140 " _retained_filler: [" PTR_FORMAT "," PTR_FORMAT ")\n", |
140 " _retained_filler: [" PTR_FORMAT "," PTR_FORMAT ")\n", |
141 _bottom, _top, _end, _hard_end, |
141 _bottom, _top, _end, _hard_end, |
142 "FT"[_retained], _retained_filler.start(), _retained_filler.end()); |
142 "FT"[_retained], _retained_filler.start(), _retained_filler.end()); |
143 } |
143 } |
144 #endif // !PRODUCT |
144 #endif // !PRODUCT |
145 |
|
146 const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords = |
|
147 MIN2(CardTableModRefBS::par_chunk_heapword_alignment(), |
|
148 ((size_t)Generation::GenGrain)/HeapWordSize); |
|
149 const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes = |
|
150 MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize, |
|
151 (size_t)Generation::GenGrain); |
|
152 |
|
153 ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz, |
|
154 BlockOffsetSharedArray* bsa) : |
|
155 ParGCAllocBuffer(word_sz), |
|
156 _bsa(bsa), |
|
157 _bt(bsa, MemRegion(_bottom, _hard_end)), |
|
158 _true_end(_hard_end) |
|
159 {} |
|
160 |
|
161 // The buffer comes with its own BOT, with a shared (obviously) underlying |
|
162 // BlockOffsetSharedArray. We manipulate this BOT in the normal way |
|
163 // as we would for any contiguous space. However, on occasion we |
|
164 // need to do some buffer surgery at the extremities before we |
|
165 // start using the body of the buffer for allocations. Such surgery |
|
166 // (as explained elsewhere) is to prevent allocation on a card that |
|
167 // is in the process of being walked concurrently by another GC thread. |
|
168 // When such surgery happens at a point that is far removed (to the |
|
169 // right of the current allocation point, top), we use the "contig" |
|
170 // parameter below to directly manipulate the shared array without |
|
171 // modifying the _next_threshold state in the BOT. |
|
172 void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr, |
|
173 bool contig) { |
|
174 CollectedHeap::fill_with_object(mr); |
|
175 if (contig) { |
|
176 _bt.alloc_block(mr.start(), mr.end()); |
|
177 } else { |
|
178 _bt.BlockOffsetArray::alloc_block(mr.start(), mr.end()); |
|
179 } |
|
180 } |
|
181 |
|
182 HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) { |
|
183 HeapWord* res = NULL; |
|
184 if (_true_end > _hard_end) { |
|
185 assert((HeapWord*)align_size_down(intptr_t(_hard_end), |
|
186 ChunkSizeInBytes) == _hard_end, |
|
187 "or else _true_end should be equal to _hard_end"); |
|
188 assert(_retained, "or else _true_end should be equal to _hard_end"); |
|
189 assert(_retained_filler.end() <= _top, "INVARIANT"); |
|
190 CollectedHeap::fill_with_object(_retained_filler); |
|
191 if (_top < _hard_end) { |
|
192 fill_region_with_block(MemRegion(_top, _hard_end), true); |
|
193 } |
|
194 HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords); |
|
195 _retained_filler = MemRegion(_hard_end, FillerHeaderSize); |
|
196 _bt.alloc_block(_retained_filler.start(), _retained_filler.word_size()); |
|
197 _top = _retained_filler.end(); |
|
198 _hard_end = next_hard_end; |
|
199 _end = _hard_end - AlignmentReserve; |
|
200 res = ParGCAllocBuffer::allocate(word_sz); |
|
201 if (res != NULL) { |
|
202 _bt.alloc_block(res, word_sz); |
|
203 } |
|
204 } |
|
205 return res; |
|
206 } |
|
207 |
|
208 void |
|
209 ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) { |
|
210 ParGCAllocBuffer::undo_allocation(obj, word_sz); |
|
211 // This may back us up beyond the previous threshold, so reset. |
|
212 _bt.set_region(MemRegion(_top, _hard_end)); |
|
213 _bt.initialize_threshold(); |
|
214 } |
|
215 |
|
216 void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) { |
|
217 assert(!retain || end_of_gc, "Can only retain at GC end."); |
|
218 if (_retained) { |
|
219 // We're about to make the retained_filler into a block. |
|
220 _bt.BlockOffsetArray::alloc_block(_retained_filler.start(), |
|
221 _retained_filler.end()); |
|
222 } |
|
223 // Reset _hard_end to _true_end (and update _end) |
|
224 if (retain && _hard_end != NULL) { |
|
225 assert(_hard_end <= _true_end, "Invariant."); |
|
226 _hard_end = _true_end; |
|
227 _end = MAX2(_top, _hard_end - AlignmentReserve); |
|
228 assert(_end <= _hard_end, "Invariant."); |
|
229 } |
|
230 _true_end = _hard_end; |
|
231 HeapWord* pre_top = _top; |
|
232 |
|
233 ParGCAllocBuffer::retire(end_of_gc, retain); |
|
234 // Now any old _retained_filler is cut back to size, the free part is |
|
235 // filled with a filler object, and top is past the header of that |
|
236 // object. |
|
237 |
|
238 if (retain && _top < _end) { |
|
239 assert(end_of_gc && retain, "Or else retain should be false."); |
|
240 // If the lab does not start on a card boundary, we don't want to |
|
241 // allocate onto that card, since that might lead to concurrent |
|
242 // allocation and card scanning, which we don't support. So we fill |
|
243 // the first card with a garbage object. |
|
244 size_t first_card_index = _bsa->index_for(pre_top); |
|
245 HeapWord* first_card_start = _bsa->address_for_index(first_card_index); |
|
246 if (first_card_start < pre_top) { |
|
247 HeapWord* second_card_start = |
|
248 _bsa->inc_by_region_size(first_card_start); |
|
249 |
|
250 // Ensure enough room to fill with the smallest block |
|
251 second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve); |
|
252 |
|
253 // If the end is already in the first card, don't go beyond it! |
|
254 // Or if the remainder is too small for a filler object, gobble it up. |
|
255 if (_hard_end < second_card_start || |
|
256 pointer_delta(_hard_end, second_card_start) < AlignmentReserve) { |
|
257 second_card_start = _hard_end; |
|
258 } |
|
259 if (pre_top < second_card_start) { |
|
260 MemRegion first_card_suffix(pre_top, second_card_start); |
|
261 fill_region_with_block(first_card_suffix, true); |
|
262 } |
|
263 pre_top = second_card_start; |
|
264 _top = pre_top; |
|
265 _end = MAX2(_top, _hard_end - AlignmentReserve); |
|
266 } |
|
267 |
|
268 // If the lab does not end on a card boundary, we don't want to |
|
269 // allocate onto that card, since that might lead to concurrent |
|
270 // allocation and card scanning, which we don't support. So we fill |
|
271 // the last card with a garbage object. |
|
272 size_t last_card_index = _bsa->index_for(_hard_end); |
|
273 HeapWord* last_card_start = _bsa->address_for_index(last_card_index); |
|
274 if (last_card_start < _hard_end) { |
|
275 |
|
276 // Ensure enough room to fill with the smallest block |
|
277 last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve); |
|
278 |
|
279 // If the top is already in the last card, don't go back beyond it! |
|
280 // Or if the remainder is too small for a filler object, gobble it up. |
|
281 if (_top > last_card_start || |
|
282 pointer_delta(last_card_start, _top) < AlignmentReserve) { |
|
283 last_card_start = _top; |
|
284 } |
|
285 if (last_card_start < _hard_end) { |
|
286 MemRegion last_card_prefix(last_card_start, _hard_end); |
|
287 fill_region_with_block(last_card_prefix, false); |
|
288 } |
|
289 _hard_end = last_card_start; |
|
290 _end = MAX2(_top, _hard_end - AlignmentReserve); |
|
291 _true_end = _hard_end; |
|
292 assert(_end <= _hard_end, "Invariant."); |
|
293 } |
|
294 |
|
295 // At this point: |
|
296 // 1) we had a filler object from the original top to hard_end. |
|
297 // 2) We've filled in any partial cards at the front and back. |
|
298 if (pre_top < _hard_end) { |
|
299 // Now we can reset the _bt to do allocation in the given area. |
|
300 MemRegion new_filler(pre_top, _hard_end); |
|
301 fill_region_with_block(new_filler, false); |
|
302 _top = pre_top + ParGCAllocBuffer::FillerHeaderSize; |
|
303 // If there's no space left, don't retain. |
|
304 if (_top >= _end) { |
|
305 _retained = false; |
|
306 invalidate(); |
|
307 return; |
|
308 } |
|
309 _retained_filler = MemRegion(pre_top, _top); |
|
310 _bt.set_region(MemRegion(_top, _hard_end)); |
|
311 _bt.initialize_threshold(); |
|
312 assert(_bt.threshold() > _top, "initialize_threshold failed!"); |
|
313 |
|
314 // There may be other reasons for queries into the middle of the |
|
315 // filler object. When such queries are done in parallel with |
|
316 // allocation, bad things can happen, if the query involves object |
|
317 // iteration. So we ensure that such queries do not involve object |
|
318 // iteration, by putting another filler object on the boundaries of |
|
319 // such queries. One such is the object spanning a parallel card |
|
320 // chunk boundary. |
|
321 |
|
322 // "chunk_boundary" is the address of the first chunk boundary less |
|
323 // than "hard_end". |
|
324 HeapWord* chunk_boundary = |
|
325 (HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes); |
|
326 assert(chunk_boundary < _hard_end, "Or else above did not work."); |
|
327 assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve, |
|
328 "Consequence of last card handling above."); |
|
329 |
|
330 if (_top <= chunk_boundary) { |
|
331 assert(_true_end == _hard_end, "Invariant."); |
|
332 while (_top <= chunk_boundary) { |
|
333 assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve, |
|
334 "Consequence of last card handling above."); |
|
335 _bt.BlockOffsetArray::alloc_block(chunk_boundary, _hard_end); |
|
336 CollectedHeap::fill_with_object(chunk_boundary, _hard_end); |
|
337 _hard_end = chunk_boundary; |
|
338 chunk_boundary -= ChunkSizeInWords; |
|
339 } |
|
340 _end = _hard_end - AlignmentReserve; |
|
341 assert(_top <= _end, "Invariant."); |
|
342 // Now reset the initial filler chunk so it doesn't overlap with |
|
343 // the one(s) inserted above. |
|
344 MemRegion new_filler(pre_top, _hard_end); |
|
345 fill_region_with_block(new_filler, false); |
|
346 } |
|
347 } else { |
|
348 _retained = false; |
|
349 invalidate(); |
|
350 } |
|
351 } else { |
|
352 assert(!end_of_gc || |
|
353 (!_retained && _true_end == _hard_end), "Checking."); |
|
354 } |
|
355 assert(_end <= _hard_end, "Invariant."); |
|
356 assert(_top < _end || _top == _hard_end, "Invariant"); |
|
357 } |
|