Tue, 10 May 2011 00:33:21 -0700
6883834: ParNew: assert(!_g->to()->is_in_reserved(obj),"Scanning field twice?") with LargeObjects tests
Summary: Fixed process_chunk_boundaries(), used for parallel card scanning when using ParNew/CMS, so as to prevent double-scanning, or worse, non-scanning of imprecisely marked objects exceeding parallel chunk size. Made some sizing parameters for parallel card scanning diagnostic, disabled ParallelGCRetainPLAB, and elaborated and clarified some comments.
Reviewed-by: stefank, johnc
1 /*
2 * Copyright (c) 2001, 2010, 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 */
25 #include "precompiled.hpp"
26 #include "gc_implementation/parNew/parGCAllocBuffer.hpp"
27 #include "memory/sharedHeap.hpp"
28 #include "oops/arrayOop.hpp"
29 #include "oops/oop.inline.hpp"
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 }
43 size_t ParGCAllocBuffer::FillerHeaderSize;
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;
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 }
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 }
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() {
91 assert(ResizePLAB, "Not set");
92 if (_allocated == 0) {
93 assert(_unused == 0, "Inconsistency in PLAB stats");
94 _allocated = 1;
95 }
96 double wasted_frac = (double)_unused/(double)_allocated;
97 size_t target_refills = (size_t)((wasted_frac*TargetSurvivorRatio)/
98 TargetPLABWastePct);
99 if (target_refills == 0) {
100 target_refills = 1;
101 }
102 _used = _allocated - _wasted - _unused;
103 size_t plab_sz = _used/(target_refills*ParallelGCThreads);
104 if (PrintPLAB) gclog_or_tty->print(" (plab_sz = %d ", plab_sz);
105 // Take historical weighted average
106 _filter.sample(plab_sz);
107 // Clip from above and below, and align to object boundary
108 plab_sz = MAX2(min_size(), (size_t)_filter.average());
109 plab_sz = MIN2(max_size(), plab_sz);
110 plab_sz = align_object_size(plab_sz);
111 // Latch the result
112 if (PrintPLAB) gclog_or_tty->print(" desired_plab_sz = %d) ", plab_sz);
113 if (ResizePLAB) {
114 _desired_plab_sz = plab_sz;
115 }
116 // Now clear the accumulators for next round:
117 // note this needs to be fixed in the case where we
118 // are retaining across scavenges. FIX ME !!! XXX
119 _allocated = 0;
120 _wasted = 0;
121 _unused = 0;
122 }
124 #ifndef PRODUCT
125 void ParGCAllocBuffer::print() {
126 gclog_or_tty->print("parGCAllocBuffer: _bottom: %p _top: %p _end: %p _hard_end: %p"
127 "_retained: %c _retained_filler: [%p,%p)\n",
128 _bottom, _top, _end, _hard_end,
129 "FT"[_retained], _retained_filler.start(), _retained_filler.end());
130 }
131 #endif // !PRODUCT
133 const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords =
134 MIN2(CardTableModRefBS::par_chunk_heapword_alignment(),
135 ((size_t)Generation::GenGrain)/HeapWordSize);
136 const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes =
137 MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize,
138 (size_t)Generation::GenGrain);
140 ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz,
141 BlockOffsetSharedArray* bsa) :
142 ParGCAllocBuffer(word_sz),
143 _bsa(bsa),
144 _bt(bsa, MemRegion(_bottom, _hard_end)),
145 _true_end(_hard_end)
146 {}
148 // The buffer comes with its own BOT, with a shared (obviously) underlying
149 // BlockOffsetSharedArray. We manipulate this BOT in the normal way
150 // as we would for any contiguous space. However, on accasion we
151 // need to do some buffer surgery at the extremities before we
152 // start using the body of the buffer for allocations. Such surgery
153 // (as explained elsewhere) is to prevent allocation on a card that
154 // is in the process of being walked concurrently by another GC thread.
155 // When such surgery happens at a point that is far removed (to the
156 // right of the current allocation point, top), we use the "contig"
157 // parameter below to directly manipulate the shared array without
158 // modifying the _next_threshold state in the BOT.
159 void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr,
160 bool contig) {
161 CollectedHeap::fill_with_object(mr);
162 if (contig) {
163 _bt.alloc_block(mr.start(), mr.end());
164 } else {
165 _bt.BlockOffsetArray::alloc_block(mr.start(), mr.end());
166 }
167 }
169 HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) {
170 HeapWord* res = NULL;
171 if (_true_end > _hard_end) {
172 assert((HeapWord*)align_size_down(intptr_t(_hard_end),
173 ChunkSizeInBytes) == _hard_end,
174 "or else _true_end should be equal to _hard_end");
175 assert(_retained, "or else _true_end should be equal to _hard_end");
176 assert(_retained_filler.end() <= _top, "INVARIANT");
177 CollectedHeap::fill_with_object(_retained_filler);
178 if (_top < _hard_end) {
179 fill_region_with_block(MemRegion(_top, _hard_end), true);
180 }
181 HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords);
182 _retained_filler = MemRegion(_hard_end, FillerHeaderSize);
183 _bt.alloc_block(_retained_filler.start(), _retained_filler.word_size());
184 _top = _retained_filler.end();
185 _hard_end = next_hard_end;
186 _end = _hard_end - AlignmentReserve;
187 res = ParGCAllocBuffer::allocate(word_sz);
188 if (res != NULL) {
189 _bt.alloc_block(res, word_sz);
190 }
191 }
192 return res;
193 }
195 void
196 ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) {
197 ParGCAllocBuffer::undo_allocation(obj, word_sz);
198 // This may back us up beyond the previous threshold, so reset.
199 _bt.set_region(MemRegion(_top, _hard_end));
200 _bt.initialize_threshold();
201 }
203 void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) {
204 assert(!retain || end_of_gc, "Can only retain at GC end.");
205 if (_retained) {
206 // We're about to make the retained_filler into a block.
207 _bt.BlockOffsetArray::alloc_block(_retained_filler.start(),
208 _retained_filler.end());
209 }
210 // Reset _hard_end to _true_end (and update _end)
211 if (retain && _hard_end != NULL) {
212 assert(_hard_end <= _true_end, "Invariant.");
213 _hard_end = _true_end;
214 _end = MAX2(_top, _hard_end - AlignmentReserve);
215 assert(_end <= _hard_end, "Invariant.");
216 }
217 _true_end = _hard_end;
218 HeapWord* pre_top = _top;
220 ParGCAllocBuffer::retire(end_of_gc, retain);
221 // Now any old _retained_filler is cut back to size, the free part is
222 // filled with a filler object, and top is past the header of that
223 // object.
225 if (retain && _top < _end) {
226 assert(end_of_gc && retain, "Or else retain should be false.");
227 // If the lab does not start on a card boundary, we don't want to
228 // allocate onto that card, since that might lead to concurrent
229 // allocation and card scanning, which we don't support. So we fill
230 // the first card with a garbage object.
231 size_t first_card_index = _bsa->index_for(pre_top);
232 HeapWord* first_card_start = _bsa->address_for_index(first_card_index);
233 if (first_card_start < pre_top) {
234 HeapWord* second_card_start =
235 _bsa->inc_by_region_size(first_card_start);
237 // Ensure enough room to fill with the smallest block
238 second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve);
240 // If the end is already in the first card, don't go beyond it!
241 // Or if the remainder is too small for a filler object, gobble it up.
242 if (_hard_end < second_card_start ||
243 pointer_delta(_hard_end, second_card_start) < AlignmentReserve) {
244 second_card_start = _hard_end;
245 }
246 if (pre_top < second_card_start) {
247 MemRegion first_card_suffix(pre_top, second_card_start);
248 fill_region_with_block(first_card_suffix, true);
249 }
250 pre_top = second_card_start;
251 _top = pre_top;
252 _end = MAX2(_top, _hard_end - AlignmentReserve);
253 }
255 // If the lab does not end on a card boundary, we don't want to
256 // allocate onto that card, since that might lead to concurrent
257 // allocation and card scanning, which we don't support. So we fill
258 // the last card with a garbage object.
259 size_t last_card_index = _bsa->index_for(_hard_end);
260 HeapWord* last_card_start = _bsa->address_for_index(last_card_index);
261 if (last_card_start < _hard_end) {
263 // Ensure enough room to fill with the smallest block
264 last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve);
266 // If the top is already in the last card, don't go back beyond it!
267 // Or if the remainder is too small for a filler object, gobble it up.
268 if (_top > last_card_start ||
269 pointer_delta(last_card_start, _top) < AlignmentReserve) {
270 last_card_start = _top;
271 }
272 if (last_card_start < _hard_end) {
273 MemRegion last_card_prefix(last_card_start, _hard_end);
274 fill_region_with_block(last_card_prefix, false);
275 }
276 _hard_end = last_card_start;
277 _end = MAX2(_top, _hard_end - AlignmentReserve);
278 _true_end = _hard_end;
279 assert(_end <= _hard_end, "Invariant.");
280 }
282 // At this point:
283 // 1) we had a filler object from the original top to hard_end.
284 // 2) We've filled in any partial cards at the front and back.
285 if (pre_top < _hard_end) {
286 // Now we can reset the _bt to do allocation in the given area.
287 MemRegion new_filler(pre_top, _hard_end);
288 fill_region_with_block(new_filler, false);
289 _top = pre_top + ParGCAllocBuffer::FillerHeaderSize;
290 // If there's no space left, don't retain.
291 if (_top >= _end) {
292 _retained = false;
293 invalidate();
294 return;
295 }
296 _retained_filler = MemRegion(pre_top, _top);
297 _bt.set_region(MemRegion(_top, _hard_end));
298 _bt.initialize_threshold();
299 assert(_bt.threshold() > _top, "initialize_threshold failed!");
301 // There may be other reasons for queries into the middle of the
302 // filler object. When such queries are done in parallel with
303 // allocation, bad things can happen, if the query involves object
304 // iteration. So we ensure that such queries do not involve object
305 // iteration, by putting another filler object on the boundaries of
306 // such queries. One such is the object spanning a parallel card
307 // chunk boundary.
309 // "chunk_boundary" is the address of the first chunk boundary less
310 // than "hard_end".
311 HeapWord* chunk_boundary =
312 (HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes);
313 assert(chunk_boundary < _hard_end, "Or else above did not work.");
314 assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve,
315 "Consequence of last card handling above.");
317 if (_top <= chunk_boundary) {
318 assert(_true_end == _hard_end, "Invariant.");
319 while (_top <= chunk_boundary) {
320 assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve,
321 "Consequence of last card handling above.");
322 _bt.BlockOffsetArray::alloc_block(chunk_boundary, _hard_end);
323 CollectedHeap::fill_with_object(chunk_boundary, _hard_end);
324 _hard_end = chunk_boundary;
325 chunk_boundary -= ChunkSizeInWords;
326 }
327 _end = _hard_end - AlignmentReserve;
328 assert(_top <= _end, "Invariant.");
329 // Now reset the initial filler chunk so it doesn't overlap with
330 // the one(s) inserted above.
331 MemRegion new_filler(pre_top, _hard_end);
332 fill_region_with_block(new_filler, false);
333 }
334 } else {
335 _retained = false;
336 invalidate();
337 }
338 } else {
339 assert(!end_of_gc ||
340 (!_retained && _true_end == _hard_end), "Checking.");
341 }
342 assert(_end <= _hard_end, "Invariant.");
343 assert(_top < _end || _top == _hard_end, "Invariant");
344 }