1 /* 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc_implementation/shared/parGCAllocBuffer.hpp" 27 #include "memory/sharedHeap.hpp" 28 #include "oops/arrayOop.hpp" 29 #include "oops/oop.inline.hpp" 30 31 ParGCAllocBuffer::ParGCAllocBuffer(size_t desired_plab_sz_) : 32 _word_sz(desired_plab_sz_), _bottom(NULL), _top(NULL), 33 _end(NULL), _hard_end(NULL), 34 _retained(false), _retained_filler(), 35 _allocated(0), _wasted(0) 36 { 37 assert (min_size() > AlignmentReserve, "Inconsistency!"); 38 // arrayOopDesc::header_size depends on command line initialization. 39 FillerHeaderSize = align_object_size(arrayOopDesc::header_size(T_INT)); 40 AlignmentReserve = oopDesc::header_size() > MinObjAlignment ? FillerHeaderSize : 0; 41 } 42 43 size_t ParGCAllocBuffer::FillerHeaderSize; 44 45 // If the minimum object size is greater than MinObjAlignment, we can 46 // end up with a shard at the end of the buffer that's smaller than 47 // the smallest object. We can't allow that because the buffer must 48 // look like it's full of objects when we retire it, so we make 49 // sure we have enough space for a filler int array object. 50 size_t ParGCAllocBuffer::AlignmentReserve; 51 52 void ParGCAllocBuffer::retire(bool end_of_gc, bool retain) { 53 assert(!retain || end_of_gc, "Can only retain at GC end."); 54 if (_retained) { 55 // If the buffer had been retained shorten the previous filler object. 56 assert(_retained_filler.end() <= _top, "INVARIANT"); 57 CollectedHeap::fill_with_object(_retained_filler); 58 // Wasted space book-keeping, otherwise (normally) done in invalidate() 59 _wasted += _retained_filler.word_size(); 60 _retained = false; 61 } 62 assert(!end_of_gc || !_retained, "At this point, end_of_gc ==> !_retained."); 63 if (_top < _hard_end) { 64 CollectedHeap::fill_with_object(_top, _hard_end); 65 if (!retain) { 66 invalidate(); 67 } else { 68 // Is there wasted space we'd like to retain for the next GC? 69 if (pointer_delta(_end, _top) > FillerHeaderSize) { 70 _retained = true; 71 _retained_filler = MemRegion(_top, FillerHeaderSize); 72 _top = _top + FillerHeaderSize; 73 } else { 74 invalidate(); 75 } 76 } 77 } 78 } 79 80 void ParGCAllocBuffer::flush_stats(PLABStats* stats) { 81 assert(ResizePLAB, "Wasted work"); 82 stats->add_allocated(_allocated); 83 stats->add_wasted(_wasted); 84 stats->add_unused(pointer_delta(_end, _top)); 85 } 86 87 // Compute desired plab size and latch result for later 88 // use. This should be called once at the end of parallel 89 // scavenge; it clears the sensor accumulators. 90 void PLABStats::adjust_desired_plab_sz(uint no_of_gc_workers) { 91 assert(ResizePLAB, "Not set"); 92 if (_allocated == 0) { 93 assert(_unused == 0, 94 err_msg("Inconsistency in PLAB stats: " 95 "_allocated: "SIZE_FORMAT", " 96 "_wasted: "SIZE_FORMAT", " 97 "_unused: "SIZE_FORMAT", " 98 "_used : "SIZE_FORMAT, 99 _allocated, _wasted, _unused, _used)); 100 101 _allocated = 1; 102 } 103 double wasted_frac = (double)_unused/(double)_allocated; 104 size_t target_refills = (size_t)((wasted_frac*TargetSurvivorRatio)/ 105 TargetPLABWastePct); 106 if (target_refills == 0) { 107 target_refills = 1; 108 } 109 _used = _allocated - _wasted - _unused; 110 size_t plab_sz = _used/(target_refills*no_of_gc_workers); 111 if (PrintPLAB) gclog_or_tty->print(" (plab_sz = %d ", plab_sz); 112 // Take historical weighted average 113 _filter.sample(plab_sz); 114 // Clip from above and below, and align to object boundary 115 plab_sz = MAX2(min_size(), (size_t)_filter.average()); 116 plab_sz = MIN2(max_size(), plab_sz); 117 plab_sz = align_object_size(plab_sz); 118 // Latch the result 119 if (PrintPLAB) gclog_or_tty->print(" desired_plab_sz = %d) ", plab_sz); 120 _desired_plab_sz = plab_sz; 121 // Now clear the accumulators for next round: 122 // note this needs to be fixed in the case where we 123 // are retaining across scavenges. FIX ME !!! XXX 124 _allocated = 0; 125 _wasted = 0; 126 _unused = 0; 127 } 128 129 #ifndef PRODUCT 130 void ParGCAllocBuffer::print() { 131 gclog_or_tty->print("parGCAllocBuffer: _bottom: %p _top: %p _end: %p _hard_end: %p" 132 "_retained: %c _retained_filler: [%p,%p)\n", 133 _bottom, _top, _end, _hard_end, 134 "FT"[_retained], _retained_filler.start(), _retained_filler.end()); 135 } 136 #endif // !PRODUCT 137 138 const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords = 139 MIN2(CardTableModRefBS::par_chunk_heapword_alignment(), 140 ((size_t)Generation::GenGrain)/HeapWordSize); 141 const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes = 142 MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize, 143 (size_t)Generation::GenGrain); 144 145 ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz, 146 BlockOffsetSharedArray* bsa) : 147 ParGCAllocBuffer(word_sz), 148 _bsa(bsa), 149 _bt(bsa, MemRegion(_bottom, _hard_end)), 150 _true_end(_hard_end) 151 {} 152 153 // The buffer comes with its own BOT, with a shared (obviously) underlying 154 // BlockOffsetSharedArray. We manipulate this BOT in the normal way 155 // as we would for any contiguous space. However, on accasion we 156 // need to do some buffer surgery at the extremities before we 157 // start using the body of the buffer for allocations. Such surgery 158 // (as explained elsewhere) is to prevent allocation on a card that 159 // is in the process of being walked concurrently by another GC thread. 160 // When such surgery happens at a point that is far removed (to the 161 // right of the current allocation point, top), we use the "contig" 162 // parameter below to directly manipulate the shared array without 163 // modifying the _next_threshold state in the BOT. 164 void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr, 165 bool contig) { 166 CollectedHeap::fill_with_object(mr); 167 if (contig) { 168 _bt.alloc_block(mr.start(), mr.end()); 169 } else { 170 _bt.BlockOffsetArray::alloc_block(mr.start(), mr.end()); 171 } 172 } 173 174 HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) { 175 HeapWord* res = NULL; 176 if (_true_end > _hard_end) { 177 assert((HeapWord*)align_size_down(intptr_t(_hard_end), 178 ChunkSizeInBytes) == _hard_end, 179 "or else _true_end should be equal to _hard_end"); 180 assert(_retained, "or else _true_end should be equal to _hard_end"); 181 assert(_retained_filler.end() <= _top, "INVARIANT"); 182 CollectedHeap::fill_with_object(_retained_filler); 183 if (_top < _hard_end) { 184 fill_region_with_block(MemRegion(_top, _hard_end), true); 185 } 186 HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords); 187 _retained_filler = MemRegion(_hard_end, FillerHeaderSize); 188 _bt.alloc_block(_retained_filler.start(), _retained_filler.word_size()); 189 _top = _retained_filler.end(); 190 _hard_end = next_hard_end; 191 _end = _hard_end - AlignmentReserve; 192 res = ParGCAllocBuffer::allocate(word_sz); 193 if (res != NULL) { 194 _bt.alloc_block(res, word_sz); 195 } 196 } 197 return res; 198 } 199 200 void 201 ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) { 202 ParGCAllocBuffer::undo_allocation(obj, word_sz); 203 // This may back us up beyond the previous threshold, so reset. 204 _bt.set_region(MemRegion(_top, _hard_end)); 205 _bt.initialize_threshold(); 206 } 207 208 void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) { 209 assert(!retain || end_of_gc, "Can only retain at GC end."); 210 if (_retained) { 211 // We're about to make the retained_filler into a block. 212 _bt.BlockOffsetArray::alloc_block(_retained_filler.start(), 213 _retained_filler.end()); 214 } 215 // Reset _hard_end to _true_end (and update _end) 216 if (retain && _hard_end != NULL) { 217 assert(_hard_end <= _true_end, "Invariant."); 218 _hard_end = _true_end; 219 _end = MAX2(_top, _hard_end - AlignmentReserve); 220 assert(_end <= _hard_end, "Invariant."); 221 } 222 _true_end = _hard_end; 223 HeapWord* pre_top = _top; 224 225 ParGCAllocBuffer::retire(end_of_gc, retain); 226 // Now any old _retained_filler is cut back to size, the free part is 227 // filled with a filler object, and top is past the header of that 228 // object. 229 230 if (retain && _top < _end) { 231 assert(end_of_gc && retain, "Or else retain should be false."); 232 // If the lab does not start on a card boundary, we don't want to 233 // allocate onto that card, since that might lead to concurrent 234 // allocation and card scanning, which we don't support. So we fill 235 // the first card with a garbage object. 236 size_t first_card_index = _bsa->index_for(pre_top); 237 HeapWord* first_card_start = _bsa->address_for_index(first_card_index); 238 if (first_card_start < pre_top) { 239 HeapWord* second_card_start = 240 _bsa->inc_by_region_size(first_card_start); 241 242 // Ensure enough room to fill with the smallest block 243 second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve); 244 245 // If the end is already in the first card, don't go beyond it! 246 // Or if the remainder is too small for a filler object, gobble it up. 247 if (_hard_end < second_card_start || 248 pointer_delta(_hard_end, second_card_start) < AlignmentReserve) { 249 second_card_start = _hard_end; 250 } 251 if (pre_top < second_card_start) { 252 MemRegion first_card_suffix(pre_top, second_card_start); 253 fill_region_with_block(first_card_suffix, true); 254 } 255 pre_top = second_card_start; 256 _top = pre_top; 257 _end = MAX2(_top, _hard_end - AlignmentReserve); 258 } 259 260 // If the lab does not end on a card boundary, we don't want to 261 // allocate onto that card, since that might lead to concurrent 262 // allocation and card scanning, which we don't support. So we fill 263 // the last card with a garbage object. 264 size_t last_card_index = _bsa->index_for(_hard_end); 265 HeapWord* last_card_start = _bsa->address_for_index(last_card_index); 266 if (last_card_start < _hard_end) { 267 268 // Ensure enough room to fill with the smallest block 269 last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve); 270 271 // If the top is already in the last card, don't go back beyond it! 272 // Or if the remainder is too small for a filler object, gobble it up. 273 if (_top > last_card_start || 274 pointer_delta(last_card_start, _top) < AlignmentReserve) { 275 last_card_start = _top; 276 } 277 if (last_card_start < _hard_end) { 278 MemRegion last_card_prefix(last_card_start, _hard_end); 279 fill_region_with_block(last_card_prefix, false); 280 } 281 _hard_end = last_card_start; 282 _end = MAX2(_top, _hard_end - AlignmentReserve); 283 _true_end = _hard_end; 284 assert(_end <= _hard_end, "Invariant."); 285 } 286 287 // At this point: 288 // 1) we had a filler object from the original top to hard_end. 289 // 2) We've filled in any partial cards at the front and back. 290 if (pre_top < _hard_end) { 291 // Now we can reset the _bt to do allocation in the given area. 292 MemRegion new_filler(pre_top, _hard_end); 293 fill_region_with_block(new_filler, false); 294 _top = pre_top + ParGCAllocBuffer::FillerHeaderSize; 295 // If there's no space left, don't retain. 296 if (_top >= _end) { 297 _retained = false; 298 invalidate(); 299 return; 300 } 301 _retained_filler = MemRegion(pre_top, _top); 302 _bt.set_region(MemRegion(_top, _hard_end)); 303 _bt.initialize_threshold(); 304 assert(_bt.threshold() > _top, "initialize_threshold failed!"); 305 306 // There may be other reasons for queries into the middle of the 307 // filler object. When such queries are done in parallel with 308 // allocation, bad things can happen, if the query involves object 309 // iteration. So we ensure that such queries do not involve object 310 // iteration, by putting another filler object on the boundaries of 311 // such queries. One such is the object spanning a parallel card 312 // chunk boundary. 313 314 // "chunk_boundary" is the address of the first chunk boundary less 315 // than "hard_end". 316 HeapWord* chunk_boundary = 317 (HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes); 318 assert(chunk_boundary < _hard_end, "Or else above did not work."); 319 assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve, 320 "Consequence of last card handling above."); 321 322 if (_top <= chunk_boundary) { 323 assert(_true_end == _hard_end, "Invariant."); 324 while (_top <= chunk_boundary) { 325 assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve, 326 "Consequence of last card handling above."); 327 _bt.BlockOffsetArray::alloc_block(chunk_boundary, _hard_end); 328 CollectedHeap::fill_with_object(chunk_boundary, _hard_end); 329 _hard_end = chunk_boundary; 330 chunk_boundary -= ChunkSizeInWords; 331 } 332 _end = _hard_end - AlignmentReserve; 333 assert(_top <= _end, "Invariant."); 334 // Now reset the initial filler chunk so it doesn't overlap with 335 // the one(s) inserted above. 336 MemRegion new_filler(pre_top, _hard_end); 337 fill_region_with_block(new_filler, false); 338 } 339 } else { 340 _retained = false; 341 invalidate(); 342 } 343 } else { 344 assert(!end_of_gc || 345 (!_retained && _true_end == _hard_end), "Checking."); 346 } 347 assert(_end <= _hard_end, "Invariant."); 348 assert(_top < _end || _top == _hard_end, "Invariant"); 349 }