1 /*
   2  * Copyright (c) 2002, 2016, 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  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc/parallel/gcTaskManager.hpp"
  27 #include "gc/parallel/mutableSpace.hpp"
  28 #include "gc/parallel/parallelScavengeHeap.hpp"
  29 #include "gc/parallel/psOldGen.hpp"
  30 #include "gc/parallel/psPromotionManager.inline.hpp"
  31 #include "gc/parallel/psScavenge.inline.hpp"
  32 #include "gc/shared/gcTrace.hpp"
  33 #include "gc/shared/preservedMarks.inline.hpp"
  34 #include "gc/shared/taskqueue.inline.hpp"
  35 #include "logging/log.hpp"
  36 #include "memory/allocation.inline.hpp"
  37 #include "memory/memRegion.hpp"
  38 #include "memory/padded.inline.hpp"
  39 #include "memory/resourceArea.hpp"
  40 #include "oops/instanceKlass.inline.hpp"
  41 #include "oops/instanceMirrorKlass.inline.hpp"
  42 #include "oops/objArrayKlass.inline.hpp"
  43 #include "oops/oop.inline.hpp"
  44 
  45 PaddedEnd<PSPromotionManager>* PSPromotionManager::_manager_array = NULL;
  46 OopStarTaskQueueSet*           PSPromotionManager::_stack_array_depth = NULL;
  47 PreservedMarksSet*             PSPromotionManager::_preserved_marks_set = NULL;
  48 PSOldGen*                      PSPromotionManager::_old_gen = NULL;
  49 MutableSpace*                  PSPromotionManager::_young_space = NULL;
  50 
  51 void PSPromotionManager::initialize() {
  52   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
  53 
  54   _old_gen = heap->old_gen();
  55   _young_space = heap->young_gen()->to_space();
  56 
  57   const uint promotion_manager_num = ParallelGCThreads + 1;
  58 
  59   // To prevent false sharing, we pad the PSPromotionManagers
  60   // and make sure that the first instance starts at a cache line.
  61   assert(_manager_array == NULL, "Attempt to initialize twice");
  62   _manager_array = PaddedArray<PSPromotionManager, mtGC>::create_unfreeable(promotion_manager_num);
  63   guarantee(_manager_array != NULL, "Could not initialize promotion manager");
  64 
  65   _stack_array_depth = new OopStarTaskQueueSet(ParallelGCThreads);
  66   guarantee(_stack_array_depth != NULL, "Could not initialize promotion manager");
  67 
  68   // Create and register the PSPromotionManager(s) for the worker threads.
  69   for(uint i=0; i<ParallelGCThreads; i++) {
  70     stack_array_depth()->register_queue(i, _manager_array[i].claimed_stack_depth());
  71   }
  72   // The VMThread gets its own PSPromotionManager, which is not available
  73   // for work stealing.
  74 
  75   assert(_preserved_marks_set == NULL, "Attempt to initialize twice");
  76   _preserved_marks_set = new PreservedMarksSet(true /* in_c_heap */);
  77   guarantee(_preserved_marks_set != NULL, "Could not initialize preserved marks set");
  78   _preserved_marks_set->init(promotion_manager_num);
  79   for (uint i = 0; i < promotion_manager_num; i += 1) {
  80     _manager_array[i].register_preserved_marks(_preserved_marks_set->get(i));
  81   }
  82 }
  83 
  84 // Helper functions to get around the circular dependency between
  85 // psScavenge.inline.hpp and psPromotionManager.inline.hpp.
  86 bool PSPromotionManager::should_scavenge(oop* p, bool check_to_space) {
  87   return PSScavenge::should_scavenge(p, check_to_space);
  88 }
  89 bool PSPromotionManager::should_scavenge(narrowOop* p, bool check_to_space) {
  90   return PSScavenge::should_scavenge(p, check_to_space);
  91 }
  92 
  93 PSPromotionManager* PSPromotionManager::gc_thread_promotion_manager(uint index) {
  94   assert(index < ParallelGCThreads, "index out of range");
  95   assert(_manager_array != NULL, "Sanity");
  96   return &_manager_array[index];
  97 }
  98 
  99 PSPromotionManager* PSPromotionManager::vm_thread_promotion_manager() {
 100   assert(_manager_array != NULL, "Sanity");
 101   return &_manager_array[ParallelGCThreads];
 102 }
 103 
 104 void PSPromotionManager::pre_scavenge() {
 105   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 106 
 107   _preserved_marks_set->assert_empty();
 108   _young_space = heap->young_gen()->to_space();
 109 
 110   for(uint i=0; i<ParallelGCThreads+1; i++) {
 111     manager_array(i)->reset();
 112   }
 113 }
 114 
 115 bool PSPromotionManager::post_scavenge(YoungGCTracer& gc_tracer) {
 116   bool promotion_failure_occurred = false;
 117 
 118   TASKQUEUE_STATS_ONLY(print_taskqueue_stats());
 119   for (uint i = 0; i < ParallelGCThreads + 1; i++) {
 120     PSPromotionManager* manager = manager_array(i);
 121     assert(manager->claimed_stack_depth()->is_empty(), "should be empty");
 122     if (manager->_promotion_failed_info.has_failed()) {
 123       gc_tracer.report_promotion_failed(manager->_promotion_failed_info);
 124       promotion_failure_occurred = true;
 125     }
 126     manager->flush_labs();
 127   }
 128   if (!promotion_failure_occurred) {
 129     // If there was no promotion failure, the preserved mark stacks
 130     // should be empty.
 131     _preserved_marks_set->assert_empty();
 132   }
 133   return promotion_failure_occurred;
 134 }
 135 
 136 #if TASKQUEUE_STATS
 137 void
 138 PSPromotionManager::print_local_stats(outputStream* const out, uint i) const {
 139   #define FMT " " SIZE_FORMAT_W(10)
 140   out->print_cr("%3u" FMT FMT FMT FMT, i, _masked_pushes, _masked_steals,
 141                 _arrays_chunked, _array_chunks_processed);
 142   #undef FMT
 143 }
 144 
 145 static const char* const pm_stats_hdr[] = {
 146   "    --------masked-------     arrays      array",
 147   "thr       push      steal    chunked     chunks",
 148   "--- ---------- ---------- ---------- ----------"
 149 };
 150 
 151 void
 152 PSPromotionManager::print_taskqueue_stats() {
 153   if (!log_develop_is_enabled(Trace, gc, task, stats)) {
 154     return;
 155   }
 156   Log(gc, task, stats) log;
 157   ResourceMark rm;
 158   outputStream* out = log.trace_stream();
 159   out->print_cr("== GC Tasks Stats, GC %3d",
 160                 ParallelScavengeHeap::heap()->total_collections());
 161 
 162   TaskQueueStats totals;
 163   out->print("thr "); TaskQueueStats::print_header(1, out); out->cr();
 164   out->print("--- "); TaskQueueStats::print_header(2, out); out->cr();
 165   for (uint i = 0; i < ParallelGCThreads + 1; ++i) {
 166     TaskQueueStats& next = manager_array(i)->_claimed_stack_depth.stats;
 167     out->print("%3d ", i); next.print(out); out->cr();
 168     totals += next;
 169   }
 170   out->print("tot "); totals.print(out); out->cr();
 171 
 172   const uint hlines = sizeof(pm_stats_hdr) / sizeof(pm_stats_hdr[0]);
 173   for (uint i = 0; i < hlines; ++i) out->print_cr("%s", pm_stats_hdr[i]);
 174   for (uint i = 0; i < ParallelGCThreads + 1; ++i) {
 175     manager_array(i)->print_local_stats(out, i);
 176   }
 177 }
 178 
 179 void
 180 PSPromotionManager::reset_stats() {
 181   claimed_stack_depth()->stats.reset();
 182   _masked_pushes = _masked_steals = 0;
 183   _arrays_chunked = _array_chunks_processed = 0;
 184 }
 185 #endif // TASKQUEUE_STATS
 186 
 187 PSPromotionManager::PSPromotionManager() {
 188   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 189 
 190   // We set the old lab's start array.
 191   _old_lab.set_start_array(old_gen()->start_array());
 192 
 193   uint queue_size;
 194   claimed_stack_depth()->initialize();
 195   queue_size = claimed_stack_depth()->max_elems();
 196 
 197   _totally_drain = (ParallelGCThreads == 1) || (GCDrainStackTargetSize == 0);
 198   if (_totally_drain) {
 199     _target_stack_size = 0;
 200   } else {
 201     // don't let the target stack size to be more than 1/4 of the entries
 202     _target_stack_size = (uint) MIN2((uint) GCDrainStackTargetSize,
 203                                      (uint) (queue_size / 4));
 204   }
 205 
 206   _array_chunk_size = ParGCArrayScanChunk;
 207   // let's choose 1.5x the chunk size
 208   _min_array_size_for_chunking = 3 * _array_chunk_size / 2;
 209 
 210   _preserved_marks = NULL;
 211 
 212   reset();
 213 }
 214 
 215 void PSPromotionManager::reset() {
 216   assert(stacks_empty(), "reset of non-empty stack");
 217 
 218   // We need to get an assert in here to make sure the labs are always flushed.
 219 
 220   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 221 
 222   // Do not prefill the LAB's, save heap wastage!
 223   HeapWord* lab_base = young_space()->top();
 224   _young_lab.initialize(MemRegion(lab_base, (size_t)0));
 225   _young_gen_is_full = false;
 226 
 227   lab_base = old_gen()->object_space()->top();
 228   _old_lab.initialize(MemRegion(lab_base, (size_t)0));
 229   _old_gen_is_full = false;
 230 
 231   _promotion_failed_info.reset();
 232 
 233   TASKQUEUE_STATS_ONLY(reset_stats());
 234 }
 235 
 236 void PSPromotionManager::register_preserved_marks(PreservedMarks* preserved_marks) {
 237   assert(_preserved_marks == NULL, "do not set it twice");
 238   _preserved_marks = preserved_marks;
 239 }
 240 
 241 class ParRestoreGCTask : public GCTask {
 242 private:
 243   const uint _id;
 244   PreservedMarksSet* const _preserved_marks_set;
 245   volatile size_t* const _total_size_addr;
 246 
 247 public:
 248   virtual char* name() {
 249     return (char*) "preserved mark restoration task";
 250   }
 251 
 252   virtual void do_it(GCTaskManager* manager, uint which){
 253     _preserved_marks_set->get(_id)->restore_and_increment(_total_size_addr);
 254   }
 255 
 256   ParRestoreGCTask(uint id,
 257                    PreservedMarksSet* preserved_marks_set,
 258                    volatile size_t* total_size_addr)
 259     : _id(id),
 260       _preserved_marks_set(preserved_marks_set),
 261       _total_size_addr(total_size_addr) { }
 262 };
 263 
 264 class PSRestorePreservedMarksTaskExecutor : public RestorePreservedMarksTaskExecutor {
 265 private:
 266   GCTaskManager* _gc_task_manager;
 267 
 268 public:
 269   PSRestorePreservedMarksTaskExecutor(GCTaskManager* gc_task_manager)
 270       : _gc_task_manager(gc_task_manager) { }
 271 
 272   void restore(PreservedMarksSet* preserved_marks_set,
 273                volatile size_t* total_size_addr) {
 274     // GCTask / GCTaskQueue are ResourceObjs
 275     ResourceMark rm;
 276 
 277     GCTaskQueue* q = GCTaskQueue::create();
 278     for (uint i = 0; i < preserved_marks_set->num(); i += 1) {
 279       q->enqueue(new ParRestoreGCTask(i, preserved_marks_set, total_size_addr));
 280     }
 281     _gc_task_manager->execute_and_wait(q);
 282   }
 283 };
 284 
 285 void PSPromotionManager::restore_preserved_marks() {
 286   PSRestorePreservedMarksTaskExecutor task_executor(PSScavenge::gc_task_manager());
 287   _preserved_marks_set->restore(&task_executor);
 288 }
 289 
 290 void PSPromotionManager::drain_stacks_depth(bool totally_drain) {
 291   totally_drain = totally_drain || _totally_drain;
 292 
 293 #ifdef ASSERT
 294   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 295   MutableSpace* to_space = heap->young_gen()->to_space();
 296   MutableSpace* old_space = heap->old_gen()->object_space();
 297 #endif /* ASSERT */
 298 
 299   OopStarTaskQueue* const tq = claimed_stack_depth();
 300   do {
 301     StarTask p;
 302 
 303     // Drain overflow stack first, so other threads can steal from
 304     // claimed stack while we work.
 305     while (tq->pop_overflow(p)) {
 306       process_popped_location_depth(p);
 307     }
 308 
 309     if (totally_drain) {
 310       while (tq->pop_local(p)) {
 311         process_popped_location_depth(p);
 312       }
 313     } else {
 314       while (tq->size() > _target_stack_size && tq->pop_local(p)) {
 315         process_popped_location_depth(p);
 316       }
 317     }
 318   } while (totally_drain && !tq->taskqueue_empty() || !tq->overflow_empty());
 319 
 320   assert(!totally_drain || tq->taskqueue_empty(), "Sanity");
 321   assert(totally_drain || tq->size() <= _target_stack_size, "Sanity");
 322   assert(tq->overflow_empty(), "Sanity");
 323 }
 324 
 325 void PSPromotionManager::flush_labs() {
 326   assert(stacks_empty(), "Attempt to flush lab with live stack");
 327 
 328   // If either promotion lab fills up, we can flush the
 329   // lab but not refill it, so check first.
 330   assert(!_young_lab.is_flushed() || _young_gen_is_full, "Sanity");
 331   if (!_young_lab.is_flushed())
 332     _young_lab.flush();
 333 
 334   assert(!_old_lab.is_flushed() || _old_gen_is_full, "Sanity");
 335   if (!_old_lab.is_flushed())
 336     _old_lab.flush();
 337 
 338   // Let PSScavenge know if we overflowed
 339   if (_young_gen_is_full) {
 340     PSScavenge::set_survivor_overflow(true);
 341   }
 342 }
 343 
 344 template <class T> void PSPromotionManager::process_array_chunk_work(
 345                                                  oop obj,
 346                                                  int start, int end) {
 347   assert(start <= end, "invariant");
 348   T* const base      = (T*)objArrayOop(obj)->base();
 349   T* p               = base + start;
 350   T* const chunk_end = base + end;
 351   while (p < chunk_end) {
 352     if (PSScavenge::should_scavenge(p)) {
 353       claim_or_forward_depth(p);
 354     }
 355     ++p;
 356   }
 357 }
 358 
 359 void PSPromotionManager::process_array_chunk(oop old) {
 360   assert(PSChunkLargeArrays, "invariant");
 361   assert(old->is_objArray(), "invariant");
 362   assert(old->is_forwarded(), "invariant");
 363 
 364   TASKQUEUE_STATS_ONLY(++_array_chunks_processed);
 365 
 366   oop const obj = old->forwardee();
 367 
 368   int start;
 369   int const end = arrayOop(old)->length();
 370   if (end > (int) _min_array_size_for_chunking) {
 371     // we'll chunk more
 372     start = end - _array_chunk_size;
 373     assert(start > 0, "invariant");
 374     arrayOop(old)->set_length(start);
 375     push_depth(mask_chunked_array_oop(old));
 376     TASKQUEUE_STATS_ONLY(++_masked_pushes);
 377   } else {
 378     // this is the final chunk for this array
 379     start = 0;
 380     int const actual_length = arrayOop(obj)->length();
 381     arrayOop(old)->set_length(actual_length);
 382   }
 383 
 384   if (UseCompressedOops) {
 385     process_array_chunk_work<narrowOop>(obj, start, end);
 386   } else {
 387     process_array_chunk_work<oop>(obj, start, end);
 388   }
 389 }
 390 
 391 class PushContentsClosure : public ExtendedOopClosure {
 392   PSPromotionManager* _pm;
 393  public:
 394   PushContentsClosure(PSPromotionManager* pm) : _pm(pm) {}
 395 
 396   template <typename T> void do_oop_nv(T* p) {
 397     if (PSScavenge::should_scavenge(p)) {
 398       _pm->claim_or_forward_depth(p);
 399     }
 400   }
 401 
 402   virtual void do_oop(oop* p)       { do_oop_nv(p); }
 403   virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
 404 
 405   // Don't use the oop verification code in the oop_oop_iterate framework.
 406   debug_only(virtual bool should_verify_oops() { return false; })
 407 };
 408 
 409 void InstanceKlass::oop_ps_push_contents(oop obj, PSPromotionManager* pm) {
 410   PushContentsClosure cl(pm);
 411   oop_oop_iterate_oop_maps_reverse<true>(obj, &cl);
 412 }
 413 
 414 void InstanceMirrorKlass::oop_ps_push_contents(oop obj, PSPromotionManager* pm) {
 415     // Note that we don't have to follow the mirror -> klass pointer, since all
 416     // klasses that are dirty will be scavenged when we iterate over the
 417     // ClassLoaderData objects.
 418 
 419   InstanceKlass::oop_ps_push_contents(obj, pm);
 420 
 421   PushContentsClosure cl(pm);
 422   oop_oop_iterate_statics<true>(obj, &cl);
 423 }
 424 
 425 void InstanceClassLoaderKlass::oop_ps_push_contents(oop obj, PSPromotionManager* pm) {
 426   InstanceKlass::oop_ps_push_contents(obj, pm);
 427 
 428   // This is called by the young collector. It will already have taken care of
 429   // all class loader data. So, we don't have to follow the class loader ->
 430   // class loader data link.
 431 }
 432 
 433 template <class T>
 434 static void oop_ps_push_contents_specialized(oop obj, InstanceRefKlass *klass, PSPromotionManager* pm) {
 435   T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj);
 436   if (PSScavenge::should_scavenge(referent_addr)) {
 437     ReferenceProcessor* rp = PSScavenge::reference_processor();
 438     if (rp->discover_reference(obj, klass->reference_type())) {
 439       // reference already enqueued, referent and next will be traversed later
 440       klass->InstanceKlass::oop_ps_push_contents(obj, pm);
 441       return;
 442     } else {
 443       // treat referent as normal oop
 444       pm->claim_or_forward_depth(referent_addr);
 445     }
 446   }
 447   // Treat discovered as normal oop, if ref is not "active",
 448   // i.e. if next is non-NULL.
 449   T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj);
 450   T  next_oop = oopDesc::load_heap_oop(next_addr);
 451   if (!oopDesc::is_null(next_oop)) { // i.e. ref is not "active"
 452     T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj);
 453     log_develop_trace(gc, ref)("   Process discovered as normal " PTR_FORMAT, p2i(discovered_addr));
 454     if (PSScavenge::should_scavenge(discovered_addr)) {
 455       pm->claim_or_forward_depth(discovered_addr);
 456     }
 457   }
 458   // Treat next as normal oop;  next is a link in the reference queue.
 459   if (PSScavenge::should_scavenge(next_addr)) {
 460     pm->claim_or_forward_depth(next_addr);
 461   }
 462   klass->InstanceKlass::oop_ps_push_contents(obj, pm);
 463 }
 464 
 465 void InstanceRefKlass::oop_ps_push_contents(oop obj, PSPromotionManager* pm) {
 466   if (UseCompressedOops) {
 467     oop_ps_push_contents_specialized<narrowOop>(obj, this, pm);
 468   } else {
 469     oop_ps_push_contents_specialized<oop>(obj, this, pm);
 470   }
 471 }
 472 
 473 void ObjArrayKlass::oop_ps_push_contents(oop obj, PSPromotionManager* pm) {
 474   assert(obj->is_objArray(), "obj must be obj array");
 475   PushContentsClosure cl(pm);
 476   oop_oop_iterate_elements<true>(objArrayOop(obj), &cl);
 477 }
 478 
 479 void TypeArrayKlass::oop_ps_push_contents(oop obj, PSPromotionManager* pm) {
 480   assert(obj->is_typeArray(),"must be a type array");
 481   ShouldNotReachHere();
 482 }
 483 
 484 oop PSPromotionManager::oop_promotion_failed(oop obj, markOop obj_mark) {
 485   assert(_old_gen_is_full || PromotionFailureALot, "Sanity");
 486 
 487   // Attempt to CAS in the header.
 488   // This tests if the header is still the same as when
 489   // this started.  If it is the same (i.e., no forwarding
 490   // pointer has been installed), then this thread owns
 491   // it.
 492   if (obj->cas_forward_to(obj, obj_mark, memory_order_relaxed)) {
 493     // We won any races, we "own" this object.
 494     assert(obj == obj->forwardee(), "Sanity");
 495 
 496     _promotion_failed_info.register_copy_failure(obj->size());
 497 
 498     push_contents(obj);
 499 
 500     _preserved_marks->push_if_necessary(obj, obj_mark);
 501   }  else {
 502     // We lost, someone else "owns" this object
 503     guarantee(obj->is_forwarded(), "Object must be forwarded if the cas failed.");
 504 
 505     // No unallocation to worry about.
 506     obj = obj->forwardee();
 507   }
 508 
 509   log_develop_trace(gc, scavenge)("{promotion-failure %s " PTR_FORMAT " (%d)}", obj->klass()->internal_name(), p2i(obj), obj->size());
 510 
 511   return obj;
 512 }