Mercurial > hg > graal-compiler
annotate src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp @ 237:1fdb98a17101
6716785: implicit null checks not triggering with CompressedOops
Summary: allocate alignment-sized page(s) below java heap so that memory accesses at heap_base+1page give signal and cause an implicit null check
Reviewed-by: kvn, jmasa, phh, jcoomes
| author | coleenp |
|---|---|
| date | Sat, 19 Jul 2008 17:38:22 -0400 |
| parents | 9c2ecc2ffb12 |
| children | 850fdf70db2b |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 196 | 2 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 #include "incls/_precompiled.incl" | |
| 26 #include "incls/_psParallelCompact.cpp.incl" | |
| 27 | |
| 28 #include <math.h> | |
| 29 | |
| 30 // All sizes are in HeapWords. | |
| 31 const size_t ParallelCompactData::Log2ChunkSize = 9; // 512 words | |
| 32 const size_t ParallelCompactData::ChunkSize = (size_t)1 << Log2ChunkSize; | |
| 33 const size_t ParallelCompactData::ChunkSizeBytes = ChunkSize << LogHeapWordSize; | |
| 34 const size_t ParallelCompactData::ChunkSizeOffsetMask = ChunkSize - 1; | |
| 35 const size_t ParallelCompactData::ChunkAddrOffsetMask = ChunkSizeBytes - 1; | |
| 36 const size_t ParallelCompactData::ChunkAddrMask = ~ChunkAddrOffsetMask; | |
| 37 | |
| 38 // 32-bit: 128 words covers 4 bitmap words | |
| 39 // 64-bit: 128 words covers 2 bitmap words | |
| 40 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words | |
| 41 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; | |
| 42 const size_t ParallelCompactData::BlockOffsetMask = BlockSize - 1; | |
| 43 const size_t ParallelCompactData::BlockMask = ~BlockOffsetMask; | |
| 44 | |
| 45 const size_t ParallelCompactData::BlocksPerChunk = ChunkSize / BlockSize; | |
| 46 | |
| 47 const ParallelCompactData::ChunkData::chunk_sz_t | |
| 48 ParallelCompactData::ChunkData::dc_shift = 27; | |
| 49 | |
| 50 const ParallelCompactData::ChunkData::chunk_sz_t | |
| 51 ParallelCompactData::ChunkData::dc_mask = ~0U << dc_shift; | |
| 52 | |
| 53 const ParallelCompactData::ChunkData::chunk_sz_t | |
| 54 ParallelCompactData::ChunkData::dc_one = 0x1U << dc_shift; | |
| 55 | |
| 56 const ParallelCompactData::ChunkData::chunk_sz_t | |
| 57 ParallelCompactData::ChunkData::los_mask = ~dc_mask; | |
| 58 | |
| 59 const ParallelCompactData::ChunkData::chunk_sz_t | |
| 60 ParallelCompactData::ChunkData::dc_claimed = 0x8U << dc_shift; | |
| 61 | |
| 62 const ParallelCompactData::ChunkData::chunk_sz_t | |
| 63 ParallelCompactData::ChunkData::dc_completed = 0xcU << dc_shift; | |
| 64 | |
| 65 #ifdef ASSERT | |
| 66 short ParallelCompactData::BlockData::_cur_phase = 0; | |
| 67 #endif | |
| 68 | |
| 69 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; | |
| 70 bool PSParallelCompact::_print_phases = false; | |
| 71 | |
| 72 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; | |
| 73 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL; | |
| 74 | |
| 75 double PSParallelCompact::_dwl_mean; | |
| 76 double PSParallelCompact::_dwl_std_dev; | |
| 77 double PSParallelCompact::_dwl_first_term; | |
| 78 double PSParallelCompact::_dwl_adjustment; | |
| 79 #ifdef ASSERT | |
| 80 bool PSParallelCompact::_dwl_initialized = false; | |
| 81 #endif // #ifdef ASSERT | |
| 82 | |
| 83 #ifdef VALIDATE_MARK_SWEEP | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
84 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL; |
| 0 | 85 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL; |
| 86 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL; | |
| 87 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL; | |
| 88 size_t PSParallelCompact::_live_oops_index = 0; | |
| 89 size_t PSParallelCompact::_live_oops_index_at_perm = 0; | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
90 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
91 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL; |
| 0 | 92 bool PSParallelCompact::_pointer_tracking = false; |
| 93 bool PSParallelCompact::_root_tracking = true; | |
| 94 | |
| 95 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL; | |
| 96 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL; | |
| 97 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL; | |
| 98 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL; | |
| 99 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL; | |
| 100 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL; | |
| 101 #endif | |
| 102 | |
| 103 // XXX beg - verification code; only works while we also mark in object headers | |
| 104 static void | |
| 105 verify_mark_bitmap(ParMarkBitMap& _mark_bitmap) | |
| 106 { | |
| 107 ParallelScavengeHeap* heap = PSParallelCompact::gc_heap(); | |
| 108 | |
| 109 PSPermGen* perm_gen = heap->perm_gen(); | |
| 110 PSOldGen* old_gen = heap->old_gen(); | |
| 111 PSYoungGen* young_gen = heap->young_gen(); | |
| 112 | |
| 113 MutableSpace* perm_space = perm_gen->object_space(); | |
| 114 MutableSpace* old_space = old_gen->object_space(); | |
| 115 MutableSpace* eden_space = young_gen->eden_space(); | |
| 116 MutableSpace* from_space = young_gen->from_space(); | |
| 117 MutableSpace* to_space = young_gen->to_space(); | |
| 118 | |
| 119 // 'from_space' here is the survivor space at the lower address. | |
| 120 if (to_space->bottom() < from_space->bottom()) { | |
| 121 from_space = to_space; | |
| 122 to_space = young_gen->from_space(); | |
| 123 } | |
| 124 | |
| 125 HeapWord* boundaries[12]; | |
| 126 unsigned int bidx = 0; | |
| 127 const unsigned int bidx_max = sizeof(boundaries) / sizeof(boundaries[0]); | |
| 128 | |
| 129 boundaries[0] = perm_space->bottom(); | |
| 130 boundaries[1] = perm_space->top(); | |
| 131 boundaries[2] = old_space->bottom(); | |
| 132 boundaries[3] = old_space->top(); | |
| 133 boundaries[4] = eden_space->bottom(); | |
| 134 boundaries[5] = eden_space->top(); | |
| 135 boundaries[6] = from_space->bottom(); | |
| 136 boundaries[7] = from_space->top(); | |
| 137 boundaries[8] = to_space->bottom(); | |
| 138 boundaries[9] = to_space->top(); | |
| 139 boundaries[10] = to_space->end(); | |
| 140 boundaries[11] = to_space->end(); | |
| 141 | |
| 142 BitMap::idx_t beg_bit = 0; | |
| 143 BitMap::idx_t end_bit; | |
| 144 BitMap::idx_t tmp_bit; | |
| 145 const BitMap::idx_t last_bit = _mark_bitmap.size(); | |
| 146 do { | |
| 147 HeapWord* addr = _mark_bitmap.bit_to_addr(beg_bit); | |
| 148 if (_mark_bitmap.is_marked(beg_bit)) { | |
| 149 oop obj = (oop)addr; | |
| 150 assert(obj->is_gc_marked(), "obj header is not marked"); | |
| 151 end_bit = _mark_bitmap.find_obj_end(beg_bit, last_bit); | |
| 152 const size_t size = _mark_bitmap.obj_size(beg_bit, end_bit); | |
| 153 assert(size == (size_t)obj->size(), "end bit wrong?"); | |
| 154 beg_bit = _mark_bitmap.find_obj_beg(beg_bit + 1, last_bit); | |
| 155 assert(beg_bit > end_bit, "bit set in middle of an obj"); | |
| 156 } else { | |
| 157 if (addr >= boundaries[bidx] && addr < boundaries[bidx + 1]) { | |
| 158 // a dead object in the current space. | |
| 159 oop obj = (oop)addr; | |
| 160 end_bit = _mark_bitmap.addr_to_bit(addr + obj->size()); | |
| 161 assert(!obj->is_gc_marked(), "obj marked in header, not in bitmap"); | |
| 162 tmp_bit = beg_bit + 1; | |
| 163 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, end_bit); | |
| 164 assert(beg_bit == end_bit, "beg bit set in unmarked obj"); | |
| 165 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, end_bit); | |
| 166 assert(beg_bit == end_bit, "end bit set in unmarked obj"); | |
| 167 } else if (addr < boundaries[bidx + 2]) { | |
| 168 // addr is between top in the current space and bottom in the next. | |
| 169 end_bit = beg_bit + pointer_delta(boundaries[bidx + 2], addr); | |
| 170 tmp_bit = beg_bit; | |
| 171 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, end_bit); | |
| 172 assert(beg_bit == end_bit, "beg bit set above top"); | |
| 173 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, end_bit); | |
| 174 assert(beg_bit == end_bit, "end bit set above top"); | |
| 175 bidx += 2; | |
| 176 } else if (bidx < bidx_max - 2) { | |
| 177 bidx += 2; // ??? | |
| 178 } else { | |
| 179 tmp_bit = beg_bit; | |
| 180 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, last_bit); | |
| 181 assert(beg_bit == last_bit, "beg bit set outside heap"); | |
| 182 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, last_bit); | |
| 183 assert(beg_bit == last_bit, "end bit set outside heap"); | |
| 184 } | |
| 185 } | |
| 186 } while (beg_bit < last_bit); | |
| 187 } | |
| 188 // XXX end - verification code; only works while we also mark in object headers | |
| 189 | |
| 190 #ifndef PRODUCT | |
| 191 const char* PSParallelCompact::space_names[] = { | |
| 192 "perm", "old ", "eden", "from", "to " | |
| 193 }; | |
| 194 | |
| 195 void PSParallelCompact::print_chunk_ranges() | |
| 196 { | |
| 197 tty->print_cr("space bottom top end new_top"); | |
| 198 tty->print_cr("------ ---------- ---------- ---------- ----------"); | |
| 199 | |
| 200 for (unsigned int id = 0; id < last_space_id; ++id) { | |
| 201 const MutableSpace* space = _space_info[id].space(); | |
| 202 tty->print_cr("%u %s " | |
| 203 SIZE_FORMAT_W("10") " " SIZE_FORMAT_W("10") " " | |
| 204 SIZE_FORMAT_W("10") " " SIZE_FORMAT_W("10") " ", | |
| 205 id, space_names[id], | |
| 206 summary_data().addr_to_chunk_idx(space->bottom()), | |
| 207 summary_data().addr_to_chunk_idx(space->top()), | |
| 208 summary_data().addr_to_chunk_idx(space->end()), | |
| 209 summary_data().addr_to_chunk_idx(_space_info[id].new_top())); | |
| 210 } | |
| 211 } | |
| 212 | |
| 213 void | |
| 214 print_generic_summary_chunk(size_t i, const ParallelCompactData::ChunkData* c) | |
| 215 { | |
| 216 #define CHUNK_IDX_FORMAT SIZE_FORMAT_W("7") | |
| 217 #define CHUNK_DATA_FORMAT SIZE_FORMAT_W("5") | |
| 218 | |
| 219 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 220 size_t dci = c->destination() ? sd.addr_to_chunk_idx(c->destination()) : 0; | |
| 221 tty->print_cr(CHUNK_IDX_FORMAT " " PTR_FORMAT " " | |
| 222 CHUNK_IDX_FORMAT " " PTR_FORMAT " " | |
| 223 CHUNK_DATA_FORMAT " " CHUNK_DATA_FORMAT " " | |
| 224 CHUNK_DATA_FORMAT " " CHUNK_IDX_FORMAT " %d", | |
| 225 i, c->data_location(), dci, c->destination(), | |
| 226 c->partial_obj_size(), c->live_obj_size(), | |
| 227 c->data_size(), c->source_chunk(), c->destination_count()); | |
| 228 | |
| 229 #undef CHUNK_IDX_FORMAT | |
| 230 #undef CHUNK_DATA_FORMAT | |
| 231 } | |
| 232 | |
| 233 void | |
| 234 print_generic_summary_data(ParallelCompactData& summary_data, | |
| 235 HeapWord* const beg_addr, | |
| 236 HeapWord* const end_addr) | |
| 237 { | |
| 238 size_t total_words = 0; | |
| 239 size_t i = summary_data.addr_to_chunk_idx(beg_addr); | |
| 240 const size_t last = summary_data.addr_to_chunk_idx(end_addr); | |
| 241 HeapWord* pdest = 0; | |
| 242 | |
| 243 while (i <= last) { | |
| 244 ParallelCompactData::ChunkData* c = summary_data.chunk(i); | |
| 245 if (c->data_size() != 0 || c->destination() != pdest) { | |
| 246 print_generic_summary_chunk(i, c); | |
| 247 total_words += c->data_size(); | |
| 248 pdest = c->destination(); | |
| 249 } | |
| 250 ++i; | |
| 251 } | |
| 252 | |
| 253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); | |
| 254 } | |
| 255 | |
| 256 void | |
| 257 print_generic_summary_data(ParallelCompactData& summary_data, | |
| 258 SpaceInfo* space_info) | |
| 259 { | |
| 260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { | |
| 261 const MutableSpace* space = space_info[id].space(); | |
| 262 print_generic_summary_data(summary_data, space->bottom(), | |
| 263 MAX2(space->top(), space_info[id].new_top())); | |
| 264 } | |
| 265 } | |
| 266 | |
| 267 void | |
| 268 print_initial_summary_chunk(size_t i, | |
| 269 const ParallelCompactData::ChunkData* c, | |
| 270 bool newline = true) | |
| 271 { | |
| 272 tty->print(SIZE_FORMAT_W("5") " " PTR_FORMAT " " | |
| 273 SIZE_FORMAT_W("5") " " SIZE_FORMAT_W("5") " " | |
| 274 SIZE_FORMAT_W("5") " " SIZE_FORMAT_W("5") " %d", | |
| 275 i, c->destination(), | |
| 276 c->partial_obj_size(), c->live_obj_size(), | |
| 277 c->data_size(), c->source_chunk(), c->destination_count()); | |
| 278 if (newline) tty->cr(); | |
| 279 } | |
| 280 | |
| 281 void | |
| 282 print_initial_summary_data(ParallelCompactData& summary_data, | |
| 283 const MutableSpace* space) { | |
| 284 if (space->top() == space->bottom()) { | |
| 285 return; | |
| 286 } | |
| 287 | |
| 288 const size_t chunk_size = ParallelCompactData::ChunkSize; | |
| 289 HeapWord* const top_aligned_up = summary_data.chunk_align_up(space->top()); | |
| 290 const size_t end_chunk = summary_data.addr_to_chunk_idx(top_aligned_up); | |
| 291 const ParallelCompactData::ChunkData* c = summary_data.chunk(end_chunk - 1); | |
| 292 HeapWord* end_addr = c->destination() + c->data_size(); | |
| 293 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); | |
| 294 | |
| 295 // Print (and count) the full chunks at the beginning of the space. | |
| 296 size_t full_chunk_count = 0; | |
| 297 size_t i = summary_data.addr_to_chunk_idx(space->bottom()); | |
| 298 while (i < end_chunk && summary_data.chunk(i)->data_size() == chunk_size) { | |
| 299 print_initial_summary_chunk(i, summary_data.chunk(i)); | |
| 300 ++full_chunk_count; | |
| 301 ++i; | |
| 302 } | |
| 303 | |
| 304 size_t live_to_right = live_in_space - full_chunk_count * chunk_size; | |
| 305 | |
| 306 double max_reclaimed_ratio = 0.0; | |
| 307 size_t max_reclaimed_ratio_chunk = 0; | |
| 308 size_t max_dead_to_right = 0; | |
| 309 size_t max_live_to_right = 0; | |
| 310 | |
| 311 // Print the 'reclaimed ratio' for chunks while there is something live in the | |
| 312 // chunk or to the right of it. The remaining chunks are empty (and | |
| 313 // uninteresting), and computing the ratio will result in division by 0. | |
| 314 while (i < end_chunk && live_to_right > 0) { | |
| 315 c = summary_data.chunk(i); | |
| 316 HeapWord* const chunk_addr = summary_data.chunk_to_addr(i); | |
| 317 const size_t used_to_right = pointer_delta(space->top(), chunk_addr); | |
| 318 const size_t dead_to_right = used_to_right - live_to_right; | |
| 319 const double reclaimed_ratio = double(dead_to_right) / live_to_right; | |
| 320 | |
| 321 if (reclaimed_ratio > max_reclaimed_ratio) { | |
| 322 max_reclaimed_ratio = reclaimed_ratio; | |
| 323 max_reclaimed_ratio_chunk = i; | |
| 324 max_dead_to_right = dead_to_right; | |
| 325 max_live_to_right = live_to_right; | |
| 326 } | |
| 327 | |
| 328 print_initial_summary_chunk(i, c, false); | |
| 329 tty->print_cr(" %12.10f " SIZE_FORMAT_W("10") " " SIZE_FORMAT_W("10"), | |
| 330 reclaimed_ratio, dead_to_right, live_to_right); | |
| 331 | |
| 332 live_to_right -= c->data_size(); | |
| 333 ++i; | |
| 334 } | |
| 335 | |
| 336 // Any remaining chunks are empty. Print one more if there is one. | |
| 337 if (i < end_chunk) { | |
| 338 print_initial_summary_chunk(i, summary_data.chunk(i)); | |
| 339 } | |
| 340 | |
| 341 tty->print_cr("max: " SIZE_FORMAT_W("4") " d2r=" SIZE_FORMAT_W("10") " " | |
| 342 "l2r=" SIZE_FORMAT_W("10") " max_ratio=%14.12f", | |
| 343 max_reclaimed_ratio_chunk, max_dead_to_right, | |
| 344 max_live_to_right, max_reclaimed_ratio); | |
| 345 } | |
| 346 | |
| 347 void | |
| 348 print_initial_summary_data(ParallelCompactData& summary_data, | |
| 349 SpaceInfo* space_info) { | |
| 350 unsigned int id = PSParallelCompact::perm_space_id; | |
| 351 const MutableSpace* space; | |
| 352 do { | |
| 353 space = space_info[id].space(); | |
| 354 print_initial_summary_data(summary_data, space); | |
| 355 } while (++id < PSParallelCompact::eden_space_id); | |
| 356 | |
| 357 do { | |
| 358 space = space_info[id].space(); | |
| 359 print_generic_summary_data(summary_data, space->bottom(), space->top()); | |
| 360 } while (++id < PSParallelCompact::last_space_id); | |
| 361 } | |
| 362 #endif // #ifndef PRODUCT | |
| 363 | |
| 364 #ifdef ASSERT | |
| 365 size_t add_obj_count; | |
| 366 size_t add_obj_size; | |
| 367 size_t mark_bitmap_count; | |
| 368 size_t mark_bitmap_size; | |
| 369 #endif // #ifdef ASSERT | |
| 370 | |
| 371 ParallelCompactData::ParallelCompactData() | |
| 372 { | |
| 373 _region_start = 0; | |
| 374 | |
| 375 _chunk_vspace = 0; | |
| 376 _chunk_data = 0; | |
| 377 _chunk_count = 0; | |
| 378 | |
| 379 _block_vspace = 0; | |
| 380 _block_data = 0; | |
| 381 _block_count = 0; | |
| 382 } | |
| 383 | |
| 384 bool ParallelCompactData::initialize(MemRegion covered_region) | |
| 385 { | |
| 386 _region_start = covered_region.start(); | |
| 387 const size_t region_size = covered_region.word_size(); | |
| 388 DEBUG_ONLY(_region_end = _region_start + region_size;) | |
| 389 | |
| 390 assert(chunk_align_down(_region_start) == _region_start, | |
| 391 "region start not aligned"); | |
| 392 assert((region_size & ChunkSizeOffsetMask) == 0, | |
| 393 "region size not a multiple of ChunkSize"); | |
| 394 | |
| 395 bool result = initialize_chunk_data(region_size); | |
| 396 | |
| 397 // Initialize the block data if it will be used for updating pointers, or if | |
| 398 // this is a debug build. | |
| 399 if (!UseParallelOldGCChunkPointerCalc || trueInDebug) { | |
| 400 result = result && initialize_block_data(region_size); | |
| 401 } | |
| 402 | |
| 403 return result; | |
| 404 } | |
| 405 | |
| 406 PSVirtualSpace* | |
| 407 ParallelCompactData::create_vspace(size_t count, size_t element_size) | |
| 408 { | |
| 409 const size_t raw_bytes = count * element_size; | |
| 410 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); | |
| 411 const size_t granularity = os::vm_allocation_granularity(); | |
| 412 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity)); | |
| 413 | |
| 414 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : | |
| 415 MAX2(page_sz, granularity); | |
|
79
82db0859acbe
6642862: Code cache allocation fails with large pages after 6588638
jcoomes
parents:
0
diff
changeset
|
416 ReservedSpace rs(bytes, rs_align, rs_align > 0); |
| 0 | 417 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), |
| 418 rs.size()); | |
| 419 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); | |
| 420 if (vspace != 0) { | |
| 421 if (vspace->expand_by(bytes)) { | |
| 422 return vspace; | |
| 423 } | |
| 424 delete vspace; | |
|
237
1fdb98a17101
6716785: implicit null checks not triggering with CompressedOops
coleenp
parents:
235
diff
changeset
|
425 // Release memory reserved in the space. |
|
1fdb98a17101
6716785: implicit null checks not triggering with CompressedOops
coleenp
parents:
235
diff
changeset
|
426 rs.release(); |
| 0 | 427 } |
| 428 | |
| 429 return 0; | |
| 430 } | |
| 431 | |
| 432 bool ParallelCompactData::initialize_chunk_data(size_t region_size) | |
| 433 { | |
| 434 const size_t count = (region_size + ChunkSizeOffsetMask) >> Log2ChunkSize; | |
| 435 _chunk_vspace = create_vspace(count, sizeof(ChunkData)); | |
| 436 if (_chunk_vspace != 0) { | |
| 437 _chunk_data = (ChunkData*)_chunk_vspace->reserved_low_addr(); | |
| 438 _chunk_count = count; | |
| 439 return true; | |
| 440 } | |
| 441 return false; | |
| 442 } | |
| 443 | |
| 444 bool ParallelCompactData::initialize_block_data(size_t region_size) | |
| 445 { | |
| 446 const size_t count = (region_size + BlockOffsetMask) >> Log2BlockSize; | |
| 447 _block_vspace = create_vspace(count, sizeof(BlockData)); | |
| 448 if (_block_vspace != 0) { | |
| 449 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); | |
| 450 _block_count = count; | |
| 451 return true; | |
| 452 } | |
| 453 return false; | |
| 454 } | |
| 455 | |
| 456 void ParallelCompactData::clear() | |
| 457 { | |
| 458 if (_block_data) { | |
| 459 memset(_block_data, 0, _block_vspace->committed_size()); | |
| 460 } | |
| 461 memset(_chunk_data, 0, _chunk_vspace->committed_size()); | |
| 462 } | |
| 463 | |
| 464 void ParallelCompactData::clear_range(size_t beg_chunk, size_t end_chunk) { | |
| 465 assert(beg_chunk <= _chunk_count, "beg_chunk out of range"); | |
| 466 assert(end_chunk <= _chunk_count, "end_chunk out of range"); | |
| 467 assert(ChunkSize % BlockSize == 0, "ChunkSize not a multiple of BlockSize"); | |
| 468 | |
| 469 const size_t chunk_cnt = end_chunk - beg_chunk; | |
| 470 | |
| 471 if (_block_data) { | |
| 472 const size_t blocks_per_chunk = ChunkSize / BlockSize; | |
| 473 const size_t beg_block = beg_chunk * blocks_per_chunk; | |
| 474 const size_t block_cnt = chunk_cnt * blocks_per_chunk; | |
| 475 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); | |
| 476 } | |
| 477 memset(_chunk_data + beg_chunk, 0, chunk_cnt * sizeof(ChunkData)); | |
| 478 } | |
| 479 | |
| 480 HeapWord* ParallelCompactData::partial_obj_end(size_t chunk_idx) const | |
| 481 { | |
| 482 const ChunkData* cur_cp = chunk(chunk_idx); | |
| 483 const ChunkData* const end_cp = chunk(chunk_count() - 1); | |
| 484 | |
| 485 HeapWord* result = chunk_to_addr(chunk_idx); | |
| 486 if (cur_cp < end_cp) { | |
| 487 do { | |
| 488 result += cur_cp->partial_obj_size(); | |
| 489 } while (cur_cp->partial_obj_size() == ChunkSize && ++cur_cp < end_cp); | |
| 490 } | |
| 491 return result; | |
| 492 } | |
| 493 | |
| 494 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) | |
| 495 { | |
| 496 const size_t obj_ofs = pointer_delta(addr, _region_start); | |
| 497 const size_t beg_chunk = obj_ofs >> Log2ChunkSize; | |
| 498 const size_t end_chunk = (obj_ofs + len - 1) >> Log2ChunkSize; | |
| 499 | |
| 500 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) | |
| 501 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) | |
| 502 | |
| 503 if (beg_chunk == end_chunk) { | |
| 504 // All in one chunk. | |
| 505 _chunk_data[beg_chunk].add_live_obj(len); | |
| 506 return; | |
| 507 } | |
| 508 | |
| 509 // First chunk. | |
| 510 const size_t beg_ofs = chunk_offset(addr); | |
| 511 _chunk_data[beg_chunk].add_live_obj(ChunkSize - beg_ofs); | |
| 512 | |
| 513 klassOop klass = ((oop)addr)->klass(); | |
| 514 // Middle chunks--completely spanned by this object. | |
| 515 for (size_t chunk = beg_chunk + 1; chunk < end_chunk; ++chunk) { | |
| 516 _chunk_data[chunk].set_partial_obj_size(ChunkSize); | |
| 517 _chunk_data[chunk].set_partial_obj_addr(addr); | |
| 518 } | |
| 519 | |
| 520 // Last chunk. | |
| 521 const size_t end_ofs = chunk_offset(addr + len - 1); | |
| 522 _chunk_data[end_chunk].set_partial_obj_size(end_ofs + 1); | |
| 523 _chunk_data[end_chunk].set_partial_obj_addr(addr); | |
| 524 } | |
| 525 | |
| 526 void | |
| 527 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) | |
| 528 { | |
| 529 assert(chunk_offset(beg) == 0, "not ChunkSize aligned"); | |
| 530 assert(chunk_offset(end) == 0, "not ChunkSize aligned"); | |
| 531 | |
| 532 size_t cur_chunk = addr_to_chunk_idx(beg); | |
| 533 const size_t end_chunk = addr_to_chunk_idx(end); | |
| 534 HeapWord* addr = beg; | |
| 535 while (cur_chunk < end_chunk) { | |
| 536 _chunk_data[cur_chunk].set_destination(addr); | |
| 537 _chunk_data[cur_chunk].set_destination_count(0); | |
| 538 _chunk_data[cur_chunk].set_source_chunk(cur_chunk); | |
| 539 _chunk_data[cur_chunk].set_data_location(addr); | |
| 540 | |
| 541 // Update live_obj_size so the chunk appears completely full. | |
| 542 size_t live_size = ChunkSize - _chunk_data[cur_chunk].partial_obj_size(); | |
| 543 _chunk_data[cur_chunk].set_live_obj_size(live_size); | |
| 544 | |
| 545 ++cur_chunk; | |
| 546 addr += ChunkSize; | |
| 547 } | |
| 548 } | |
| 549 | |
| 550 bool ParallelCompactData::summarize(HeapWord* target_beg, HeapWord* target_end, | |
| 551 HeapWord* source_beg, HeapWord* source_end, | |
| 552 HeapWord** target_next, | |
| 553 HeapWord** source_next) { | |
| 554 // This is too strict. | |
| 555 // assert(chunk_offset(source_beg) == 0, "not ChunkSize aligned"); | |
| 556 | |
| 557 if (TraceParallelOldGCSummaryPhase) { | |
| 558 tty->print_cr("tb=" PTR_FORMAT " te=" PTR_FORMAT " " | |
| 559 "sb=" PTR_FORMAT " se=" PTR_FORMAT " " | |
| 560 "tn=" PTR_FORMAT " sn=" PTR_FORMAT, | |
| 561 target_beg, target_end, | |
| 562 source_beg, source_end, | |
| 563 target_next != 0 ? *target_next : (HeapWord*) 0, | |
| 564 source_next != 0 ? *source_next : (HeapWord*) 0); | |
| 565 } | |
| 566 | |
| 567 size_t cur_chunk = addr_to_chunk_idx(source_beg); | |
| 568 const size_t end_chunk = addr_to_chunk_idx(chunk_align_up(source_end)); | |
| 569 | |
| 570 HeapWord *dest_addr = target_beg; | |
| 571 while (cur_chunk < end_chunk) { | |
| 572 size_t words = _chunk_data[cur_chunk].data_size(); | |
| 573 | |
| 574 #if 1 | |
| 575 assert(pointer_delta(target_end, dest_addr) >= words, | |
| 576 "source region does not fit into target region"); | |
| 577 #else | |
| 578 // XXX - need some work on the corner cases here. If the chunk does not | |
| 579 // fit, then must either make sure any partial_obj from the chunk fits, or | |
| 580 // 'undo' the initial part of the partial_obj that is in the previous chunk. | |
| 581 if (dest_addr + words >= target_end) { | |
| 582 // Let the caller know where to continue. | |
| 583 *target_next = dest_addr; | |
| 584 *source_next = chunk_to_addr(cur_chunk); | |
| 585 return false; | |
| 586 } | |
| 587 #endif // #if 1 | |
| 588 | |
| 589 _chunk_data[cur_chunk].set_destination(dest_addr); | |
| 590 | |
| 591 // Set the destination_count for cur_chunk, and if necessary, update | |
| 592 // source_chunk for a destination chunk. The source_chunk field is updated | |
| 593 // if cur_chunk is the first (left-most) chunk to be copied to a destination | |
| 594 // chunk. | |
| 595 // | |
| 596 // The destination_count calculation is a bit subtle. A chunk that has data | |
| 597 // that compacts into itself does not count itself as a destination. This | |
| 598 // maintains the invariant that a zero count means the chunk is available | |
| 599 // and can be claimed and then filled. | |
| 600 if (words > 0) { | |
| 601 HeapWord* const last_addr = dest_addr + words - 1; | |
| 602 const size_t dest_chunk_1 = addr_to_chunk_idx(dest_addr); | |
| 603 const size_t dest_chunk_2 = addr_to_chunk_idx(last_addr); | |
| 604 #if 0 | |
| 605 // Initially assume that the destination chunks will be the same and | |
| 606 // adjust the value below if necessary. Under this assumption, if | |
| 607 // cur_chunk == dest_chunk_2, then cur_chunk will be compacted completely | |
| 608 // into itself. | |
| 609 uint destination_count = cur_chunk == dest_chunk_2 ? 0 : 1; | |
| 610 if (dest_chunk_1 != dest_chunk_2) { | |
| 611 // Destination chunks differ; adjust destination_count. | |
| 612 destination_count += 1; | |
| 613 // Data from cur_chunk will be copied to the start of dest_chunk_2. | |
| 614 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk); | |
| 615 } else if (chunk_offset(dest_addr) == 0) { | |
| 616 // Data from cur_chunk will be copied to the start of the destination | |
| 617 // chunk. | |
| 618 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk); | |
| 619 } | |
| 620 #else | |
| 621 // Initially assume that the destination chunks will be different and | |
| 622 // adjust the value below if necessary. Under this assumption, if | |
| 623 // cur_chunk == dest_chunk2, then cur_chunk will be compacted partially | |
| 624 // into dest_chunk_1 and partially into itself. | |
| 625 uint destination_count = cur_chunk == dest_chunk_2 ? 1 : 2; | |
| 626 if (dest_chunk_1 != dest_chunk_2) { | |
| 627 // Data from cur_chunk will be copied to the start of dest_chunk_2. | |
| 628 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk); | |
| 629 } else { | |
| 630 // Destination chunks are the same; adjust destination_count. | |
| 631 destination_count -= 1; | |
| 632 if (chunk_offset(dest_addr) == 0) { | |
| 633 // Data from cur_chunk will be copied to the start of the destination | |
| 634 // chunk. | |
| 635 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk); | |
| 636 } | |
| 637 } | |
| 638 #endif // #if 0 | |
| 639 | |
| 640 _chunk_data[cur_chunk].set_destination_count(destination_count); | |
| 641 _chunk_data[cur_chunk].set_data_location(chunk_to_addr(cur_chunk)); | |
| 642 dest_addr += words; | |
| 643 } | |
| 644 | |
| 645 ++cur_chunk; | |
| 646 } | |
| 647 | |
| 648 *target_next = dest_addr; | |
| 649 return true; | |
| 650 } | |
| 651 | |
| 652 bool ParallelCompactData::partial_obj_ends_in_block(size_t block_index) { | |
| 653 HeapWord* block_addr = block_to_addr(block_index); | |
| 654 HeapWord* block_end_addr = block_addr + BlockSize; | |
| 655 size_t chunk_index = addr_to_chunk_idx(block_addr); | |
| 656 HeapWord* partial_obj_end_addr = partial_obj_end(chunk_index); | |
| 657 | |
| 658 // An object that ends at the end of the block, ends | |
| 659 // in the block (the last word of the object is to | |
| 660 // the left of the end). | |
| 661 if ((block_addr < partial_obj_end_addr) && | |
| 662 (partial_obj_end_addr <= block_end_addr)) { | |
| 663 return true; | |
| 664 } | |
| 665 | |
| 666 return false; | |
| 667 } | |
| 668 | |
| 669 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { | |
| 670 HeapWord* result = NULL; | |
| 671 if (UseParallelOldGCChunkPointerCalc) { | |
| 672 result = chunk_calc_new_pointer(addr); | |
| 673 } else { | |
| 674 result = block_calc_new_pointer(addr); | |
| 675 } | |
| 676 return result; | |
| 677 } | |
| 678 | |
| 679 // This method is overly complicated (expensive) to be called | |
| 680 // for every reference. | |
| 681 // Try to restructure this so that a NULL is returned if | |
| 682 // the object is dead. But don't wast the cycles to explicitly check | |
| 683 // that it is dead since only live objects should be passed in. | |
| 684 | |
| 685 HeapWord* ParallelCompactData::chunk_calc_new_pointer(HeapWord* addr) { | |
| 686 assert(addr != NULL, "Should detect NULL oop earlier"); | |
| 687 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); | |
| 688 #ifdef ASSERT | |
| 689 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { | |
| 690 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); | |
| 691 } | |
| 692 #endif | |
| 693 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); | |
| 694 | |
| 695 // Chunk covering the object. | |
| 696 size_t chunk_index = addr_to_chunk_idx(addr); | |
| 697 const ChunkData* const chunk_ptr = chunk(chunk_index); | |
| 698 HeapWord* const chunk_addr = chunk_align_down(addr); | |
| 699 | |
| 700 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object"); | |
| 701 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check"); | |
| 702 | |
| 703 HeapWord* result = chunk_ptr->destination(); | |
| 704 | |
| 705 // If all the data in the chunk is live, then the new location of the object | |
| 706 // can be calculated from the destination of the chunk plus the offset of the | |
| 707 // object in the chunk. | |
| 708 if (chunk_ptr->data_size() == ChunkSize) { | |
| 709 result += pointer_delta(addr, chunk_addr); | |
| 710 return result; | |
| 711 } | |
| 712 | |
| 713 // The new location of the object is | |
| 714 // chunk destination + | |
| 715 // size of the partial object extending onto the chunk + | |
| 716 // sizes of the live objects in the Chunk that are to the left of addr | |
| 717 const size_t partial_obj_size = chunk_ptr->partial_obj_size(); | |
| 718 HeapWord* const search_start = chunk_addr + partial_obj_size; | |
| 719 | |
| 720 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); | |
| 721 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); | |
| 722 | |
| 723 result += partial_obj_size + live_to_left; | |
| 724 assert(result <= addr, "object cannot move to the right"); | |
| 725 return result; | |
| 726 } | |
| 727 | |
| 728 HeapWord* ParallelCompactData::block_calc_new_pointer(HeapWord* addr) { | |
| 729 assert(addr != NULL, "Should detect NULL oop earlier"); | |
| 730 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); | |
| 731 #ifdef ASSERT | |
| 732 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { | |
| 733 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); | |
| 734 } | |
| 735 #endif | |
| 736 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); | |
| 737 | |
| 738 // Chunk covering the object. | |
| 739 size_t chunk_index = addr_to_chunk_idx(addr); | |
| 740 const ChunkData* const chunk_ptr = chunk(chunk_index); | |
| 741 HeapWord* const chunk_addr = chunk_align_down(addr); | |
| 742 | |
| 743 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object"); | |
| 744 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check"); | |
| 745 | |
| 746 HeapWord* result = chunk_ptr->destination(); | |
| 747 | |
| 748 // If all the data in the chunk is live, then the new location of the object | |
| 749 // can be calculated from the destination of the chunk plus the offset of the | |
| 750 // object in the chunk. | |
| 751 if (chunk_ptr->data_size() == ChunkSize) { | |
| 752 result += pointer_delta(addr, chunk_addr); | |
| 753 return result; | |
| 754 } | |
| 755 | |
| 756 // The new location of the object is | |
| 757 // chunk destination + | |
| 758 // block offset + | |
| 759 // sizes of the live objects in the Block that are to the left of addr | |
| 760 const size_t block_offset = addr_to_block_ptr(addr)->offset(); | |
| 761 HeapWord* const search_start = chunk_addr + block_offset; | |
| 762 | |
| 763 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); | |
| 764 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); | |
| 765 | |
| 766 result += block_offset + live_to_left; | |
| 767 assert(result <= addr, "object cannot move to the right"); | |
| 768 assert(result == chunk_calc_new_pointer(addr), "Should match"); | |
| 769 return result; | |
| 770 } | |
| 771 | |
| 772 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) { | |
| 773 klassOop updated_klass; | |
| 774 if (PSParallelCompact::should_update_klass(old_klass)) { | |
| 775 updated_klass = (klassOop) calc_new_pointer(old_klass); | |
| 776 } else { | |
| 777 updated_klass = old_klass; | |
| 778 } | |
| 779 | |
| 780 return updated_klass; | |
| 781 } | |
| 782 | |
| 783 #ifdef ASSERT | |
| 784 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) | |
| 785 { | |
| 786 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); | |
| 787 const size_t* const end = (const size_t*)vspace->committed_high_addr(); | |
| 788 for (const size_t* p = beg; p < end; ++p) { | |
| 789 assert(*p == 0, "not zero"); | |
| 790 } | |
| 791 } | |
| 792 | |
| 793 void ParallelCompactData::verify_clear() | |
| 794 { | |
| 795 verify_clear(_chunk_vspace); | |
| 796 verify_clear(_block_vspace); | |
| 797 } | |
| 798 #endif // #ifdef ASSERT | |
| 799 | |
| 800 #ifdef NOT_PRODUCT | |
| 801 ParallelCompactData::ChunkData* debug_chunk(size_t chunk_index) { | |
| 802 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 803 return sd.chunk(chunk_index); | |
| 804 } | |
| 805 #endif | |
| 806 | |
| 807 elapsedTimer PSParallelCompact::_accumulated_time; | |
| 808 unsigned int PSParallelCompact::_total_invocations = 0; | |
| 809 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; | |
| 810 jlong PSParallelCompact::_time_of_last_gc = 0; | |
| 811 CollectorCounters* PSParallelCompact::_counters = NULL; | |
| 812 ParMarkBitMap PSParallelCompact::_mark_bitmap; | |
| 813 ParallelCompactData PSParallelCompact::_summary_data; | |
| 814 | |
| 815 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
816 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
817 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
818 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
819 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
820 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
821 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
822 |
| 0 | 823 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true); |
| 824 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false); | |
| 825 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
826 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
827 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
828 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
829 void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
830 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
831 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
832 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } |
| 0 | 833 |
| 834 void PSParallelCompact::post_initialize() { | |
| 835 ParallelScavengeHeap* heap = gc_heap(); | |
| 836 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); | |
| 837 | |
| 838 MemRegion mr = heap->reserved_region(); | |
| 839 _ref_processor = ReferenceProcessor::create_ref_processor( | |
| 840 mr, // span | |
| 841 true, // atomic_discovery | |
| 842 true, // mt_discovery | |
| 843 &_is_alive_closure, | |
| 844 ParallelGCThreads, | |
| 845 ParallelRefProcEnabled); | |
| 846 _counters = new CollectorCounters("PSParallelCompact", 1); | |
| 847 | |
| 848 // Initialize static fields in ParCompactionManager. | |
| 849 ParCompactionManager::initialize(mark_bitmap()); | |
| 850 } | |
| 851 | |
| 852 bool PSParallelCompact::initialize() { | |
| 853 ParallelScavengeHeap* heap = gc_heap(); | |
| 854 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); | |
| 855 MemRegion mr = heap->reserved_region(); | |
| 856 | |
| 857 // Was the old gen get allocated successfully? | |
| 858 if (!heap->old_gen()->is_allocated()) { | |
| 859 return false; | |
| 860 } | |
| 861 | |
| 862 initialize_space_info(); | |
| 863 initialize_dead_wood_limiter(); | |
| 864 | |
| 865 if (!_mark_bitmap.initialize(mr)) { | |
| 866 vm_shutdown_during_initialization("Unable to allocate bit map for " | |
| 867 "parallel garbage collection for the requested heap size."); | |
| 868 return false; | |
| 869 } | |
| 870 | |
| 871 if (!_summary_data.initialize(mr)) { | |
| 872 vm_shutdown_during_initialization("Unable to allocate tables for " | |
| 873 "parallel garbage collection for the requested heap size."); | |
| 874 return false; | |
| 875 } | |
| 876 | |
| 877 return true; | |
| 878 } | |
| 879 | |
| 880 void PSParallelCompact::initialize_space_info() | |
| 881 { | |
| 882 memset(&_space_info, 0, sizeof(_space_info)); | |
| 883 | |
| 884 ParallelScavengeHeap* heap = gc_heap(); | |
| 885 PSYoungGen* young_gen = heap->young_gen(); | |
| 886 MutableSpace* perm_space = heap->perm_gen()->object_space(); | |
| 887 | |
| 888 _space_info[perm_space_id].set_space(perm_space); | |
| 889 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); | |
| 890 _space_info[eden_space_id].set_space(young_gen->eden_space()); | |
| 891 _space_info[from_space_id].set_space(young_gen->from_space()); | |
| 892 _space_info[to_space_id].set_space(young_gen->to_space()); | |
| 893 | |
| 894 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array()); | |
| 895 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); | |
| 896 | |
| 897 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top()); | |
| 898 if (TraceParallelOldGCDensePrefix) { | |
| 899 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT, | |
| 900 _space_info[perm_space_id].min_dense_prefix()); | |
| 901 } | |
| 902 } | |
| 903 | |
| 904 void PSParallelCompact::initialize_dead_wood_limiter() | |
| 905 { | |
| 906 const size_t max = 100; | |
| 907 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; | |
| 908 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; | |
| 909 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); | |
| 910 DEBUG_ONLY(_dwl_initialized = true;) | |
| 911 _dwl_adjustment = normal_distribution(1.0); | |
| 912 } | |
| 913 | |
| 914 // Simple class for storing info about the heap at the start of GC, to be used | |
| 915 // after GC for comparison/printing. | |
| 916 class PreGCValues { | |
| 917 public: | |
| 918 PreGCValues() { } | |
| 919 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } | |
| 920 | |
| 921 void fill(ParallelScavengeHeap* heap) { | |
| 922 _heap_used = heap->used(); | |
| 923 _young_gen_used = heap->young_gen()->used_in_bytes(); | |
| 924 _old_gen_used = heap->old_gen()->used_in_bytes(); | |
| 925 _perm_gen_used = heap->perm_gen()->used_in_bytes(); | |
| 926 }; | |
| 927 | |
| 928 size_t heap_used() const { return _heap_used; } | |
| 929 size_t young_gen_used() const { return _young_gen_used; } | |
| 930 size_t old_gen_used() const { return _old_gen_used; } | |
| 931 size_t perm_gen_used() const { return _perm_gen_used; } | |
| 932 | |
| 933 private: | |
| 934 size_t _heap_used; | |
| 935 size_t _young_gen_used; | |
| 936 size_t _old_gen_used; | |
| 937 size_t _perm_gen_used; | |
| 938 }; | |
| 939 | |
| 940 void | |
| 941 PSParallelCompact::clear_data_covering_space(SpaceId id) | |
| 942 { | |
| 943 // At this point, top is the value before GC, new_top() is the value that will | |
| 944 // be set at the end of GC. The marking bitmap is cleared to top; nothing | |
| 945 // should be marked above top. The summary data is cleared to the larger of | |
| 946 // top & new_top. | |
| 947 MutableSpace* const space = _space_info[id].space(); | |
| 948 HeapWord* const bot = space->bottom(); | |
| 949 HeapWord* const top = space->top(); | |
| 950 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); | |
| 951 | |
| 952 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); | |
| 953 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); | |
| 954 _mark_bitmap.clear_range(beg_bit, end_bit); | |
| 955 | |
| 956 const size_t beg_chunk = _summary_data.addr_to_chunk_idx(bot); | |
| 957 const size_t end_chunk = | |
| 958 _summary_data.addr_to_chunk_idx(_summary_data.chunk_align_up(max_top)); | |
| 959 _summary_data.clear_range(beg_chunk, end_chunk); | |
| 960 } | |
| 961 | |
| 962 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) | |
| 963 { | |
| 964 // Update the from & to space pointers in space_info, since they are swapped | |
| 965 // at each young gen gc. Do the update unconditionally (even though a | |
| 966 // promotion failure does not swap spaces) because an unknown number of minor | |
| 967 // collections will have swapped the spaces an unknown number of times. | |
| 968 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); | |
| 969 ParallelScavengeHeap* heap = gc_heap(); | |
| 970 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); | |
| 971 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); | |
| 972 | |
| 973 pre_gc_values->fill(heap); | |
| 974 | |
| 975 ParCompactionManager::reset(); | |
| 976 NOT_PRODUCT(_mark_bitmap.reset_counters()); | |
| 977 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) | |
| 978 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) | |
| 979 | |
| 980 // Increment the invocation count | |
|
139
c0492d52d55b
6539517: CR 6186200 should be extended to perm gen allocation to prevent spurious OOM's from perm gen
apetrusenko
parents:
0
diff
changeset
|
981 heap->increment_total_collections(true); |
| 0 | 982 |
| 983 // We need to track unique mark sweep invocations as well. | |
| 984 _total_invocations++; | |
| 985 | |
| 986 if (PrintHeapAtGC) { | |
| 987 Universe::print_heap_before_gc(); | |
| 988 } | |
| 989 | |
| 990 // Fill in TLABs | |
| 991 heap->accumulate_statistics_all_tlabs(); | |
| 992 heap->ensure_parsability(true); // retire TLABs | |
| 993 | |
| 994 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { | |
| 995 HandleMark hm; // Discard invalid handles created during verification | |
| 996 gclog_or_tty->print(" VerifyBeforeGC:"); | |
| 997 Universe::verify(true); | |
| 998 } | |
| 999 | |
| 1000 // Verify object start arrays | |
| 1001 if (VerifyObjectStartArray && | |
| 1002 VerifyBeforeGC) { | |
| 1003 heap->old_gen()->verify_object_start_array(); | |
| 1004 heap->perm_gen()->verify_object_start_array(); | |
| 1005 } | |
| 1006 | |
| 1007 DEBUG_ONLY(mark_bitmap()->verify_clear();) | |
| 1008 DEBUG_ONLY(summary_data().verify_clear();) | |
| 210 | 1009 |
| 1010 // Have worker threads release resources the next time they run a task. | |
| 1011 gc_task_manager()->release_all_resources(); | |
| 0 | 1012 } |
| 1013 | |
| 1014 void PSParallelCompact::post_compact() | |
| 1015 { | |
| 1016 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); | |
| 1017 | |
| 1018 // Clear the marking bitmap and summary data and update top() in each space. | |
| 1019 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { | |
| 1020 clear_data_covering_space(SpaceId(id)); | |
| 1021 _space_info[id].space()->set_top(_space_info[id].new_top()); | |
| 1022 } | |
| 1023 | |
| 1024 MutableSpace* const eden_space = _space_info[eden_space_id].space(); | |
| 1025 MutableSpace* const from_space = _space_info[from_space_id].space(); | |
| 1026 MutableSpace* const to_space = _space_info[to_space_id].space(); | |
| 1027 | |
| 1028 ParallelScavengeHeap* heap = gc_heap(); | |
| 1029 bool eden_empty = eden_space->is_empty(); | |
| 1030 if (!eden_empty) { | |
| 1031 eden_empty = absorb_live_data_from_eden(heap->size_policy(), | |
| 1032 heap->young_gen(), heap->old_gen()); | |
| 1033 } | |
| 1034 | |
| 1035 // Update heap occupancy information which is used as input to the soft ref | |
| 1036 // clearing policy at the next gc. | |
| 1037 Universe::update_heap_info_at_gc(); | |
| 1038 | |
| 1039 bool young_gen_empty = eden_empty && from_space->is_empty() && | |
| 1040 to_space->is_empty(); | |
| 1041 | |
| 1042 BarrierSet* bs = heap->barrier_set(); | |
| 1043 if (bs->is_a(BarrierSet::ModRef)) { | |
| 1044 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; | |
| 1045 MemRegion old_mr = heap->old_gen()->reserved(); | |
| 1046 MemRegion perm_mr = heap->perm_gen()->reserved(); | |
| 1047 assert(perm_mr.end() <= old_mr.start(), "Generations out of order"); | |
| 1048 | |
| 1049 if (young_gen_empty) { | |
| 1050 modBS->clear(MemRegion(perm_mr.start(), old_mr.end())); | |
| 1051 } else { | |
| 1052 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end())); | |
| 1053 } | |
| 1054 } | |
| 1055 | |
| 1056 Threads::gc_epilogue(); | |
| 1057 CodeCache::gc_epilogue(); | |
| 1058 | |
| 1059 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); | |
| 1060 | |
| 1061 ref_processor()->enqueue_discovered_references(NULL); | |
| 1062 | |
| 1063 // Update time of last GC | |
| 1064 reset_millis_since_last_gc(); | |
| 1065 } | |
| 1066 | |
| 1067 HeapWord* | |
| 1068 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, | |
| 1069 bool maximum_compaction) | |
| 1070 { | |
| 1071 const size_t chunk_size = ParallelCompactData::ChunkSize; | |
| 1072 const ParallelCompactData& sd = summary_data(); | |
| 1073 | |
| 1074 const MutableSpace* const space = _space_info[id].space(); | |
| 1075 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top()); | |
| 1076 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(space->bottom()); | |
| 1077 const ChunkData* const end_cp = sd.addr_to_chunk_ptr(top_aligned_up); | |
| 1078 | |
| 1079 // Skip full chunks at the beginning of the space--they are necessarily part | |
| 1080 // of the dense prefix. | |
| 1081 size_t full_count = 0; | |
| 1082 const ChunkData* cp; | |
| 1083 for (cp = beg_cp; cp < end_cp && cp->data_size() == chunk_size; ++cp) { | |
| 1084 ++full_count; | |
| 1085 } | |
| 1086 | |
| 1087 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); | |
| 1088 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; | |
| 1089 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; | |
| 1090 if (maximum_compaction || cp == end_cp || interval_ended) { | |
| 1091 _maximum_compaction_gc_num = total_invocations(); | |
| 1092 return sd.chunk_to_addr(cp); | |
| 1093 } | |
| 1094 | |
| 1095 HeapWord* const new_top = _space_info[id].new_top(); | |
| 1096 const size_t space_live = pointer_delta(new_top, space->bottom()); | |
| 1097 const size_t space_used = space->used_in_words(); | |
| 1098 const size_t space_capacity = space->capacity_in_words(); | |
| 1099 | |
| 1100 const double cur_density = double(space_live) / space_capacity; | |
| 1101 const double deadwood_density = | |
| 1102 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; | |
| 1103 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); | |
| 1104 | |
| 1105 if (TraceParallelOldGCDensePrefix) { | |
| 1106 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, | |
| 1107 cur_density, deadwood_density, deadwood_goal); | |
| 1108 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " | |
| 1109 "space_cap=" SIZE_FORMAT, | |
| 1110 space_live, space_used, | |
| 1111 space_capacity); | |
| 1112 } | |
| 1113 | |
| 1114 // XXX - Use binary search? | |
| 1115 HeapWord* dense_prefix = sd.chunk_to_addr(cp); | |
| 1116 const ChunkData* full_cp = cp; | |
| 1117 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(space->top() - 1); | |
| 1118 while (cp < end_cp) { | |
| 1119 HeapWord* chunk_destination = cp->destination(); | |
| 1120 const size_t cur_deadwood = pointer_delta(dense_prefix, chunk_destination); | |
| 1121 if (TraceParallelOldGCDensePrefix && Verbose) { | |
| 1122 tty->print_cr("c#=" SIZE_FORMAT_W("04") " dst=" PTR_FORMAT " " | |
| 1123 "dp=" SIZE_FORMAT_W("08") " " "cdw=" SIZE_FORMAT_W("08"), | |
| 1124 sd.chunk(cp), chunk_destination, | |
| 1125 dense_prefix, cur_deadwood); | |
| 1126 } | |
| 1127 | |
| 1128 if (cur_deadwood >= deadwood_goal) { | |
| 1129 // Found the chunk that has the correct amount of deadwood to the left. | |
| 1130 // This typically occurs after crossing a fairly sparse set of chunks, so | |
| 1131 // iterate backwards over those sparse chunks, looking for the chunk that | |
| 1132 // has the lowest density of live objects 'to the right.' | |
| 1133 size_t space_to_left = sd.chunk(cp) * chunk_size; | |
| 1134 size_t live_to_left = space_to_left - cur_deadwood; | |
| 1135 size_t space_to_right = space_capacity - space_to_left; | |
| 1136 size_t live_to_right = space_live - live_to_left; | |
| 1137 double density_to_right = double(live_to_right) / space_to_right; | |
| 1138 while (cp > full_cp) { | |
| 1139 --cp; | |
| 1140 const size_t prev_chunk_live_to_right = live_to_right - cp->data_size(); | |
| 1141 const size_t prev_chunk_space_to_right = space_to_right + chunk_size; | |
| 1142 double prev_chunk_density_to_right = | |
| 1143 double(prev_chunk_live_to_right) / prev_chunk_space_to_right; | |
| 1144 if (density_to_right <= prev_chunk_density_to_right) { | |
| 1145 return dense_prefix; | |
| 1146 } | |
| 1147 if (TraceParallelOldGCDensePrefix && Verbose) { | |
| 1148 tty->print_cr("backing up from c=" SIZE_FORMAT_W("4") " d2r=%10.8f " | |
| 1149 "pc_d2r=%10.8f", sd.chunk(cp), density_to_right, | |
| 1150 prev_chunk_density_to_right); | |
| 1151 } | |
| 1152 dense_prefix -= chunk_size; | |
| 1153 live_to_right = prev_chunk_live_to_right; | |
| 1154 space_to_right = prev_chunk_space_to_right; | |
| 1155 density_to_right = prev_chunk_density_to_right; | |
| 1156 } | |
| 1157 return dense_prefix; | |
| 1158 } | |
| 1159 | |
| 1160 dense_prefix += chunk_size; | |
| 1161 ++cp; | |
| 1162 } | |
| 1163 | |
| 1164 return dense_prefix; | |
| 1165 } | |
| 1166 | |
| 1167 #ifndef PRODUCT | |
| 1168 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, | |
| 1169 const SpaceId id, | |
| 1170 const bool maximum_compaction, | |
| 1171 HeapWord* const addr) | |
| 1172 { | |
| 1173 const size_t chunk_idx = summary_data().addr_to_chunk_idx(addr); | |
| 1174 ChunkData* const cp = summary_data().chunk(chunk_idx); | |
| 1175 const MutableSpace* const space = _space_info[id].space(); | |
| 1176 HeapWord* const new_top = _space_info[id].new_top(); | |
| 1177 | |
| 1178 const size_t space_live = pointer_delta(new_top, space->bottom()); | |
| 1179 const size_t dead_to_left = pointer_delta(addr, cp->destination()); | |
| 1180 const size_t space_cap = space->capacity_in_words(); | |
| 1181 const double dead_to_left_pct = double(dead_to_left) / space_cap; | |
| 1182 const size_t live_to_right = new_top - cp->destination(); | |
| 1183 const size_t dead_to_right = space->top() - addr - live_to_right; | |
| 1184 | |
| 1185 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W("05") " " | |
| 1186 "spl=" SIZE_FORMAT " " | |
| 1187 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " | |
| 1188 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT | |
| 1189 " ratio=%10.8f", | |
| 1190 algorithm, addr, chunk_idx, | |
| 1191 space_live, | |
| 1192 dead_to_left, dead_to_left_pct, | |
| 1193 dead_to_right, live_to_right, | |
| 1194 double(dead_to_right) / live_to_right); | |
| 1195 } | |
| 1196 #endif // #ifndef PRODUCT | |
| 1197 | |
| 1198 // Return a fraction indicating how much of the generation can be treated as | |
| 1199 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution | |
| 1200 // based on the density of live objects in the generation to determine a limit, | |
| 1201 // which is then adjusted so the return value is min_percent when the density is | |
| 1202 // 1. | |
| 1203 // | |
| 1204 // The following table shows some return values for a different values of the | |
| 1205 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and | |
| 1206 // min_percent is 1. | |
| 1207 // | |
| 1208 // fraction allowed as dead wood | |
| 1209 // ----------------------------------------------------------------- | |
| 1210 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 | |
| 1211 // ------- ---------- ---------- ---------- ---------- ---------- ---------- | |
| 1212 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 | |
| 1213 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 | |
| 1214 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 | |
| 1215 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 | |
| 1216 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 | |
| 1217 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 | |
| 1218 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 | |
| 1219 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 | |
| 1220 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 | |
| 1221 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 | |
| 1222 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 | |
| 1223 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 | |
| 1224 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 | |
| 1225 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 | |
| 1226 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 | |
| 1227 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 | |
| 1228 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 | |
| 1229 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 | |
| 1230 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 | |
| 1231 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 | |
| 1232 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 | |
| 1233 | |
| 1234 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) | |
| 1235 { | |
| 1236 assert(_dwl_initialized, "uninitialized"); | |
| 1237 | |
| 1238 // The raw limit is the value of the normal distribution at x = density. | |
| 1239 const double raw_limit = normal_distribution(density); | |
| 1240 | |
| 1241 // Adjust the raw limit so it becomes the minimum when the density is 1. | |
| 1242 // | |
| 1243 // First subtract the adjustment value (which is simply the precomputed value | |
| 1244 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. | |
| 1245 // Then add the minimum value, so the minimum is returned when the density is | |
| 1246 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. | |
| 1247 const double min = double(min_percent) / 100.0; | |
| 1248 const double limit = raw_limit - _dwl_adjustment + min; | |
| 1249 return MAX2(limit, 0.0); | |
| 1250 } | |
| 1251 | |
| 1252 ParallelCompactData::ChunkData* | |
| 1253 PSParallelCompact::first_dead_space_chunk(const ChunkData* beg, | |
| 1254 const ChunkData* end) | |
| 1255 { | |
| 1256 const size_t chunk_size = ParallelCompactData::ChunkSize; | |
| 1257 ParallelCompactData& sd = summary_data(); | |
| 1258 size_t left = sd.chunk(beg); | |
| 1259 size_t right = end > beg ? sd.chunk(end) - 1 : left; | |
| 1260 | |
| 1261 // Binary search. | |
| 1262 while (left < right) { | |
| 1263 // Equivalent to (left + right) / 2, but does not overflow. | |
| 1264 const size_t middle = left + (right - left) / 2; | |
| 1265 ChunkData* const middle_ptr = sd.chunk(middle); | |
| 1266 HeapWord* const dest = middle_ptr->destination(); | |
| 1267 HeapWord* const addr = sd.chunk_to_addr(middle); | |
| 1268 assert(dest != NULL, "sanity"); | |
| 1269 assert(dest <= addr, "must move left"); | |
| 1270 | |
| 1271 if (middle > left && dest < addr) { | |
| 1272 right = middle - 1; | |
| 1273 } else if (middle < right && middle_ptr->data_size() == chunk_size) { | |
| 1274 left = middle + 1; | |
| 1275 } else { | |
| 1276 return middle_ptr; | |
| 1277 } | |
| 1278 } | |
| 1279 return sd.chunk(left); | |
| 1280 } | |
| 1281 | |
| 1282 ParallelCompactData::ChunkData* | |
| 1283 PSParallelCompact::dead_wood_limit_chunk(const ChunkData* beg, | |
| 1284 const ChunkData* end, | |
| 1285 size_t dead_words) | |
| 1286 { | |
| 1287 ParallelCompactData& sd = summary_data(); | |
| 1288 size_t left = sd.chunk(beg); | |
| 1289 size_t right = end > beg ? sd.chunk(end) - 1 : left; | |
| 1290 | |
| 1291 // Binary search. | |
| 1292 while (left < right) { | |
| 1293 // Equivalent to (left + right) / 2, but does not overflow. | |
| 1294 const size_t middle = left + (right - left) / 2; | |
| 1295 ChunkData* const middle_ptr = sd.chunk(middle); | |
| 1296 HeapWord* const dest = middle_ptr->destination(); | |
| 1297 HeapWord* const addr = sd.chunk_to_addr(middle); | |
| 1298 assert(dest != NULL, "sanity"); | |
| 1299 assert(dest <= addr, "must move left"); | |
| 1300 | |
| 1301 const size_t dead_to_left = pointer_delta(addr, dest); | |
| 1302 if (middle > left && dead_to_left > dead_words) { | |
| 1303 right = middle - 1; | |
| 1304 } else if (middle < right && dead_to_left < dead_words) { | |
| 1305 left = middle + 1; | |
| 1306 } else { | |
| 1307 return middle_ptr; | |
| 1308 } | |
| 1309 } | |
| 1310 return sd.chunk(left); | |
| 1311 } | |
| 1312 | |
| 1313 // The result is valid during the summary phase, after the initial summarization | |
| 1314 // of each space into itself, and before final summarization. | |
| 1315 inline double | |
| 1316 PSParallelCompact::reclaimed_ratio(const ChunkData* const cp, | |
| 1317 HeapWord* const bottom, | |
| 1318 HeapWord* const top, | |
| 1319 HeapWord* const new_top) | |
| 1320 { | |
| 1321 ParallelCompactData& sd = summary_data(); | |
| 1322 | |
| 1323 assert(cp != NULL, "sanity"); | |
| 1324 assert(bottom != NULL, "sanity"); | |
| 1325 assert(top != NULL, "sanity"); | |
| 1326 assert(new_top != NULL, "sanity"); | |
| 1327 assert(top >= new_top, "summary data problem?"); | |
| 1328 assert(new_top > bottom, "space is empty; should not be here"); | |
| 1329 assert(new_top >= cp->destination(), "sanity"); | |
| 1330 assert(top >= sd.chunk_to_addr(cp), "sanity"); | |
| 1331 | |
| 1332 HeapWord* const destination = cp->destination(); | |
| 1333 const size_t dense_prefix_live = pointer_delta(destination, bottom); | |
| 1334 const size_t compacted_region_live = pointer_delta(new_top, destination); | |
| 1335 const size_t compacted_region_used = pointer_delta(top, sd.chunk_to_addr(cp)); | |
| 1336 const size_t reclaimable = compacted_region_used - compacted_region_live; | |
| 1337 | |
| 1338 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; | |
| 1339 return double(reclaimable) / divisor; | |
| 1340 } | |
| 1341 | |
| 1342 // Return the address of the end of the dense prefix, a.k.a. the start of the | |
| 1343 // compacted region. The address is always on a chunk boundary. | |
| 1344 // | |
| 1345 // Completely full chunks at the left are skipped, since no compaction can occur | |
| 1346 // in those chunks. Then the maximum amount of dead wood to allow is computed, | |
| 1347 // based on the density (amount live / capacity) of the generation; the chunk | |
| 1348 // with approximately that amount of dead space to the left is identified as the | |
| 1349 // limit chunk. Chunks between the last completely full chunk and the limit | |
| 1350 // chunk are scanned and the one that has the best (maximum) reclaimed_ratio() | |
| 1351 // is selected. | |
| 1352 HeapWord* | |
| 1353 PSParallelCompact::compute_dense_prefix(const SpaceId id, | |
| 1354 bool maximum_compaction) | |
| 1355 { | |
| 1356 const size_t chunk_size = ParallelCompactData::ChunkSize; | |
| 1357 const ParallelCompactData& sd = summary_data(); | |
| 1358 | |
| 1359 const MutableSpace* const space = _space_info[id].space(); | |
| 1360 HeapWord* const top = space->top(); | |
| 1361 HeapWord* const top_aligned_up = sd.chunk_align_up(top); | |
| 1362 HeapWord* const new_top = _space_info[id].new_top(); | |
| 1363 HeapWord* const new_top_aligned_up = sd.chunk_align_up(new_top); | |
| 1364 HeapWord* const bottom = space->bottom(); | |
| 1365 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(bottom); | |
| 1366 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up); | |
| 1367 const ChunkData* const new_top_cp = sd.addr_to_chunk_ptr(new_top_aligned_up); | |
| 1368 | |
| 1369 // Skip full chunks at the beginning of the space--they are necessarily part | |
| 1370 // of the dense prefix. | |
| 1371 const ChunkData* const full_cp = first_dead_space_chunk(beg_cp, new_top_cp); | |
| 1372 assert(full_cp->destination() == sd.chunk_to_addr(full_cp) || | |
| 1373 space->is_empty(), "no dead space allowed to the left"); | |
| 1374 assert(full_cp->data_size() < chunk_size || full_cp == new_top_cp - 1, | |
| 1375 "chunk must have dead space"); | |
| 1376 | |
| 1377 // The gc number is saved whenever a maximum compaction is done, and used to | |
| 1378 // determine when the maximum compaction interval has expired. This avoids | |
| 1379 // successive max compactions for different reasons. | |
| 1380 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); | |
| 1381 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; | |
| 1382 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || | |
| 1383 total_invocations() == HeapFirstMaximumCompactionCount; | |
| 1384 if (maximum_compaction || full_cp == top_cp || interval_ended) { | |
| 1385 _maximum_compaction_gc_num = total_invocations(); | |
| 1386 return sd.chunk_to_addr(full_cp); | |
| 1387 } | |
| 1388 | |
| 1389 const size_t space_live = pointer_delta(new_top, bottom); | |
| 1390 const size_t space_used = space->used_in_words(); | |
| 1391 const size_t space_capacity = space->capacity_in_words(); | |
| 1392 | |
| 1393 const double density = double(space_live) / double(space_capacity); | |
| 1394 const size_t min_percent_free = | |
| 1395 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio; | |
| 1396 const double limiter = dead_wood_limiter(density, min_percent_free); | |
| 1397 const size_t dead_wood_max = space_used - space_live; | |
| 1398 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), | |
| 1399 dead_wood_max); | |
| 1400 | |
| 1401 if (TraceParallelOldGCDensePrefix) { | |
| 1402 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " | |
| 1403 "space_cap=" SIZE_FORMAT, | |
| 1404 space_live, space_used, | |
| 1405 space_capacity); | |
| 1406 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " | |
| 1407 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, | |
| 1408 density, min_percent_free, limiter, | |
| 1409 dead_wood_max, dead_wood_limit); | |
| 1410 } | |
| 1411 | |
| 1412 // Locate the chunk with the desired amount of dead space to the left. | |
| 1413 const ChunkData* const limit_cp = | |
| 1414 dead_wood_limit_chunk(full_cp, top_cp, dead_wood_limit); | |
| 1415 | |
| 1416 // Scan from the first chunk with dead space to the limit chunk and find the | |
| 1417 // one with the best (largest) reclaimed ratio. | |
| 1418 double best_ratio = 0.0; | |
| 1419 const ChunkData* best_cp = full_cp; | |
| 1420 for (const ChunkData* cp = full_cp; cp < limit_cp; ++cp) { | |
| 1421 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); | |
| 1422 if (tmp_ratio > best_ratio) { | |
| 1423 best_cp = cp; | |
| 1424 best_ratio = tmp_ratio; | |
| 1425 } | |
| 1426 } | |
| 1427 | |
| 1428 #if 0 | |
| 1429 // Something to consider: if the chunk with the best ratio is 'close to' the | |
| 1430 // first chunk w/free space, choose the first chunk with free space | |
| 1431 // ("first-free"). The first-free chunk is usually near the start of the | |
| 1432 // heap, which means we are copying most of the heap already, so copy a bit | |
| 1433 // more to get complete compaction. | |
| 1434 if (pointer_delta(best_cp, full_cp, sizeof(ChunkData)) < 4) { | |
| 1435 _maximum_compaction_gc_num = total_invocations(); | |
| 1436 best_cp = full_cp; | |
| 1437 } | |
| 1438 #endif // #if 0 | |
| 1439 | |
| 1440 return sd.chunk_to_addr(best_cp); | |
| 1441 } | |
| 1442 | |
| 1443 void PSParallelCompact::summarize_spaces_quick() | |
| 1444 { | |
| 1445 for (unsigned int i = 0; i < last_space_id; ++i) { | |
| 1446 const MutableSpace* space = _space_info[i].space(); | |
| 1447 bool result = _summary_data.summarize(space->bottom(), space->end(), | |
| 1448 space->bottom(), space->top(), | |
| 1449 _space_info[i].new_top_addr()); | |
| 1450 assert(result, "should never fail"); | |
| 1451 _space_info[i].set_dense_prefix(space->bottom()); | |
| 1452 } | |
| 1453 } | |
| 1454 | |
| 1455 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) | |
| 1456 { | |
| 1457 HeapWord* const dense_prefix_end = dense_prefix(id); | |
| 1458 const ChunkData* chunk = _summary_data.addr_to_chunk_ptr(dense_prefix_end); | |
| 1459 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); | |
| 1460 if (dead_space_crosses_boundary(chunk, dense_prefix_bit)) { | |
| 1461 // Only enough dead space is filled so that any remaining dead space to the | |
| 1462 // left is larger than the minimum filler object. (The remainder is filled | |
| 1463 // during the copy/update phase.) | |
| 1464 // | |
| 1465 // The size of the dead space to the right of the boundary is not a | |
| 1466 // concern, since compaction will be able to use whatever space is | |
| 1467 // available. | |
| 1468 // | |
| 1469 // Here '||' is the boundary, 'x' represents a don't care bit and a box | |
| 1470 // surrounds the space to be filled with an object. | |
| 1471 // | |
| 1472 // In the 32-bit VM, each bit represents two 32-bit words: | |
| 1473 // +---+ | |
| 1474 // a) beg_bits: ... x x x | 0 | || 0 x x ... | |
| 1475 // end_bits: ... x x x | 0 | || 0 x x ... | |
| 1476 // +---+ | |
| 1477 // | |
| 1478 // In the 64-bit VM, each bit represents one 64-bit word: | |
| 1479 // +------------+ | |
| 1480 // b) beg_bits: ... x x x | 0 || 0 | x x ... | |
| 1481 // end_bits: ... x x 1 | 0 || 0 | x x ... | |
| 1482 // +------------+ | |
| 1483 // +-------+ | |
| 1484 // c) beg_bits: ... x x | 0 0 | || 0 x x ... | |
| 1485 // end_bits: ... x 1 | 0 0 | || 0 x x ... | |
| 1486 // +-------+ | |
| 1487 // +-----------+ | |
| 1488 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... | |
| 1489 // end_bits: ... 1 | 0 0 0 | || 0 x x ... | |
| 1490 // +-----------+ | |
| 1491 // +-------+ | |
| 1492 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... | |
| 1493 // end_bits: ... 0 0 | 0 0 | || 0 x x ... | |
| 1494 // +-------+ | |
| 1495 | |
| 1496 // Initially assume case a, c or e will apply. | |
| 1497 size_t obj_len = (size_t)oopDesc::header_size(); | |
| 1498 HeapWord* obj_beg = dense_prefix_end - obj_len; | |
| 1499 | |
| 1500 #ifdef _LP64 | |
| 1501 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { | |
| 1502 // Case b above. | |
| 1503 obj_beg = dense_prefix_end - 1; | |
| 1504 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && | |
| 1505 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { | |
| 1506 // Case d above. | |
| 1507 obj_beg = dense_prefix_end - 3; | |
| 1508 obj_len = 3; | |
| 1509 } | |
| 1510 #endif // #ifdef _LP64 | |
| 1511 | |
| 1512 MemRegion region(obj_beg, obj_len); | |
| 1513 SharedHeap::fill_region_with_object(region); | |
| 1514 _mark_bitmap.mark_obj(obj_beg, obj_len); | |
| 1515 _summary_data.add_obj(obj_beg, obj_len); | |
| 1516 assert(start_array(id) != NULL, "sanity"); | |
| 1517 start_array(id)->allocate_block(obj_beg); | |
| 1518 } | |
| 1519 } | |
| 1520 | |
| 1521 void | |
| 1522 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) | |
| 1523 { | |
| 1524 assert(id < last_space_id, "id out of range"); | |
| 1525 | |
| 1526 const MutableSpace* space = _space_info[id].space(); | |
| 1527 HeapWord** new_top_addr = _space_info[id].new_top_addr(); | |
| 1528 | |
| 1529 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); | |
| 1530 _space_info[id].set_dense_prefix(dense_prefix_end); | |
| 1531 | |
| 1532 #ifndef PRODUCT | |
| 1533 if (TraceParallelOldGCDensePrefix) { | |
| 1534 print_dense_prefix_stats("ratio", id, maximum_compaction, dense_prefix_end); | |
| 1535 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); | |
| 1536 print_dense_prefix_stats("density", id, maximum_compaction, addr); | |
| 1537 } | |
| 1538 #endif // #ifndef PRODUCT | |
| 1539 | |
| 1540 // If dead space crosses the dense prefix boundary, it is (at least partially) | |
| 1541 // filled with a dummy object, marked live and added to the summary data. | |
| 1542 // This simplifies the copy/update phase and must be done before the final | |
| 1543 // locations of objects are determined, to prevent leaving a fragment of dead | |
| 1544 // space that is too small to fill with an object. | |
| 1545 if (!maximum_compaction && dense_prefix_end != space->bottom()) { | |
| 1546 fill_dense_prefix_end(id); | |
| 1547 } | |
| 1548 | |
| 1549 // Compute the destination of each Chunk, and thus each object. | |
| 1550 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); | |
| 1551 _summary_data.summarize(dense_prefix_end, space->end(), | |
| 1552 dense_prefix_end, space->top(), | |
| 1553 new_top_addr); | |
| 1554 | |
| 1555 if (TraceParallelOldGCSummaryPhase) { | |
| 1556 const size_t chunk_size = ParallelCompactData::ChunkSize; | |
| 1557 const size_t dp_chunk = _summary_data.addr_to_chunk_idx(dense_prefix_end); | |
| 1558 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); | |
| 1559 const HeapWord* nt_aligned_up = _summary_data.chunk_align_up(*new_top_addr); | |
| 1560 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); | |
| 1561 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " | |
| 1562 "dp_chunk=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " | |
| 1563 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, | |
| 1564 id, space->capacity_in_words(), dense_prefix_end, | |
| 1565 dp_chunk, dp_words / chunk_size, | |
| 1566 cr_words / chunk_size, *new_top_addr); | |
| 1567 } | |
| 1568 } | |
| 1569 | |
| 1570 void PSParallelCompact::summary_phase(ParCompactionManager* cm, | |
| 1571 bool maximum_compaction) | |
| 1572 { | |
| 1573 EventMark m("2 summarize"); | |
| 1574 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); | |
| 1575 // trace("2"); | |
| 1576 | |
| 1577 #ifdef ASSERT | |
| 1578 if (VerifyParallelOldWithMarkSweep && | |
| 1579 (PSParallelCompact::total_invocations() % | |
| 1580 VerifyParallelOldWithMarkSweepInterval) == 0) { | |
| 1581 verify_mark_bitmap(_mark_bitmap); | |
| 1582 } | |
| 1583 if (TraceParallelOldGCMarkingPhase) { | |
| 1584 tty->print_cr("add_obj_count=" SIZE_FORMAT " " | |
| 1585 "add_obj_bytes=" SIZE_FORMAT, | |
| 1586 add_obj_count, add_obj_size * HeapWordSize); | |
| 1587 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " | |
| 1588 "mark_bitmap_bytes=" SIZE_FORMAT, | |
| 1589 mark_bitmap_count, mark_bitmap_size * HeapWordSize); | |
| 1590 } | |
| 1591 #endif // #ifdef ASSERT | |
| 1592 | |
| 1593 // Quick summarization of each space into itself, to see how much is live. | |
| 1594 summarize_spaces_quick(); | |
| 1595 | |
| 1596 if (TraceParallelOldGCSummaryPhase) { | |
| 1597 tty->print_cr("summary_phase: after summarizing each space to self"); | |
| 1598 Universe::print(); | |
| 1599 NOT_PRODUCT(print_chunk_ranges()); | |
| 1600 if (Verbose) { | |
| 1601 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); | |
| 1602 } | |
| 1603 } | |
| 1604 | |
| 1605 // The amount of live data that will end up in old space (assuming it fits). | |
| 1606 size_t old_space_total_live = 0; | |
| 1607 unsigned int id; | |
| 1608 for (id = old_space_id; id < last_space_id; ++id) { | |
| 1609 old_space_total_live += pointer_delta(_space_info[id].new_top(), | |
| 1610 _space_info[id].space()->bottom()); | |
| 1611 } | |
| 1612 | |
| 1613 const MutableSpace* old_space = _space_info[old_space_id].space(); | |
| 1614 if (old_space_total_live > old_space->capacity_in_words()) { | |
| 1615 // XXX - should also try to expand | |
| 1616 maximum_compaction = true; | |
| 1617 } else if (!UseParallelOldGCDensePrefix) { | |
| 1618 maximum_compaction = true; | |
| 1619 } | |
| 1620 | |
| 1621 // Permanent and Old generations. | |
| 1622 summarize_space(perm_space_id, maximum_compaction); | |
| 1623 summarize_space(old_space_id, maximum_compaction); | |
| 1624 | |
| 1625 // Summarize the remaining spaces (those in the young gen) into old space. If | |
| 1626 // the live data from a space doesn't fit, the existing summarization is left | |
| 1627 // intact, so the data is compacted down within the space itself. | |
| 1628 HeapWord** new_top_addr = _space_info[old_space_id].new_top_addr(); | |
| 1629 HeapWord* const target_space_end = old_space->end(); | |
| 1630 for (id = eden_space_id; id < last_space_id; ++id) { | |
| 1631 const MutableSpace* space = _space_info[id].space(); | |
| 1632 const size_t live = pointer_delta(_space_info[id].new_top(), | |
| 1633 space->bottom()); | |
| 1634 const size_t available = pointer_delta(target_space_end, *new_top_addr); | |
| 1635 if (live <= available) { | |
| 1636 // All the live data will fit. | |
| 1637 if (TraceParallelOldGCSummaryPhase) { | |
| 1638 tty->print_cr("summarizing %d into old_space @ " PTR_FORMAT, | |
| 1639 id, *new_top_addr); | |
| 1640 } | |
| 1641 _summary_data.summarize(*new_top_addr, target_space_end, | |
| 1642 space->bottom(), space->top(), | |
| 1643 new_top_addr); | |
| 1644 | |
| 1645 // Reset the new_top value for the space. | |
| 1646 _space_info[id].set_new_top(space->bottom()); | |
| 1647 | |
| 1648 // Clear the source_chunk field for each chunk in the space. | |
| 1649 ChunkData* beg_chunk = _summary_data.addr_to_chunk_ptr(space->bottom()); | |
| 1650 ChunkData* end_chunk = _summary_data.addr_to_chunk_ptr(space->top() - 1); | |
| 1651 while (beg_chunk <= end_chunk) { | |
| 1652 beg_chunk->set_source_chunk(0); | |
| 1653 ++beg_chunk; | |
| 1654 } | |
| 1655 } | |
| 1656 } | |
| 1657 | |
| 1658 // Fill in the block data after any changes to the chunks have | |
| 1659 // been made. | |
| 1660 #ifdef ASSERT | |
| 1661 summarize_blocks(cm, perm_space_id); | |
| 1662 summarize_blocks(cm, old_space_id); | |
| 1663 #else | |
| 1664 if (!UseParallelOldGCChunkPointerCalc) { | |
| 1665 summarize_blocks(cm, perm_space_id); | |
| 1666 summarize_blocks(cm, old_space_id); | |
| 1667 } | |
| 1668 #endif | |
| 1669 | |
| 1670 if (TraceParallelOldGCSummaryPhase) { | |
| 1671 tty->print_cr("summary_phase: after final summarization"); | |
| 1672 Universe::print(); | |
| 1673 NOT_PRODUCT(print_chunk_ranges()); | |
| 1674 if (Verbose) { | |
| 1675 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); | |
| 1676 } | |
| 1677 } | |
| 1678 } | |
| 1679 | |
| 1680 // Fill in the BlockData. | |
| 1681 // Iterate over the spaces and within each space iterate over | |
| 1682 // the chunks and fill in the BlockData for each chunk. | |
| 1683 | |
| 1684 void PSParallelCompact::summarize_blocks(ParCompactionManager* cm, | |
| 1685 SpaceId first_compaction_space_id) { | |
| 1686 #if 0 | |
| 1687 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(1);) | |
| 1688 for (SpaceId cur_space_id = first_compaction_space_id; | |
| 1689 cur_space_id != last_space_id; | |
| 1690 cur_space_id = next_compaction_space_id(cur_space_id)) { | |
| 1691 // Iterate over the chunks in the space | |
| 1692 size_t start_chunk_index = | |
| 1693 _summary_data.addr_to_chunk_idx(space(cur_space_id)->bottom()); | |
| 1694 BitBlockUpdateClosure bbu(mark_bitmap(), | |
| 1695 cm, | |
| 1696 start_chunk_index); | |
| 1697 // Iterate over blocks. | |
| 1698 for (size_t chunk_index = start_chunk_index; | |
| 1699 chunk_index < _summary_data.chunk_count() && | |
| 1700 _summary_data.chunk_to_addr(chunk_index) < space(cur_space_id)->top(); | |
| 1701 chunk_index++) { | |
| 1702 | |
| 1703 // Reset the closure for the new chunk. Note that the closure | |
| 1704 // maintains some data that does not get reset for each chunk | |
| 1705 // so a new instance of the closure is no appropriate. | |
| 1706 bbu.reset_chunk(chunk_index); | |
| 1707 | |
| 1708 // Start the iteration with the first live object. This | |
| 1709 // may return the end of the chunk. That is acceptable since | |
| 1710 // it will properly limit the iterations. | |
| 1711 ParMarkBitMap::idx_t left_offset = mark_bitmap()->addr_to_bit( | |
| 1712 _summary_data.first_live_or_end_in_chunk(chunk_index)); | |
| 1713 | |
| 1714 // End the iteration at the end of the chunk. | |
| 1715 HeapWord* chunk_addr = _summary_data.chunk_to_addr(chunk_index); | |
| 1716 HeapWord* chunk_end = chunk_addr + ParallelCompactData::ChunkSize; | |
| 1717 ParMarkBitMap::idx_t right_offset = | |
| 1718 mark_bitmap()->addr_to_bit(chunk_end); | |
| 1719 | |
| 1720 // Blocks that have not objects starting in them can be | |
| 1721 // skipped because their data will never be used. | |
| 1722 if (left_offset < right_offset) { | |
| 1723 | |
| 1724 // Iterate through the objects in the chunk. | |
| 1725 ParMarkBitMap::idx_t last_offset = | |
| 1726 mark_bitmap()->pair_iterate(&bbu, left_offset, right_offset); | |
| 1727 | |
| 1728 // If last_offset is less than right_offset, then the iterations | |
| 1729 // terminated while it was looking for an end bit. "last_offset" | |
| 1730 // is then the offset for the last start bit. In this situation | |
| 1731 // the "offset" field for the next block to the right (_cur_block + 1) | |
| 1732 // will not have been update although there may be live data | |
| 1733 // to the left of the chunk. | |
| 1734 | |
| 1735 size_t cur_block_plus_1 = bbu.cur_block() + 1; | |
| 1736 HeapWord* cur_block_plus_1_addr = | |
| 1737 _summary_data.block_to_addr(bbu.cur_block()) + | |
| 1738 ParallelCompactData::BlockSize; | |
| 1739 HeapWord* last_offset_addr = mark_bitmap()->bit_to_addr(last_offset); | |
| 1740 #if 1 // This code works. The else doesn't but should. Why does it? | |
| 1741 // The current block (cur_block()) has already been updated. | |
| 1742 // The last block that may need to be updated is either the | |
| 1743 // next block (current block + 1) or the block where the | |
| 1744 // last object starts (which can be greater than the | |
| 1745 // next block if there were no objects found in intervening | |
| 1746 // blocks). | |
| 1747 size_t last_block = | |
| 1748 MAX2(bbu.cur_block() + 1, | |
| 1749 _summary_data.addr_to_block_idx(last_offset_addr)); | |
| 1750 #else | |
| 1751 // The current block has already been updated. The only block | |
| 1752 // that remains to be updated is the block where the last | |
| 1753 // object in the chunk starts. | |
| 1754 size_t last_block = _summary_data.addr_to_block_idx(last_offset_addr); | |
| 1755 #endif | |
| 1756 assert_bit_is_start(last_offset); | |
| 1757 assert((last_block == _summary_data.block_count()) || | |
| 1758 (_summary_data.block(last_block)->raw_offset() == 0), | |
| 1759 "Should not have been set"); | |
| 1760 // Is the last block still in the current chunk? If still | |
| 1761 // in this chunk, update the last block (the counting that | |
| 1762 // included the current block is meant for the offset of the last | |
| 1763 // block). If not in this chunk, do nothing. Should not | |
| 1764 // update a block in the next chunk. | |
| 1765 if (ParallelCompactData::chunk_contains_block(bbu.chunk_index(), | |
| 1766 last_block)) { | |
| 1767 if (last_offset < right_offset) { | |
| 1768 // The last object started in this chunk but ends beyond | |
| 1769 // this chunk. Update the block for this last object. | |
| 1770 assert(mark_bitmap()->is_marked(last_offset), "Should be marked"); | |
| 1771 // No end bit was found. The closure takes care of | |
| 1772 // the cases where | |
| 1773 // an objects crosses over into the next block | |
| 1774 // an objects starts and ends in the next block | |
| 1775 // It does not handle the case where an object is | |
| 1776 // the first object in a later block and extends | |
| 1777 // past the end of the chunk (i.e., the closure | |
| 1778 // only handles complete objects that are in the range | |
| 1779 // it is given). That object is handed back here | |
| 1780 // for any special consideration necessary. | |
| 1781 // | |
| 1782 // Is the first bit in the last block a start or end bit? | |
| 1783 // | |
| 1784 // If the partial object ends in the last block L, | |
| 1785 // then the 1st bit in L may be an end bit. | |
| 1786 // | |
| 1787 // Else does the last object start in a block after the current | |
| 1788 // block? A block AA will already have been updated if an | |
| 1789 // object ends in the next block AA+1. An object found to end in | |
| 1790 // the AA+1 is the trigger that updates AA. Objects are being | |
| 1791 // counted in the current block for updaing a following | |
| 1792 // block. An object may start in later block | |
| 1793 // block but may extend beyond the last block in the chunk. | |
| 1794 // Updates are only done when the end of an object has been | |
| 1795 // found. If the last object (covered by block L) starts | |
| 1796 // beyond the current block, then no object ends in L (otherwise | |
| 1797 // L would be the current block). So the first bit in L is | |
| 1798 // a start bit. | |
| 1799 // | |
| 1800 // Else the last objects start in the current block and ends | |
| 1801 // beyond the chunk. The current block has already been | |
| 1802 // updated and there is no later block (with an object | |
| 1803 // starting in it) that needs to be updated. | |
| 1804 // | |
| 1805 if (_summary_data.partial_obj_ends_in_block(last_block)) { | |
| 1806 _summary_data.block(last_block)->set_end_bit_offset( | |
| 1807 bbu.live_data_left()); | |
| 1808 } else if (last_offset_addr >= cur_block_plus_1_addr) { | |
| 1809 // The start of the object is on a later block | |
| 1810 // (to the right of the current block and there are no | |
| 1811 // complete live objects to the left of this last object | |
| 1812 // within the chunk. | |
| 1813 // The first bit in the block is for the start of the | |
| 1814 // last object. | |
| 1815 _summary_data.block(last_block)->set_start_bit_offset( | |
| 1816 bbu.live_data_left()); | |
| 1817 } else { | |
| 1818 // The start of the last object was found in | |
| 1819 // the current chunk (which has already | |
| 1820 // been updated). | |
| 1821 assert(bbu.cur_block() == | |
| 1822 _summary_data.addr_to_block_idx(last_offset_addr), | |
| 1823 "Should be a block already processed"); | |
| 1824 } | |
| 1825 #ifdef ASSERT | |
| 1826 // Is there enough block information to find this object? | |
| 1827 // The destination of the chunk has not been set so the | |
| 1828 // values returned by calc_new_pointer() and | |
| 1829 // block_calc_new_pointer() will only be | |
| 1830 // offsets. But they should agree. | |
| 1831 HeapWord* moved_obj_with_chunks = | |
| 1832 _summary_data.chunk_calc_new_pointer(last_offset_addr); | |
| 1833 HeapWord* moved_obj_with_blocks = | |
| 1834 _summary_data.calc_new_pointer(last_offset_addr); | |
| 1835 assert(moved_obj_with_chunks == moved_obj_with_blocks, | |
| 1836 "Block calculation is wrong"); | |
| 1837 #endif | |
| 1838 } else if (last_block < _summary_data.block_count()) { | |
| 1839 // Iterations ended looking for a start bit (but | |
| 1840 // did not run off the end of the block table). | |
| 1841 _summary_data.block(last_block)->set_start_bit_offset( | |
| 1842 bbu.live_data_left()); | |
| 1843 } | |
| 1844 } | |
| 1845 #ifdef ASSERT | |
| 1846 // Is there enough block information to find this object? | |
| 1847 HeapWord* left_offset_addr = mark_bitmap()->bit_to_addr(left_offset); | |
| 1848 HeapWord* moved_obj_with_chunks = | |
| 1849 _summary_data.calc_new_pointer(left_offset_addr); | |
| 1850 HeapWord* moved_obj_with_blocks = | |
| 1851 _summary_data.calc_new_pointer(left_offset_addr); | |
| 1852 assert(moved_obj_with_chunks == moved_obj_with_blocks, | |
| 1853 "Block calculation is wrong"); | |
| 1854 #endif | |
| 1855 | |
| 1856 // Is there another block after the end of this chunk? | |
| 1857 #ifdef ASSERT | |
| 1858 if (last_block < _summary_data.block_count()) { | |
| 1859 // No object may have been found in a block. If that | |
| 1860 // block is at the end of the chunk, the iteration will | |
| 1861 // terminate without incrementing the current block so | |
| 1862 // that the current block is not the last block in the | |
| 1863 // chunk. That situation precludes asserting that the | |
| 1864 // current block is the last block in the chunk. Assert | |
| 1865 // the lesser condition that the current block does not | |
| 1866 // exceed the chunk. | |
| 1867 assert(_summary_data.block_to_addr(last_block) <= | |
| 1868 (_summary_data.chunk_to_addr(chunk_index) + | |
| 1869 ParallelCompactData::ChunkSize), | |
| 1870 "Chunk and block inconsistency"); | |
| 1871 assert(last_offset <= right_offset, "Iteration over ran end"); | |
| 1872 } | |
| 1873 #endif | |
| 1874 } | |
| 1875 #ifdef ASSERT | |
| 1876 if (PrintGCDetails && Verbose) { | |
| 1877 if (_summary_data.chunk(chunk_index)->partial_obj_size() == 1) { | |
| 1878 size_t first_block = | |
| 1879 chunk_index / ParallelCompactData::BlocksPerChunk; | |
| 1880 gclog_or_tty->print_cr("first_block " PTR_FORMAT | |
| 1881 " _offset " PTR_FORMAT | |
| 1882 "_first_is_start_bit %d", | |
| 1883 first_block, | |
| 1884 _summary_data.block(first_block)->raw_offset(), | |
| 1885 _summary_data.block(first_block)->first_is_start_bit()); | |
| 1886 } | |
| 1887 } | |
| 1888 #endif | |
| 1889 } | |
| 1890 } | |
| 1891 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(16);) | |
| 1892 #endif // #if 0 | |
| 1893 } | |
| 1894 | |
| 1895 // This method should contain all heap-specific policy for invoking a full | |
| 1896 // collection. invoke_no_policy() will only attempt to compact the heap; it | |
| 1897 // will do nothing further. If we need to bail out for policy reasons, scavenge | |
| 1898 // before full gc, or any other specialized behavior, it needs to be added here. | |
| 1899 // | |
| 1900 // Note that this method should only be called from the vm_thread while at a | |
| 1901 // safepoint. | |
| 1902 void PSParallelCompact::invoke(bool maximum_heap_compaction) { | |
| 1903 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); | |
| 1904 assert(Thread::current() == (Thread*)VMThread::vm_thread(), | |
| 1905 "should be in vm thread"); | |
| 1906 ParallelScavengeHeap* heap = gc_heap(); | |
| 1907 GCCause::Cause gc_cause = heap->gc_cause(); | |
| 1908 assert(!heap->is_gc_active(), "not reentrant"); | |
| 1909 | |
| 1910 PSAdaptiveSizePolicy* policy = heap->size_policy(); | |
| 1911 | |
| 1912 // Before each allocation/collection attempt, find out from the | |
| 1913 // policy object if GCs are, on the whole, taking too long. If so, | |
| 1914 // bail out without attempting a collection. The exceptions are | |
| 1915 // for explicitly requested GC's. | |
| 1916 if (!policy->gc_time_limit_exceeded() || | |
| 1917 GCCause::is_user_requested_gc(gc_cause) || | |
| 1918 GCCause::is_serviceability_requested_gc(gc_cause)) { | |
| 1919 IsGCActiveMark mark; | |
| 1920 | |
| 1921 if (ScavengeBeforeFullGC) { | |
| 1922 PSScavenge::invoke_no_policy(); | |
| 1923 } | |
| 1924 | |
| 1925 PSParallelCompact::invoke_no_policy(maximum_heap_compaction); | |
| 1926 } | |
| 1927 } | |
| 1928 | |
| 1929 bool ParallelCompactData::chunk_contains(size_t chunk_index, HeapWord* addr) { | |
| 1930 size_t addr_chunk_index = addr_to_chunk_idx(addr); | |
| 1931 return chunk_index == addr_chunk_index; | |
| 1932 } | |
| 1933 | |
| 1934 bool ParallelCompactData::chunk_contains_block(size_t chunk_index, | |
| 1935 size_t block_index) { | |
| 1936 size_t first_block_in_chunk = chunk_index * BlocksPerChunk; | |
| 1937 size_t last_block_in_chunk = (chunk_index + 1) * BlocksPerChunk - 1; | |
| 1938 | |
| 1939 return (first_block_in_chunk <= block_index) && | |
| 1940 (block_index <= last_block_in_chunk); | |
| 1941 } | |
| 1942 | |
| 1943 // This method contains no policy. You should probably | |
| 1944 // be calling invoke() instead. | |
| 1945 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { | |
| 1946 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); | |
| 1947 assert(ref_processor() != NULL, "Sanity"); | |
| 1948 | |
|
139
c0492d52d55b
6539517: CR 6186200 should be extended to perm gen allocation to prevent spurious OOM's from perm gen
apetrusenko
parents:
0
diff
changeset
|
1949 if (GC_locker::check_active_before_gc()) { |
| 0 | 1950 return; |
| 1951 } | |
| 1952 | |
| 1953 TimeStamp marking_start; | |
| 1954 TimeStamp compaction_start; | |
| 1955 TimeStamp collection_exit; | |
| 1956 | |
| 1957 ParallelScavengeHeap* heap = gc_heap(); | |
| 1958 GCCause::Cause gc_cause = heap->gc_cause(); | |
| 1959 PSYoungGen* young_gen = heap->young_gen(); | |
| 1960 PSOldGen* old_gen = heap->old_gen(); | |
| 1961 PSPermGen* perm_gen = heap->perm_gen(); | |
| 1962 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); | |
| 1963 | |
| 1964 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; | |
| 1965 | |
| 1966 // Make sure data structures are sane, make the heap parsable, and do other | |
| 1967 // miscellaneous bookkeeping. | |
| 1968 PreGCValues pre_gc_values; | |
| 1969 pre_compact(&pre_gc_values); | |
| 1970 | |
| 210 | 1971 // Get the compaction manager reserved for the VM thread. |
| 1972 ParCompactionManager* const vmthread_cm = | |
| 1973 ParCompactionManager::manager_array(gc_task_manager()->workers()); | |
| 1974 | |
| 0 | 1975 // Place after pre_compact() where the number of invocations is incremented. |
| 1976 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); | |
| 1977 | |
| 1978 { | |
| 1979 ResourceMark rm; | |
| 1980 HandleMark hm; | |
| 1981 | |
| 1982 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc; | |
| 1983 | |
| 1984 // This is useful for debugging but don't change the output the | |
| 1985 // the customer sees. | |
| 1986 const char* gc_cause_str = "Full GC"; | |
| 1987 if (is_system_gc && PrintGCDetails) { | |
| 1988 gc_cause_str = "Full GC (System)"; | |
| 1989 } | |
| 1990 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); | |
| 1991 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
| 1992 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty); | |
| 1993 TraceCollectorStats tcs(counters()); | |
| 1994 TraceMemoryManagerStats tms(true /* Full GC */); | |
| 1995 | |
| 1996 if (TraceGen1Time) accumulated_time()->start(); | |
| 1997 | |
| 1998 // Let the size policy know we're starting | |
| 1999 size_policy->major_collection_begin(); | |
| 2000 | |
| 2001 // When collecting the permanent generation methodOops may be moving, | |
| 2002 // so we either have to flush all bcp data or convert it into bci. | |
| 2003 CodeCache::gc_prologue(); | |
| 2004 Threads::gc_prologue(); | |
| 2005 | |
| 2006 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); | |
| 2007 COMPILER2_PRESENT(DerivedPointerTable::clear()); | |
| 2008 | |
| 2009 ref_processor()->enable_discovery(); | |
| 2010 | |
| 2011 bool marked_for_unloading = false; | |
| 2012 | |
| 2013 marking_start.update(); | |
| 210 | 2014 marking_phase(vmthread_cm, maximum_heap_compaction); |
| 0 | 2015 |
| 2016 #ifndef PRODUCT | |
| 2017 if (TraceParallelOldGCMarkingPhase) { | |
| 2018 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " | |
| 2019 "cas_by_another %d", | |
| 2020 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), | |
| 2021 mark_bitmap()->cas_by_another()); | |
| 2022 } | |
| 2023 #endif // #ifndef PRODUCT | |
| 2024 | |
| 2025 #ifdef ASSERT | |
| 2026 if (VerifyParallelOldWithMarkSweep && | |
| 2027 (PSParallelCompact::total_invocations() % | |
| 2028 VerifyParallelOldWithMarkSweepInterval) == 0) { | |
| 2029 gclog_or_tty->print_cr("Verify marking with mark_sweep_phase1()"); | |
| 2030 if (PrintGCDetails && Verbose) { | |
| 2031 gclog_or_tty->print_cr("mark_sweep_phase1:"); | |
| 2032 } | |
| 2033 // Clear the discovered lists so that discovered objects | |
| 2034 // don't look like they have been discovered twice. | |
| 2035 ref_processor()->clear_discovered_references(); | |
| 2036 | |
| 2037 PSMarkSweep::allocate_stacks(); | |
| 2038 MemRegion mr = Universe::heap()->reserved_region(); | |
| 2039 PSMarkSweep::ref_processor()->enable_discovery(); | |
| 2040 PSMarkSweep::mark_sweep_phase1(maximum_heap_compaction); | |
| 2041 } | |
| 2042 #endif | |
| 2043 | |
| 2044 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc; | |
| 210 | 2045 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); |
| 0 | 2046 |
| 2047 #ifdef ASSERT | |
| 2048 if (VerifyParallelOldWithMarkSweep && | |
| 2049 (PSParallelCompact::total_invocations() % | |
| 2050 VerifyParallelOldWithMarkSweepInterval) == 0) { | |
| 2051 if (PrintGCDetails && Verbose) { | |
| 2052 gclog_or_tty->print_cr("mark_sweep_phase2:"); | |
| 2053 } | |
| 2054 PSMarkSweep::mark_sweep_phase2(); | |
| 2055 } | |
| 2056 #endif | |
| 2057 | |
| 2058 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); | |
| 2059 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); | |
| 2060 | |
| 2061 // adjust_roots() updates Universe::_intArrayKlassObj which is | |
| 2062 // needed by the compaction for filling holes in the dense prefix. | |
| 2063 adjust_roots(); | |
| 2064 | |
| 2065 #ifdef ASSERT | |
| 2066 if (VerifyParallelOldWithMarkSweep && | |
| 2067 (PSParallelCompact::total_invocations() % | |
| 2068 VerifyParallelOldWithMarkSweepInterval) == 0) { | |
| 2069 // Do a separate verify phase so that the verify | |
| 2070 // code can use the the forwarding pointers to | |
| 2071 // check the new pointer calculation. The restore_marks() | |
| 2072 // has to be done before the real compact. | |
| 210 | 2073 vmthread_cm->set_action(ParCompactionManager::VerifyUpdate); |
| 2074 compact_perm(vmthread_cm); | |
| 2075 compact_serial(vmthread_cm); | |
| 2076 vmthread_cm->set_action(ParCompactionManager::ResetObjects); | |
| 2077 compact_perm(vmthread_cm); | |
| 2078 compact_serial(vmthread_cm); | |
| 2079 vmthread_cm->set_action(ParCompactionManager::UpdateAndCopy); | |
| 0 | 2080 |
| 2081 // For debugging only | |
| 2082 PSMarkSweep::restore_marks(); | |
| 2083 PSMarkSweep::deallocate_stacks(); | |
| 2084 } | |
| 2085 #endif | |
| 2086 | |
| 2087 compaction_start.update(); | |
| 2088 // Does the perm gen always have to be done serially because | |
| 2089 // klasses are used in the update of an object? | |
| 210 | 2090 compact_perm(vmthread_cm); |
| 0 | 2091 |
| 2092 if (UseParallelOldGCCompacting) { | |
| 2093 compact(); | |
| 2094 } else { | |
| 210 | 2095 compact_serial(vmthread_cm); |
| 0 | 2096 } |
| 2097 | |
| 2098 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be | |
| 2099 // done before resizing. | |
| 2100 post_compact(); | |
| 2101 | |
| 2102 // Let the size policy know we're done | |
| 2103 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); | |
| 2104 | |
| 2105 if (UseAdaptiveSizePolicy) { | |
| 2106 if (PrintAdaptiveSizePolicy) { | |
| 2107 gclog_or_tty->print("AdaptiveSizeStart: "); | |
| 2108 gclog_or_tty->stamp(); | |
| 2109 gclog_or_tty->print_cr(" collection: %d ", | |
| 2110 heap->total_collections()); | |
| 2111 if (Verbose) { | |
| 2112 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d" | |
| 2113 " perm_gen_capacity: %d ", | |
| 2114 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(), | |
| 2115 perm_gen->capacity_in_bytes()); | |
| 2116 } | |
| 2117 } | |
| 2118 | |
| 2119 // Don't check if the size_policy is ready here. Let | |
| 2120 // the size_policy check that internally. | |
| 2121 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && | |
| 2122 ((gc_cause != GCCause::_java_lang_system_gc) || | |
| 2123 UseAdaptiveSizePolicyWithSystemGC)) { | |
| 2124 // Calculate optimal free space amounts | |
| 2125 assert(young_gen->max_size() > | |
| 2126 young_gen->from_space()->capacity_in_bytes() + | |
| 2127 young_gen->to_space()->capacity_in_bytes(), | |
| 2128 "Sizes of space in young gen are out-of-bounds"); | |
| 2129 size_t max_eden_size = young_gen->max_size() - | |
| 2130 young_gen->from_space()->capacity_in_bytes() - | |
| 2131 young_gen->to_space()->capacity_in_bytes(); | |
| 2132 size_policy->compute_generation_free_space(young_gen->used_in_bytes(), | |
| 2133 young_gen->eden_space()->used_in_bytes(), | |
| 2134 old_gen->used_in_bytes(), | |
| 2135 perm_gen->used_in_bytes(), | |
| 2136 young_gen->eden_space()->capacity_in_bytes(), | |
| 2137 old_gen->max_gen_size(), | |
| 2138 max_eden_size, | |
| 2139 true /* full gc*/, | |
| 2140 gc_cause); | |
| 2141 | |
| 2142 heap->resize_old_gen(size_policy->calculated_old_free_size_in_bytes()); | |
| 2143 | |
| 2144 // Don't resize the young generation at an major collection. A | |
| 2145 // desired young generation size may have been calculated but | |
| 2146 // resizing the young generation complicates the code because the | |
| 2147 // resizing of the old generation may have moved the boundary | |
| 2148 // between the young generation and the old generation. Let the | |
| 2149 // young generation resizing happen at the minor collections. | |
| 2150 } | |
| 2151 if (PrintAdaptiveSizePolicy) { | |
| 2152 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", | |
| 2153 heap->total_collections()); | |
| 2154 } | |
| 2155 } | |
| 2156 | |
| 2157 if (UsePerfData) { | |
| 2158 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); | |
| 2159 counters->update_counters(); | |
| 2160 counters->update_old_capacity(old_gen->capacity_in_bytes()); | |
| 2161 counters->update_young_capacity(young_gen->capacity_in_bytes()); | |
| 2162 } | |
| 2163 | |
| 2164 heap->resize_all_tlabs(); | |
| 2165 | |
| 2166 // We collected the perm gen, so we'll resize it here. | |
| 2167 perm_gen->compute_new_size(pre_gc_values.perm_gen_used()); | |
| 2168 | |
| 2169 if (TraceGen1Time) accumulated_time()->stop(); | |
| 2170 | |
| 2171 if (PrintGC) { | |
| 2172 if (PrintGCDetails) { | |
| 2173 // No GC timestamp here. This is after GC so it would be confusing. | |
| 2174 young_gen->print_used_change(pre_gc_values.young_gen_used()); | |
| 2175 old_gen->print_used_change(pre_gc_values.old_gen_used()); | |
| 2176 heap->print_heap_change(pre_gc_values.heap_used()); | |
| 2177 // Print perm gen last (print_heap_change() excludes the perm gen). | |
| 2178 perm_gen->print_used_change(pre_gc_values.perm_gen_used()); | |
| 2179 } else { | |
| 2180 heap->print_heap_change(pre_gc_values.heap_used()); | |
| 2181 } | |
| 2182 } | |
| 2183 | |
| 2184 // Track memory usage and detect low memory | |
| 2185 MemoryService::track_memory_usage(); | |
| 2186 heap->update_counters(); | |
| 2187 | |
| 2188 if (PrintGCDetails) { | |
| 2189 if (size_policy->print_gc_time_limit_would_be_exceeded()) { | |
| 2190 if (size_policy->gc_time_limit_exceeded()) { | |
| 2191 gclog_or_tty->print_cr(" GC time is exceeding GCTimeLimit " | |
| 2192 "of %d%%", GCTimeLimit); | |
| 2193 } else { | |
| 2194 gclog_or_tty->print_cr(" GC time would exceed GCTimeLimit " | |
| 2195 "of %d%%", GCTimeLimit); | |
| 2196 } | |
| 2197 } | |
| 2198 size_policy->set_print_gc_time_limit_would_be_exceeded(false); | |
| 2199 } | |
| 2200 } | |
| 2201 | |
| 2202 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { | |
| 2203 HandleMark hm; // Discard invalid handles created during verification | |
| 2204 gclog_or_tty->print(" VerifyAfterGC:"); | |
| 2205 Universe::verify(false); | |
| 2206 } | |
| 2207 | |
| 2208 // Re-verify object start arrays | |
| 2209 if (VerifyObjectStartArray && | |
| 2210 VerifyAfterGC) { | |
| 2211 old_gen->verify_object_start_array(); | |
| 2212 perm_gen->verify_object_start_array(); | |
| 2213 } | |
| 2214 | |
| 2215 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); | |
| 2216 | |
| 2217 collection_exit.update(); | |
| 2218 | |
| 2219 if (PrintHeapAtGC) { | |
| 2220 Universe::print_heap_after_gc(); | |
| 2221 } | |
| 2222 if (PrintGCTaskTimeStamps) { | |
| 2223 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " | |
| 2224 INT64_FORMAT, | |
| 2225 marking_start.ticks(), compaction_start.ticks(), | |
| 2226 collection_exit.ticks()); | |
| 2227 gc_task_manager()->print_task_time_stamps(); | |
| 2228 } | |
| 2229 } | |
| 2230 | |
| 2231 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, | |
| 2232 PSYoungGen* young_gen, | |
| 2233 PSOldGen* old_gen) { | |
| 2234 MutableSpace* const eden_space = young_gen->eden_space(); | |
| 2235 assert(!eden_space->is_empty(), "eden must be non-empty"); | |
| 2236 assert(young_gen->virtual_space()->alignment() == | |
| 2237 old_gen->virtual_space()->alignment(), "alignments do not match"); | |
| 2238 | |
| 2239 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { | |
| 2240 return false; | |
| 2241 } | |
| 2242 | |
| 2243 // Both generations must be completely committed. | |
| 2244 if (young_gen->virtual_space()->uncommitted_size() != 0) { | |
| 2245 return false; | |
| 2246 } | |
| 2247 if (old_gen->virtual_space()->uncommitted_size() != 0) { | |
| 2248 return false; | |
| 2249 } | |
| 2250 | |
| 2251 // Figure out how much to take from eden. Include the average amount promoted | |
| 2252 // in the total; otherwise the next young gen GC will simply bail out to a | |
| 2253 // full GC. | |
| 2254 const size_t alignment = old_gen->virtual_space()->alignment(); | |
| 2255 const size_t eden_used = eden_space->used_in_bytes(); | |
| 2256 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); | |
| 2257 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); | |
| 2258 const size_t eden_capacity = eden_space->capacity_in_bytes(); | |
| 2259 | |
| 2260 if (absorb_size >= eden_capacity) { | |
| 2261 return false; // Must leave some space in eden. | |
| 2262 } | |
| 2263 | |
| 2264 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; | |
| 2265 if (new_young_size < young_gen->min_gen_size()) { | |
| 2266 return false; // Respect young gen minimum size. | |
| 2267 } | |
| 2268 | |
| 2269 if (TraceAdaptiveGCBoundary && Verbose) { | |
| 2270 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " | |
| 2271 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " | |
| 2272 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " | |
| 2273 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", | |
| 2274 absorb_size / K, | |
| 2275 eden_capacity / K, (eden_capacity - absorb_size) / K, | |
| 2276 young_gen->from_space()->used_in_bytes() / K, | |
| 2277 young_gen->to_space()->used_in_bytes() / K, | |
| 2278 young_gen->capacity_in_bytes() / K, new_young_size / K); | |
| 2279 } | |
| 2280 | |
| 2281 // Fill the unused part of the old gen. | |
| 2282 MutableSpace* const old_space = old_gen->object_space(); | |
| 2283 MemRegion old_gen_unused(old_space->top(), old_space->end()); | |
| 2284 if (!old_gen_unused.is_empty()) { | |
| 2285 SharedHeap::fill_region_with_object(old_gen_unused); | |
| 2286 } | |
| 2287 | |
| 2288 // Take the live data from eden and set both top and end in the old gen to | |
| 2289 // eden top. (Need to set end because reset_after_change() mangles the region | |
| 2290 // from end to virtual_space->high() in debug builds). | |
| 2291 HeapWord* const new_top = eden_space->top(); | |
| 2292 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), | |
| 2293 absorb_size); | |
| 2294 young_gen->reset_after_change(); | |
| 2295 old_space->set_top(new_top); | |
| 2296 old_space->set_end(new_top); | |
| 2297 old_gen->reset_after_change(); | |
| 2298 | |
| 2299 // Update the object start array for the filler object and the data from eden. | |
| 2300 ObjectStartArray* const start_array = old_gen->start_array(); | |
| 2301 HeapWord* const start = old_gen_unused.start(); | |
| 2302 for (HeapWord* addr = start; addr < new_top; addr += oop(addr)->size()) { | |
| 2303 start_array->allocate_block(addr); | |
| 2304 } | |
| 2305 | |
| 2306 // Could update the promoted average here, but it is not typically updated at | |
| 2307 // full GCs and the value to use is unclear. Something like | |
| 2308 // | |
| 2309 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. | |
| 2310 | |
| 2311 size_policy->set_bytes_absorbed_from_eden(absorb_size); | |
| 2312 return true; | |
| 2313 } | |
| 2314 | |
| 2315 GCTaskManager* const PSParallelCompact::gc_task_manager() { | |
| 2316 assert(ParallelScavengeHeap::gc_task_manager() != NULL, | |
| 2317 "shouldn't return NULL"); | |
| 2318 return ParallelScavengeHeap::gc_task_manager(); | |
| 2319 } | |
| 2320 | |
| 2321 void PSParallelCompact::marking_phase(ParCompactionManager* cm, | |
| 2322 bool maximum_heap_compaction) { | |
| 2323 // Recursively traverse all live objects and mark them | |
| 2324 EventMark m("1 mark object"); | |
| 2325 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); | |
| 2326 | |
| 2327 ParallelScavengeHeap* heap = gc_heap(); | |
| 2328 uint parallel_gc_threads = heap->gc_task_manager()->workers(); | |
| 2329 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array(); | |
| 2330 ParallelTaskTerminator terminator(parallel_gc_threads, qset); | |
| 2331 | |
| 2332 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); | |
| 2333 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); | |
| 2334 | |
| 2335 { | |
| 2336 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); | |
| 2337 | |
| 2338 GCTaskQueue* q = GCTaskQueue::create(); | |
| 2339 | |
| 2340 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); | |
| 2341 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); | |
| 2342 // We scan the thread roots in parallel | |
| 2343 Threads::create_thread_roots_marking_tasks(q); | |
| 2344 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); | |
| 2345 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); | |
| 2346 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); | |
| 2347 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); | |
| 2348 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); | |
| 2349 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols)); | |
| 2350 | |
| 2351 if (parallel_gc_threads > 1) { | |
| 2352 for (uint j = 0; j < parallel_gc_threads; j++) { | |
| 2353 q->enqueue(new StealMarkingTask(&terminator)); | |
| 2354 } | |
| 2355 } | |
| 2356 | |
| 2357 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); | |
| 2358 q->enqueue(fin); | |
| 2359 | |
| 2360 gc_task_manager()->add_list(q); | |
| 2361 | |
| 2362 fin->wait_for(); | |
| 2363 | |
| 2364 // We have to release the barrier tasks! | |
| 2365 WaitForBarrierGCTask::destroy(fin); | |
| 2366 } | |
| 2367 | |
| 2368 // Process reference objects found during marking | |
| 2369 { | |
| 2370 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); | |
| 2371 ReferencePolicy *soft_ref_policy; | |
| 2372 if (maximum_heap_compaction) { | |
| 2373 soft_ref_policy = new AlwaysClearPolicy(); | |
| 2374 } else { | |
| 2375 #ifdef COMPILER2 | |
| 2376 soft_ref_policy = new LRUMaxHeapPolicy(); | |
| 2377 #else | |
| 2378 soft_ref_policy = new LRUCurrentHeapPolicy(); | |
| 2379 #endif // COMPILER2 | |
| 2380 } | |
| 2381 assert(soft_ref_policy != NULL, "No soft reference policy"); | |
| 2382 if (ref_processor()->processing_is_mt()) { | |
| 2383 RefProcTaskExecutor task_executor; | |
| 2384 ref_processor()->process_discovered_references( | |
| 2385 soft_ref_policy, is_alive_closure(), &mark_and_push_closure, | |
| 2386 &follow_stack_closure, &task_executor); | |
| 2387 } else { | |
| 2388 ref_processor()->process_discovered_references( | |
| 2389 soft_ref_policy, is_alive_closure(), &mark_and_push_closure, | |
| 2390 &follow_stack_closure, NULL); | |
| 2391 } | |
| 2392 } | |
| 2393 | |
| 2394 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); | |
| 2395 // Follow system dictionary roots and unload classes. | |
| 2396 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); | |
| 2397 | |
| 2398 // Follow code cache roots. | |
| 2399 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure, | |
| 2400 purged_class); | |
| 2401 follow_stack(cm); // Flush marking stack. | |
| 2402 | |
| 2403 // Update subklass/sibling/implementor links of live klasses | |
| 2404 // revisit_klass_stack is used in follow_weak_klass_links(). | |
| 2405 follow_weak_klass_links(cm); | |
| 2406 | |
| 2407 // Visit symbol and interned string tables and delete unmarked oops | |
| 2408 SymbolTable::unlink(is_alive_closure()); | |
| 2409 StringTable::unlink(is_alive_closure()); | |
| 2410 | |
| 2411 assert(cm->marking_stack()->size() == 0, "stack should be empty by now"); | |
| 2412 assert(cm->overflow_stack()->is_empty(), "stack should be empty by now"); | |
| 2413 } | |
| 2414 | |
| 2415 // This should be moved to the shared markSweep code! | |
| 2416 class PSAlwaysTrueClosure: public BoolObjectClosure { | |
| 2417 public: | |
| 2418 void do_object(oop p) { ShouldNotReachHere(); } | |
| 2419 bool do_object_b(oop p) { return true; } | |
| 2420 }; | |
| 2421 static PSAlwaysTrueClosure always_true; | |
| 2422 | |
| 2423 void PSParallelCompact::adjust_roots() { | |
| 2424 // Adjust the pointers to reflect the new locations | |
| 2425 EventMark m("3 adjust roots"); | |
| 2426 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); | |
| 2427 | |
| 2428 // General strong roots. | |
| 2429 Universe::oops_do(adjust_root_pointer_closure()); | |
| 2430 ReferenceProcessor::oops_do(adjust_root_pointer_closure()); | |
| 2431 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles | |
| 2432 Threads::oops_do(adjust_root_pointer_closure()); | |
| 2433 ObjectSynchronizer::oops_do(adjust_root_pointer_closure()); | |
| 2434 FlatProfiler::oops_do(adjust_root_pointer_closure()); | |
| 2435 Management::oops_do(adjust_root_pointer_closure()); | |
| 2436 JvmtiExport::oops_do(adjust_root_pointer_closure()); | |
| 2437 // SO_AllClasses | |
| 2438 SystemDictionary::oops_do(adjust_root_pointer_closure()); | |
| 2439 vmSymbols::oops_do(adjust_root_pointer_closure()); | |
| 2440 | |
| 2441 // Now adjust pointers in remaining weak roots. (All of which should | |
| 2442 // have been cleared if they pointed to non-surviving objects.) | |
| 2443 // Global (weak) JNI handles | |
| 2444 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure()); | |
| 2445 | |
| 2446 CodeCache::oops_do(adjust_pointer_closure()); | |
| 2447 SymbolTable::oops_do(adjust_root_pointer_closure()); | |
| 2448 StringTable::oops_do(adjust_root_pointer_closure()); | |
| 2449 ref_processor()->weak_oops_do(adjust_root_pointer_closure()); | |
| 2450 // Roots were visited so references into the young gen in roots | |
| 2451 // may have been scanned. Process them also. | |
| 2452 // Should the reference processor have a span that excludes | |
| 2453 // young gen objects? | |
| 2454 PSScavenge::reference_processor()->weak_oops_do( | |
| 2455 adjust_root_pointer_closure()); | |
| 2456 } | |
| 2457 | |
| 2458 void PSParallelCompact::compact_perm(ParCompactionManager* cm) { | |
| 2459 EventMark m("4 compact perm"); | |
| 2460 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty); | |
| 2461 // trace("4"); | |
| 2462 | |
| 2463 gc_heap()->perm_gen()->start_array()->reset(); | |
| 2464 move_and_update(cm, perm_space_id); | |
| 2465 } | |
| 2466 | |
| 2467 void PSParallelCompact::enqueue_chunk_draining_tasks(GCTaskQueue* q, | |
| 2468 uint parallel_gc_threads) { | |
| 2469 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); | |
| 2470 | |
| 2471 const unsigned int task_count = MAX2(parallel_gc_threads, 1U); | |
| 2472 for (unsigned int j = 0; j < task_count; j++) { | |
| 2473 q->enqueue(new DrainStacksCompactionTask()); | |
| 2474 } | |
| 2475 | |
| 2476 // Find all chunks that are available (can be filled immediately) and | |
| 2477 // distribute them to the thread stacks. The iteration is done in reverse | |
| 2478 // order (high to low) so the chunks will be removed in ascending order. | |
| 2479 | |
| 2480 const ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 2481 | |
| 2482 size_t fillable_chunks = 0; // A count for diagnostic purposes. | |
| 2483 unsigned int which = 0; // The worker thread number. | |
| 2484 | |
| 2485 for (unsigned int id = to_space_id; id > perm_space_id; --id) { | |
| 2486 SpaceInfo* const space_info = _space_info + id; | |
| 2487 MutableSpace* const space = space_info->space(); | |
| 2488 HeapWord* const new_top = space_info->new_top(); | |
| 2489 | |
| 2490 const size_t beg_chunk = sd.addr_to_chunk_idx(space_info->dense_prefix()); | |
| 2491 const size_t end_chunk = sd.addr_to_chunk_idx(sd.chunk_align_up(new_top)); | |
| 2492 assert(end_chunk > 0, "perm gen cannot be empty"); | |
| 2493 | |
| 2494 for (size_t cur = end_chunk - 1; cur >= beg_chunk; --cur) { | |
| 2495 if (sd.chunk(cur)->claim_unsafe()) { | |
| 2496 ParCompactionManager* cm = ParCompactionManager::manager_array(which); | |
| 2497 cm->save_for_processing(cur); | |
| 2498 | |
| 2499 if (TraceParallelOldGCCompactionPhase && Verbose) { | |
| 2500 const size_t count_mod_8 = fillable_chunks & 7; | |
| 2501 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); | |
| 2502 gclog_or_tty->print(" " SIZE_FORMAT_W("7"), cur); | |
| 2503 if (count_mod_8 == 7) gclog_or_tty->cr(); | |
| 2504 } | |
| 2505 | |
| 2506 NOT_PRODUCT(++fillable_chunks;) | |
| 2507 | |
| 2508 // Assign chunks to threads in round-robin fashion. | |
| 2509 if (++which == task_count) { | |
| 2510 which = 0; | |
| 2511 } | |
| 2512 } | |
| 2513 } | |
| 2514 } | |
| 2515 | |
| 2516 if (TraceParallelOldGCCompactionPhase) { | |
| 2517 if (Verbose && (fillable_chunks & 7) != 0) gclog_or_tty->cr(); | |
| 2518 gclog_or_tty->print_cr("%u initially fillable chunks", fillable_chunks); | |
| 2519 } | |
| 2520 } | |
| 2521 | |
| 2522 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 | |
| 2523 | |
| 2524 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, | |
| 2525 uint parallel_gc_threads) { | |
| 2526 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); | |
| 2527 | |
| 2528 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 2529 | |
| 2530 // Iterate over all the spaces adding tasks for updating | |
| 2531 // chunks in the dense prefix. Assume that 1 gc thread | |
| 2532 // will work on opening the gaps and the remaining gc threads | |
| 2533 // will work on the dense prefix. | |
| 2534 SpaceId space_id = old_space_id; | |
| 2535 while (space_id != last_space_id) { | |
| 2536 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); | |
| 2537 const MutableSpace* const space = _space_info[space_id].space(); | |
| 2538 | |
| 2539 if (dense_prefix_end == space->bottom()) { | |
| 2540 // There is no dense prefix for this space. | |
| 2541 space_id = next_compaction_space_id(space_id); | |
| 2542 continue; | |
| 2543 } | |
| 2544 | |
| 2545 // The dense prefix is before this chunk. | |
| 2546 size_t chunk_index_end_dense_prefix = | |
| 2547 sd.addr_to_chunk_idx(dense_prefix_end); | |
| 2548 ChunkData* const dense_prefix_cp = sd.chunk(chunk_index_end_dense_prefix); | |
| 2549 assert(dense_prefix_end == space->end() || | |
| 2550 dense_prefix_cp->available() || | |
| 2551 dense_prefix_cp->claimed(), | |
| 2552 "The chunk after the dense prefix should always be ready to fill"); | |
| 2553 | |
| 2554 size_t chunk_index_start = sd.addr_to_chunk_idx(space->bottom()); | |
| 2555 | |
| 2556 // Is there dense prefix work? | |
| 2557 size_t total_dense_prefix_chunks = | |
| 2558 chunk_index_end_dense_prefix - chunk_index_start; | |
| 2559 // How many chunks of the dense prefix should be given to | |
| 2560 // each thread? | |
| 2561 if (total_dense_prefix_chunks > 0) { | |
| 2562 uint tasks_for_dense_prefix = 1; | |
| 2563 if (UseParallelDensePrefixUpdate) { | |
| 2564 if (total_dense_prefix_chunks <= | |
| 2565 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { | |
| 2566 // Don't over partition. This assumes that | |
| 2567 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value | |
| 2568 // so there are not many chunks to process. | |
| 2569 tasks_for_dense_prefix = parallel_gc_threads; | |
| 2570 } else { | |
| 2571 // Over partition | |
| 2572 tasks_for_dense_prefix = parallel_gc_threads * | |
| 2573 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; | |
| 2574 } | |
| 2575 } | |
| 2576 size_t chunks_per_thread = total_dense_prefix_chunks / | |
| 2577 tasks_for_dense_prefix; | |
| 2578 // Give each thread at least 1 chunk. | |
| 2579 if (chunks_per_thread == 0) { | |
| 2580 chunks_per_thread = 1; | |
| 2581 } | |
| 2582 | |
| 2583 for (uint k = 0; k < tasks_for_dense_prefix; k++) { | |
| 2584 if (chunk_index_start >= chunk_index_end_dense_prefix) { | |
| 2585 break; | |
| 2586 } | |
| 2587 // chunk_index_end is not processed | |
| 2588 size_t chunk_index_end = MIN2(chunk_index_start + chunks_per_thread, | |
| 2589 chunk_index_end_dense_prefix); | |
| 2590 q->enqueue(new UpdateDensePrefixTask( | |
| 2591 space_id, | |
| 2592 chunk_index_start, | |
| 2593 chunk_index_end)); | |
| 2594 chunk_index_start = chunk_index_end; | |
| 2595 } | |
| 2596 } | |
| 2597 // This gets any part of the dense prefix that did not | |
| 2598 // fit evenly. | |
| 2599 if (chunk_index_start < chunk_index_end_dense_prefix) { | |
| 2600 q->enqueue(new UpdateDensePrefixTask( | |
| 2601 space_id, | |
| 2602 chunk_index_start, | |
| 2603 chunk_index_end_dense_prefix)); | |
| 2604 } | |
| 2605 space_id = next_compaction_space_id(space_id); | |
| 2606 } // End tasks for dense prefix | |
| 2607 } | |
| 2608 | |
| 2609 void PSParallelCompact::enqueue_chunk_stealing_tasks( | |
| 2610 GCTaskQueue* q, | |
| 2611 ParallelTaskTerminator* terminator_ptr, | |
| 2612 uint parallel_gc_threads) { | |
| 2613 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); | |
| 2614 | |
| 2615 // Once a thread has drained it's stack, it should try to steal chunks from | |
| 2616 // other threads. | |
| 2617 if (parallel_gc_threads > 1) { | |
| 2618 for (uint j = 0; j < parallel_gc_threads; j++) { | |
| 2619 q->enqueue(new StealChunkCompactionTask(terminator_ptr)); | |
| 2620 } | |
| 2621 } | |
| 2622 } | |
| 2623 | |
| 2624 void PSParallelCompact::compact() { | |
| 2625 EventMark m("5 compact"); | |
| 2626 // trace("5"); | |
| 2627 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); | |
| 2628 | |
| 2629 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); | |
| 2630 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); | |
| 2631 PSOldGen* old_gen = heap->old_gen(); | |
| 2632 old_gen->start_array()->reset(); | |
| 2633 uint parallel_gc_threads = heap->gc_task_manager()->workers(); | |
| 2634 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array(); | |
| 2635 ParallelTaskTerminator terminator(parallel_gc_threads, qset); | |
| 2636 | |
| 2637 GCTaskQueue* q = GCTaskQueue::create(); | |
| 2638 enqueue_chunk_draining_tasks(q, parallel_gc_threads); | |
| 2639 enqueue_dense_prefix_tasks(q, parallel_gc_threads); | |
| 2640 enqueue_chunk_stealing_tasks(q, &terminator, parallel_gc_threads); | |
| 2641 | |
| 2642 { | |
| 2643 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); | |
| 2644 | |
| 2645 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); | |
| 2646 q->enqueue(fin); | |
| 2647 | |
| 2648 gc_task_manager()->add_list(q); | |
| 2649 | |
| 2650 fin->wait_for(); | |
| 2651 | |
| 2652 // We have to release the barrier tasks! | |
| 2653 WaitForBarrierGCTask::destroy(fin); | |
| 2654 | |
| 2655 #ifdef ASSERT | |
| 2656 // Verify that all chunks have been processed before the deferred updates. | |
| 2657 // Note that perm_space_id is skipped; this type of verification is not | |
| 2658 // valid until the perm gen is compacted by chunks. | |
| 2659 for (unsigned int id = old_space_id; id < last_space_id; ++id) { | |
| 2660 verify_complete(SpaceId(id)); | |
| 2661 } | |
| 2662 #endif | |
| 2663 } | |
| 2664 | |
| 2665 { | |
| 2666 // Update the deferred objects, if any. Any compaction manager can be used. | |
| 2667 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); | |
| 2668 ParCompactionManager* cm = ParCompactionManager::manager_array(0); | |
| 2669 for (unsigned int id = old_space_id; id < last_space_id; ++id) { | |
| 2670 update_deferred_objects(cm, SpaceId(id)); | |
| 2671 } | |
| 2672 } | |
| 2673 } | |
| 2674 | |
| 2675 #ifdef ASSERT | |
| 2676 void PSParallelCompact::verify_complete(SpaceId space_id) { | |
| 2677 // All Chunks between space bottom() to new_top() should be marked as filled | |
| 2678 // and all Chunks between new_top() and top() should be available (i.e., | |
| 2679 // should have been emptied). | |
| 2680 ParallelCompactData& sd = summary_data(); | |
| 2681 SpaceInfo si = _space_info[space_id]; | |
| 2682 HeapWord* new_top_addr = sd.chunk_align_up(si.new_top()); | |
| 2683 HeapWord* old_top_addr = sd.chunk_align_up(si.space()->top()); | |
| 2684 const size_t beg_chunk = sd.addr_to_chunk_idx(si.space()->bottom()); | |
| 2685 const size_t new_top_chunk = sd.addr_to_chunk_idx(new_top_addr); | |
| 2686 const size_t old_top_chunk = sd.addr_to_chunk_idx(old_top_addr); | |
| 2687 | |
| 2688 bool issued_a_warning = false; | |
| 2689 | |
| 2690 size_t cur_chunk; | |
| 2691 for (cur_chunk = beg_chunk; cur_chunk < new_top_chunk; ++cur_chunk) { | |
| 2692 const ChunkData* const c = sd.chunk(cur_chunk); | |
| 2693 if (!c->completed()) { | |
| 2694 warning("chunk " SIZE_FORMAT " not filled: " | |
| 2695 "destination_count=" SIZE_FORMAT, | |
| 2696 cur_chunk, c->destination_count()); | |
| 2697 issued_a_warning = true; | |
| 2698 } | |
| 2699 } | |
| 2700 | |
| 2701 for (cur_chunk = new_top_chunk; cur_chunk < old_top_chunk; ++cur_chunk) { | |
| 2702 const ChunkData* const c = sd.chunk(cur_chunk); | |
| 2703 if (!c->available()) { | |
| 2704 warning("chunk " SIZE_FORMAT " not empty: " | |
| 2705 "destination_count=" SIZE_FORMAT, | |
| 2706 cur_chunk, c->destination_count()); | |
| 2707 issued_a_warning = true; | |
| 2708 } | |
| 2709 } | |
| 2710 | |
| 2711 if (issued_a_warning) { | |
| 2712 print_chunk_ranges(); | |
| 2713 } | |
| 2714 } | |
| 2715 #endif // #ifdef ASSERT | |
| 2716 | |
| 2717 void PSParallelCompact::compact_serial(ParCompactionManager* cm) { | |
| 2718 EventMark m("5 compact serial"); | |
| 2719 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty); | |
| 2720 | |
| 2721 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); | |
| 2722 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); | |
| 2723 | |
| 2724 PSYoungGen* young_gen = heap->young_gen(); | |
| 2725 PSOldGen* old_gen = heap->old_gen(); | |
| 2726 | |
| 2727 old_gen->start_array()->reset(); | |
| 2728 old_gen->move_and_update(cm); | |
| 2729 young_gen->move_and_update(cm); | |
| 2730 } | |
| 2731 | |
| 2732 | |
| 2733 void PSParallelCompact::follow_stack(ParCompactionManager* cm) { | |
| 2734 while(!cm->overflow_stack()->is_empty()) { | |
| 2735 oop obj = cm->overflow_stack()->pop(); | |
| 2736 obj->follow_contents(cm); | |
| 2737 } | |
| 2738 | |
| 2739 oop obj; | |
| 2740 // obj is a reference!!! | |
| 2741 while (cm->marking_stack()->pop_local(obj)) { | |
| 2742 // It would be nice to assert about the type of objects we might | |
| 2743 // pop, but they can come from anywhere, unfortunately. | |
| 2744 obj->follow_contents(cm); | |
| 2745 } | |
| 2746 } | |
| 2747 | |
| 2748 void | |
| 2749 PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) { | |
| 2750 // All klasses on the revisit stack are marked at this point. | |
| 2751 // Update and follow all subklass, sibling and implementor links. | |
| 2752 for (uint i = 0; i < ParallelGCThreads+1; i++) { | |
| 2753 ParCompactionManager* cm = ParCompactionManager::manager_array(i); | |
| 2754 KeepAliveClosure keep_alive_closure(cm); | |
| 2755 for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) { | |
| 2756 cm->revisit_klass_stack()->at(i)->follow_weak_klass_links( | |
| 2757 is_alive_closure(), | |
| 2758 &keep_alive_closure); | |
| 2759 } | |
| 2760 follow_stack(cm); | |
| 2761 } | |
| 2762 } | |
| 2763 | |
| 2764 void | |
| 2765 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) { | |
| 2766 cm->revisit_klass_stack()->push(k); | |
| 2767 } | |
| 2768 | |
| 2769 #ifdef VALIDATE_MARK_SWEEP | |
| 2770 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
2771 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) { |
| 0 | 2772 if (!ValidateMarkSweep) |
| 2773 return; | |
| 2774 | |
| 2775 if (!isroot) { | |
| 2776 if (_pointer_tracking) { | |
| 2777 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer"); | |
| 2778 _adjusted_pointers->remove(p); | |
| 2779 } | |
| 2780 } else { | |
| 2781 ptrdiff_t index = _root_refs_stack->find(p); | |
| 2782 if (index != -1) { | |
| 2783 int l = _root_refs_stack->length(); | |
| 2784 if (l > 0 && l - 1 != index) { | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
2785 void* last = _root_refs_stack->pop(); |
| 0 | 2786 assert(last != p, "should be different"); |
| 2787 _root_refs_stack->at_put(index, last); | |
| 2788 } else { | |
| 2789 _root_refs_stack->remove(p); | |
| 2790 } | |
| 2791 } | |
| 2792 } | |
| 2793 } | |
| 2794 | |
| 2795 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
2796 void PSParallelCompact::check_adjust_pointer(void* p) { |
| 0 | 2797 _adjusted_pointers->push(p); |
| 2798 } | |
| 2799 | |
| 2800 | |
| 2801 class AdjusterTracker: public OopClosure { | |
| 2802 public: | |
| 2803 AdjusterTracker() {}; | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
2804 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
79
diff
changeset
|
2805 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); } |
| 0 | 2806 }; |
| 2807 | |
| 2808 | |
| 2809 void PSParallelCompact::track_interior_pointers(oop obj) { | |
| 2810 if (ValidateMarkSweep) { | |
| 2811 _adjusted_pointers->clear(); | |
| 2812 _pointer_tracking = true; | |
| 2813 | |
| 2814 AdjusterTracker checker; | |
| 2815 obj->oop_iterate(&checker); | |
| 2816 } | |
| 2817 } | |
| 2818 | |
| 2819 | |
| 2820 void PSParallelCompact::check_interior_pointers() { | |
| 2821 if (ValidateMarkSweep) { | |
| 2822 _pointer_tracking = false; | |
| 2823 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers"); | |
| 2824 } | |
| 2825 } | |
| 2826 | |
| 2827 | |
| 2828 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) { | |
| 2829 if (ValidateMarkSweep) { | |
| 2830 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops"); | |
| 2831 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0; | |
| 2832 } | |
| 2833 } | |
| 2834 | |
| 2835 | |
| 2836 void PSParallelCompact::register_live_oop(oop p, size_t size) { | |
| 2837 if (ValidateMarkSweep) { | |
| 2838 _live_oops->push(p); | |
| 2839 _live_oops_size->push(size); | |
| 2840 _live_oops_index++; | |
| 2841 } | |
| 2842 } | |
| 2843 | |
| 2844 void PSParallelCompact::validate_live_oop(oop p, size_t size) { | |
| 2845 if (ValidateMarkSweep) { | |
| 2846 oop obj = _live_oops->at((int)_live_oops_index); | |
| 2847 guarantee(obj == p, "should be the same object"); | |
| 2848 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size"); | |
| 2849 _live_oops_index++; | |
| 2850 } | |
| 2851 } | |
| 2852 | |
| 2853 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size, | |
| 2854 HeapWord* compaction_top) { | |
| 2855 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top), | |
| 2856 "should be moved to forwarded location"); | |
| 2857 if (ValidateMarkSweep) { | |
| 2858 PSParallelCompact::validate_live_oop(oop(q), size); | |
| 2859 _live_oops_moved_to->push(oop(compaction_top)); | |
| 2860 } | |
| 2861 if (RecordMarkSweepCompaction) { | |
| 2862 _cur_gc_live_oops->push(q); | |
| 2863 _cur_gc_live_oops_moved_to->push(compaction_top); | |
| 2864 _cur_gc_live_oops_size->push(size); | |
| 2865 } | |
| 2866 } | |
| 2867 | |
| 2868 | |
| 2869 void PSParallelCompact::compaction_complete() { | |
| 2870 if (RecordMarkSweepCompaction) { | |
| 2871 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops; | |
| 2872 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to; | |
| 2873 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size; | |
| 2874 | |
| 2875 _cur_gc_live_oops = _last_gc_live_oops; | |
| 2876 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to; | |
| 2877 _cur_gc_live_oops_size = _last_gc_live_oops_size; | |
| 2878 _last_gc_live_oops = _tmp_live_oops; | |
| 2879 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to; | |
| 2880 _last_gc_live_oops_size = _tmp_live_oops_size; | |
| 2881 } | |
| 2882 } | |
| 2883 | |
| 2884 | |
| 2885 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) { | |
| 2886 if (!RecordMarkSweepCompaction) { | |
| 2887 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled"); | |
| 2888 return; | |
| 2889 } | |
| 2890 | |
| 2891 if (_last_gc_live_oops == NULL) { | |
| 2892 tty->print_cr("No compaction information gathered yet"); | |
| 2893 return; | |
| 2894 } | |
| 2895 | |
| 2896 for (int i = 0; i < _last_gc_live_oops->length(); i++) { | |
| 2897 HeapWord* old_oop = _last_gc_live_oops->at(i); | |
| 2898 size_t sz = _last_gc_live_oops_size->at(i); | |
| 2899 if (old_oop <= q && q < (old_oop + sz)) { | |
| 2900 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i); | |
| 2901 size_t offset = (q - old_oop); | |
| 2902 tty->print_cr("Address " PTR_FORMAT, q); | |
| 2903 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset); | |
| 2904 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset); | |
| 2905 return; | |
| 2906 } | |
| 2907 } | |
| 2908 | |
| 2909 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q); | |
| 2910 } | |
| 2911 #endif //VALIDATE_MARK_SWEEP | |
| 2912 | |
| 2913 // Update interior oops in the ranges of chunks [beg_chunk, end_chunk). | |
| 2914 void | |
| 2915 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, | |
| 2916 SpaceId space_id, | |
| 2917 size_t beg_chunk, | |
| 2918 size_t end_chunk) { | |
| 2919 ParallelCompactData& sd = summary_data(); | |
| 2920 ParMarkBitMap* const mbm = mark_bitmap(); | |
| 2921 | |
| 2922 HeapWord* beg_addr = sd.chunk_to_addr(beg_chunk); | |
| 2923 HeapWord* const end_addr = sd.chunk_to_addr(end_chunk); | |
| 2924 assert(beg_chunk <= end_chunk, "bad chunk range"); | |
| 2925 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); | |
| 2926 | |
| 2927 #ifdef ASSERT | |
| 2928 // Claim the chunks to avoid triggering an assert when they are marked as | |
| 2929 // filled. | |
| 2930 for (size_t claim_chunk = beg_chunk; claim_chunk < end_chunk; ++claim_chunk) { | |
| 2931 assert(sd.chunk(claim_chunk)->claim_unsafe(), "claim() failed"); | |
| 2932 } | |
| 2933 #endif // #ifdef ASSERT | |
| 2934 | |
| 2935 if (beg_addr != space(space_id)->bottom()) { | |
| 2936 // Find the first live object or block of dead space that *starts* in this | |
| 2937 // range of chunks. If a partial object crosses onto the chunk, skip it; it | |
| 2938 // will be marked for 'deferred update' when the object head is processed. | |
| 2939 // If dead space crosses onto the chunk, it is also skipped; it will be | |
| 2940 // filled when the prior chunk is processed. If neither of those apply, the | |
| 2941 // first word in the chunk is the start of a live object or dead space. | |
| 2942 assert(beg_addr > space(space_id)->bottom(), "sanity"); | |
| 2943 const ChunkData* const cp = sd.chunk(beg_chunk); | |
| 2944 if (cp->partial_obj_size() != 0) { | |
| 2945 beg_addr = sd.partial_obj_end(beg_chunk); | |
| 2946 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { | |
| 2947 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); | |
| 2948 } | |
| 2949 } | |
| 2950 | |
| 2951 if (beg_addr < end_addr) { | |
| 2952 // A live object or block of dead space starts in this range of Chunks. | |
| 2953 HeapWord* const dense_prefix_end = dense_prefix(space_id); | |
| 2954 | |
| 2955 // Create closures and iterate. | |
| 2956 UpdateOnlyClosure update_closure(mbm, cm, space_id); | |
| 2957 FillClosure fill_closure(cm, space_id); | |
| 2958 ParMarkBitMap::IterationStatus status; | |
| 2959 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, | |
| 2960 dense_prefix_end); | |
| 2961 if (status == ParMarkBitMap::incomplete) { | |
| 2962 update_closure.do_addr(update_closure.source()); | |
| 2963 } | |
| 2964 } | |
| 2965 | |
| 2966 // Mark the chunks as filled. | |
| 2967 ChunkData* const beg_cp = sd.chunk(beg_chunk); | |
| 2968 ChunkData* const end_cp = sd.chunk(end_chunk); | |
| 2969 for (ChunkData* cp = beg_cp; cp < end_cp; ++cp) { | |
| 2970 cp->set_completed(); | |
| 2971 } | |
| 2972 } | |
| 2973 | |
| 2974 // Return the SpaceId for the space containing addr. If addr is not in the | |
| 2975 // heap, last_space_id is returned. In debug mode it expects the address to be | |
| 2976 // in the heap and asserts such. | |
| 2977 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { | |
| 2978 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); | |
| 2979 | |
| 2980 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { | |
| 2981 if (_space_info[id].space()->contains(addr)) { | |
| 2982 return SpaceId(id); | |
| 2983 } | |
| 2984 } | |
| 2985 | |
| 2986 assert(false, "no space contains the addr"); | |
| 2987 return last_space_id; | |
| 2988 } | |
| 2989 | |
| 2990 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, | |
| 2991 SpaceId id) { | |
| 2992 assert(id < last_space_id, "bad space id"); | |
| 2993 | |
| 2994 ParallelCompactData& sd = summary_data(); | |
| 2995 const SpaceInfo* const space_info = _space_info + id; | |
| 2996 ObjectStartArray* const start_array = space_info->start_array(); | |
| 2997 | |
| 2998 const MutableSpace* const space = space_info->space(); | |
| 2999 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); | |
| 3000 HeapWord* const beg_addr = space_info->dense_prefix(); | |
| 3001 HeapWord* const end_addr = sd.chunk_align_up(space_info->new_top()); | |
| 3002 | |
| 3003 const ChunkData* const beg_chunk = sd.addr_to_chunk_ptr(beg_addr); | |
| 3004 const ChunkData* const end_chunk = sd.addr_to_chunk_ptr(end_addr); | |
| 3005 const ChunkData* cur_chunk; | |
| 3006 for (cur_chunk = beg_chunk; cur_chunk < end_chunk; ++cur_chunk) { | |
| 3007 HeapWord* const addr = cur_chunk->deferred_obj_addr(); | |
| 3008 if (addr != NULL) { | |
| 3009 if (start_array != NULL) { | |
| 3010 start_array->allocate_block(addr); | |
| 3011 } | |
| 3012 oop(addr)->update_contents(cm); | |
| 3013 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); | |
| 3014 } | |
| 3015 } | |
| 3016 } | |
| 3017 | |
| 3018 // Skip over count live words starting from beg, and return the address of the | |
| 3019 // next live word. Unless marked, the word corresponding to beg is assumed to | |
| 3020 // be dead. Callers must either ensure beg does not correspond to the middle of | |
| 3021 // an object, or account for those live words in some other way. Callers must | |
| 3022 // also ensure that there are enough live words in the range [beg, end) to skip. | |
| 3023 HeapWord* | |
| 3024 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) | |
| 3025 { | |
| 3026 assert(count > 0, "sanity"); | |
| 3027 | |
| 3028 ParMarkBitMap* m = mark_bitmap(); | |
| 3029 idx_t bits_to_skip = m->words_to_bits(count); | |
| 3030 idx_t cur_beg = m->addr_to_bit(beg); | |
| 3031 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); | |
| 3032 | |
| 3033 do { | |
| 3034 cur_beg = m->find_obj_beg(cur_beg, search_end); | |
| 3035 idx_t cur_end = m->find_obj_end(cur_beg, search_end); | |
| 3036 const size_t obj_bits = cur_end - cur_beg + 1; | |
| 3037 if (obj_bits > bits_to_skip) { | |
| 3038 return m->bit_to_addr(cur_beg + bits_to_skip); | |
| 3039 } | |
| 3040 bits_to_skip -= obj_bits; | |
| 3041 cur_beg = cur_end + 1; | |
| 3042 } while (bits_to_skip > 0); | |
| 3043 | |
| 3044 // Skipping the desired number of words landed just past the end of an object. | |
| 3045 // Find the start of the next object. | |
| 3046 cur_beg = m->find_obj_beg(cur_beg, search_end); | |
| 3047 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); | |
| 3048 return m->bit_to_addr(cur_beg); | |
| 3049 } | |
| 3050 | |
| 3051 HeapWord* | |
| 3052 PSParallelCompact::first_src_addr(HeapWord* const dest_addr, | |
| 3053 size_t src_chunk_idx) | |
| 3054 { | |
| 3055 ParMarkBitMap* const bitmap = mark_bitmap(); | |
| 3056 const ParallelCompactData& sd = summary_data(); | |
| 3057 const size_t ChunkSize = ParallelCompactData::ChunkSize; | |
| 3058 | |
| 3059 assert(sd.is_chunk_aligned(dest_addr), "not aligned"); | |
| 3060 | |
| 3061 const ChunkData* const src_chunk_ptr = sd.chunk(src_chunk_idx); | |
| 3062 const size_t partial_obj_size = src_chunk_ptr->partial_obj_size(); | |
| 3063 HeapWord* const src_chunk_destination = src_chunk_ptr->destination(); | |
| 3064 | |
| 3065 assert(dest_addr >= src_chunk_destination, "wrong src chunk"); | |
| 3066 assert(src_chunk_ptr->data_size() > 0, "src chunk cannot be empty"); | |
| 3067 | |
| 3068 HeapWord* const src_chunk_beg = sd.chunk_to_addr(src_chunk_idx); | |
| 3069 HeapWord* const src_chunk_end = src_chunk_beg + ChunkSize; | |
| 3070 | |
| 3071 HeapWord* addr = src_chunk_beg; | |
| 3072 if (dest_addr == src_chunk_destination) { | |
| 3073 // Return the first live word in the source chunk. | |
| 3074 if (partial_obj_size == 0) { | |
| 3075 addr = bitmap->find_obj_beg(addr, src_chunk_end); | |
| 3076 assert(addr < src_chunk_end, "no objects start in src chunk"); | |
| 3077 } | |
| 3078 return addr; | |
| 3079 } | |
| 3080 | |
| 3081 // Must skip some live data. | |
| 3082 size_t words_to_skip = dest_addr - src_chunk_destination; | |
| 3083 assert(src_chunk_ptr->data_size() > words_to_skip, "wrong src chunk"); | |
| 3084 | |
| 3085 if (partial_obj_size >= words_to_skip) { | |
| 3086 // All the live words to skip are part of the partial object. | |
| 3087 addr += words_to_skip; | |
| 3088 if (partial_obj_size == words_to_skip) { | |
| 3089 // Find the first live word past the partial object. | |
| 3090 addr = bitmap->find_obj_beg(addr, src_chunk_end); | |
| 3091 assert(addr < src_chunk_end, "wrong src chunk"); | |
| 3092 } | |
| 3093 return addr; | |
| 3094 } | |
| 3095 | |
| 3096 // Skip over the partial object (if any). | |
| 3097 if (partial_obj_size != 0) { | |
| 3098 words_to_skip -= partial_obj_size; | |
| 3099 addr += partial_obj_size; | |
| 3100 } | |
| 3101 | |
| 3102 // Skip over live words due to objects that start in the chunk. | |
| 3103 addr = skip_live_words(addr, src_chunk_end, words_to_skip); | |
| 3104 assert(addr < src_chunk_end, "wrong src chunk"); | |
| 3105 return addr; | |
| 3106 } | |
| 3107 | |
| 3108 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, | |
| 3109 size_t beg_chunk, | |
| 3110 HeapWord* end_addr) | |
| 3111 { | |
| 3112 ParallelCompactData& sd = summary_data(); | |
| 3113 ChunkData* const beg = sd.chunk(beg_chunk); | |
| 3114 HeapWord* const end_addr_aligned_up = sd.chunk_align_up(end_addr); | |
| 3115 ChunkData* const end = sd.addr_to_chunk_ptr(end_addr_aligned_up); | |
| 3116 size_t cur_idx = beg_chunk; | |
| 3117 for (ChunkData* cur = beg; cur < end; ++cur, ++cur_idx) { | |
| 3118 assert(cur->data_size() > 0, "chunk must have live data"); | |
| 3119 cur->decrement_destination_count(); | |
| 3120 if (cur_idx <= cur->source_chunk() && cur->available() && cur->claim()) { | |
| 3121 cm->save_for_processing(cur_idx); | |
| 3122 } | |
| 3123 } | |
| 3124 } | |
| 3125 | |
| 3126 size_t PSParallelCompact::next_src_chunk(MoveAndUpdateClosure& closure, | |
| 3127 SpaceId& src_space_id, | |
| 3128 HeapWord*& src_space_top, | |
| 3129 HeapWord* end_addr) | |
| 3130 { | |
| 3131 typedef ParallelCompactData::ChunkData ChunkData; | |
| 3132 | |
| 3133 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 3134 const size_t chunk_size = ParallelCompactData::ChunkSize; | |
| 3135 | |
| 3136 size_t src_chunk_idx = 0; | |
| 3137 | |
| 3138 // Skip empty chunks (if any) up to the top of the space. | |
| 3139 HeapWord* const src_aligned_up = sd.chunk_align_up(end_addr); | |
| 3140 ChunkData* src_chunk_ptr = sd.addr_to_chunk_ptr(src_aligned_up); | |
| 3141 HeapWord* const top_aligned_up = sd.chunk_align_up(src_space_top); | |
| 3142 const ChunkData* const top_chunk_ptr = sd.addr_to_chunk_ptr(top_aligned_up); | |
| 3143 while (src_chunk_ptr < top_chunk_ptr && src_chunk_ptr->data_size() == 0) { | |
| 3144 ++src_chunk_ptr; | |
| 3145 } | |
| 3146 | |
| 3147 if (src_chunk_ptr < top_chunk_ptr) { | |
| 3148 // The next source chunk is in the current space. Update src_chunk_idx and | |
| 3149 // the source address to match src_chunk_ptr. | |
| 3150 src_chunk_idx = sd.chunk(src_chunk_ptr); | |
| 3151 HeapWord* const src_chunk_addr = sd.chunk_to_addr(src_chunk_idx); | |
| 3152 if (src_chunk_addr > closure.source()) { | |
| 3153 closure.set_source(src_chunk_addr); | |
| 3154 } | |
| 3155 return src_chunk_idx; | |
| 3156 } | |
| 3157 | |
| 3158 // Switch to a new source space and find the first non-empty chunk. | |
| 3159 unsigned int space_id = src_space_id + 1; | |
| 3160 assert(space_id < last_space_id, "not enough spaces"); | |
| 3161 | |
| 3162 HeapWord* const destination = closure.destination(); | |
| 3163 | |
| 3164 do { | |
| 3165 MutableSpace* space = _space_info[space_id].space(); | |
| 3166 HeapWord* const bottom = space->bottom(); | |
| 3167 const ChunkData* const bottom_cp = sd.addr_to_chunk_ptr(bottom); | |
| 3168 | |
| 3169 // Iterate over the spaces that do not compact into themselves. | |
| 3170 if (bottom_cp->destination() != bottom) { | |
| 3171 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top()); | |
| 3172 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up); | |
| 3173 | |
| 3174 for (const ChunkData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { | |
| 3175 if (src_cp->live_obj_size() > 0) { | |
| 3176 // Found it. | |
| 3177 assert(src_cp->destination() == destination, | |
| 3178 "first live obj in the space must match the destination"); | |
| 3179 assert(src_cp->partial_obj_size() == 0, | |
| 3180 "a space cannot begin with a partial obj"); | |
| 3181 | |
| 3182 src_space_id = SpaceId(space_id); | |
| 3183 src_space_top = space->top(); | |
| 3184 const size_t src_chunk_idx = sd.chunk(src_cp); | |
| 3185 closure.set_source(sd.chunk_to_addr(src_chunk_idx)); | |
| 3186 return src_chunk_idx; | |
| 3187 } else { | |
| 3188 assert(src_cp->data_size() == 0, "sanity"); | |
| 3189 } | |
| 3190 } | |
| 3191 } | |
| 3192 } while (++space_id < last_space_id); | |
| 3193 | |
| 3194 assert(false, "no source chunk was found"); | |
| 3195 return 0; | |
| 3196 } | |
| 3197 | |
| 3198 void PSParallelCompact::fill_chunk(ParCompactionManager* cm, size_t chunk_idx) | |
| 3199 { | |
| 3200 typedef ParMarkBitMap::IterationStatus IterationStatus; | |
| 3201 const size_t ChunkSize = ParallelCompactData::ChunkSize; | |
| 3202 ParMarkBitMap* const bitmap = mark_bitmap(); | |
| 3203 ParallelCompactData& sd = summary_data(); | |
| 3204 ChunkData* const chunk_ptr = sd.chunk(chunk_idx); | |
| 3205 | |
| 3206 // Get the items needed to construct the closure. | |
| 3207 HeapWord* dest_addr = sd.chunk_to_addr(chunk_idx); | |
| 3208 SpaceId dest_space_id = space_id(dest_addr); | |
| 3209 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); | |
| 3210 HeapWord* new_top = _space_info[dest_space_id].new_top(); | |
| 3211 assert(dest_addr < new_top, "sanity"); | |
| 3212 const size_t words = MIN2(pointer_delta(new_top, dest_addr), ChunkSize); | |
| 3213 | |
| 3214 // Get the source chunk and related info. | |
| 3215 size_t src_chunk_idx = chunk_ptr->source_chunk(); | |
| 3216 SpaceId src_space_id = space_id(sd.chunk_to_addr(src_chunk_idx)); | |
| 3217 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); | |
| 3218 | |
| 3219 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); | |
| 3220 closure.set_source(first_src_addr(dest_addr, src_chunk_idx)); | |
| 3221 | |
| 3222 // Adjust src_chunk_idx to prepare for decrementing destination counts (the | |
| 3223 // destination count is not decremented when a chunk is copied to itself). | |
| 3224 if (src_chunk_idx == chunk_idx) { | |
| 3225 src_chunk_idx += 1; | |
| 3226 } | |
| 3227 | |
| 3228 if (bitmap->is_unmarked(closure.source())) { | |
| 3229 // The first source word is in the middle of an object; copy the remainder | |
| 3230 // of the object or as much as will fit. The fact that pointer updates were | |
| 3231 // deferred will be noted when the object header is processed. | |
| 3232 HeapWord* const old_src_addr = closure.source(); | |
| 3233 closure.copy_partial_obj(); | |
| 3234 if (closure.is_full()) { | |
| 3235 decrement_destination_counts(cm, src_chunk_idx, closure.source()); | |
| 3236 chunk_ptr->set_deferred_obj_addr(NULL); | |
| 3237 chunk_ptr->set_completed(); | |
| 3238 return; | |
| 3239 } | |
| 3240 | |
| 3241 HeapWord* const end_addr = sd.chunk_align_down(closure.source()); | |
| 3242 if (sd.chunk_align_down(old_src_addr) != end_addr) { | |
| 3243 // The partial object was copied from more than one source chunk. | |
| 3244 decrement_destination_counts(cm, src_chunk_idx, end_addr); | |
| 3245 | |
| 3246 // Move to the next source chunk, possibly switching spaces as well. All | |
| 3247 // args except end_addr may be modified. | |
| 3248 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top, | |
| 3249 end_addr); | |
| 3250 } | |
| 3251 } | |
| 3252 | |
| 3253 do { | |
| 3254 HeapWord* const cur_addr = closure.source(); | |
| 3255 HeapWord* const end_addr = MIN2(sd.chunk_align_up(cur_addr + 1), | |
| 3256 src_space_top); | |
| 3257 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); | |
| 3258 | |
| 3259 if (status == ParMarkBitMap::incomplete) { | |
| 3260 // The last obj that starts in the source chunk does not end in the chunk. | |
| 3261 assert(closure.source() < end_addr, "sanity") | |
| 3262 HeapWord* const obj_beg = closure.source(); | |
| 3263 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), | |
| 3264 src_space_top); | |
| 3265 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); | |
| 3266 if (obj_end < range_end) { | |
| 3267 // The end was found; the entire object will fit. | |
| 3268 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); | |
| 3269 assert(status != ParMarkBitMap::would_overflow, "sanity"); | |
| 3270 } else { | |
| 3271 // The end was not found; the object will not fit. | |
| 3272 assert(range_end < src_space_top, "obj cannot cross space boundary"); | |
| 3273 status = ParMarkBitMap::would_overflow; | |
| 3274 } | |
| 3275 } | |
| 3276 | |
| 3277 if (status == ParMarkBitMap::would_overflow) { | |
| 3278 // The last object did not fit. Note that interior oop updates were | |
| 3279 // deferred, then copy enough of the object to fill the chunk. | |
| 3280 chunk_ptr->set_deferred_obj_addr(closure.destination()); | |
| 3281 status = closure.copy_until_full(); // copies from closure.source() | |
| 3282 | |
| 3283 decrement_destination_counts(cm, src_chunk_idx, closure.source()); | |
| 3284 chunk_ptr->set_completed(); | |
| 3285 return; | |
| 3286 } | |
| 3287 | |
| 3288 if (status == ParMarkBitMap::full) { | |
| 3289 decrement_destination_counts(cm, src_chunk_idx, closure.source()); | |
| 3290 chunk_ptr->set_deferred_obj_addr(NULL); | |
| 3291 chunk_ptr->set_completed(); | |
| 3292 return; | |
| 3293 } | |
| 3294 | |
| 3295 decrement_destination_counts(cm, src_chunk_idx, end_addr); | |
| 3296 | |
| 3297 // Move to the next source chunk, possibly switching spaces as well. All | |
| 3298 // args except end_addr may be modified. | |
| 3299 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top, | |
| 3300 end_addr); | |
| 3301 } while (true); | |
| 3302 } | |
| 3303 | |
| 3304 void | |
| 3305 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { | |
| 3306 const MutableSpace* sp = space(space_id); | |
| 3307 if (sp->is_empty()) { | |
| 3308 return; | |
| 3309 } | |
| 3310 | |
| 3311 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 3312 ParMarkBitMap* const bitmap = mark_bitmap(); | |
| 3313 HeapWord* const dp_addr = dense_prefix(space_id); | |
| 3314 HeapWord* beg_addr = sp->bottom(); | |
| 3315 HeapWord* end_addr = sp->top(); | |
| 3316 | |
| 3317 #ifdef ASSERT | |
| 3318 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); | |
| 3319 if (cm->should_verify_only()) { | |
| 3320 VerifyUpdateClosure verify_update(cm, sp); | |
| 3321 bitmap->iterate(&verify_update, beg_addr, end_addr); | |
| 3322 return; | |
| 3323 } | |
| 3324 | |
| 3325 if (cm->should_reset_only()) { | |
| 3326 ResetObjectsClosure reset_objects(cm); | |
| 3327 bitmap->iterate(&reset_objects, beg_addr, end_addr); | |
| 3328 return; | |
| 3329 } | |
| 3330 #endif | |
| 3331 | |
| 3332 const size_t beg_chunk = sd.addr_to_chunk_idx(beg_addr); | |
| 3333 const size_t dp_chunk = sd.addr_to_chunk_idx(dp_addr); | |
| 3334 if (beg_chunk < dp_chunk) { | |
| 3335 update_and_deadwood_in_dense_prefix(cm, space_id, beg_chunk, dp_chunk); | |
| 3336 } | |
| 3337 | |
| 3338 // The destination of the first live object that starts in the chunk is one | |
| 3339 // past the end of the partial object entering the chunk (if any). | |
| 3340 HeapWord* const dest_addr = sd.partial_obj_end(dp_chunk); | |
| 3341 HeapWord* const new_top = _space_info[space_id].new_top(); | |
| 3342 assert(new_top >= dest_addr, "bad new_top value"); | |
| 3343 const size_t words = pointer_delta(new_top, dest_addr); | |
| 3344 | |
| 3345 if (words > 0) { | |
| 3346 ObjectStartArray* start_array = _space_info[space_id].start_array(); | |
| 3347 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); | |
| 3348 | |
| 3349 ParMarkBitMap::IterationStatus status; | |
| 3350 status = bitmap->iterate(&closure, dest_addr, end_addr); | |
| 3351 assert(status == ParMarkBitMap::full, "iteration not complete"); | |
| 3352 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, | |
| 3353 "live objects skipped because closure is full"); | |
| 3354 } | |
| 3355 } | |
| 3356 | |
| 3357 jlong PSParallelCompact::millis_since_last_gc() { | |
| 3358 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc; | |
| 3359 // XXX See note in genCollectedHeap::millis_since_last_gc(). | |
| 3360 if (ret_val < 0) { | |
| 3361 NOT_PRODUCT(warning("time warp: %d", ret_val);) | |
| 3362 return 0; | |
| 3363 } | |
| 3364 return ret_val; | |
| 3365 } | |
| 3366 | |
| 3367 void PSParallelCompact::reset_millis_since_last_gc() { | |
| 3368 _time_of_last_gc = os::javaTimeMillis(); | |
| 3369 } | |
| 3370 | |
| 3371 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() | |
| 3372 { | |
| 3373 if (source() != destination()) { | |
| 3374 assert(source() > destination(), "must copy to the left"); | |
| 3375 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); | |
| 3376 } | |
| 3377 update_state(words_remaining()); | |
| 3378 assert(is_full(), "sanity"); | |
| 3379 return ParMarkBitMap::full; | |
| 3380 } | |
| 3381 | |
| 3382 void MoveAndUpdateClosure::copy_partial_obj() | |
| 3383 { | |
| 3384 size_t words = words_remaining(); | |
| 3385 | |
| 3386 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); | |
| 3387 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); | |
| 3388 if (end_addr < range_end) { | |
| 3389 words = bitmap()->obj_size(source(), end_addr); | |
| 3390 } | |
| 3391 | |
| 3392 // This test is necessary; if omitted, the pointer updates to a partial object | |
| 3393 // that crosses the dense prefix boundary could be overwritten. | |
| 3394 if (source() != destination()) { | |
| 3395 assert(source() > destination(), "must copy to the left"); | |
| 3396 Copy::aligned_conjoint_words(source(), destination(), words); | |
| 3397 } | |
| 3398 update_state(words); | |
| 3399 } | |
| 3400 | |
| 3401 ParMarkBitMapClosure::IterationStatus | |
| 3402 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { | |
| 3403 assert(destination() != NULL, "sanity"); | |
| 3404 assert(bitmap()->obj_size(addr) == words, "bad size"); | |
| 3405 | |
| 3406 _source = addr; | |
| 3407 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == | |
| 3408 destination(), "wrong destination"); | |
| 3409 | |
| 3410 if (words > words_remaining()) { | |
| 3411 return ParMarkBitMap::would_overflow; | |
| 3412 } | |
| 3413 | |
| 3414 // The start_array must be updated even if the object is not moving. | |
| 3415 if (_start_array != NULL) { | |
| 3416 _start_array->allocate_block(destination()); | |
| 3417 } | |
| 3418 | |
| 3419 if (destination() != source()) { | |
| 3420 assert(destination() < source(), "must copy to the left"); | |
| 3421 Copy::aligned_conjoint_words(source(), destination(), words); | |
| 3422 } | |
| 3423 | |
| 3424 oop moved_oop = (oop) destination(); | |
| 3425 moved_oop->update_contents(compaction_manager()); | |
| 3426 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); | |
| 3427 | |
| 3428 update_state(words); | |
| 3429 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); | |
| 3430 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; | |
| 3431 } | |
| 3432 | |
| 3433 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, | |
| 3434 ParCompactionManager* cm, | |
| 3435 PSParallelCompact::SpaceId space_id) : | |
| 3436 ParMarkBitMapClosure(mbm, cm), | |
| 3437 _space_id(space_id), | |
| 3438 _start_array(PSParallelCompact::start_array(space_id)) | |
| 3439 { | |
| 3440 } | |
| 3441 | |
| 3442 // Updates the references in the object to their new values. | |
| 3443 ParMarkBitMapClosure::IterationStatus | |
| 3444 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { | |
| 3445 do_addr(addr); | |
| 3446 return ParMarkBitMap::incomplete; | |
| 3447 } | |
| 3448 | |
| 3449 BitBlockUpdateClosure::BitBlockUpdateClosure(ParMarkBitMap* mbm, | |
| 3450 ParCompactionManager* cm, | |
| 3451 size_t chunk_index) : | |
| 3452 ParMarkBitMapClosure(mbm, cm), | |
| 3453 _live_data_left(0), | |
| 3454 _cur_block(0) { | |
| 3455 _chunk_start = | |
| 3456 PSParallelCompact::summary_data().chunk_to_addr(chunk_index); | |
| 3457 _chunk_end = | |
| 3458 PSParallelCompact::summary_data().chunk_to_addr(chunk_index) + | |
| 3459 ParallelCompactData::ChunkSize; | |
| 3460 _chunk_index = chunk_index; | |
| 3461 _cur_block = | |
| 3462 PSParallelCompact::summary_data().addr_to_block_idx(_chunk_start); | |
| 3463 } | |
| 3464 | |
| 3465 bool BitBlockUpdateClosure::chunk_contains_cur_block() { | |
| 3466 return ParallelCompactData::chunk_contains_block(_chunk_index, _cur_block); | |
| 3467 } | |
| 3468 | |
| 3469 void BitBlockUpdateClosure::reset_chunk(size_t chunk_index) { | |
| 3470 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(7);) | |
| 3471 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 3472 _chunk_index = chunk_index; | |
| 3473 _live_data_left = 0; | |
| 3474 _chunk_start = sd.chunk_to_addr(chunk_index); | |
| 3475 _chunk_end = sd.chunk_to_addr(chunk_index) + ParallelCompactData::ChunkSize; | |
| 3476 | |
| 3477 // The first block in this chunk | |
| 3478 size_t first_block = sd.addr_to_block_idx(_chunk_start); | |
| 3479 size_t partial_live_size = sd.chunk(chunk_index)->partial_obj_size(); | |
| 3480 | |
| 3481 // Set the offset to 0. By definition it should have that value | |
| 3482 // but it may have been written while processing an earlier chunk. | |
| 3483 if (partial_live_size == 0) { | |
| 3484 // No live object extends onto the chunk. The first bit | |
| 3485 // in the bit map for the first chunk must be a start bit. | |
| 3486 // Although there may not be any marked bits, it is safe | |
| 3487 // to set it as a start bit. | |
| 3488 sd.block(first_block)->set_start_bit_offset(0); | |
| 3489 sd.block(first_block)->set_first_is_start_bit(true); | |
| 3490 } else if (sd.partial_obj_ends_in_block(first_block)) { | |
| 3491 sd.block(first_block)->set_end_bit_offset(0); | |
| 3492 sd.block(first_block)->set_first_is_start_bit(false); | |
| 3493 } else { | |
| 3494 // The partial object extends beyond the first block. | |
| 3495 // There is no object starting in the first block | |
| 3496 // so the offset and bit parity are not needed. | |
| 3497 // Set the the bit parity to start bit so assertions | |
| 3498 // work when not bit is found. | |
| 3499 sd.block(first_block)->set_end_bit_offset(0); | |
| 3500 sd.block(first_block)->set_first_is_start_bit(false); | |
| 3501 } | |
| 3502 _cur_block = first_block; | |
| 3503 #ifdef ASSERT | |
| 3504 if (sd.block(first_block)->first_is_start_bit()) { | |
| 3505 assert(!sd.partial_obj_ends_in_block(first_block), | |
| 3506 "Partial object cannot end in first block"); | |
| 3507 } | |
| 3508 | |
| 3509 if (PrintGCDetails && Verbose) { | |
| 3510 if (partial_live_size == 1) { | |
| 3511 gclog_or_tty->print_cr("first_block " PTR_FORMAT | |
| 3512 " _offset " PTR_FORMAT | |
| 3513 " _first_is_start_bit %d", | |
| 3514 first_block, | |
| 3515 sd.block(first_block)->raw_offset(), | |
| 3516 sd.block(first_block)->first_is_start_bit()); | |
| 3517 } | |
| 3518 } | |
| 3519 #endif | |
| 3520 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(17);) | |
| 3521 } | |
| 3522 | |
| 3523 // This method is called when a object has been found (both beginning | |
| 3524 // and end of the object) in the range of iteration. This method is | |
| 3525 // calculating the words of live data to the left of a block. That live | |
| 3526 // data includes any object starting to the left of the block (i.e., | |
| 3527 // the live-data-to-the-left of block AAA will include the full size | |
| 3528 // of any object entering AAA). | |
| 3529 | |
| 3530 ParMarkBitMapClosure::IterationStatus | |
| 3531 BitBlockUpdateClosure::do_addr(HeapWord* addr, size_t words) { | |
| 3532 // add the size to the block data. | |
| 3533 HeapWord* obj = addr; | |
| 3534 ParallelCompactData& sd = PSParallelCompact::summary_data(); | |
| 3535 | |
| 3536 assert(bitmap()->obj_size(obj) == words, "bad size"); | |
| 3537 assert(_chunk_start <= obj, "object is not in chunk"); | |
| 3538 assert(obj + words <= _chunk_end, "object is not in chunk"); | |
| 3539 | |
| 3540 // Update the live data to the left | |
| 3541 size_t prev_live_data_left = _live_data_left; | |
| 3542 _live_data_left = _live_data_left + words; | |
| 3543 | |
| 3544 // Is this object in the current block. | |
| 3545 size_t block_of_obj = sd.addr_to_block_idx(obj); | |
| 3546 size_t block_of_obj_last = sd.addr_to_block_idx(obj + words - 1); | |
| 3547 HeapWord* block_of_obj_last_addr = sd.block_to_addr(block_of_obj_last); | |
| 3548 if (_cur_block < block_of_obj) { | |
| 3549 | |
| 3550 // | |
| 3551 // No object crossed the block boundary and this object was found | |
| 3552 // on the other side of the block boundary. Update the offset for | |
| 3553 // the new block with the data size that does not include this object. | |
| 3554 // | |
| 3555 // The first bit in block_of_obj is a start bit except in the | |
| 3556 // case where the partial object for the chunk extends into | |
| 3557 // this block. | |
| 3558 if (sd.partial_obj_ends_in_block(block_of_obj)) { | |
| 3559 sd.block(block_of_obj)->set_end_bit_offset(prev_live_data_left); | |
| 3560 } else { | |
| 3561 sd.block(block_of_obj)->set_start_bit_offset(prev_live_data_left); | |
| 3562 } | |
| 3563 | |
| 3564 // Does this object pass beyond the its block? | |
| 3565 if (block_of_obj < block_of_obj_last) { | |
| 3566 // Object crosses block boundary. Two blocks need to be udpated: | |
| 3567 // the current block where the object started | |
| 3568 // the block where the object ends | |
| 3569 // | |
| 3570 // The offset for blocks with no objects starting in them | |
| 3571 // (e.g., blocks between _cur_block and block_of_obj_last) | |
| 3572 // should not be needed. | |
| 3573 // Note that block_of_obj_last may be in another chunk. If so, | |
| 3574 // it should be overwritten later. This is a problem (writting | |
| 3575 // into a block in a later chunk) for parallel execution. | |
| 3576 assert(obj < block_of_obj_last_addr, | |
| 3577 "Object should start in previous block"); | |
| 3578 | |
| 3579 // obj is crossing into block_of_obj_last so the first bit | |
| 3580 // is and end bit. | |
| 3581 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left); | |
| 3582 | |
| 3583 _cur_block = block_of_obj_last; | |
| 3584 } else { | |
| 3585 // _first_is_start_bit has already been set correctly | |
| 3586 // in the if-then-else above so don't reset it here. | |
| 3587 _cur_block = block_of_obj; | |
| 3588 } | |
| 3589 } else { | |
| 3590 // The current block only changes if the object extends beyound | |
| 3591 // the block it starts in. | |
| 3592 // | |
| 3593 // The object starts in the current block. | |
| 3594 // Does this object pass beyond the end of it? | |
| 3595 if (block_of_obj < block_of_obj_last) { | |
| 3596 // Object crosses block boundary. | |
| 3597 // See note above on possible blocks between block_of_obj and | |
| 3598 // block_of_obj_last | |
| 3599 assert(obj < block_of_obj_last_addr, | |
| 3600 "Object should start in previous block"); | |
| 3601 | |
| 3602 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left); | |
| 3603 | |
| 3604 _cur_block = block_of_obj_last; | |
| 3605 } | |
| 3606 } | |
| 3607 | |
| 3608 // Return incomplete if there are more blocks to be done. | |
| 3609 if (chunk_contains_cur_block()) { | |
| 3610 return ParMarkBitMap::incomplete; | |
| 3611 } | |
| 3612 return ParMarkBitMap::complete; | |
| 3613 } | |
| 3614 | |
| 3615 // Verify the new location using the forwarding pointer | |
| 3616 // from MarkSweep::mark_sweep_phase2(). Set the mark_word | |
| 3617 // to the initial value. | |
| 3618 ParMarkBitMapClosure::IterationStatus | |
| 3619 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) { | |
| 3620 // The second arg (words) is not used. | |
| 3621 oop obj = (oop) addr; | |
| 3622 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer(); | |
| 3623 HeapWord* new_pointer = summary_data().calc_new_pointer(obj); | |
| 3624 if (forwarding_ptr == NULL) { | |
| 3625 // The object is dead or not moving. | |
| 3626 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj), | |
| 3627 "Object liveness is wrong."); | |
| 3628 return ParMarkBitMap::incomplete; | |
| 3629 } | |
| 3630 assert(UseParallelOldGCDensePrefix || | |
| 3631 (HeapMaximumCompactionInterval > 1) || | |
| 3632 (MarkSweepAlwaysCompactCount > 1) || | |
| 3633 (forwarding_ptr == new_pointer), | |
| 3634 "Calculation of new location is incorrect"); | |
| 3635 return ParMarkBitMap::incomplete; | |
| 3636 } | |
| 3637 | |
| 3638 // Reset objects modified for debug checking. | |
| 3639 ParMarkBitMapClosure::IterationStatus | |
| 3640 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) { | |
| 3641 // The second arg (words) is not used. | |
| 3642 oop obj = (oop) addr; | |
| 3643 obj->init_mark(); | |
| 3644 return ParMarkBitMap::incomplete; | |
| 3645 } | |
| 3646 | |
| 3647 // Prepare for compaction. This method is executed once | |
| 3648 // (i.e., by a single thread) before compaction. | |
| 3649 // Save the updated location of the intArrayKlassObj for | |
| 3650 // filling holes in the dense prefix. | |
| 3651 void PSParallelCompact::compact_prologue() { | |
| 3652 _updated_int_array_klass_obj = (klassOop) | |
| 3653 summary_data().calc_new_pointer(Universe::intArrayKlassObj()); | |
| 3654 } | |
| 3655 | |
| 3656 // The initial implementation of this method created a field | |
| 3657 // _next_compaction_space_id in SpaceInfo and initialized | |
| 3658 // that field in SpaceInfo::initialize_space_info(). That | |
| 3659 // required that _next_compaction_space_id be declared a | |
| 3660 // SpaceId in SpaceInfo and that would have required that | |
| 3661 // either SpaceId be declared in a separate class or that | |
| 3662 // it be declared in SpaceInfo. It didn't seem consistent | |
| 3663 // to declare it in SpaceInfo (didn't really fit logically). | |
| 3664 // Alternatively, defining a separate class to define SpaceId | |
| 3665 // seem excessive. This implementation is simple and localizes | |
| 3666 // the knowledge. | |
| 3667 | |
| 3668 PSParallelCompact::SpaceId | |
| 3669 PSParallelCompact::next_compaction_space_id(SpaceId id) { | |
| 3670 assert(id < last_space_id, "id out of range"); | |
| 3671 switch (id) { | |
| 3672 case perm_space_id : | |
| 3673 return last_space_id; | |
| 3674 case old_space_id : | |
| 3675 return eden_space_id; | |
| 3676 case eden_space_id : | |
| 3677 return from_space_id; | |
| 3678 case from_space_id : | |
| 3679 return to_space_id; | |
| 3680 case to_space_id : | |
| 3681 return last_space_id; | |
| 3682 default: | |
| 3683 assert(false, "Bad space id"); | |
| 3684 return last_space_id; | |
| 3685 } | |
| 3686 } | |
| 3687 | |
| 3688 // Here temporarily for debugging | |
| 3689 #ifdef ASSERT | |
| 3690 size_t ParallelCompactData::block_idx(BlockData* block) { | |
| 3691 size_t index = pointer_delta(block, | |
| 3692 PSParallelCompact::summary_data()._block_data, sizeof(BlockData)); | |
| 3693 return index; | |
| 3694 } | |
| 3695 #endif |