Tue, 08 Aug 2017 15:57:29 +0800
merge
1 /*
2 * Copyright (c) 2001, 2014, 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/shared/parGCAllocBuffer.hpp"
27 #include "memory/sharedHeap.hpp"
28 #include "oops/arrayOop.hpp"
29 #include "oops/oop.inline.hpp"
31 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
33 ParGCAllocBuffer::ParGCAllocBuffer(size_t desired_plab_sz_) :
34 _word_sz(desired_plab_sz_), _bottom(NULL), _top(NULL),
35 _end(NULL), _hard_end(NULL),
36 _retained(false), _retained_filler(),
37 _allocated(0), _wasted(0)
38 {
39 assert (min_size() > AlignmentReserve, "Inconsistency!");
40 // arrayOopDesc::header_size depends on command line initialization.
41 FillerHeaderSize = align_object_size(arrayOopDesc::header_size(T_INT));
42 AlignmentReserve = oopDesc::header_size() > MinObjAlignment ? FillerHeaderSize : 0;
43 }
45 size_t ParGCAllocBuffer::FillerHeaderSize;
47 // If the minimum object size is greater than MinObjAlignment, we can
48 // end up with a shard at the end of the buffer that's smaller than
49 // the smallest object. We can't allow that because the buffer must
50 // look like it's full of objects when we retire it, so we make
51 // sure we have enough space for a filler int array object.
52 size_t ParGCAllocBuffer::AlignmentReserve;
54 void ParGCAllocBuffer::retire(bool end_of_gc, bool retain) {
55 assert(!retain || end_of_gc, "Can only retain at GC end.");
56 if (_retained) {
57 // If the buffer had been retained shorten the previous filler object.
58 assert(_retained_filler.end() <= _top, "INVARIANT");
59 CollectedHeap::fill_with_object(_retained_filler);
60 // Wasted space book-keeping, otherwise (normally) done in invalidate()
61 _wasted += _retained_filler.word_size();
62 _retained = false;
63 }
64 assert(!end_of_gc || !_retained, "At this point, end_of_gc ==> !_retained.");
65 if (_top < _hard_end) {
66 CollectedHeap::fill_with_object(_top, _hard_end);
67 if (!retain) {
68 invalidate();
69 } else {
70 // Is there wasted space we'd like to retain for the next GC?
71 if (pointer_delta(_end, _top) > FillerHeaderSize) {
72 _retained = true;
73 _retained_filler = MemRegion(_top, FillerHeaderSize);
74 _top = _top + FillerHeaderSize;
75 } else {
76 invalidate();
77 }
78 }
79 }
80 }
82 void ParGCAllocBuffer::flush_stats(PLABStats* stats) {
83 assert(ResizePLAB, "Wasted work");
84 stats->add_allocated(_allocated);
85 stats->add_wasted(_wasted);
86 stats->add_unused(pointer_delta(_end, _top));
87 }
89 // Compute desired plab size and latch result for later
90 // use. This should be called once at the end of parallel
91 // scavenge; it clears the sensor accumulators.
92 void PLABStats::adjust_desired_plab_sz(uint no_of_gc_workers) {
93 assert(ResizePLAB, "Not set");
95 assert(is_object_aligned(max_size()) && min_size() <= max_size(),
96 "PLAB clipping computation may be incorrect");
98 if (_allocated == 0) {
99 assert(_unused == 0,
100 err_msg("Inconsistency in PLAB stats: "
101 "_allocated: "SIZE_FORMAT", "
102 "_wasted: "SIZE_FORMAT", "
103 "_unused: "SIZE_FORMAT", "
104 "_used : "SIZE_FORMAT,
105 _allocated, _wasted, _unused, _used));
107 _allocated = 1;
108 }
109 double wasted_frac = (double)_unused/(double)_allocated;
110 size_t target_refills = (size_t)((wasted_frac*TargetSurvivorRatio)/
111 TargetPLABWastePct);
112 if (target_refills == 0) {
113 target_refills = 1;
114 }
115 _used = _allocated - _wasted - _unused;
116 size_t plab_sz = _used/(target_refills*no_of_gc_workers);
117 if (PrintPLAB) gclog_or_tty->print(" (plab_sz = %d ", plab_sz);
118 // Take historical weighted average
119 _filter.sample(plab_sz);
120 // Clip from above and below, and align to object boundary
121 plab_sz = MAX2(min_size(), (size_t)_filter.average());
122 plab_sz = MIN2(max_size(), plab_sz);
123 plab_sz = align_object_size(plab_sz);
124 // Latch the result
125 if (PrintPLAB) gclog_or_tty->print(" desired_plab_sz = %d) ", plab_sz);
126 _desired_plab_sz = plab_sz;
127 // Now clear the accumulators for next round:
128 // note this needs to be fixed in the case where we
129 // are retaining across scavenges. FIX ME !!! XXX
130 _allocated = 0;
131 _wasted = 0;
132 _unused = 0;
133 }
135 #ifndef PRODUCT
136 void ParGCAllocBuffer::print() {
137 gclog_or_tty->print("parGCAllocBuffer: _bottom: %p _top: %p _end: %p _hard_end: %p"
138 "_retained: %c _retained_filler: [%p,%p)\n",
139 _bottom, _top, _end, _hard_end,
140 "FT"[_retained], _retained_filler.start(), _retained_filler.end());
141 }
142 #endif // !PRODUCT
144 const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords =
145 MIN2(CardTableModRefBS::par_chunk_heapword_alignment(),
146 ((size_t)Generation::GenGrain)/HeapWordSize);
147 const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes =
148 MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize,
149 (size_t)Generation::GenGrain);
151 ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz,
152 BlockOffsetSharedArray* bsa) :
153 ParGCAllocBuffer(word_sz),
154 _bsa(bsa),
155 _bt(bsa, MemRegion(_bottom, _hard_end)),
156 _true_end(_hard_end)
157 {}
159 // The buffer comes with its own BOT, with a shared (obviously) underlying
160 // BlockOffsetSharedArray. We manipulate this BOT in the normal way
161 // as we would for any contiguous space. However, on accasion we
162 // need to do some buffer surgery at the extremities before we
163 // start using the body of the buffer for allocations. Such surgery
164 // (as explained elsewhere) is to prevent allocation on a card that
165 // is in the process of being walked concurrently by another GC thread.
166 // When such surgery happens at a point that is far removed (to the
167 // right of the current allocation point, top), we use the "contig"
168 // parameter below to directly manipulate the shared array without
169 // modifying the _next_threshold state in the BOT.
170 void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr,
171 bool contig) {
172 CollectedHeap::fill_with_object(mr);
173 if (contig) {
174 _bt.alloc_block(mr.start(), mr.end());
175 } else {
176 _bt.BlockOffsetArray::alloc_block(mr.start(), mr.end());
177 }
178 }
180 HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) {
181 HeapWord* res = NULL;
182 if (_true_end > _hard_end) {
183 assert((HeapWord*)align_size_down(intptr_t(_hard_end),
184 ChunkSizeInBytes) == _hard_end,
185 "or else _true_end should be equal to _hard_end");
186 assert(_retained, "or else _true_end should be equal to _hard_end");
187 assert(_retained_filler.end() <= _top, "INVARIANT");
188 CollectedHeap::fill_with_object(_retained_filler);
189 if (_top < _hard_end) {
190 fill_region_with_block(MemRegion(_top, _hard_end), true);
191 }
192 HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords);
193 _retained_filler = MemRegion(_hard_end, FillerHeaderSize);
194 _bt.alloc_block(_retained_filler.start(), _retained_filler.word_size());
195 _top = _retained_filler.end();
196 _hard_end = next_hard_end;
197 _end = _hard_end - AlignmentReserve;
198 res = ParGCAllocBuffer::allocate(word_sz);
199 if (res != NULL) {
200 _bt.alloc_block(res, word_sz);
201 }
202 }
203 return res;
204 }
206 void
207 ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) {
208 ParGCAllocBuffer::undo_allocation(obj, word_sz);
209 // This may back us up beyond the previous threshold, so reset.
210 _bt.set_region(MemRegion(_top, _hard_end));
211 _bt.initialize_threshold();
212 }
214 void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) {
215 assert(!retain || end_of_gc, "Can only retain at GC end.");
216 if (_retained) {
217 // We're about to make the retained_filler into a block.
218 _bt.BlockOffsetArray::alloc_block(_retained_filler.start(),
219 _retained_filler.end());
220 }
221 // Reset _hard_end to _true_end (and update _end)
222 if (retain && _hard_end != NULL) {
223 assert(_hard_end <= _true_end, "Invariant.");
224 _hard_end = _true_end;
225 _end = MAX2(_top, _hard_end - AlignmentReserve);
226 assert(_end <= _hard_end, "Invariant.");
227 }
228 _true_end = _hard_end;
229 HeapWord* pre_top = _top;
231 ParGCAllocBuffer::retire(end_of_gc, retain);
232 // Now any old _retained_filler is cut back to size, the free part is
233 // filled with a filler object, and top is past the header of that
234 // object.
236 if (retain && _top < _end) {
237 assert(end_of_gc && retain, "Or else retain should be false.");
238 // If the lab does not start on a card boundary, we don't want to
239 // allocate onto that card, since that might lead to concurrent
240 // allocation and card scanning, which we don't support. So we fill
241 // the first card with a garbage object.
242 size_t first_card_index = _bsa->index_for(pre_top);
243 HeapWord* first_card_start = _bsa->address_for_index(first_card_index);
244 if (first_card_start < pre_top) {
245 HeapWord* second_card_start =
246 _bsa->inc_by_region_size(first_card_start);
248 // Ensure enough room to fill with the smallest block
249 second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve);
251 // If the end is already in the first card, don't go beyond it!
252 // Or if the remainder is too small for a filler object, gobble it up.
253 if (_hard_end < second_card_start ||
254 pointer_delta(_hard_end, second_card_start) < AlignmentReserve) {
255 second_card_start = _hard_end;
256 }
257 if (pre_top < second_card_start) {
258 MemRegion first_card_suffix(pre_top, second_card_start);
259 fill_region_with_block(first_card_suffix, true);
260 }
261 pre_top = second_card_start;
262 _top = pre_top;
263 _end = MAX2(_top, _hard_end - AlignmentReserve);
264 }
266 // If the lab does not end on a card boundary, we don't want to
267 // allocate onto that card, since that might lead to concurrent
268 // allocation and card scanning, which we don't support. So we fill
269 // the last card with a garbage object.
270 size_t last_card_index = _bsa->index_for(_hard_end);
271 HeapWord* last_card_start = _bsa->address_for_index(last_card_index);
272 if (last_card_start < _hard_end) {
274 // Ensure enough room to fill with the smallest block
275 last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve);
277 // If the top is already in the last card, don't go back beyond it!
278 // Or if the remainder is too small for a filler object, gobble it up.
279 if (_top > last_card_start ||
280 pointer_delta(last_card_start, _top) < AlignmentReserve) {
281 last_card_start = _top;
282 }
283 if (last_card_start < _hard_end) {
284 MemRegion last_card_prefix(last_card_start, _hard_end);
285 fill_region_with_block(last_card_prefix, false);
286 }
287 _hard_end = last_card_start;
288 _end = MAX2(_top, _hard_end - AlignmentReserve);
289 _true_end = _hard_end;
290 assert(_end <= _hard_end, "Invariant.");
291 }
293 // At this point:
294 // 1) we had a filler object from the original top to hard_end.
295 // 2) We've filled in any partial cards at the front and back.
296 if (pre_top < _hard_end) {
297 // Now we can reset the _bt to do allocation in the given area.
298 MemRegion new_filler(pre_top, _hard_end);
299 fill_region_with_block(new_filler, false);
300 _top = pre_top + ParGCAllocBuffer::FillerHeaderSize;
301 // If there's no space left, don't retain.
302 if (_top >= _end) {
303 _retained = false;
304 invalidate();
305 return;
306 }
307 _retained_filler = MemRegion(pre_top, _top);
308 _bt.set_region(MemRegion(_top, _hard_end));
309 _bt.initialize_threshold();
310 assert(_bt.threshold() > _top, "initialize_threshold failed!");
312 // There may be other reasons for queries into the middle of the
313 // filler object. When such queries are done in parallel with
314 // allocation, bad things can happen, if the query involves object
315 // iteration. So we ensure that such queries do not involve object
316 // iteration, by putting another filler object on the boundaries of
317 // such queries. One such is the object spanning a parallel card
318 // chunk boundary.
320 // "chunk_boundary" is the address of the first chunk boundary less
321 // than "hard_end".
322 HeapWord* chunk_boundary =
323 (HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes);
324 assert(chunk_boundary < _hard_end, "Or else above did not work.");
325 assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve,
326 "Consequence of last card handling above.");
328 if (_top <= chunk_boundary) {
329 assert(_true_end == _hard_end, "Invariant.");
330 while (_top <= chunk_boundary) {
331 assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve,
332 "Consequence of last card handling above.");
333 _bt.BlockOffsetArray::alloc_block(chunk_boundary, _hard_end);
334 CollectedHeap::fill_with_object(chunk_boundary, _hard_end);
335 _hard_end = chunk_boundary;
336 chunk_boundary -= ChunkSizeInWords;
337 }
338 _end = _hard_end - AlignmentReserve;
339 assert(_top <= _end, "Invariant.");
340 // Now reset the initial filler chunk so it doesn't overlap with
341 // the one(s) inserted above.
342 MemRegion new_filler(pre_top, _hard_end);
343 fill_region_with_block(new_filler, false);
344 }
345 } else {
346 _retained = false;
347 invalidate();
348 }
349 } else {
350 assert(!end_of_gc ||
351 (!_retained && _true_end == _hard_end), "Checking.");
352 }
353 assert(_end <= _hard_end, "Invariant.");
354 assert(_top < _end || _top == _hard_end, "Invariant");
355 }