code that can change
2031 2031 // CollectorPolicy::_should_clear_all_soft_refs.
2032 2032 ClearedAllSoftRefs casr(maximum_heap_compaction,
2033 2033 heap->collector_policy());
2034 2034
2035 2035 if (ZapUnusedHeapArea) {
2036 2036 // Save information needed to minimize mangling
2037 2037 heap->record_gen_tops_before_GC();
2038 2038 }
2039 2039
2040 2040 heap->pre_full_gc_dump(&_gc_timer);
2041 2041
2042 2042 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2043 2043
2044 2044 // Make sure data structures are sane, make the heap parsable, and do other
2045 2045 // miscellaneous bookkeeping.
2046 2046 PreGCValues pre_gc_values;
2047 2047 pre_compact(&pre_gc_values);
2048 2048
2049 2049 // Get the compaction manager reserved for the VM thread.
2050 2050 ParCompactionManager* const vmthread_cm =
2051 2051 ParCompactionManager::manager_array(gc_task_manager()->workers());
2052 2052
2053 2053 // Place after pre_compact() where the number of invocations is incremented.
2054 2054 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
↓ open down ↓ |
2054 lines elided |
↑ open up ↑ |
2055 2055
2056 2056 {
2057 2057 ResourceMark rm;
2058 2058 HandleMark hm;
2059 2059
2060 2060 // Set the number of GC threads to be used in this collection
2061 2061 gc_task_manager()->set_active_gang();
2062 2062 gc_task_manager()->task_idle_workers();
2063 2063 heap->set_par_threads(gc_task_manager()->active_workers());
2064 2064
2065 - gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2066 2065 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2067 2066 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
2068 2067 TraceCollectorStats tcs(counters());
2069 2068 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2070 2069
2071 2070 if (TraceGen1Time) accumulated_time()->start();
2072 2071
2073 2072 // Let the size policy know we're starting
2074 2073 size_policy->major_collection_begin();
2075 2074
2076 2075 // When collecting the permanent generation methodOops may be moving,
2077 2076 // so we either have to flush all bcp data or convert it into bci.
2078 2077 CodeCache::gc_prologue();
2079 2078 Threads::gc_prologue();
2080 2079
2081 2080 COMPILER2_PRESENT(DerivedPointerTable::clear());
2082 2081
2083 2082 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2084 2083 ref_processor()->setup_policy(maximum_heap_compaction);
2085 2084
2086 2085 bool marked_for_unloading = false;
2087 2086
2088 2087 marking_start.update();
2089 2088 marking_phase(vmthread_cm, maximum_heap_compaction);
2090 2089
2091 2090 #ifndef PRODUCT
2092 2091 if (TraceParallelOldGCMarkingPhase) {
2093 2092 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2094 2093 "cas_by_another %d",
2095 2094 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2096 2095 mark_bitmap()->cas_by_another());
2097 2096 }
2098 2097 #endif // #ifndef PRODUCT
2099 2098
2100 2099 bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2101 2100 && gc_cause == GCCause::_java_lang_system_gc;
2102 2101 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2103 2102
2104 2103 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2105 2104 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2106 2105
2107 2106 // adjust_roots() updates Universe::_intArrayKlassObj which is
2108 2107 // needed by the compaction for filling holes in the dense prefix.
2109 2108 adjust_roots();
2110 2109
2111 2110 compaction_start.update();
2112 2111 // Does the perm gen always have to be done serially because
2113 2112 // klasses are used in the update of an object?
2114 2113 compact_perm(vmthread_cm);
2115 2114
2116 2115 compact();
2117 2116
2118 2117 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2119 2118 // done before resizing.
2120 2119 post_compact();
2121 2120
2122 2121 // Let the size policy know we're done
2123 2122 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2124 2123
2125 2124 if (UseAdaptiveSizePolicy) {
2126 2125 if (PrintAdaptiveSizePolicy) {
2127 2126 gclog_or_tty->print("AdaptiveSizeStart: ");
2128 2127 gclog_or_tty->stamp();
2129 2128 gclog_or_tty->print_cr(" collection: %d ",
2130 2129 heap->total_collections());
2131 2130 if (Verbose) {
2132 2131 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2133 2132 " perm_gen_capacity: %d ",
2134 2133 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2135 2134 perm_gen->capacity_in_bytes());
2136 2135 }
2137 2136 }
2138 2137
2139 2138 // Don't check if the size_policy is ready here. Let
2140 2139 // the size_policy check that internally.
2141 2140 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2142 2141 ((gc_cause != GCCause::_java_lang_system_gc) ||
2143 2142 UseAdaptiveSizePolicyWithSystemGC)) {
2144 2143 // Calculate optimal free space amounts
2145 2144 assert(young_gen->max_size() >
2146 2145 young_gen->from_space()->capacity_in_bytes() +
2147 2146 young_gen->to_space()->capacity_in_bytes(),
2148 2147 "Sizes of space in young gen are out-of-bounds");
2149 2148 size_t max_eden_size = young_gen->max_size() -
2150 2149 young_gen->from_space()->capacity_in_bytes() -
2151 2150 young_gen->to_space()->capacity_in_bytes();
2152 2151 size_policy->compute_generation_free_space(
2153 2152 young_gen->used_in_bytes(),
2154 2153 young_gen->eden_space()->used_in_bytes(),
2155 2154 old_gen->used_in_bytes(),
2156 2155 perm_gen->used_in_bytes(),
2157 2156 young_gen->eden_space()->capacity_in_bytes(),
2158 2157 old_gen->max_gen_size(),
2159 2158 max_eden_size,
2160 2159 true /* full gc*/,
2161 2160 gc_cause,
2162 2161 heap->collector_policy());
2163 2162
2164 2163 heap->resize_old_gen(
2165 2164 size_policy->calculated_old_free_size_in_bytes());
2166 2165
2167 2166 // Don't resize the young generation at an major collection. A
2168 2167 // desired young generation size may have been calculated but
2169 2168 // resizing the young generation complicates the code because the
2170 2169 // resizing of the old generation may have moved the boundary
2171 2170 // between the young generation and the old generation. Let the
2172 2171 // young generation resizing happen at the minor collections.
2173 2172 }
2174 2173 if (PrintAdaptiveSizePolicy) {
2175 2174 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2176 2175 heap->total_collections());
2177 2176 }
2178 2177 }
2179 2178
2180 2179 if (UsePerfData) {
2181 2180 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2182 2181 counters->update_counters();
2183 2182 counters->update_old_capacity(old_gen->capacity_in_bytes());
2184 2183 counters->update_young_capacity(young_gen->capacity_in_bytes());
2185 2184 }
2186 2185
2187 2186 heap->resize_all_tlabs();
2188 2187
2189 2188 // We collected the perm gen, so we'll resize it here.
2190 2189 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2191 2190
2192 2191 if (TraceGen1Time) accumulated_time()->stop();
2193 2192
2194 2193 if (PrintGC) {
2195 2194 if (PrintGCDetails) {
2196 2195 // No GC timestamp here. This is after GC so it would be confusing.
2197 2196 young_gen->print_used_change(pre_gc_values.young_gen_used());
2198 2197 old_gen->print_used_change(pre_gc_values.old_gen_used());
2199 2198 heap->print_heap_change(pre_gc_values.heap_used());
2200 2199 // Print perm gen last (print_heap_change() excludes the perm gen).
2201 2200 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2202 2201 } else {
2203 2202 heap->print_heap_change(pre_gc_values.heap_used());
2204 2203 }
2205 2204 }
2206 2205
2207 2206 // Track memory usage and detect low memory
2208 2207 MemoryService::track_memory_usage();
2209 2208 heap->update_counters();
2210 2209 gc_task_manager()->release_idle_workers();
2211 2210 }
2212 2211
2213 2212 #ifdef ASSERT
2214 2213 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2215 2214 ParCompactionManager* const cm =
2216 2215 ParCompactionManager::manager_array(int(i));
2217 2216 assert(cm->marking_stack()->is_empty(), "should be empty");
2218 2217 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2219 2218 assert(cm->revisit_klass_stack()->is_empty(), "should be empty");
2220 2219 }
2221 2220 #endif // ASSERT
2222 2221
2223 2222 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2224 2223 HandleMark hm; // Discard invalid handles created during verification
2225 2224 gclog_or_tty->print(" VerifyAfterGC:");
2226 2225 Universe::verify();
2227 2226 }
2228 2227
2229 2228 // Re-verify object start arrays
2230 2229 if (VerifyObjectStartArray &&
2231 2230 VerifyAfterGC) {
2232 2231 old_gen->verify_object_start_array();
2233 2232 perm_gen->verify_object_start_array();
2234 2233 }
2235 2234
2236 2235 if (ZapUnusedHeapArea) {
2237 2236 old_gen->object_space()->check_mangled_unused_area_complete();
2238 2237 perm_gen->object_space()->check_mangled_unused_area_complete();
2239 2238 }
2240 2239
2241 2240 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2242 2241
2243 2242 collection_exit.update();
2244 2243
2245 2244 heap->print_heap_after_gc();
2246 2245 heap->trace_heap_after_gc(&_gc_tracer);
2247 2246
2248 2247 if (PrintGCTaskTimeStamps) {
2249 2248 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2250 2249 INT64_FORMAT,
2251 2250 marking_start.ticks(), compaction_start.ticks(),
2252 2251 collection_exit.ticks());
2253 2252 gc_task_manager()->print_task_time_stamps();
2254 2253 }
2255 2254
2256 2255 heap->post_full_gc_dump(&_gc_timer);
2257 2256
2258 2257 #ifdef TRACESPINNING
2259 2258 ParallelTaskTerminator::print_termination_counts();
2260 2259 #endif
2261 2260
2262 2261 _gc_timer.register_gc_end(os::elapsed_counter());
2263 2262
2264 2263 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2265 2264 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2266 2265
2267 2266 return true;
2268 2267 }
2269 2268
2270 2269 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2271 2270 PSYoungGen* young_gen,
2272 2271 PSOldGen* old_gen) {
2273 2272 MutableSpace* const eden_space = young_gen->eden_space();
2274 2273 assert(!eden_space->is_empty(), "eden must be non-empty");
2275 2274 assert(young_gen->virtual_space()->alignment() ==
2276 2275 old_gen->virtual_space()->alignment(), "alignments do not match");
2277 2276
2278 2277 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2279 2278 return false;
2280 2279 }
2281 2280
2282 2281 // Both generations must be completely committed.
2283 2282 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2284 2283 return false;
2285 2284 }
2286 2285 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2287 2286 return false;
2288 2287 }
2289 2288
2290 2289 // Figure out how much to take from eden. Include the average amount promoted
2291 2290 // in the total; otherwise the next young gen GC will simply bail out to a
2292 2291 // full GC.
2293 2292 const size_t alignment = old_gen->virtual_space()->alignment();
2294 2293 const size_t eden_used = eden_space->used_in_bytes();
2295 2294 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2296 2295 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2297 2296 const size_t eden_capacity = eden_space->capacity_in_bytes();
2298 2297
2299 2298 if (absorb_size >= eden_capacity) {
2300 2299 return false; // Must leave some space in eden.
2301 2300 }
2302 2301
2303 2302 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2304 2303 if (new_young_size < young_gen->min_gen_size()) {
2305 2304 return false; // Respect young gen minimum size.
2306 2305 }
2307 2306
2308 2307 if (TraceAdaptiveGCBoundary && Verbose) {
2309 2308 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2310 2309 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2311 2310 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2312 2311 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2313 2312 absorb_size / K,
2314 2313 eden_capacity / K, (eden_capacity - absorb_size) / K,
2315 2314 young_gen->from_space()->used_in_bytes() / K,
2316 2315 young_gen->to_space()->used_in_bytes() / K,
2317 2316 young_gen->capacity_in_bytes() / K, new_young_size / K);
2318 2317 }
2319 2318
2320 2319 // Fill the unused part of the old gen.
2321 2320 MutableSpace* const old_space = old_gen->object_space();
2322 2321 HeapWord* const unused_start = old_space->top();
2323 2322 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2324 2323
2325 2324 if (unused_words > 0) {
2326 2325 if (unused_words < CollectedHeap::min_fill_size()) {
2327 2326 return false; // If the old gen cannot be filled, must give up.
2328 2327 }
2329 2328 CollectedHeap::fill_with_objects(unused_start, unused_words);
2330 2329 }
2331 2330
2332 2331 // Take the live data from eden and set both top and end in the old gen to
2333 2332 // eden top. (Need to set end because reset_after_change() mangles the region
2334 2333 // from end to virtual_space->high() in debug builds).
2335 2334 HeapWord* const new_top = eden_space->top();
2336 2335 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2337 2336 absorb_size);
2338 2337 young_gen->reset_after_change();
2339 2338 old_space->set_top(new_top);
2340 2339 old_space->set_end(new_top);
2341 2340 old_gen->reset_after_change();
2342 2341
2343 2342 // Update the object start array for the filler object and the data from eden.
2344 2343 ObjectStartArray* const start_array = old_gen->start_array();
2345 2344 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2346 2345 start_array->allocate_block(p);
2347 2346 }
2348 2347
2349 2348 // Could update the promoted average here, but it is not typically updated at
2350 2349 // full GCs and the value to use is unclear. Something like
2351 2350 //
2352 2351 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2353 2352
2354 2353 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2355 2354 return true;
2356 2355 }
2357 2356
2358 2357 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2359 2358 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2360 2359 "shouldn't return NULL");
2361 2360 return ParallelScavengeHeap::gc_task_manager();
2362 2361 }
2363 2362
2364 2363 void PSParallelCompact::marking_phase(ParCompactionManager* cm, bool maximum_heap_compaction) {
2365 2364 // Recursively traverse all live objects and mark them
2366 2365 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer);
2367 2366
2368 2367 ParallelScavengeHeap* heap = gc_heap();
2369 2368 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2370 2369 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2371 2370 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2372 2371 ParallelTaskTerminator terminator(active_gc_threads, qset);
2373 2372
2374 2373 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2375 2374 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2376 2375
2377 2376 {
2378 2377 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer);
2379 2378
2380 2379 ParallelScavengeHeap::ParStrongRootsScope psrs;
2381 2380
2382 2381 GCTaskQueue* q = GCTaskQueue::create();
2383 2382
2384 2383 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2385 2384 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2386 2385 // We scan the thread roots in parallel
2387 2386 Threads::create_thread_roots_marking_tasks(q);
2388 2387 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2389 2388 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2390 2389 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2391 2390 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2392 2391 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2393 2392 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2394 2393
2395 2394 if (active_gc_threads > 1) {
2396 2395 for (uint j = 0; j < active_gc_threads; j++) {
2397 2396 q->enqueue(new StealMarkingTask(&terminator));
2398 2397 }
2399 2398 }
2400 2399
2401 2400 gc_task_manager()->execute_and_wait(q);
2402 2401 }
2403 2402
2404 2403 // Process reference objects found during marking
2405 2404 {
2406 2405 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer);
2407 2406
2408 2407 ReferenceProcessorStats stats;
2409 2408 if (ref_processor()->processing_is_mt()) {
2410 2409 RefProcTaskExecutor task_executor;
2411 2410 stats = ref_processor()->process_discovered_references(
2412 2411 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2413 2412 &task_executor, &_gc_timer);
2414 2413 } else {
2415 2414 stats = ref_processor()->process_discovered_references(
2416 2415 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2417 2416 &_gc_timer);
2418 2417 }
2419 2418
2420 2419 _gc_tracer.report_gc_reference_stats(stats);
2421 2420 }
2422 2421
2423 2422 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer);
2424 2423
2425 2424 // Follow system dictionary roots and unload classes.
2426 2425 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2427 2426
2428 2427 // Follow code cache roots.
2429 2428 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2430 2429 purged_class);
2431 2430 cm->follow_marking_stacks(); // Flush marking stack.
2432 2431
2433 2432 // Update subklass/sibling/implementor links of live klasses
2434 2433 // revisit_klass_stack is used in follow_weak_klass_links().
2435 2434 follow_weak_klass_links();
2436 2435
2437 2436 // Revisit memoized MDO's and clear any unmarked weak refs
2438 2437 follow_mdo_weak_refs();
2439 2438
2440 2439 // Visit interned string tables and delete unmarked oops
2441 2440 StringTable::unlink(is_alive_closure());
2442 2441 // Clean up unreferenced symbols in symbol table.
2443 2442 SymbolTable::unlink();
2444 2443
2445 2444 assert(cm->marking_stacks_empty(), "marking stacks should be empty");
2446 2445 _gc_tracer.report_object_count_after_gc(is_alive_closure());
2447 2446 }
2448 2447
2449 2448 // This should be moved to the shared markSweep code!
2450 2449 class PSAlwaysTrueClosure: public BoolObjectClosure {
2451 2450 public:
2452 2451 void do_object(oop p) { ShouldNotReachHere(); }
2453 2452 bool do_object_b(oop p) { return true; }
2454 2453 };
2455 2454 static PSAlwaysTrueClosure always_true;
2456 2455
2457 2456 void PSParallelCompact::adjust_roots() {
2458 2457 // Adjust the pointers to reflect the new locations
2459 2458 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer);
2460 2459
2461 2460 // General strong roots.
2462 2461 Universe::oops_do(adjust_root_pointer_closure());
2463 2462 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
2464 2463 Threads::oops_do(adjust_root_pointer_closure(), NULL);
2465 2464 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2466 2465 FlatProfiler::oops_do(adjust_root_pointer_closure());
2467 2466 Management::oops_do(adjust_root_pointer_closure());
2468 2467 JvmtiExport::oops_do(adjust_root_pointer_closure());
2469 2468 // SO_AllClasses
2470 2469 SystemDictionary::oops_do(adjust_root_pointer_closure());
2471 2470
2472 2471 // Now adjust pointers in remaining weak roots. (All of which should
2473 2472 // have been cleared if they pointed to non-surviving objects.)
2474 2473 // Global (weak) JNI handles
2475 2474 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2476 2475
2477 2476 CodeCache::oops_do(adjust_pointer_closure());
2478 2477 StringTable::oops_do(adjust_root_pointer_closure());
2479 2478 ref_processor()->weak_oops_do(adjust_root_pointer_closure());
2480 2479 // Roots were visited so references into the young gen in roots
2481 2480 // may have been scanned. Process them also.
2482 2481 // Should the reference processor have a span that excludes
2483 2482 // young gen objects?
2484 2483 PSScavenge::reference_processor()->weak_oops_do(
2485 2484 adjust_root_pointer_closure());
2486 2485 }
2487 2486
2488 2487 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2489 2488 GCTraceTime tm("compact perm gen", print_phases(), true, &_gc_timer);
2490 2489 // trace("4");
2491 2490
2492 2491 gc_heap()->perm_gen()->start_array()->reset();
2493 2492 move_and_update(cm, perm_space_id);
2494 2493 }
2495 2494
2496 2495 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2497 2496 uint parallel_gc_threads)
2498 2497 {
2499 2498 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer);
2500 2499
2501 2500 // Find the threads that are active
2502 2501 unsigned int which = 0;
2503 2502
2504 2503 const uint task_count = MAX2(parallel_gc_threads, 1U);
2505 2504 for (uint j = 0; j < task_count; j++) {
2506 2505 q->enqueue(new DrainStacksCompactionTask(j));
2507 2506 ParCompactionManager::verify_region_list_empty(j);
2508 2507 // Set the region stacks variables to "no" region stack values
2509 2508 // so that they will be recognized and needing a region stack
2510 2509 // in the stealing tasks if they do not get one by executing
2511 2510 // a draining stack.
2512 2511 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2513 2512 cm->set_region_stack(NULL);
2514 2513 cm->set_region_stack_index((uint)max_uintx);
2515 2514 }
2516 2515 ParCompactionManager::reset_recycled_stack_index();
2517 2516
2518 2517 // Find all regions that are available (can be filled immediately) and
2519 2518 // distribute them to the thread stacks. The iteration is done in reverse
2520 2519 // order (high to low) so the regions will be removed in ascending order.
2521 2520
2522 2521 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2523 2522
2524 2523 size_t fillable_regions = 0; // A count for diagnostic purposes.
2525 2524 // A region index which corresponds to the tasks created above.
2526 2525 // "which" must be 0 <= which < task_count
2527 2526
2528 2527 which = 0;
2529 2528 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2530 2529 SpaceInfo* const space_info = _space_info + id;
2531 2530 MutableSpace* const space = space_info->space();
2532 2531 HeapWord* const new_top = space_info->new_top();
2533 2532
2534 2533 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2535 2534 const size_t end_region =
2536 2535 sd.addr_to_region_idx(sd.region_align_up(new_top));
2537 2536 assert(end_region > 0, "perm gen cannot be empty");
2538 2537
2539 2538 for (size_t cur = end_region - 1; cur >= beg_region; --cur) {
2540 2539 if (sd.region(cur)->claim_unsafe()) {
2541 2540 ParCompactionManager::region_list_push(which, cur);
2542 2541
2543 2542 if (TraceParallelOldGCCompactionPhase && Verbose) {
2544 2543 const size_t count_mod_8 = fillable_regions & 7;
2545 2544 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2546 2545 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2547 2546 if (count_mod_8 == 7) gclog_or_tty->cr();
2548 2547 }
2549 2548
2550 2549 NOT_PRODUCT(++fillable_regions;)
2551 2550
2552 2551 // Assign regions to tasks in round-robin fashion.
2553 2552 if (++which == task_count) {
2554 2553 assert(which <= parallel_gc_threads,
2555 2554 "Inconsistent number of workers");
2556 2555 which = 0;
2557 2556 }
2558 2557 }
2559 2558 }
2560 2559 }
2561 2560
2562 2561 if (TraceParallelOldGCCompactionPhase) {
2563 2562 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2564 2563 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2565 2564 }
2566 2565 }
2567 2566
2568 2567 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2569 2568
2570 2569 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2571 2570 uint parallel_gc_threads) {
2572 2571 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer);
2573 2572
2574 2573 ParallelCompactData& sd = PSParallelCompact::summary_data();
2575 2574
2576 2575 // Iterate over all the spaces adding tasks for updating
2577 2576 // regions in the dense prefix. Assume that 1 gc thread
2578 2577 // will work on opening the gaps and the remaining gc threads
2579 2578 // will work on the dense prefix.
2580 2579 unsigned int space_id;
2581 2580 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2582 2581 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2583 2582 const MutableSpace* const space = _space_info[space_id].space();
2584 2583
2585 2584 if (dense_prefix_end == space->bottom()) {
2586 2585 // There is no dense prefix for this space.
2587 2586 continue;
2588 2587 }
2589 2588
2590 2589 // The dense prefix is before this region.
2591 2590 size_t region_index_end_dense_prefix =
2592 2591 sd.addr_to_region_idx(dense_prefix_end);
2593 2592 RegionData* const dense_prefix_cp =
2594 2593 sd.region(region_index_end_dense_prefix);
2595 2594 assert(dense_prefix_end == space->end() ||
2596 2595 dense_prefix_cp->available() ||
2597 2596 dense_prefix_cp->claimed(),
2598 2597 "The region after the dense prefix should always be ready to fill");
2599 2598
2600 2599 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2601 2600
2602 2601 // Is there dense prefix work?
2603 2602 size_t total_dense_prefix_regions =
2604 2603 region_index_end_dense_prefix - region_index_start;
2605 2604 // How many regions of the dense prefix should be given to
2606 2605 // each thread?
2607 2606 if (total_dense_prefix_regions > 0) {
2608 2607 uint tasks_for_dense_prefix = 1;
2609 2608 if (total_dense_prefix_regions <=
2610 2609 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2611 2610 // Don't over partition. This assumes that
2612 2611 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2613 2612 // so there are not many regions to process.
2614 2613 tasks_for_dense_prefix = parallel_gc_threads;
2615 2614 } else {
2616 2615 // Over partition
2617 2616 tasks_for_dense_prefix = parallel_gc_threads *
2618 2617 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2619 2618 }
2620 2619 size_t regions_per_thread = total_dense_prefix_regions /
2621 2620 tasks_for_dense_prefix;
2622 2621 // Give each thread at least 1 region.
2623 2622 if (regions_per_thread == 0) {
2624 2623 regions_per_thread = 1;
2625 2624 }
2626 2625
2627 2626 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2628 2627 if (region_index_start >= region_index_end_dense_prefix) {
2629 2628 break;
2630 2629 }
2631 2630 // region_index_end is not processed
2632 2631 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2633 2632 region_index_end_dense_prefix);
2634 2633 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2635 2634 region_index_start,
2636 2635 region_index_end));
2637 2636 region_index_start = region_index_end;
2638 2637 }
2639 2638 }
2640 2639 // This gets any part of the dense prefix that did not
2641 2640 // fit evenly.
2642 2641 if (region_index_start < region_index_end_dense_prefix) {
2643 2642 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2644 2643 region_index_start,
2645 2644 region_index_end_dense_prefix));
2646 2645 }
2647 2646 }
2648 2647 }
2649 2648
2650 2649 void PSParallelCompact::enqueue_region_stealing_tasks(
2651 2650 GCTaskQueue* q,
2652 2651 ParallelTaskTerminator* terminator_ptr,
2653 2652 uint parallel_gc_threads) {
2654 2653 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer);
2655 2654
2656 2655 // Once a thread has drained it's stack, it should try to steal regions from
2657 2656 // other threads.
2658 2657 if (parallel_gc_threads > 1) {
2659 2658 for (uint j = 0; j < parallel_gc_threads; j++) {
2660 2659 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2661 2660 }
2662 2661 }
2663 2662 }
2664 2663
2665 2664 void PSParallelCompact::compact() {
2666 2665 // trace("5");
2667 2666 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer);
2668 2667
2669 2668 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2670 2669 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2671 2670 PSOldGen* old_gen = heap->old_gen();
2672 2671 old_gen->start_array()->reset();
2673 2672 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2674 2673 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2675 2674 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2676 2675 ParallelTaskTerminator terminator(active_gc_threads, qset);
2677 2676
2678 2677 GCTaskQueue* q = GCTaskQueue::create();
2679 2678 enqueue_region_draining_tasks(q, active_gc_threads);
2680 2679 enqueue_dense_prefix_tasks(q, active_gc_threads);
2681 2680 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2682 2681
2683 2682 {
2684 2683 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer);
2685 2684
2686 2685 gc_task_manager()->execute_and_wait(q);
2687 2686
2688 2687 #ifdef ASSERT
2689 2688 // Verify that all regions have been processed before the deferred updates.
2690 2689 // Note that perm_space_id is skipped; this type of verification is not
2691 2690 // valid until the perm gen is compacted by regions.
2692 2691 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2693 2692 verify_complete(SpaceId(id));
2694 2693 }
2695 2694 #endif
2696 2695 }
2697 2696
2698 2697 {
2699 2698 // Update the deferred objects, if any. Any compaction manager can be used.
2700 2699 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer);
2701 2700 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2702 2701 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2703 2702 update_deferred_objects(cm, SpaceId(id));
2704 2703 }
2705 2704 }
2706 2705 }
2707 2706
2708 2707 #ifdef ASSERT
2709 2708 void PSParallelCompact::verify_complete(SpaceId space_id) {
2710 2709 // All Regions between space bottom() to new_top() should be marked as filled
2711 2710 // and all Regions between new_top() and top() should be available (i.e.,
2712 2711 // should have been emptied).
2713 2712 ParallelCompactData& sd = summary_data();
2714 2713 SpaceInfo si = _space_info[space_id];
2715 2714 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2716 2715 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2717 2716 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2718 2717 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2719 2718 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2720 2719
2721 2720 bool issued_a_warning = false;
2722 2721
2723 2722 size_t cur_region;
2724 2723 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2725 2724 const RegionData* const c = sd.region(cur_region);
2726 2725 if (!c->completed()) {
2727 2726 warning("region " SIZE_FORMAT " not filled: "
2728 2727 "destination_count=" SIZE_FORMAT,
2729 2728 cur_region, c->destination_count());
2730 2729 issued_a_warning = true;
2731 2730 }
2732 2731 }
2733 2732
2734 2733 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2735 2734 const RegionData* const c = sd.region(cur_region);
2736 2735 if (!c->available()) {
2737 2736 warning("region " SIZE_FORMAT " not empty: "
2738 2737 "destination_count=" SIZE_FORMAT,
2739 2738 cur_region, c->destination_count());
2740 2739 issued_a_warning = true;
2741 2740 }
2742 2741 }
2743 2742
2744 2743 if (issued_a_warning) {
2745 2744 print_region_ranges();
2746 2745 }
2747 2746 }
2748 2747 #endif // #ifdef ASSERT
2749 2748
2750 2749 void
2751 2750 PSParallelCompact::follow_weak_klass_links() {
2752 2751 // All klasses on the revisit stack are marked at this point.
2753 2752 // Update and follow all subklass, sibling and implementor links.
2754 2753 // Check all the stacks here even if not all the workers are active.
2755 2754 // There is no accounting which indicates which stacks might have
2756 2755 // contents to be followed.
2757 2756 if (PrintRevisitStats) {
2758 2757 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2759 2758 SystemDictionary::number_of_classes());
2760 2759 }
2761 2760 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2762 2761 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2763 2762 KeepAliveClosure keep_alive_closure(cm);
2764 2763 Stack<Klass*, mtGC>* const rks = cm->revisit_klass_stack();
2765 2764 if (PrintRevisitStats) {
2766 2765 gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT,
2767 2766 i, rks->size());
2768 2767 }
2769 2768 while (!rks->is_empty()) {
2770 2769 Klass* const k = rks->pop();
2771 2770 k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure);
2772 2771 }
2773 2772
2774 2773 cm->follow_marking_stacks();
2775 2774 }
2776 2775 }
2777 2776
2778 2777 void
2779 2778 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2780 2779 cm->revisit_klass_stack()->push(k);
2781 2780 }
2782 2781
2783 2782 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) {
2784 2783 cm->revisit_mdo_stack()->push(p);
2785 2784 }
2786 2785
2787 2786 void PSParallelCompact::follow_mdo_weak_refs() {
2788 2787 // All strongly reachable oops have been marked at this point;
2789 2788 // we can visit and clear any weak references from MDO's which
2790 2789 // we memoized during the strong marking phase.
2791 2790 if (PrintRevisitStats) {
2792 2791 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2793 2792 SystemDictionary::number_of_classes());
2794 2793 }
2795 2794 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2796 2795 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2797 2796 Stack<DataLayout*, mtGC>* rms = cm->revisit_mdo_stack();
2798 2797 if (PrintRevisitStats) {
2799 2798 gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT,
2800 2799 i, rms->size());
2801 2800 }
2802 2801 while (!rms->is_empty()) {
2803 2802 rms->pop()->follow_weak_refs(is_alive_closure());
2804 2803 }
2805 2804
2806 2805 cm->follow_marking_stacks();
2807 2806 }
2808 2807 }
2809 2808
2810 2809
2811 2810 #ifdef VALIDATE_MARK_SWEEP
2812 2811
2813 2812 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2814 2813 if (!ValidateMarkSweep)
2815 2814 return;
2816 2815
2817 2816 if (!isroot) {
2818 2817 if (_pointer_tracking) {
2819 2818 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2820 2819 _adjusted_pointers->remove(p);
2821 2820 }
2822 2821 } else {
2823 2822 ptrdiff_t index = _root_refs_stack->find(p);
2824 2823 if (index != -1) {
2825 2824 int l = _root_refs_stack->length();
2826 2825 if (l > 0 && l - 1 != index) {
2827 2826 void* last = _root_refs_stack->pop();
2828 2827 assert(last != p, "should be different");
2829 2828 _root_refs_stack->at_put(index, last);
2830 2829 } else {
2831 2830 _root_refs_stack->remove(p);
2832 2831 }
2833 2832 }
2834 2833 }
2835 2834 }
2836 2835
2837 2836
2838 2837 void PSParallelCompact::check_adjust_pointer(void* p) {
2839 2838 _adjusted_pointers->push(p);
2840 2839 }
2841 2840
2842 2841
2843 2842 class AdjusterTracker: public OopClosure {
2844 2843 public:
2845 2844 AdjusterTracker() {};
2846 2845 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
2847 2846 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); }
2848 2847 };
2849 2848
2850 2849
2851 2850 void PSParallelCompact::track_interior_pointers(oop obj) {
2852 2851 if (ValidateMarkSweep) {
2853 2852 _adjusted_pointers->clear();
2854 2853 _pointer_tracking = true;
2855 2854
2856 2855 AdjusterTracker checker;
2857 2856 obj->oop_iterate(&checker);
2858 2857 }
2859 2858 }
2860 2859
2861 2860
2862 2861 void PSParallelCompact::check_interior_pointers() {
2863 2862 if (ValidateMarkSweep) {
2864 2863 _pointer_tracking = false;
2865 2864 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2866 2865 }
2867 2866 }
2868 2867
2869 2868
2870 2869 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2871 2870 if (ValidateMarkSweep) {
2872 2871 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2873 2872 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2874 2873 }
2875 2874 }
2876 2875
2877 2876
2878 2877 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2879 2878 if (ValidateMarkSweep) {
2880 2879 _live_oops->push(p);
2881 2880 _live_oops_size->push(size);
2882 2881 _live_oops_index++;
2883 2882 }
2884 2883 }
2885 2884
2886 2885 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2887 2886 if (ValidateMarkSweep) {
2888 2887 oop obj = _live_oops->at((int)_live_oops_index);
2889 2888 guarantee(obj == p, "should be the same object");
2890 2889 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2891 2890 _live_oops_index++;
2892 2891 }
2893 2892 }
2894 2893
2895 2894 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2896 2895 HeapWord* compaction_top) {
2897 2896 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2898 2897 "should be moved to forwarded location");
2899 2898 if (ValidateMarkSweep) {
2900 2899 PSParallelCompact::validate_live_oop(oop(q), size);
2901 2900 _live_oops_moved_to->push(oop(compaction_top));
2902 2901 }
2903 2902 if (RecordMarkSweepCompaction) {
2904 2903 _cur_gc_live_oops->push(q);
2905 2904 _cur_gc_live_oops_moved_to->push(compaction_top);
2906 2905 _cur_gc_live_oops_size->push(size);
2907 2906 }
2908 2907 }
2909 2908
2910 2909
2911 2910 void PSParallelCompact::compaction_complete() {
2912 2911 if (RecordMarkSweepCompaction) {
2913 2912 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
2914 2913 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2915 2914 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
2916 2915
2917 2916 _cur_gc_live_oops = _last_gc_live_oops;
2918 2917 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
2919 2918 _cur_gc_live_oops_size = _last_gc_live_oops_size;
2920 2919 _last_gc_live_oops = _tmp_live_oops;
2921 2920 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2922 2921 _last_gc_live_oops_size = _tmp_live_oops_size;
2923 2922 }
2924 2923 }
2925 2924
2926 2925
2927 2926 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2928 2927 if (!RecordMarkSweepCompaction) {
2929 2928 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2930 2929 return;
2931 2930 }
2932 2931
2933 2932 if (_last_gc_live_oops == NULL) {
2934 2933 tty->print_cr("No compaction information gathered yet");
2935 2934 return;
2936 2935 }
2937 2936
2938 2937 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2939 2938 HeapWord* old_oop = _last_gc_live_oops->at(i);
2940 2939 size_t sz = _last_gc_live_oops_size->at(i);
2941 2940 if (old_oop <= q && q < (old_oop + sz)) {
2942 2941 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2943 2942 size_t offset = (q - old_oop);
2944 2943 tty->print_cr("Address " PTR_FORMAT, q);
2945 2944 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2946 2945 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2947 2946 return;
2948 2947 }
2949 2948 }
2950 2949
2951 2950 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2952 2951 }
2953 2952 #endif //VALIDATE_MARK_SWEEP
2954 2953
2955 2954 // Update interior oops in the ranges of regions [beg_region, end_region).
2956 2955 void
2957 2956 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2958 2957 SpaceId space_id,
2959 2958 size_t beg_region,
2960 2959 size_t end_region) {
2961 2960 ParallelCompactData& sd = summary_data();
2962 2961 ParMarkBitMap* const mbm = mark_bitmap();
2963 2962
2964 2963 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2965 2964 HeapWord* const end_addr = sd.region_to_addr(end_region);
2966 2965 assert(beg_region <= end_region, "bad region range");
2967 2966 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2968 2967
2969 2968 #ifdef ASSERT
2970 2969 // Claim the regions to avoid triggering an assert when they are marked as
2971 2970 // filled.
2972 2971 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2973 2972 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2974 2973 }
2975 2974 #endif // #ifdef ASSERT
2976 2975
2977 2976 if (beg_addr != space(space_id)->bottom()) {
2978 2977 // Find the first live object or block of dead space that *starts* in this
2979 2978 // range of regions. If a partial object crosses onto the region, skip it;
2980 2979 // it will be marked for 'deferred update' when the object head is
2981 2980 // processed. If dead space crosses onto the region, it is also skipped; it
2982 2981 // will be filled when the prior region is processed. If neither of those
2983 2982 // apply, the first word in the region is the start of a live object or dead
2984 2983 // space.
2985 2984 assert(beg_addr > space(space_id)->bottom(), "sanity");
2986 2985 const RegionData* const cp = sd.region(beg_region);
2987 2986 if (cp->partial_obj_size() != 0) {
2988 2987 beg_addr = sd.partial_obj_end(beg_region);
2989 2988 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2990 2989 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2991 2990 }
2992 2991 }
2993 2992
2994 2993 if (beg_addr < end_addr) {
2995 2994 // A live object or block of dead space starts in this range of Regions.
2996 2995 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2997 2996
2998 2997 // Create closures and iterate.
2999 2998 UpdateOnlyClosure update_closure(mbm, cm, space_id);
3000 2999 FillClosure fill_closure(cm, space_id);
3001 3000 ParMarkBitMap::IterationStatus status;
3002 3001 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
3003 3002 dense_prefix_end);
3004 3003 if (status == ParMarkBitMap::incomplete) {
3005 3004 update_closure.do_addr(update_closure.source());
3006 3005 }
3007 3006 }
3008 3007
3009 3008 // Mark the regions as filled.
3010 3009 RegionData* const beg_cp = sd.region(beg_region);
3011 3010 RegionData* const end_cp = sd.region(end_region);
3012 3011 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
3013 3012 cp->set_completed();
3014 3013 }
3015 3014 }
3016 3015
3017 3016 // Return the SpaceId for the space containing addr. If addr is not in the
3018 3017 // heap, last_space_id is returned. In debug mode it expects the address to be
3019 3018 // in the heap and asserts such.
3020 3019 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
3021 3020 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
3022 3021
3023 3022 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
3024 3023 if (_space_info[id].space()->contains(addr)) {
3025 3024 return SpaceId(id);
3026 3025 }
3027 3026 }
3028 3027
3029 3028 assert(false, "no space contains the addr");
3030 3029 return last_space_id;
3031 3030 }
3032 3031
3033 3032 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
3034 3033 SpaceId id) {
3035 3034 assert(id < last_space_id, "bad space id");
3036 3035
3037 3036 ParallelCompactData& sd = summary_data();
3038 3037 const SpaceInfo* const space_info = _space_info + id;
3039 3038 ObjectStartArray* const start_array = space_info->start_array();
3040 3039
3041 3040 const MutableSpace* const space = space_info->space();
3042 3041 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
3043 3042 HeapWord* const beg_addr = space_info->dense_prefix();
3044 3043 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
3045 3044
3046 3045 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
3047 3046 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
3048 3047 const RegionData* cur_region;
3049 3048 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
3050 3049 HeapWord* const addr = cur_region->deferred_obj_addr();
3051 3050 if (addr != NULL) {
3052 3051 if (start_array != NULL) {
3053 3052 start_array->allocate_block(addr);
3054 3053 }
3055 3054 oop(addr)->update_contents(cm);
3056 3055 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3057 3056 }
3058 3057 }
3059 3058 }
3060 3059
3061 3060 // Skip over count live words starting from beg, and return the address of the
3062 3061 // next live word. Unless marked, the word corresponding to beg is assumed to
3063 3062 // be dead. Callers must either ensure beg does not correspond to the middle of
3064 3063 // an object, or account for those live words in some other way. Callers must
3065 3064 // also ensure that there are enough live words in the range [beg, end) to skip.
3066 3065 HeapWord*
3067 3066 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3068 3067 {
3069 3068 assert(count > 0, "sanity");
3070 3069
3071 3070 ParMarkBitMap* m = mark_bitmap();
3072 3071 idx_t bits_to_skip = m->words_to_bits(count);
3073 3072 idx_t cur_beg = m->addr_to_bit(beg);
3074 3073 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3075 3074
3076 3075 do {
3077 3076 cur_beg = m->find_obj_beg(cur_beg, search_end);
3078 3077 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3079 3078 const size_t obj_bits = cur_end - cur_beg + 1;
3080 3079 if (obj_bits > bits_to_skip) {
3081 3080 return m->bit_to_addr(cur_beg + bits_to_skip);
3082 3081 }
3083 3082 bits_to_skip -= obj_bits;
3084 3083 cur_beg = cur_end + 1;
3085 3084 } while (bits_to_skip > 0);
3086 3085
3087 3086 // Skipping the desired number of words landed just past the end of an object.
3088 3087 // Find the start of the next object.
3089 3088 cur_beg = m->find_obj_beg(cur_beg, search_end);
3090 3089 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3091 3090 return m->bit_to_addr(cur_beg);
3092 3091 }
3093 3092
3094 3093 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3095 3094 SpaceId src_space_id,
3096 3095 size_t src_region_idx)
3097 3096 {
3098 3097 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
3099 3098
3100 3099 const SplitInfo& split_info = _space_info[src_space_id].split_info();
3101 3100 if (split_info.dest_region_addr() == dest_addr) {
3102 3101 // The partial object ending at the split point contains the first word to
3103 3102 // be copied to dest_addr.
3104 3103 return split_info.first_src_addr();
3105 3104 }
3106 3105
3107 3106 const ParallelCompactData& sd = summary_data();
3108 3107 ParMarkBitMap* const bitmap = mark_bitmap();
3109 3108 const size_t RegionSize = ParallelCompactData::RegionSize;
3110 3109
3111 3110 assert(sd.is_region_aligned(dest_addr), "not aligned");
3112 3111 const RegionData* const src_region_ptr = sd.region(src_region_idx);
3113 3112 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
3114 3113 HeapWord* const src_region_destination = src_region_ptr->destination();
3115 3114
3116 3115 assert(dest_addr >= src_region_destination, "wrong src region");
3117 3116 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
3118 3117
3119 3118 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
3120 3119 HeapWord* const src_region_end = src_region_beg + RegionSize;
3121 3120
3122 3121 HeapWord* addr = src_region_beg;
3123 3122 if (dest_addr == src_region_destination) {
3124 3123 // Return the first live word in the source region.
3125 3124 if (partial_obj_size == 0) {
3126 3125 addr = bitmap->find_obj_beg(addr, src_region_end);
3127 3126 assert(addr < src_region_end, "no objects start in src region");
3128 3127 }
3129 3128 return addr;
3130 3129 }
3131 3130
3132 3131 // Must skip some live data.
3133 3132 size_t words_to_skip = dest_addr - src_region_destination;
3134 3133 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
3135 3134
3136 3135 if (partial_obj_size >= words_to_skip) {
3137 3136 // All the live words to skip are part of the partial object.
3138 3137 addr += words_to_skip;
3139 3138 if (partial_obj_size == words_to_skip) {
3140 3139 // Find the first live word past the partial object.
3141 3140 addr = bitmap->find_obj_beg(addr, src_region_end);
3142 3141 assert(addr < src_region_end, "wrong src region");
3143 3142 }
3144 3143 return addr;
3145 3144 }
3146 3145
3147 3146 // Skip over the partial object (if any).
3148 3147 if (partial_obj_size != 0) {
3149 3148 words_to_skip -= partial_obj_size;
3150 3149 addr += partial_obj_size;
3151 3150 }
3152 3151
3153 3152 // Skip over live words due to objects that start in the region.
3154 3153 addr = skip_live_words(addr, src_region_end, words_to_skip);
3155 3154 assert(addr < src_region_end, "wrong src region");
3156 3155 return addr;
3157 3156 }
3158 3157
3159 3158 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3160 3159 SpaceId src_space_id,
3161 3160 size_t beg_region,
3162 3161 HeapWord* end_addr)
3163 3162 {
3164 3163 ParallelCompactData& sd = summary_data();
3165 3164
3166 3165 #ifdef ASSERT
3167 3166 MutableSpace* const src_space = _space_info[src_space_id].space();
3168 3167 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3169 3168 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3170 3169 "src_space_id does not match beg_addr");
3171 3170 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3172 3171 "src_space_id does not match end_addr");
3173 3172 #endif // #ifdef ASSERT
3174 3173
3175 3174 RegionData* const beg = sd.region(beg_region);
3176 3175 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3177 3176
3178 3177 // Regions up to new_top() are enqueued if they become available.
3179 3178 HeapWord* const new_top = _space_info[src_space_id].new_top();
3180 3179 RegionData* const enqueue_end =
3181 3180 sd.addr_to_region_ptr(sd.region_align_up(new_top));
3182 3181
3183 3182 for (RegionData* cur = beg; cur < end; ++cur) {
3184 3183 assert(cur->data_size() > 0, "region must have live data");
3185 3184 cur->decrement_destination_count();
3186 3185 if (cur < enqueue_end && cur->available() && cur->claim()) {
3187 3186 cm->push_region(sd.region(cur));
3188 3187 }
3189 3188 }
3190 3189 }
3191 3190
3192 3191 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3193 3192 SpaceId& src_space_id,
3194 3193 HeapWord*& src_space_top,
3195 3194 HeapWord* end_addr)
3196 3195 {
3197 3196 typedef ParallelCompactData::RegionData RegionData;
3198 3197
3199 3198 ParallelCompactData& sd = PSParallelCompact::summary_data();
3200 3199 const size_t region_size = ParallelCompactData::RegionSize;
3201 3200
3202 3201 size_t src_region_idx = 0;
3203 3202
3204 3203 // Skip empty regions (if any) up to the top of the space.
3205 3204 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3206 3205 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3207 3206 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3208 3207 const RegionData* const top_region_ptr =
3209 3208 sd.addr_to_region_ptr(top_aligned_up);
3210 3209 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3211 3210 ++src_region_ptr;
3212 3211 }
3213 3212
3214 3213 if (src_region_ptr < top_region_ptr) {
3215 3214 // The next source region is in the current space. Update src_region_idx
3216 3215 // and the source address to match src_region_ptr.
3217 3216 src_region_idx = sd.region(src_region_ptr);
3218 3217 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3219 3218 if (src_region_addr > closure.source()) {
3220 3219 closure.set_source(src_region_addr);
3221 3220 }
3222 3221 return src_region_idx;
3223 3222 }
3224 3223
3225 3224 // Switch to a new source space and find the first non-empty region.
3226 3225 unsigned int space_id = src_space_id + 1;
3227 3226 assert(space_id < last_space_id, "not enough spaces");
3228 3227
3229 3228 HeapWord* const destination = closure.destination();
3230 3229
3231 3230 do {
3232 3231 MutableSpace* space = _space_info[space_id].space();
3233 3232 HeapWord* const bottom = space->bottom();
3234 3233 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3235 3234
3236 3235 // Iterate over the spaces that do not compact into themselves.
3237 3236 if (bottom_cp->destination() != bottom) {
3238 3237 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3239 3238 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3240 3239
3241 3240 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3242 3241 if (src_cp->live_obj_size() > 0) {
3243 3242 // Found it.
3244 3243 assert(src_cp->destination() == destination,
3245 3244 "first live obj in the space must match the destination");
3246 3245 assert(src_cp->partial_obj_size() == 0,
3247 3246 "a space cannot begin with a partial obj");
3248 3247
3249 3248 src_space_id = SpaceId(space_id);
3250 3249 src_space_top = space->top();
3251 3250 const size_t src_region_idx = sd.region(src_cp);
3252 3251 closure.set_source(sd.region_to_addr(src_region_idx));
3253 3252 return src_region_idx;
3254 3253 } else {
3255 3254 assert(src_cp->data_size() == 0, "sanity");
3256 3255 }
3257 3256 }
3258 3257 }
3259 3258 } while (++space_id < last_space_id);
3260 3259
3261 3260 assert(false, "no source region was found");
3262 3261 return 0;
3263 3262 }
3264 3263
3265 3264 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3266 3265 {
3267 3266 typedef ParMarkBitMap::IterationStatus IterationStatus;
3268 3267 const size_t RegionSize = ParallelCompactData::RegionSize;
3269 3268 ParMarkBitMap* const bitmap = mark_bitmap();
3270 3269 ParallelCompactData& sd = summary_data();
3271 3270 RegionData* const region_ptr = sd.region(region_idx);
3272 3271
3273 3272 // Get the items needed to construct the closure.
3274 3273 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3275 3274 SpaceId dest_space_id = space_id(dest_addr);
3276 3275 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3277 3276 HeapWord* new_top = _space_info[dest_space_id].new_top();
3278 3277 assert(dest_addr < new_top, "sanity");
3279 3278 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3280 3279
3281 3280 // Get the source region and related info.
3282 3281 size_t src_region_idx = region_ptr->source_region();
3283 3282 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3284 3283 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3285 3284
3286 3285 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3287 3286 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3288 3287
3289 3288 // Adjust src_region_idx to prepare for decrementing destination counts (the
3290 3289 // destination count is not decremented when a region is copied to itself).
3291 3290 if (src_region_idx == region_idx) {
3292 3291 src_region_idx += 1;
3293 3292 }
3294 3293
3295 3294 if (bitmap->is_unmarked(closure.source())) {
3296 3295 // The first source word is in the middle of an object; copy the remainder
3297 3296 // of the object or as much as will fit. The fact that pointer updates were
3298 3297 // deferred will be noted when the object header is processed.
3299 3298 HeapWord* const old_src_addr = closure.source();
3300 3299 closure.copy_partial_obj();
3301 3300 if (closure.is_full()) {
3302 3301 decrement_destination_counts(cm, src_space_id, src_region_idx,
3303 3302 closure.source());
3304 3303 region_ptr->set_deferred_obj_addr(NULL);
3305 3304 region_ptr->set_completed();
3306 3305 return;
3307 3306 }
3308 3307
3309 3308 HeapWord* const end_addr = sd.region_align_down(closure.source());
3310 3309 if (sd.region_align_down(old_src_addr) != end_addr) {
3311 3310 // The partial object was copied from more than one source region.
3312 3311 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3313 3312
3314 3313 // Move to the next source region, possibly switching spaces as well. All
3315 3314 // args except end_addr may be modified.
3316 3315 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3317 3316 end_addr);
3318 3317 }
3319 3318 }
3320 3319
3321 3320 do {
3322 3321 HeapWord* const cur_addr = closure.source();
3323 3322 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3324 3323 src_space_top);
3325 3324 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3326 3325
3327 3326 if (status == ParMarkBitMap::incomplete) {
3328 3327 // The last obj that starts in the source region does not end in the
3329 3328 // region.
3330 3329 assert(closure.source() < end_addr, "sanity");
3331 3330 HeapWord* const obj_beg = closure.source();
3332 3331 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3333 3332 src_space_top);
3334 3333 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3335 3334 if (obj_end < range_end) {
3336 3335 // The end was found; the entire object will fit.
3337 3336 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3338 3337 assert(status != ParMarkBitMap::would_overflow, "sanity");
3339 3338 } else {
3340 3339 // The end was not found; the object will not fit.
3341 3340 assert(range_end < src_space_top, "obj cannot cross space boundary");
3342 3341 status = ParMarkBitMap::would_overflow;
3343 3342 }
3344 3343 }
3345 3344
3346 3345 if (status == ParMarkBitMap::would_overflow) {
3347 3346 // The last object did not fit. Note that interior oop updates were
3348 3347 // deferred, then copy enough of the object to fill the region.
3349 3348 region_ptr->set_deferred_obj_addr(closure.destination());
3350 3349 status = closure.copy_until_full(); // copies from closure.source()
3351 3350
3352 3351 decrement_destination_counts(cm, src_space_id, src_region_idx,
3353 3352 closure.source());
3354 3353 region_ptr->set_completed();
3355 3354 return;
3356 3355 }
3357 3356
3358 3357 if (status == ParMarkBitMap::full) {
3359 3358 decrement_destination_counts(cm, src_space_id, src_region_idx,
3360 3359 closure.source());
3361 3360 region_ptr->set_deferred_obj_addr(NULL);
3362 3361 region_ptr->set_completed();
3363 3362 return;
3364 3363 }
3365 3364
3366 3365 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3367 3366
3368 3367 // Move to the next source region, possibly switching spaces as well. All
3369 3368 // args except end_addr may be modified.
3370 3369 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3371 3370 end_addr);
3372 3371 } while (true);
3373 3372 }
3374 3373
3375 3374 void
3376 3375 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3377 3376 const MutableSpace* sp = space(space_id);
3378 3377 if (sp->is_empty()) {
3379 3378 return;
3380 3379 }
3381 3380
3382 3381 ParallelCompactData& sd = PSParallelCompact::summary_data();
3383 3382 ParMarkBitMap* const bitmap = mark_bitmap();
3384 3383 HeapWord* const dp_addr = dense_prefix(space_id);
3385 3384 HeapWord* beg_addr = sp->bottom();
3386 3385 HeapWord* end_addr = sp->top();
3387 3386
3388 3387 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3389 3388
3390 3389 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3391 3390 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3392 3391 if (beg_region < dp_region) {
3393 3392 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3394 3393 }
3395 3394
3396 3395 // The destination of the first live object that starts in the region is one
3397 3396 // past the end of the partial object entering the region (if any).
3398 3397 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3399 3398 HeapWord* const new_top = _space_info[space_id].new_top();
3400 3399 assert(new_top >= dest_addr, "bad new_top value");
3401 3400 const size_t words = pointer_delta(new_top, dest_addr);
3402 3401
3403 3402 if (words > 0) {
3404 3403 ObjectStartArray* start_array = _space_info[space_id].start_array();
3405 3404 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3406 3405
3407 3406 ParMarkBitMap::IterationStatus status;
3408 3407 status = bitmap->iterate(&closure, dest_addr, end_addr);
3409 3408 assert(status == ParMarkBitMap::full, "iteration not complete");
3410 3409 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3411 3410 "live objects skipped because closure is full");
3412 3411 }
3413 3412 }
3414 3413
3415 3414 jlong PSParallelCompact::millis_since_last_gc() {
3416 3415 // We need a monotonically non-deccreasing time in ms but
3417 3416 // os::javaTimeMillis() does not guarantee monotonicity.
3418 3417 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3419 3418 jlong ret_val = now - _time_of_last_gc;
3420 3419 // XXX See note in genCollectedHeap::millis_since_last_gc().
3421 3420 if (ret_val < 0) {
3422 3421 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3423 3422 return 0;
3424 3423 }
3425 3424 return ret_val;
3426 3425 }
3427 3426
3428 3427 void PSParallelCompact::reset_millis_since_last_gc() {
3429 3428 // We need a monotonically non-deccreasing time in ms but
3430 3429 // os::javaTimeMillis() does not guarantee monotonicity.
3431 3430 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3432 3431 }
3433 3432
3434 3433 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3435 3434 {
3436 3435 if (source() != destination()) {
3437 3436 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3438 3437 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3439 3438 }
3440 3439 update_state(words_remaining());
3441 3440 assert(is_full(), "sanity");
3442 3441 return ParMarkBitMap::full;
3443 3442 }
3444 3443
3445 3444 void MoveAndUpdateClosure::copy_partial_obj()
3446 3445 {
3447 3446 size_t words = words_remaining();
3448 3447
3449 3448 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3450 3449 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3451 3450 if (end_addr < range_end) {
3452 3451 words = bitmap()->obj_size(source(), end_addr);
3453 3452 }
3454 3453
3455 3454 // This test is necessary; if omitted, the pointer updates to a partial object
3456 3455 // that crosses the dense prefix boundary could be overwritten.
3457 3456 if (source() != destination()) {
3458 3457 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3459 3458 Copy::aligned_conjoint_words(source(), destination(), words);
3460 3459 }
3461 3460 update_state(words);
3462 3461 }
3463 3462
3464 3463 ParMarkBitMapClosure::IterationStatus
3465 3464 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3466 3465 assert(destination() != NULL, "sanity");
3467 3466 assert(bitmap()->obj_size(addr) == words, "bad size");
3468 3467
3469 3468 _source = addr;
3470 3469 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3471 3470 destination(), "wrong destination");
3472 3471
3473 3472 if (words > words_remaining()) {
3474 3473 return ParMarkBitMap::would_overflow;
3475 3474 }
3476 3475
3477 3476 // The start_array must be updated even if the object is not moving.
3478 3477 if (_start_array != NULL) {
3479 3478 _start_array->allocate_block(destination());
3480 3479 }
3481 3480
3482 3481 if (destination() != source()) {
3483 3482 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3484 3483 Copy::aligned_conjoint_words(source(), destination(), words);
3485 3484 }
3486 3485
3487 3486 oop moved_oop = (oop) destination();
3488 3487 moved_oop->update_contents(compaction_manager());
3489 3488 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3490 3489
3491 3490 update_state(words);
3492 3491 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3493 3492 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3494 3493 }
3495 3494
3496 3495 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3497 3496 ParCompactionManager* cm,
3498 3497 PSParallelCompact::SpaceId space_id) :
3499 3498 ParMarkBitMapClosure(mbm, cm),
3500 3499 _space_id(space_id),
3501 3500 _start_array(PSParallelCompact::start_array(space_id))
3502 3501 {
3503 3502 }
3504 3503
3505 3504 // Updates the references in the object to their new values.
3506 3505 ParMarkBitMapClosure::IterationStatus
3507 3506 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3508 3507 do_addr(addr);
3509 3508 return ParMarkBitMap::incomplete;
3510 3509 }
3511 3510
3512 3511 // Prepare for compaction. This method is executed once
3513 3512 // (i.e., by a single thread) before compaction.
3514 3513 // Save the updated location of the intArrayKlassObj for
3515 3514 // filling holes in the dense prefix.
3516 3515 void PSParallelCompact::compact_prologue() {
3517 3516 _updated_int_array_klass_obj = (klassOop)
3518 3517 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3519 3518 }
↓ open down ↓ |
1444 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX