1 /*
2 * Copyright (c) 2015, Red Hat, Inc. and/or its affiliates.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #ifndef SHARE_VM_GC_SHENANDOAH_SHENANDOAHCONCURRENTMARK_INLINE_HPP
25 #define SHARE_VM_GC_SHENANDOAH_SHENANDOAHCONCURRENTMARK_INLINE_HPP
26
27 #include "gc_implementation/shenandoah/brooksPointer.hpp"
28 #include "gc_implementation/shenandoah/shenandoahAsserts.hpp"
29 #include "gc_implementation/shenandoah/shenandoahBarrierSet.inline.hpp"
30 #include "gc_implementation/shenandoah/shenandoahConcurrentMark.hpp"
31 #include "gc_implementation/shenandoah/shenandoahHeap.inline.hpp"
32 #include "gc_implementation/shenandoah/shenandoahTaskqueue.inline.hpp"
33 #include "memory/iterator.inline.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "runtime/prefetch.inline.hpp"
36
37 template <class T, bool COUNT_LIVENESS>
38 void ShenandoahConcurrentMark::do_task(ShenandoahObjToScanQueue* q, T* cl, jushort* live_data, ShenandoahMarkTask* task) {
39 oop obj = task->obj();
40
41 shenandoah_assert_not_forwarded(NULL, obj);
42 shenandoah_assert_marked_next(NULL, obj);
43 shenandoah_assert_not_in_cset_except(NULL, obj, _heap->cancelled_concgc());
44
45 if (task->is_not_chunked()) {
46 if (COUNT_LIVENESS) count_liveness(live_data, obj);
47 if (obj->is_instance()) {
48 // Case 1: Normal oop, process as usual.
49 obj->oop_iterate(cl);
50 } else if (obj->is_objArray()) {
51 // Case 2: Object array instance and no chunk is set. Must be the first
52 // time we visit it, start the chunked processing.
53 do_chunked_array_start<T>(q, cl, obj);
54 } else {
55 // Case 3: Primitive array. Do nothing, no oops there. We use the same
56 // performance tweak TypeArrayKlass::oop_oop_iterate_impl is using:
57 // We skip iterating over the klass pointer since we know that
58 // Universe::TypeArrayKlass never moves.
59 assert (obj->is_typeArray(), "should be type array");
60 }
61 } else {
62 // Case 4: Array chunk, has sensible chunk id. Process it.
63 do_chunked_array<T>(q, cl, obj, task->chunk(), task->pow());
64 }
65 }
66
67 inline void ShenandoahConcurrentMark::count_liveness(jushort* live_data, oop obj) {
68 size_t region_idx = _heap->heap_region_index_containing(obj);
69 ShenandoahHeapRegion* region = _heap->get_region(region_idx);
70 if (!region->is_humongous_start()) {
71 assert(!region->is_humongous(), "Cannot have continuations here");
72 jushort cur = live_data[region_idx];
73 size_t size = obj->size() + BrooksPointer::word_size();
74 size_t max = (1 << (sizeof(jushort) * 8)) - 1;
75 if (size >= max) {
76 // too big, add to region data directly
77 region->increase_live_data_gc_words(size);
78 } else {
79 size_t new_val = cur + size;
80 if (new_val >= max) {
81 // overflow, flush to region data
82 region->increase_live_data_gc_words(new_val);
83 live_data[region_idx] = 0;
84 } else {
85 // still good, remember in locals
86 live_data[region_idx] = (jushort) new_val;
87 }
88 }
89 } else {
90 count_liveness_humongous(obj);
91 }
92 }
93
94 inline void ShenandoahConcurrentMark::count_liveness_humongous(oop obj) {
95 shenandoah_assert_in_correct_region(NULL, obj);
96 size_t region_idx = _heap->heap_region_index_containing(obj);
97 size_t size = obj->size() + BrooksPointer::word_size();
98 size_t num_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
99
100 for (size_t i = region_idx; i < region_idx + num_regions; i++) {
101 ShenandoahHeapRegion* chain_reg = _heap->get_region(i);
102 assert(chain_reg->is_humongous(), "Expecting a humongous region");
103 chain_reg->increase_live_data_gc_words(chain_reg->used() >> LogHeapWordSize);
104 }
105 }
106
107 template <class T>
108 inline void ShenandoahConcurrentMark::do_chunked_array_start(ShenandoahObjToScanQueue* q, T* cl, oop obj) {
109 assert(obj->is_objArray(), "expect object array");
110 objArrayOop array = objArrayOop(obj);
111 int len = array->length();
112
113 if (len <= (int) ObjArrayMarkingStride*2) {
114 // A few slices only, process directly
115 array->oop_iterate_range(cl, 0, len);
116 } else {
117 int bits = log2_long(len);
118 // Compensate for non-power-of-two arrays, cover the array in excess:
119 if (len != (1 << bits)) bits++;
120
121 // Only allow full chunks on the queue. This frees do_chunked_array() from checking from/to
122 // boundaries against array->length(), touching the array header on every chunk.
123 //
124 // To do this, we cut the prefix in full-sized chunks, and submit them on the queue.
125 // If the array is not divided in chunk sizes, then there would be an irregular tail,
126 // which we will process separately.
127
128 int last_idx = 0;
129
130 int chunk = 1;
131 int pow = bits;
132
133 // Handle overflow
134 if (pow >= 31) {
135 assert (pow == 31, "sanity");
136 pow--;
137 chunk = 2;
138 last_idx = (1 << pow);
139 bool pushed = q->push(ShenandoahMarkTask(array, 1, pow));
140 assert(pushed, "overflow queue should always succeed pushing");
141 }
142
143 // Split out tasks, as suggested in ObjArrayChunkedTask docs. Record the last
144 // successful right boundary to figure out the irregular tail.
145 while ((1 << pow) > (int)ObjArrayMarkingStride &&
146 (chunk*2 < ShenandoahMarkTask::chunk_size())) {
147 pow--;
148 int left_chunk = chunk*2 - 1;
149 int right_chunk = chunk*2;
150 int left_chunk_end = left_chunk * (1 << pow);
151 if (left_chunk_end < len) {
152 bool pushed = q->push(ShenandoahMarkTask(array, left_chunk, pow));
153 assert(pushed, "overflow queue should always succeed pushing");
154 chunk = right_chunk;
155 last_idx = left_chunk_end;
156 } else {
157 chunk = left_chunk;
158 }
159 }
160
161 // Process the irregular tail, if present
162 int from = last_idx;
163 if (from < len) {
164 array->oop_iterate_range(cl, from, len);
165 }
166 }
167 }
168
169 template <class T>
170 inline void ShenandoahConcurrentMark::do_chunked_array(ShenandoahObjToScanQueue* q, T* cl, oop obj, int chunk, int pow) {
171 assert(obj->is_objArray(), "expect object array");
172 objArrayOop array = objArrayOop(obj);
173
174 assert (ObjArrayMarkingStride > 0, "sanity");
175
176 // Split out tasks, as suggested in ObjArrayChunkedTask docs. Avoid pushing tasks that
177 // are known to start beyond the array.
178 while ((1 << pow) > (int)ObjArrayMarkingStride && (chunk*2 < ShenandoahMarkTask::chunk_size())) {
179 pow--;
180 chunk *= 2;
181 bool pushed = q->push(ShenandoahMarkTask(array, chunk - 1, pow));
182 assert(pushed, "overflow queue should always succeed pushing");
183 }
184
185 int chunk_size = 1 << pow;
186
187 int from = (chunk - 1) * chunk_size;
188 int to = chunk * chunk_size;
189
190 #ifdef ASSERT
191 int len = array->length();
192 assert (0 <= from && from < len, err_msg("from is sane: %d/%d", from, len));
193 assert (0 < to && to <= len, err_msg("to is sane: %d/%d", to, len));
194 #endif
195
196 array->oop_iterate_range(cl, from, to);
197 }
198
199 inline bool ShenandoahConcurrentMark::try_queue(ShenandoahObjToScanQueue* q, ShenandoahMarkTask &task) {
200 return (q->pop_buffer(task) ||
201 q->pop_local(task) ||
202 q->pop_overflow(task));
203 }
204
205 class ShenandoahSATBBufferClosure : public SATBBufferClosure {
206 private:
207 ShenandoahObjToScanQueue* _queue;
208 ShenandoahHeap* _heap;
209 public:
210 ShenandoahSATBBufferClosure(ShenandoahObjToScanQueue* q) :
211 _queue(q), _heap(ShenandoahHeap::heap())
212 {
213 }
214
215 void do_buffer(void** buffer, size_t size) {
216 for (size_t i = 0; i < size; ++i) {
217 oop* p = (oop*) &buffer[i];
218 ShenandoahConcurrentMark::mark_through_ref<oop, RESOLVE>(p, _heap, _queue);
219 }
220 }
221 };
222
223 inline bool ShenandoahConcurrentMark::try_draining_satb_buffer(ShenandoahObjToScanQueue *q, ShenandoahMarkTask &task) {
224 ShenandoahSATBBufferClosure cl(q);
225 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
226 bool had_refs = satb_mq_set.apply_closure_to_completed_buffer(&cl);
227 return had_refs && try_queue(q, task);
228 }
229
230 template<class T, UpdateRefsMode UPDATE_REFS>
231 inline void ShenandoahConcurrentMark::mark_through_ref(T *p, ShenandoahHeap* heap, ShenandoahObjToScanQueue* q) {
232 T o = oopDesc::load_heap_oop(p);
233 if (! oopDesc::is_null(o)) {
234 oop obj = oopDesc::decode_heap_oop_not_null(o);
235 switch (UPDATE_REFS) {
236 case NONE:
237 break;
238 case RESOLVE:
239 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
240 break;
241 case SIMPLE:
242 // We piggy-back reference updating to the marking tasks.
243 obj = heap->update_with_forwarded_not_null(p, obj);
244 break;
245 case CONCURRENT:
246 obj = heap->maybe_update_with_forwarded_not_null(p, obj);
247 break;
248 default:
249 ShouldNotReachHere();
250 }
251
252 // Note: Only when concurrently updating references can obj become NULL here.
253 // It happens when a mutator thread beats us by writing another value. In that
254 // case we don't need to do anything else.
255 if (UPDATE_REFS != CONCURRENT || !oopDesc::is_null(obj)) {
256 shenandoah_assert_not_forwarded(p, obj);
257 shenandoah_assert_not_in_cset_except(p, obj, heap->cancelled_concgc());
258
259 if (heap->mark_next(obj)) {
260 bool pushed = q->push(ShenandoahMarkTask(obj));
261 assert(pushed, "overflow queue should always succeed pushing");
262 }
263
264 shenandoah_assert_marked_next(p, obj);
265 }
266 }
267 }
268
269 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHCONCURRENTMARK_INLINE_HPP