1 /*
2 * Copyright (c) 2015, 2018, Red Hat, Inc. All rights reserved.
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/shenandoahAsserts.hpp"
28 #include "gc_implementation/shenandoah/shenandoahBarrierSet.inline.hpp"
29 #include "gc_implementation/shenandoah/shenandoahConcurrentMark.hpp"
30 #include "gc_implementation/shenandoah/shenandoahMarkingContext.inline.hpp"
31 #include "gc_implementation/shenandoah/shenandoahHeap.inline.hpp"
32 #include "gc_implementation/shenandoah/shenandoahOopClosures.inline.hpp"
33 #include "gc_implementation/shenandoah/shenandoahStringDedup.hpp"
34 #include "gc_implementation/shenandoah/shenandoahTaskqueue.inline.hpp"
35 #include "memory/iterator.inline.hpp"
36 #include "oops/oop.inline.hpp"
37 #include "runtime/prefetch.inline.hpp"
38
39 template <class T>
40 void ShenandoahConcurrentMark::do_task(ShenandoahObjToScanQueue* q, T* cl, jushort* live_data, ShenandoahMarkTask* task) {
41 oop obj = task->obj();
42
43 shenandoah_assert_not_forwarded_except(NULL, obj, _heap->is_concurrent_traversal_in_progress() && _heap->cancelled_gc());
44 shenandoah_assert_marked(NULL, obj);
45 shenandoah_assert_not_in_cset_except(NULL, obj, _heap->cancelled_gc());
46
47 if (task->is_not_chunked()) {
48 if (obj->is_instance()) {
49 // Case 1: Normal oop, process as usual.
50 obj->oop_iterate(cl);
51 } else if (obj->is_objArray()) {
52 // Case 2: Object array instance and no chunk is set. Must be the first
53 // time we visit it, start the chunked processing.
54 do_chunked_array_start<T>(q, cl, obj);
55 } else {
56 // Case 3: Primitive array. Do nothing, no oops there. We use the same
57 // performance tweak TypeArrayKlass::oop_oop_iterate_impl is using:
58 // We skip iterating over the klass pointer since we know that
59 // Universe::TypeArrayKlass never moves.
60 assert (obj->is_typeArray(), "should be type array");
61 }
62 // Count liveness the last: push the outstanding work to the queues first
63 count_liveness(live_data, obj);
64 } else {
65 // Case 4: Array chunk, has sensible chunk id. Process it.
66 do_chunked_array<T>(q, cl, obj, task->chunk(), task->pow());
67 }
68 }
69
70 inline void ShenandoahConcurrentMark::count_liveness(jushort* live_data, oop obj) {
71 size_t region_idx = _heap->heap_region_index_containing(obj);
72 ShenandoahHeapRegion* region = _heap->get_region(region_idx);
73 size_t size = obj->size();
74
75 if (!region->is_humongous_start()) {
76 assert(!region->is_humongous(), "Cannot have continuations here");
77 size_t max = (1 << (sizeof(jushort) * 8)) - 1;
78 if (size >= max) {
79 // too big, add to region data directly
80 region->increase_live_data_gc_words(size);
81 } else {
82 jushort cur = live_data[region_idx];
83 size_t new_val = cur + size;
84 if (new_val >= max) {
85 // overflow, flush to region data
86 region->increase_live_data_gc_words(new_val);
87 live_data[region_idx] = 0;
88 } else {
89 // still good, remember in locals
90 live_data[region_idx] = (jushort) new_val;
91 }
92 }
93 } else {
94 shenandoah_assert_in_correct_region(NULL, obj);
95 size_t num_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
96
97 for (size_t i = region_idx; i < region_idx + num_regions; i++) {
98 ShenandoahHeapRegion* chain_reg = _heap->get_region(i);
99 assert(chain_reg->is_humongous(), "Expecting a humongous region");
100 chain_reg->increase_live_data_gc_words(chain_reg->used() >> LogHeapWordSize);
101 }
102 }
103 }
104
105 template <class T>
106 inline void ShenandoahConcurrentMark::do_chunked_array_start(ShenandoahObjToScanQueue* q, T* cl, oop obj) {
107 assert(obj->is_objArray(), "expect object array");
108 objArrayOop array = objArrayOop(obj);
109 int len = array->length();
110
111 if (len <= (int) ObjArrayMarkingStride*2) {
112 // A few slices only, process directly
113 array->oop_iterate_range(cl, 0, len);
114 } else {
115 int bits = log2_long(len);
116 // Compensate for non-power-of-two arrays, cover the array in excess:
117 if (len != (1 << bits)) bits++;
118
119 // Only allow full chunks on the queue. This frees do_chunked_array() from checking from/to
120 // boundaries against array->length(), touching the array header on every chunk.
121 //
122 // To do this, we cut the prefix in full-sized chunks, and submit them on the queue.
123 // If the array is not divided in chunk sizes, then there would be an irregular tail,
124 // which we will process separately.
125
126 int last_idx = 0;
127
128 int chunk = 1;
129 int pow = bits;
130
131 // Handle overflow
132 if (pow >= 31) {
133 assert (pow == 31, "sanity");
134 pow--;
135 chunk = 2;
136 last_idx = (1 << pow);
137 bool pushed = q->push(ShenandoahMarkTask(array, 1, pow));
138 assert(pushed, "overflow queue should always succeed pushing");
139 }
140
141 // Split out tasks, as suggested in ObjArrayChunkedTask docs. Record the last
142 // successful right boundary to figure out the irregular tail.
143 while ((1 << pow) > (int)ObjArrayMarkingStride &&
144 (chunk*2 < ShenandoahMarkTask::chunk_size())) {
145 pow--;
146 int left_chunk = chunk*2 - 1;
147 int right_chunk = chunk*2;
148 int left_chunk_end = left_chunk * (1 << pow);
149 if (left_chunk_end < len) {
150 bool pushed = q->push(ShenandoahMarkTask(array, left_chunk, pow));
151 assert(pushed, "overflow queue should always succeed pushing");
152 chunk = right_chunk;
153 last_idx = left_chunk_end;
154 } else {
155 chunk = left_chunk;
156 }
157 }
158
159 // Process the irregular tail, if present
160 int from = last_idx;
161 if (from < len) {
162 array->oop_iterate_range(cl, from, len);
163 }
164 }
165 }
166
167 template <class T>
168 inline void ShenandoahConcurrentMark::do_chunked_array(ShenandoahObjToScanQueue* q, T* cl, oop obj, int chunk, int pow) {
169 assert(obj->is_objArray(), "expect object array");
170 objArrayOop array = objArrayOop(obj);
171
172 assert (ObjArrayMarkingStride > 0, "sanity");
173
174 // Split out tasks, as suggested in ObjArrayChunkedTask docs. Avoid pushing tasks that
175 // are known to start beyond the array.
176 while ((1 << pow) > (int)ObjArrayMarkingStride && (chunk*2 < ShenandoahMarkTask::chunk_size())) {
177 pow--;
178 chunk *= 2;
179 bool pushed = q->push(ShenandoahMarkTask(array, chunk - 1, pow));
180 assert(pushed, "overflow queue should always succeed pushing");
181 }
182
183 int chunk_size = 1 << pow;
184
185 int from = (chunk - 1) * chunk_size;
186 int to = chunk * chunk_size;
187
188 #ifdef ASSERT
189 int len = array->length();
190 assert (0 <= from && from < len, err_msg("from is sane: %d/%d", from, len));
191 assert (0 < to && to <= len, err_msg("to is sane: %d/%d", to, len));
192 #endif
193
194 array->oop_iterate_range(cl, from, to);
195 }
196
197 class ShenandoahSATBBufferClosure : public SATBBufferClosure {
198 private:
199 ShenandoahObjToScanQueue* _queue;
200 ShenandoahStrDedupQueue* _dedup_queue;
201 ShenandoahHeap* _heap;
202 ShenandoahMarkingContext* const _mark_context;
203 public:
204 ShenandoahSATBBufferClosure(ShenandoahObjToScanQueue* q, ShenandoahStrDedupQueue* dq) :
205 _queue(q),
206 _dedup_queue(dq),
207 _heap(ShenandoahHeap::heap()),
208 _mark_context(_heap->marking_context())
209 {
210 }
211
212 void do_buffer(void **buffer, size_t size) {
213 if (_heap->has_forwarded_objects()) {
214 if (ShenandoahStringDedup::is_enabled()) {
215 do_buffer_impl<RESOLVE, ENQUEUE_DEDUP>(buffer, size);
216 } else {
217 do_buffer_impl<RESOLVE, NO_DEDUP>(buffer, size);
218 }
219 } else {
220 if (ShenandoahStringDedup::is_enabled()) {
221 do_buffer_impl<NONE, ENQUEUE_DEDUP>(buffer, size);
222 } else {
223 do_buffer_impl<NONE, NO_DEDUP>(buffer, size);
224 }
225 }
226 }
227
228 template<UpdateRefsMode UPDATE_REFS, StringDedupMode STRING_DEDUP>
229 void do_buffer_impl(void **buffer, size_t size) {
230 for (size_t i = 0; i < size; ++i) {
231 oop *p = (oop *) &buffer[i];
232 ShenandoahConcurrentMark::mark_through_ref<oop, UPDATE_REFS, STRING_DEDUP>(p, _heap, _queue, _mark_context, _dedup_queue);
233 }
234 }
235 };
236
237 template<class T, UpdateRefsMode UPDATE_REFS, StringDedupMode STRING_DEDUP>
238 inline void ShenandoahConcurrentMark::mark_through_ref(T *p, ShenandoahHeap* heap, ShenandoahObjToScanQueue* q, ShenandoahMarkingContext* const mark_context, ShenandoahStrDedupQueue* dq) {
239 T o = oopDesc::load_heap_oop(p);
240 if (! oopDesc::is_null(o)) {
241 oop obj = oopDesc::decode_heap_oop_not_null(o);
242 switch (UPDATE_REFS) {
243 case NONE:
244 break;
245 case RESOLVE:
246 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
247 break;
248 case SIMPLE:
249 // We piggy-back reference updating to the marking tasks.
250 obj = heap->update_with_forwarded_not_null(p, obj);
251 break;
252 case CONCURRENT:
253 obj = heap->maybe_update_with_forwarded_not_null(p, obj);
254 break;
255 default:
256 ShouldNotReachHere();
257 }
258
259 // Note: Only when concurrently updating references can obj be different
260 // (that is, really different, not just different from-/to-space copies of the same)
261 // from the one we originally loaded. Mutator thread can beat us by writing something
262 // else into the location. In that case, we would mark through that updated value,
263 // on the off-chance it is not handled by other means (e.g. via SATB). However,
264 // if that write was NULL, we don't need to do anything else.
265 if (UPDATE_REFS != CONCURRENT || !oopDesc::is_null(obj)) {
266 shenandoah_assert_not_forwarded(p, obj);
267 shenandoah_assert_not_in_cset_except(p, obj, heap->cancelled_gc());
268
269 if (mark_context->mark(obj)) {
270 bool pushed = q->push(ShenandoahMarkTask(obj));
271 assert(pushed, "overflow queue should always succeed pushing");
272
273 if ((STRING_DEDUP == ENQUEUE_DEDUP) && ShenandoahStringDedup::is_candidate(obj)) {
274 assert(ShenandoahStringDedup::is_enabled(), "Must be enabled");
275 assert(dq != NULL, "Dedup queue not set");
276 ShenandoahStringDedup::enqueue_candidate(obj, dq);
277 }
278 }
279
280 shenandoah_assert_marked(p, obj);
281 }
282 }
283 }
284
285 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHCONCURRENTMARK_INLINE_HPP