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/shenandoah/brooksPointer.hpp"
  28 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp"
  29 #include "gc/shenandoah/shenandoahConcurrentMark.hpp"
  30 #include "memory/iterator.inline.hpp"
  31 #include "oops/oop.inline.hpp"
  32 #include "runtime/prefetch.inline.hpp"
  33 
  34 template <class T, bool COUNT_LIVENESS>
  35 void ShenandoahConcurrentMark::do_task(SCMObjToScanQueue* q, T* cl, jushort* live_data, SCMTask* task) {
  36   oop obj = task->obj();
  37 
  38   assert(obj != NULL, "expect non-null object");
  39 
  40   assert(oopDesc::unsafe_equals(obj, ShenandoahBarrierSet::resolve_oop_static_not_null(obj)), "expect forwarded obj in queue");
  41 
  42 #ifdef ASSERT
  43   if (! oopDesc::bs()->is_safe(obj)) {
  44     tty->print_cr("trying to mark obj: "PTR_FORMAT" (%s) in dirty region: ", p2i((HeapWord*) obj), BOOL_TO_STR(_heap->is_marked_next(obj)));
  45     //      _heap->heap_region_containing(obj)->print();
  46     //      _heap->print_heap_regions();
  47   }
  48 #endif
  49   assert(_heap->cancelled_concgc()
  50          || oopDesc::bs()->is_safe(obj),
  51          "we don't want to mark objects in from-space");
  52   assert(_heap->is_in(obj), "referenced objects must be in the heap. No?");
  53   assert(_heap->is_marked_next(obj), "only marked objects on task queue");
  54 
  55   if (task->is_not_chunked()) {
  56     if (COUNT_LIVENESS) count_liveness(live_data, obj);
  57     if (obj->is_instance()) {
  58       // Case 1: Normal oop, process as usual.
  59       obj->oop_iterate(cl);
  60     } else if (obj->is_objArray()) {
  61       // Case 2: Object array instance and no chunk is set. Must be the first
  62       // time we visit it, start the chunked processing.
  63       do_chunked_array_start<T>(q, cl, obj);
  64     } else {
  65       // Case 3: Primitive array. Do nothing, no oops there. We use the same
  66       // performance tweak TypeArrayKlass::oop_oop_iterate_impl is using:
  67       // We skip iterating over the klass pointer since we know that
  68       // Universe::TypeArrayKlass never moves.
  69       assert (obj->is_typeArray(), "should be type array");
  70     }
  71   } else {
  72     // Case 4: Array chunk, has sensible chunk id. Process it.
  73     do_chunked_array<T>(q, cl, obj, task->chunk(), task->pow());
  74   }
  75 }
  76 
  77 inline void ShenandoahConcurrentMark::count_liveness(jushort* live_data, oop obj) {
  78   uint region_idx = _heap->heap_region_index_containing(obj);
  79   jushort cur = live_data[region_idx];
  80   int size = obj->size() + BrooksPointer::word_size();
  81   int max = (1 << (sizeof(jushort) * 8)) - 1;
  82   if (size >= max) {
  83     // too big, add to region data directly
  84     _heap->regions()->get_fast(region_idx)->increase_live_data_words(size);
  85   } else {
  86     int new_val = cur + size;
  87     if (new_val >= max) {
  88       // overflow, flush to region data
  89       _heap->regions()->get_fast(region_idx)->increase_live_data_words(new_val);
  90       live_data[region_idx] = 0;
  91     } else {
  92       // still good, remember in locals
  93       live_data[region_idx] = (jushort) new_val;
  94     }
  95   }
  96 }
  97 
  98 template <class T>
  99 inline void ShenandoahConcurrentMark::do_chunked_array_start(SCMObjToScanQueue* q, T* cl, oop obj) {
 100   assert(obj->is_objArray(), "expect object array");
 101   objArrayOop array = objArrayOop(obj);
 102   int len = array->length();
 103 
 104   if (len <= (int) ObjArrayMarkingStride*2) {
 105     // A few slices only, process directly
 106     array->oop_iterate_range(cl, 0, len);
 107   } else {
 108     int bits = log2_long(len);
 109     // Compensate for non-power-of-two arrays, cover the array in excess:
 110     if (len != (1 << bits)) bits++;
 111 
 112     // Only allow full chunks on the queue. This frees do_chunked_array() from checking from/to
 113     // boundaries against array->length(), touching the array header on every chunk.
 114     //
 115     // To do this, we cut the prefix in full-sized chunks, and submit them on the queue.
 116     // If the array is not divided in chunk sizes, then there would be an irregular tail,
 117     // which we will process separately.
 118 
 119     int last_idx = 0;
 120 
 121     int chunk = 1;
 122     int pow = bits;
 123 
 124     // Handle overflow
 125     if (pow >= 31) {
 126       assert (pow == 31, "sanity");
 127       pow--;
 128       chunk = 2;
 129       last_idx = (1 << pow);
 130       bool pushed = q->push(SCMTask(array, 1, pow));
 131       assert(pushed, "overflow queue should always succeed pushing");
 132     }
 133 
 134     // Split out tasks, as suggested in ObjArrayChunkedTask docs. Record the last
 135     // successful right boundary to figure out the irregular tail.
 136     while ((1 << pow) > (int)ObjArrayMarkingStride &&
 137            (chunk*2 < SCMTask::chunk_size)) {
 138       pow--;
 139       int left_chunk = chunk*2 - 1;
 140       int right_chunk = chunk*2;
 141       int left_chunk_end = left_chunk * (1 << pow);
 142       if (left_chunk_end < len) {
 143         bool pushed = q->push(SCMTask(array, left_chunk, pow));
 144         assert(pushed, "overflow queue should always succeed pushing");
 145         chunk = right_chunk;
 146         last_idx = left_chunk_end;
 147       } else {
 148         chunk = left_chunk;
 149       }
 150     }
 151 
 152     // Process the irregular tail, if present
 153     int from = last_idx;
 154     if (from < len) {
 155       array->oop_iterate_range(cl, from, len);
 156     }
 157   }
 158 }
 159 
 160 template <class T>
 161 inline void ShenandoahConcurrentMark::do_chunked_array(SCMObjToScanQueue* q, T* cl, oop obj, int chunk, int pow) {
 162   assert(obj->is_objArray(), "expect object array");
 163   objArrayOop array = objArrayOop(obj);
 164 
 165   assert (ObjArrayMarkingStride > 0, "sanity");
 166 
 167   // Split out tasks, as suggested in ObjArrayChunkedTask docs. Avoid pushing tasks that
 168   // are known to start beyond the array.
 169   while ((1 << pow) > (int)ObjArrayMarkingStride && (chunk*2 < SCMTask::chunk_size)) {
 170     pow--;
 171     chunk *= 2;
 172     bool pushed = q->push(SCMTask(array, chunk - 1, pow));
 173     assert(pushed, "overflow queue should always succeed pushing");
 174   }
 175 
 176   int chunk_size = 1 << pow;
 177 
 178   int from = (chunk - 1) * chunk_size;
 179   int to = chunk * chunk_size;
 180 
 181 #ifdef ASSERT
 182   int len = array->length();
 183   assert (0 <= from && from < len, "from is sane: %d/%d", from, len);
 184   assert (0 < to && to <= len, "to is sane: %d/%d", to, len);
 185 #endif
 186 
 187   array->oop_iterate_range(cl, from, to);
 188 }
 189 
 190 inline bool ShenandoahConcurrentMark::try_queue(SCMObjToScanQueue* q, SCMTask &task) {
 191   return (q->pop_buffer(task) ||
 192           q->pop_local(task) ||
 193           q->pop_overflow(task));
 194 }
 195 
 196 class ShenandoahSATBBufferClosure : public SATBBufferClosure {
 197 private:
 198   SCMObjToScanQueue* _queue;
 199   ShenandoahHeap* _heap;
 200 public:
 201   ShenandoahSATBBufferClosure(SCMObjToScanQueue* q) :
 202     _queue(q), _heap(ShenandoahHeap::heap())
 203   {
 204   }
 205 
 206   void do_buffer(void** buffer, size_t size) {
 207     for (size_t i = 0; i < size; ++i) {
 208       oop* p = (oop*) &buffer[i];
 209       ShenandoahConcurrentMark::mark_through_ref<oop, RESOLVE, false>(p, _heap, _queue, NULL);
 210     }
 211   }
 212 };
 213 
 214 inline bool ShenandoahConcurrentMark::try_draining_satb_buffer(SCMObjToScanQueue *q, SCMTask &task) {
 215   ShenandoahSATBBufferClosure cl(q);
 216   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 217   bool had_refs = satb_mq_set.apply_closure_to_completed_buffer(&cl);
 218   return had_refs && try_queue(q, task);
 219 }
 220 
 221 template<class T, UpdateRefsMode UPDATE_REFS, bool UPDATE_MATRIX>
 222 inline void ShenandoahConcurrentMark::mark_through_ref(T *p, ShenandoahHeap* heap, SCMObjToScanQueue* q, ShenandoahConnectionMatrix* conn_matrix) {
 223   T o = oopDesc::load_heap_oop(p);
 224   if (! oopDesc::is_null(o)) {
 225     oop obj = oopDesc::decode_heap_oop_not_null(o);
 226     switch (UPDATE_REFS) {
 227     case NONE:
 228       break;
 229     case RESOLVE:
 230       obj = ShenandoahBarrierSet::resolve_oop_static_not_null(obj);
 231       break;
 232     case SIMPLE:
 233       // We piggy-back reference updating to the marking tasks.
 234       obj = heap->update_oop_ref_not_null(p, obj);
 235       break;
 236     case CONCURRENT:
 237       obj = heap->maybe_update_oop_ref_not_null(p, obj);
 238       break;
 239     default:
 240       ShouldNotReachHere();
 241     }
 242     assert(oopDesc::unsafe_equals(obj, ShenandoahBarrierSet::resolve_oop_static(obj)), "need to-space object here");
 243 
 244     // Note: Only when concurrently updating references can obj become NULL here.
 245     // It happens when a mutator thread beats us by writing another value. In that
 246     // case we don't need to do anything else.
 247     if (UPDATE_REFS != CONCURRENT || !oopDesc::is_null(obj)) {
 248       assert(!oopDesc::is_null(obj), "Must not be null here");
 249       assert(heap->is_in(obj), "We shouldn't be calling this on objects not in the heap: " PTR_FORMAT, p2i(obj));
 250       assert(oopDesc::bs()->is_safe(obj), "Only mark objects in from-space");
 251 
 252       if (UPDATE_MATRIX) {
 253         if (heap->is_in_reserved(p)) { // Could also be in CLD, when marking through metadata.
 254           uint from_idx = heap->heap_region_index_containing(p);
 255           uint to_idx = heap->heap_region_index_containing(obj);
 256           conn_matrix->set_connected(from_idx, to_idx, true);
 257         }
 258       }
 259 
 260       if (heap->mark_next(obj)) {
 261         log_develop_trace(gc, marking)("Marked obj: " PTR_FORMAT, p2i((HeapWord*) obj));
 262 
 263         bool pushed = q->push(SCMTask(obj));
 264         assert(pushed, "overflow queue should always succeed pushing");
 265       } else {
 266         log_develop_trace(gc, marking)("Failed to mark obj (already marked): " PTR_FORMAT, p2i((HeapWord*) obj));
 267         assert(heap->is_marked_next(obj), "Consistency: should be marked.");
 268       }
 269     }
 270   }
 271 }
 272 
 273 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHCONCURRENTMARK_INLINE_HPP