1 /*
   2  * Copyright (c) 2015, 2019, 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/shenandoah/shenandoahAsserts.hpp"
  28 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp"
  29 #include "gc/shenandoah/shenandoahConcurrentMark.hpp"
  30 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
  31 #include "gc/shenandoah/shenandoahStringDedup.inline.hpp"
  32 #include "gc/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>
  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_except(NULL, obj, _heap->is_concurrent_traversal_in_progress() && _heap->cancelled_gc());
  42   shenandoah_assert_marked(NULL, obj);
  43   shenandoah_assert_not_in_cset_except(NULL, obj, _heap->cancelled_gc());
  44 
  45   if (task->is_not_chunked()) {
  46     if (obj->is_instance()) {
  47       // Case 1: Normal oop, process as usual.
  48       obj->oop_iterate(cl);
  49     } else if (obj->is_objArray()) {
  50       // Case 2: Object array instance and no chunk is set. Must be the first
  51       // time we visit it, start the chunked processing.
  52       do_chunked_array_start<T>(q, cl, obj);
  53     } else {
  54       // Case 3: Primitive array. Do nothing, no oops there. We use the same
  55       // performance tweak TypeArrayKlass::oop_oop_iterate_impl is using:
  56       // We skip iterating over the klass pointer since we know that
  57       // Universe::TypeArrayKlass never moves.
  58       assert (obj->is_typeArray(), "should be type array");
  59     }
  60     // Count liveness the last: push the outstanding work to the queues first
  61     count_liveness(live_data, obj);
  62   } else {
  63     // Case 4: Array chunk, has sensible chunk id. Process it.
  64     do_chunked_array<T>(q, cl, obj, task->chunk(), task->pow());
  65   }
  66 }
  67 
  68 inline void ShenandoahConcurrentMark::count_liveness(jushort* live_data, oop obj) {
  69   size_t region_idx = _heap->heap_region_index_containing(obj);
  70   ShenandoahHeapRegion* region = _heap->get_region(region_idx);
  71   size_t size = obj->size();
  72 
  73   if (!region->is_humongous_start()) {
  74     assert(!region->is_humongous(), "Cannot have continuations here");
  75     size_t max = (1 << (sizeof(jushort) * 8)) - 1;
  76     if (size >= max) {
  77       // too big, add to region data directly
  78       region->increase_live_data_gc_words(size);
  79     } else {
  80       jushort cur = live_data[region_idx];
  81       size_t new_val = cur + size;
  82       if (new_val >= max) {
  83         // overflow, flush to region data
  84         region->increase_live_data_gc_words(new_val);
  85         live_data[region_idx] = 0;
  86       } else {
  87         // still good, remember in locals
  88         live_data[region_idx] = (jushort) new_val;
  89       }
  90     }
  91   } else {
  92     shenandoah_assert_in_correct_region(NULL, obj);
  93     size_t num_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
  94 
  95     for (size_t i = region_idx; i < region_idx + num_regions; i++) {
  96       ShenandoahHeapRegion* chain_reg = _heap->get_region(i);
  97       assert(chain_reg->is_humongous(), "Expecting a humongous region");
  98       chain_reg->increase_live_data_gc_words(chain_reg->used() >> LogHeapWordSize);
  99     }
 100   }
 101 }
 102 
 103 template <class T>
 104 inline void ShenandoahConcurrentMark::do_chunked_array_start(ShenandoahObjToScanQueue* q, T* cl, oop obj) {
 105   assert(obj->is_objArray(), "expect object array");
 106   objArrayOop array = objArrayOop(obj);
 107   int len = array->length();
 108 
 109   if (len <= (int) ObjArrayMarkingStride*2) {
 110     // A few slices only, process directly
 111     array->oop_iterate_range(cl, 0, len);
 112   } else {
 113     int bits = log2_long(len);
 114     // Compensate for non-power-of-two arrays, cover the array in excess:
 115     if (len != (1 << bits)) bits++;
 116 
 117     // Only allow full chunks on the queue. This frees do_chunked_array() from checking from/to
 118     // boundaries against array->length(), touching the array header on every chunk.
 119     //
 120     // To do this, we cut the prefix in full-sized chunks, and submit them on the queue.
 121     // If the array is not divided in chunk sizes, then there would be an irregular tail,
 122     // which we will process separately.
 123 
 124     int last_idx = 0;
 125 
 126     int chunk = 1;
 127     int pow = bits;
 128 
 129     // Handle overflow
 130     if (pow >= 31) {
 131       assert (pow == 31, "sanity");
 132       pow--;
 133       chunk = 2;
 134       last_idx = (1 << pow);
 135       bool pushed = q->push(ShenandoahMarkTask(array, 1, pow));
 136       assert(pushed, "overflow queue should always succeed pushing");
 137     }
 138 
 139     // Split out tasks, as suggested in ObjArrayChunkedTask docs. Record the last
 140     // successful right boundary to figure out the irregular tail.
 141     while ((1 << pow) > (int)ObjArrayMarkingStride &&
 142            (chunk*2 < ShenandoahMarkTask::chunk_size())) {
 143       pow--;
 144       int left_chunk = chunk*2 - 1;
 145       int right_chunk = chunk*2;
 146       int left_chunk_end = left_chunk * (1 << pow);
 147       if (left_chunk_end < len) {
 148         bool pushed = q->push(ShenandoahMarkTask(array, left_chunk, pow));
 149         assert(pushed, "overflow queue should always succeed pushing");
 150         chunk = right_chunk;
 151         last_idx = left_chunk_end;
 152       } else {
 153         chunk = left_chunk;
 154       }
 155     }
 156 
 157     // Process the irregular tail, if present
 158     int from = last_idx;
 159     if (from < len) {
 160       array->oop_iterate_range(cl, from, len);
 161     }
 162   }
 163 }
 164 
 165 template <class T>
 166 inline void ShenandoahConcurrentMark::do_chunked_array(ShenandoahObjToScanQueue* q, T* cl, oop obj, int chunk, int pow) {
 167   assert(obj->is_objArray(), "expect object array");
 168   objArrayOop array = objArrayOop(obj);
 169 
 170   assert (ObjArrayMarkingStride > 0, "sanity");
 171 
 172   // Split out tasks, as suggested in ObjArrayChunkedTask docs. Avoid pushing tasks that
 173   // are known to start beyond the array.
 174   while ((1 << pow) > (int)ObjArrayMarkingStride && (chunk*2 < ShenandoahMarkTask::chunk_size())) {
 175     pow--;
 176     chunk *= 2;
 177     bool pushed = q->push(ShenandoahMarkTask(array, chunk - 1, pow));
 178     assert(pushed, "overflow queue should always succeed pushing");
 179   }
 180 
 181   int chunk_size = 1 << pow;
 182 
 183   int from = (chunk - 1) * chunk_size;
 184   int to = chunk * chunk_size;
 185 
 186 #ifdef ASSERT
 187   int len = array->length();
 188   assert (0 <= from && from < len, "from is sane: %d/%d", from, len);
 189   assert (0 < to && to <= len, "to is sane: %d/%d", to, len);
 190 #endif
 191 
 192   array->oop_iterate_range(cl, from, to);
 193 }
 194 
 195 class ShenandoahSATBBufferClosure : public SATBBufferClosure {
 196 private:
 197   ShenandoahObjToScanQueue* _queue;
 198   ShenandoahHeap* _heap;
 199   ShenandoahMarkingContext* const _mark_context;
 200 public:
 201   ShenandoahSATBBufferClosure(ShenandoahObjToScanQueue* q) :
 202     _queue(q),
 203     _heap(ShenandoahHeap::heap()),
 204     _mark_context(_heap->marking_context())
 205   {
 206   }
 207 
 208   void do_buffer(void **buffer, size_t size) {
 209     if (_heap->has_forwarded_objects()) {
 210       if (ShenandoahStringDedup::is_enabled()) {
 211         do_buffer_impl<RESOLVE, ENQUEUE_DEDUP>(buffer, size);
 212       } else {
 213         do_buffer_impl<RESOLVE, NO_DEDUP>(buffer, size);
 214       }
 215     } else {
 216       if (ShenandoahStringDedup::is_enabled()) {
 217         do_buffer_impl<NONE, ENQUEUE_DEDUP>(buffer, size);
 218       } else {
 219         do_buffer_impl<NONE, NO_DEDUP>(buffer, size);
 220       }
 221     }
 222   }
 223 
 224   template<UpdateRefsMode UPDATE_REFS, StringDedupMode STRING_DEDUP>
 225   void do_buffer_impl(void **buffer, size_t size) {
 226     for (size_t i = 0; i < size; ++i) {
 227       oop *p = (oop *) &buffer[i];
 228       ShenandoahConcurrentMark::mark_through_ref<oop, UPDATE_REFS, STRING_DEDUP>(p, _heap, _queue, _mark_context);
 229     }
 230   }
 231 };
 232 
 233 template<class T, UpdateRefsMode UPDATE_REFS, StringDedupMode STRING_DEDUP>
 234 inline void ShenandoahConcurrentMark::mark_through_ref(T *p, ShenandoahHeap* heap, ShenandoahObjToScanQueue* q, ShenandoahMarkingContext* const mark_context) {
 235   T o = RawAccess<>::oop_load(p);
 236   if (!CompressedOops::is_null(o)) {
 237     oop obj = CompressedOops::decode_not_null(o);
 238     switch (UPDATE_REFS) {
 239     case NONE:
 240       break;
 241     case RESOLVE:
 242       obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
 243       break;
 244     case SIMPLE:
 245       // We piggy-back reference updating to the marking tasks.
 246       obj = heap->update_with_forwarded_not_null(p, obj);
 247       break;
 248     case CONCURRENT:
 249       obj = heap->maybe_update_with_forwarded_not_null(p, obj);
 250       break;
 251     default:
 252       ShouldNotReachHere();
 253     }
 254 
 255     // Note: Only when concurrently updating references can obj be different
 256     // (that is, really different, not just different from-/to-space copies of the same)
 257     // from the one we originally loaded. Mutator thread can beat us by writing something
 258     // else into the location. In that case, we would mark through that updated value,
 259     // on the off-chance it is not handled by other means (e.g. via SATB). However,
 260     // if that write was NULL, we don't need to do anything else.
 261     if (UPDATE_REFS != CONCURRENT || !CompressedOops::is_null(obj)) {
 262       shenandoah_assert_not_forwarded(p, obj);
 263       shenandoah_assert_not_in_cset_except(p, obj, heap->cancelled_gc());
 264 
 265       if (mark_context->mark(obj)) {
 266         bool pushed = q->push(ShenandoahMarkTask(obj));
 267         assert(pushed, "overflow queue should always succeed pushing");
 268 
 269         if ((STRING_DEDUP == ENQUEUE_DEDUP) && ShenandoahStringDedup::is_candidate(obj)) {
 270           assert(ShenandoahStringDedup::is_enabled(), "Must be enabled");
 271           ShenandoahStringDedup::enqueue_candidate(obj);
 272         }
 273       }
 274 
 275       shenandoah_assert_marked(p, obj);
 276     }
 277   }
 278 }
 279 
 280 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHCONCURRENTMARK_INLINE_HPP