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_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
  25 #define SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
  26 
  27 #include "classfile/javaClasses.inline.hpp"
  28 #include "gc/shared/markBitMap.inline.hpp"
  29 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
  30 #include "gc/shared/suspendibleThreadSet.hpp"
  31 #include "gc/shenandoah/shenandoahAsserts.hpp"
  32 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp"
  33 #include "gc/shenandoah/shenandoahCollectionSet.inline.hpp"
  34 #include "gc/shenandoah/shenandoahForwarding.inline.hpp"
  35 #include "gc/shenandoah/shenandoahWorkGroup.hpp"
  36 #include "gc/shenandoah/shenandoahHeap.hpp"
  37 #include "gc/shenandoah/shenandoahHeapRegionSet.inline.hpp"
  38 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
  39 #include "gc/shenandoah/shenandoahControlThread.hpp"
  40 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
  41 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  42 #include "oops/compressedOops.inline.hpp"
  43 #include "oops/oop.inline.hpp"
  44 #include "runtime/atomic.hpp"
  45 #include "runtime/prefetch.inline.hpp"
  46 #include "runtime/thread.hpp"
  47 #include "utilities/copy.hpp"
  48 #include "utilities/globalDefinitions.hpp"
  49 
  50 
  51 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() {
  52   size_t new_index = Atomic::add((size_t) 1, &_index);
  53   // get_region() provides the bounds-check and returns NULL on OOB.
  54   return _heap->get_region(new_index - 1);
  55 }
  56 
  57 inline bool ShenandoahHeap::has_forwarded_objects() const {
  58   return _gc_state.is_set(HAS_FORWARDED);
  59 }
  60 
  61 inline WorkGang* ShenandoahHeap::workers() const {
  62   return _workers;
  63 }
  64 
  65 inline WorkGang* ShenandoahHeap::get_safepoint_workers() {
  66   return _safepoint_workers;
  67 }
  68 
  69 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const {
  70   uintptr_t region_start = ((uintptr_t) addr);
  71   uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift();
  72   assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr));
  73   return index;
  74 }
  75 
  76 inline ShenandoahHeapRegion* const ShenandoahHeap::heap_region_containing(const void* addr) const {
  77   size_t index = heap_region_index_containing(addr);
  78   ShenandoahHeapRegion* const result = get_region(index);
  79   assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr));
  80   return result;
  81 }
  82 
  83 template <class T>
  84 inline oop ShenandoahHeap::update_with_forwarded_not_null(T* p, oop obj) {
  85   if (in_collection_set(obj)) {
  86     shenandoah_assert_forwarded_except(p, obj, is_full_gc_in_progress() || cancelled_gc() || is_degenerated_gc_in_progress());
  87     obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
  88     RawAccess<IS_NOT_NULL>::oop_store(p, obj);
  89   }
  90 #ifdef ASSERT
  91   else {
  92     shenandoah_assert_not_forwarded(p, obj);
  93   }
  94 #endif
  95   return obj;
  96 }
  97 
  98 template <class T>
  99 inline oop ShenandoahHeap::maybe_update_with_forwarded(T* p) {
 100   T o = RawAccess<>::oop_load(p);
 101   if (!CompressedOops::is_null(o)) {
 102     oop obj = CompressedOops::decode_not_null(o);
 103     return maybe_update_with_forwarded_not_null(p, obj);
 104   } else {
 105     return NULL;
 106   }
 107 }
 108 
 109 template <class T>
 110 inline oop ShenandoahHeap::evac_update_with_forwarded(T* p) {
 111   T o = RawAccess<>::oop_load(p);
 112   if (!CompressedOops::is_null(o)) {
 113     oop heap_oop = CompressedOops::decode_not_null(o);
 114     if (in_collection_set(heap_oop)) {
 115       oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop);
 116       if (oopDesc::equals_raw(forwarded_oop, heap_oop)) {
 117         forwarded_oop = evacuate_object(heap_oop, Thread::current());
 118       }
 119       oop prev = cas_oop(forwarded_oop, p, heap_oop);
 120       if (oopDesc::equals_raw(prev, heap_oop)) {
 121         return forwarded_oop;
 122       } else {
 123         return NULL;
 124       }
 125     }
 126     return heap_oop;
 127   } else {
 128     return NULL;
 129   }
 130 }
 131 
 132 inline oop ShenandoahHeap::cas_oop(oop n, oop* addr, oop c) {
 133   return (oop) Atomic::cmpxchg(n, addr, c);
 134 }
 135 
 136 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, oop c) {
 137   narrowOop cmp = CompressedOops::encode(c);
 138   narrowOop val = CompressedOops::encode(n);
 139   return CompressedOops::decode((narrowOop) Atomic::cmpxchg(val, addr, cmp));
 140 }
 141 
 142 template <class T>
 143 inline oop ShenandoahHeap::maybe_update_with_forwarded_not_null(T* p, oop heap_oop) {
 144   shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || is_full_gc_in_progress() || is_degenerated_gc_in_progress());
 145   shenandoah_assert_correct(p, heap_oop);
 146 
 147   if (in_collection_set(heap_oop)) {
 148     oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop);
 149     if (oopDesc::equals_raw(forwarded_oop, heap_oop)) {
 150       // E.g. during evacuation.
 151       return forwarded_oop;
 152     }
 153 
 154     shenandoah_assert_forwarded_except(p, heap_oop, is_full_gc_in_progress() || is_degenerated_gc_in_progress());
 155     shenandoah_assert_not_forwarded(p, forwarded_oop);
 156     shenandoah_assert_not_in_cset_except(p, forwarded_oop, cancelled_gc());
 157 
 158     // If this fails, another thread wrote to p before us, it will be logged in SATB and the
 159     // reference be updated later.
 160     oop witness = cas_oop(forwarded_oop, p, heap_oop);
 161 
 162     if (!oopDesc::equals_raw(witness, heap_oop)) {
 163       // CAS failed, someone had beat us to it. Normally, we would return the failure witness,
 164       // because that would be the proper write of to-space object, enforced by strong barriers.
 165       // However, there is a corner case with arraycopy. It can happen that a Java thread
 166       // beats us with an arraycopy, which first copies the array, which potentially contains
 167       // from-space refs, and only afterwards updates all from-space refs to to-space refs,
 168       // which leaves a short window where the new array elements can be from-space.
 169       // In this case, we can just resolve the result again. As we resolve, we need to consider
 170       // the contended write might have been NULL.
 171       oop result = ShenandoahBarrierSet::resolve_forwarded(witness);
 172       shenandoah_assert_not_forwarded_except(p, result, (result == NULL));
 173       shenandoah_assert_not_in_cset_except(p, result, (result == NULL) || cancelled_gc());
 174       return result;
 175     } else {
 176       // Success! We have updated with known to-space copy. We have already asserted it is sane.
 177       return forwarded_oop;
 178     }
 179   } else {
 180     shenandoah_assert_not_forwarded(p, heap_oop);
 181     return heap_oop;
 182   }
 183 }
 184 
 185 inline bool ShenandoahHeap::cancelled_gc() const {
 186   return _cancelled_gc.get() == CANCELLED;
 187 }
 188 
 189 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) {
 190   if (!sts_active) {
 191     return cancelled_gc();
 192   }
 193 
 194   jbyte prev = _cancelled_gc.cmpxchg(NOT_CANCELLED, CANCELLABLE);
 195   if (prev == CANCELLABLE || prev == NOT_CANCELLED) {
 196     if (SuspendibleThreadSet::should_yield()) {
 197       SuspendibleThreadSet::yield();
 198     }
 199 
 200     // Back to CANCELLABLE. The thread that poked NOT_CANCELLED first gets
 201     // to restore to CANCELLABLE.
 202     if (prev == CANCELLABLE) {
 203       _cancelled_gc.set(CANCELLABLE);
 204     }
 205     return false;
 206   } else {
 207     return true;
 208   }
 209 }
 210 
 211 inline void ShenandoahHeap::clear_cancelled_gc() {
 212   _cancelled_gc.set(CANCELLABLE);
 213   _oom_evac_handler.clear();
 214 }
 215 
 216 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) {
 217   assert(UseTLAB, "TLABs should be enabled");
 218 
 219   PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
 220   if (gclab == NULL) {
 221     assert(!thread->is_Java_thread() && !thread->is_Worker_thread(),
 222            "Performance: thread should have GCLAB: %s", thread->name());
 223     // No GCLABs in this thread, fallback to shared allocation
 224     return NULL;
 225   }
 226   HeapWord* obj = gclab->allocate(size);
 227   if (obj != NULL) {
 228     return obj;
 229   }
 230   // Otherwise...
 231   return allocate_from_gclab_slow(thread, size);
 232 }
 233 
 234 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) {
 235   if (ShenandoahThreadLocalData::is_oom_during_evac(Thread::current())) {
 236     // This thread went through the OOM during evac protocol and it is safe to return
 237     // the forward pointer. It must not attempt to evacuate any more.
 238     return ShenandoahBarrierSet::resolve_forwarded(p);
 239   }
 240 
 241   assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope");
 242 
 243   size_t size = p->size();
 244 
 245   assert(!heap_region_containing(p)->is_humongous(), "never evacuate humongous objects");
 246 
 247   bool alloc_from_gclab = true;
 248   HeapWord* copy = NULL;
 249 
 250 #ifdef ASSERT
 251   if (ShenandoahOOMDuringEvacALot &&
 252       (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
 253         copy = NULL;
 254   } else {
 255 #endif
 256     if (UseTLAB) {
 257       copy = allocate_from_gclab(thread, size);
 258     }
 259     if (copy == NULL) {
 260       ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size);
 261       copy = allocate_memory(req);
 262       alloc_from_gclab = false;
 263     }
 264 #ifdef ASSERT
 265   }
 266 #endif
 267 
 268   if (copy == NULL) {
 269     control_thread()->handle_alloc_failure_evac(size);
 270 
 271     _oom_evac_handler.handle_out_of_memory_during_evacuation();
 272 
 273     return ShenandoahBarrierSet::resolve_forwarded(p);
 274   }
 275 
 276   // Copy the object:
 277   Copy::aligned_disjoint_words((HeapWord*) p, copy, size);
 278 
 279   // Try to install the new forwarding pointer.
 280   oop copy_val = oop(copy);
 281   oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val);
 282   if (oopDesc::equals_raw(result, copy_val)) {
 283     // Successfully evacuated. Our copy is now the public one!
 284     shenandoah_assert_correct(NULL, copy_val);
 285     return copy_val;
 286   }  else {
 287     // Failed to evacuate. We need to deal with the object that is left behind. Since this
 288     // new allocation is certainly after TAMS, it will be considered live in the next cycle.
 289     // But if it happens to contain references to evacuated regions, those references would
 290     // not get updated for this stale copy during this cycle, and we will crash while scanning
 291     // it the next cycle.
 292     //
 293     // For GCLAB allocations, it is enough to rollback the allocation ptr. Either the next
 294     // object will overwrite this stale copy, or the filler object on LAB retirement will
 295     // do this. For non-GCLAB allocations, we have no way to retract the allocation, and
 296     // have to explicitly overwrite the copy with the filler object. With that overwrite,
 297     // we have to keep the fwdptr initialized and pointing to our (stale) copy.
 298     if (alloc_from_gclab) {
 299       ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size);
 300     } else {
 301       fill_with_object(copy, size);
 302       shenandoah_assert_correct(NULL, copy_val);
 303     }
 304     shenandoah_assert_correct(NULL, result);
 305     return result;
 306   }
 307 }
 308 
 309 template<bool RESOLVE>
 310 inline bool ShenandoahHeap::requires_marking(const void* entry) const {
 311   oop obj = oop(entry);
 312   if (RESOLVE) {
 313     obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
 314   }
 315   return !_marking_context->is_marked(obj);
 316 }
 317 
 318 template <class T>
 319 inline bool ShenandoahHeap::in_collection_set(T p) const {
 320   HeapWord* obj = (HeapWord*) p;
 321   assert(collection_set() != NULL, "Sanity");
 322   assert(is_in(obj), "should be in heap");
 323 
 324   return collection_set()->is_in(obj);
 325 }
 326 
 327 inline bool ShenandoahHeap::is_stable() const {
 328   return _gc_state.is_clear();
 329 }
 330 
 331 inline bool ShenandoahHeap::is_idle() const {
 332   return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS | TRAVERSAL);
 333 }
 334 
 335 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const {
 336   return _gc_state.is_set(MARKING);
 337 }
 338 
 339 inline bool ShenandoahHeap::is_concurrent_traversal_in_progress() const {
 340   return _gc_state.is_set(TRAVERSAL);
 341 }
 342 
 343 inline bool ShenandoahHeap::is_evacuation_in_progress() const {
 344   return _gc_state.is_set(EVACUATION);
 345 }
 346 
 347 inline bool ShenandoahHeap::is_gc_in_progress_mask(uint mask) const {
 348   return _gc_state.is_set(mask);
 349 }
 350 
 351 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const {
 352   return _degenerated_gc_in_progress.is_set();
 353 }
 354 
 355 inline bool ShenandoahHeap::is_full_gc_in_progress() const {
 356   return _full_gc_in_progress.is_set();
 357 }
 358 
 359 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const {
 360   return _full_gc_move_in_progress.is_set();
 361 }
 362 
 363 inline bool ShenandoahHeap::is_update_refs_in_progress() const {
 364   return _gc_state.is_set(UPDATEREFS);
 365 }
 366 
 367 template<class T>
 368 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) {
 369   marked_object_iterate(region, cl, region->top());
 370 }
 371 
 372 template<class T>
 373 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) {
 374   assert(! region->is_humongous_continuation(), "no humongous continuation regions here");
 375 
 376   ShenandoahMarkingContext* const ctx = complete_marking_context();
 377   assert(ctx->is_complete(), "sanity");
 378 
 379   MarkBitMap* mark_bit_map = ctx->mark_bit_map();
 380   HeapWord* tams = ctx->top_at_mark_start(region);
 381 
 382   size_t skip_bitmap_delta = 1;
 383   HeapWord* start = region->bottom();
 384   HeapWord* end = MIN2(tams, region->end());
 385 
 386   // Step 1. Scan below the TAMS based on bitmap data.
 387   HeapWord* limit_bitmap = MIN2(limit, tams);
 388 
 389   // Try to scan the initial candidate. If the candidate is above the TAMS, it would
 390   // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2.
 391   HeapWord* cb = mark_bit_map->get_next_marked_addr(start, end);
 392 
 393   intx dist = ShenandoahMarkScanPrefetch;
 394   if (dist > 0) {
 395     // Batched scan that prefetches the oop data, anticipating the access to
 396     // either header, oop field, or forwarding pointer. Not that we cannot
 397     // touch anything in oop, while it still being prefetched to get enough
 398     // time for prefetch to work. This is why we try to scan the bitmap linearly,
 399     // disregarding the object size. However, since we know forwarding pointer
 400     // preceeds the object, we can skip over it. Once we cannot trust the bitmap,
 401     // there is no point for prefetching the oop contents, as oop->size() will
 402     // touch it prematurely.
 403 
 404     // No variable-length arrays in standard C++, have enough slots to fit
 405     // the prefetch distance.
 406     static const int SLOT_COUNT = 256;
 407     guarantee(dist <= SLOT_COUNT, "adjust slot count");
 408     HeapWord* slots[SLOT_COUNT];
 409 
 410     int avail;
 411     do {
 412       avail = 0;
 413       for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) {
 414         Prefetch::read(cb, oopDesc::mark_offset_in_bytes());
 415         slots[avail++] = cb;
 416         cb += skip_bitmap_delta;
 417         if (cb < limit_bitmap) {
 418           cb = mark_bit_map->get_next_marked_addr(cb, limit_bitmap);
 419         }
 420       }
 421 
 422       for (int c = 0; c < avail; c++) {
 423         assert (slots[c] < tams,  "only objects below TAMS here: "  PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams));
 424         assert (slots[c] < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit));
 425         oop obj = oop(slots[c]);
 426         assert(oopDesc::is_oop(obj), "sanity");
 427         assert(ctx->is_marked(obj), "object expected to be marked");
 428         cl->do_object(obj);
 429       }
 430     } while (avail > 0);
 431   } else {
 432     while (cb < limit_bitmap) {
 433       assert (cb < tams,  "only objects below TAMS here: "  PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams));
 434       assert (cb < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit));
 435       oop obj = oop(cb);
 436       assert(oopDesc::is_oop(obj), "sanity");
 437       assert(ctx->is_marked(obj), "object expected to be marked");
 438       cl->do_object(obj);
 439       cb += skip_bitmap_delta;
 440       if (cb < limit_bitmap) {
 441         cb = mark_bit_map->get_next_marked_addr(cb, limit_bitmap);
 442       }
 443     }
 444   }
 445 
 446   // Step 2. Accurate size-based traversal, happens past the TAMS.
 447   // This restarts the scan at TAMS, which makes sure we traverse all objects,
 448   // regardless of what happened at Step 1.
 449   HeapWord* cs = tams;
 450   while (cs < limit) {
 451     assert (cs >= tams, "only objects past TAMS here: "   PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams));
 452     assert (cs < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit));
 453     oop obj = oop(cs);
 454     assert(oopDesc::is_oop(obj), "sanity");
 455     assert(ctx->is_marked(obj), "object expected to be marked");
 456     int size = obj->size();
 457     cl->do_object(obj);
 458     cs += size;
 459   }
 460 }
 461 
 462 template <class T>
 463 class ShenandoahObjectToOopClosure : public ObjectClosure {
 464   T* _cl;
 465 public:
 466   ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {}
 467 
 468   void do_object(oop obj) {
 469     obj->oop_iterate(_cl);
 470   }
 471 };
 472 
 473 template <class T>
 474 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure {
 475   T* _cl;
 476   MemRegion _bounds;
 477 public:
 478   ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) :
 479     _cl(cl), _bounds(bottom, top) {}
 480 
 481   void do_object(oop obj) {
 482     obj->oop_iterate(_cl, _bounds);
 483   }
 484 };
 485 
 486 template<class T>
 487 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) {
 488   if (region->is_humongous()) {
 489     HeapWord* bottom = region->bottom();
 490     if (top > bottom) {
 491       region = region->humongous_start_region();
 492       ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top);
 493       marked_object_iterate(region, &objs);
 494     }
 495   } else {
 496     ShenandoahObjectToOopClosure<T> objs(cl);
 497     marked_object_iterate(region, &objs, top);
 498   }
 499 }
 500 
 501 inline ShenandoahHeapRegion* const ShenandoahHeap::get_region(size_t region_idx) const {
 502   if (region_idx < _num_regions) {
 503     return _regions[region_idx];
 504   } else {
 505     return NULL;
 506   }
 507 }
 508 
 509 inline void ShenandoahHeap::mark_complete_marking_context() {
 510   _marking_context->mark_complete();
 511 }
 512 
 513 inline void ShenandoahHeap::mark_incomplete_marking_context() {
 514   _marking_context->mark_incomplete();
 515 }
 516 
 517 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const {
 518   assert (_marking_context->is_complete()," sanity");
 519   return _marking_context;
 520 }
 521 
 522 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const {
 523   return _marking_context;
 524 }
 525 
 526 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP