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_SHENANDOAHHEAP_INLINE_HPP 25 #define SHARE_VM_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.hpp" 34 #include "gc/shenandoah/shenandoahCollectionSet.inline.hpp" 35 #include "gc/shenandoah/shenandoahForwarding.inline.hpp" 36 #include "gc/shenandoah/shenandoahWorkGroup.hpp" 37 #include "gc/shenandoah/shenandoahHeap.hpp" 38 #include "gc/shenandoah/shenandoahHeapRegionSet.inline.hpp" 39 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp" 40 #include "gc/shenandoah/shenandoahControlThread.hpp" 41 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp" 42 #include "gc/shenandoah/shenandoahThreadLocalData.hpp" 43 #include "oops/oop.inline.hpp" 44 #include "runtime/atomic.hpp" 45 #include "runtime/interfaceSupport.inline.hpp" 46 #include "runtime/prefetch.hpp" 47 #include "runtime/prefetch.inline.hpp" 48 #include "runtime/thread.hpp" 49 #include "utilities/copy.hpp" 50 #include "utilities/globalDefinitions.hpp" 51 52 53 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() { 54 size_t new_index = Atomic::add((size_t) 1, &_index); 55 // get_region() provides the bounds-check and returns NULL on OOB. 56 return _heap->get_region(new_index - 1); 57 } 58 59 inline bool ShenandoahHeap::has_forwarded_objects() const { 60 return _gc_state.is_set(HAS_FORWARDED); 61 } 62 63 inline WorkGang* ShenandoahHeap::workers() const { 64 return _workers; 65 } 66 67 inline WorkGang* ShenandoahHeap::get_safepoint_workers() { 68 return _safepoint_workers; 69 } 70 71 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const { 72 uintptr_t region_start = ((uintptr_t) addr); 73 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift(); 74 assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr)); 75 return index; 76 } 77 78 inline ShenandoahHeapRegion* const ShenandoahHeap::heap_region_containing(const void* addr) const { 79 size_t index = heap_region_index_containing(addr); 80 ShenandoahHeapRegion* const result = get_region(index); 81 assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr)); 82 return result; 83 } 84 85 template <class T> 86 inline oop ShenandoahHeap::update_with_forwarded_not_null(T* p, oop obj) { 87 if (in_collection_set(obj)) { 88 shenandoah_assert_forwarded_except(p, obj, is_full_gc_in_progress() || cancelled_gc() || is_degenerated_gc_in_progress()); 89 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 90 RawAccess<IS_NOT_NULL>::oop_store(p, obj); 91 } 92 #ifdef ASSERT 93 else { 94 shenandoah_assert_not_forwarded(p, obj); 95 } 96 #endif 97 return obj; 98 } 99 100 template <class T> 101 inline oop ShenandoahHeap::maybe_update_with_forwarded(T* p) { 102 T o = RawAccess<>::oop_load(p); 103 if (!CompressedOops::is_null(o)) { 104 oop obj = CompressedOops::decode_not_null(o); 105 return maybe_update_with_forwarded_not_null(p, obj); 106 } else { 107 return NULL; 108 } 109 } 110 111 template <class T> 112 inline oop ShenandoahHeap::evac_update_with_forwarded(T* p) { 113 T o = RawAccess<>::oop_load(p); 114 if (!CompressedOops::is_null(o)) { 115 oop heap_oop = CompressedOops::decode_not_null(o); 116 if (in_collection_set(heap_oop)) { 117 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop); 118 if (forwarded_oop == heap_oop) { 119 forwarded_oop = evacuate_object(heap_oop, Thread::current()); 120 } 121 oop prev = cas_oop(forwarded_oop, p, heap_oop); 122 if (prev == heap_oop) { 123 return forwarded_oop; 124 } else { 125 return NULL; 126 } 127 } 128 return heap_oop; 129 } else { 130 return NULL; 131 } 132 } 133 134 inline oop ShenandoahHeap::cas_oop(oop n, oop* addr, oop c) { 135 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 136 return (oop) Atomic::cmpxchg(n, addr, c); 137 } 138 139 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, narrowOop c) { 140 narrowOop val = CompressedOops::encode(n); 141 return CompressedOops::decode((narrowOop) Atomic::cmpxchg(val, addr, c)); 142 } 143 144 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, oop c) { 145 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 146 narrowOop cmp = CompressedOops::encode(c); 147 narrowOop val = CompressedOops::encode(n); 148 return CompressedOops::decode((narrowOop) Atomic::cmpxchg(val, addr, cmp)); 149 } 150 151 template <class T> 152 inline oop ShenandoahHeap::maybe_update_with_forwarded_not_null(T* p, oop heap_oop) { 153 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || is_full_gc_in_progress() || is_degenerated_gc_in_progress()); 154 shenandoah_assert_correct(p, heap_oop); 155 156 if (in_collection_set(heap_oop)) { 157 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop); 158 if (forwarded_oop == heap_oop) { 159 // E.g. during evacuation. 160 return forwarded_oop; 161 } 162 163 shenandoah_assert_forwarded_except(p, heap_oop, is_full_gc_in_progress() || is_degenerated_gc_in_progress()); 164 shenandoah_assert_not_forwarded(p, forwarded_oop); 165 shenandoah_assert_not_in_cset_except(p, forwarded_oop, cancelled_gc()); 166 167 // If this fails, another thread wrote to p before us, it will be logged in SATB and the 168 // reference be updated later. 169 oop witness = cas_oop(forwarded_oop, p, heap_oop); 170 171 if (witness != heap_oop) { 172 // CAS failed, someone had beat us to it. Normally, we would return the failure witness, 173 // because that would be the proper write of to-space object, enforced by strong barriers. 174 // However, there is a corner case with arraycopy. It can happen that a Java thread 175 // beats us with an arraycopy, which first copies the array, which potentially contains 176 // from-space refs, and only afterwards updates all from-space refs to to-space refs, 177 // which leaves a short window where the new array elements can be from-space. 178 // In this case, we can just resolve the result again. As we resolve, we need to consider 179 // the contended write might have been NULL. 180 oop result = ShenandoahBarrierSet::resolve_forwarded(witness); 181 shenandoah_assert_not_forwarded_except(p, result, (result == NULL)); 182 shenandoah_assert_not_in_cset_except(p, result, (result == NULL) || cancelled_gc()); 183 return result; 184 } else { 185 // Success! We have updated with known to-space copy. We have already asserted it is sane. 186 return forwarded_oop; 187 } 188 } else { 189 shenandoah_assert_not_forwarded(p, heap_oop); 190 return heap_oop; 191 } 192 } 193 194 inline bool ShenandoahHeap::cancelled_gc() const { 195 return _cancelled_gc.get() == CANCELLED; 196 } 197 198 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) { 199 if (! (sts_active && ShenandoahSuspendibleWorkers)) { 200 return cancelled_gc(); 201 } 202 203 jbyte prev = _cancelled_gc.cmpxchg(NOT_CANCELLED, CANCELLABLE); 204 if (prev == CANCELLABLE || prev == NOT_CANCELLED) { 205 if (SuspendibleThreadSet::should_yield()) { 206 SuspendibleThreadSet::yield(); 207 } 208 209 // Back to CANCELLABLE. The thread that poked NOT_CANCELLED first gets 210 // to restore to CANCELLABLE. 211 if (prev == CANCELLABLE) { 212 _cancelled_gc.set(CANCELLABLE); 213 } 214 return false; 215 } else { 216 return true; 217 } 218 } 219 220 inline bool ShenandoahHeap::try_cancel_gc() { 221 while (true) { 222 jbyte prev = _cancelled_gc.cmpxchg(CANCELLED, CANCELLABLE); 223 if (prev == CANCELLABLE) return true; 224 else if (prev == CANCELLED) return false; 225 assert(ShenandoahSuspendibleWorkers, "should not get here when not using suspendible workers"); 226 assert(prev == NOT_CANCELLED, "must be NOT_CANCELLED"); 227 { 228 // We need to provide a safepoint here, otherwise we might 229 // spin forever if a SP is pending. 230 ThreadBlockInVM sp(JavaThread::current()); 231 SpinPause(); 232 } 233 } 234 } 235 236 inline void ShenandoahHeap::clear_cancelled_gc() { 237 _cancelled_gc.set(CANCELLABLE); 238 _oom_evac_handler.clear(); 239 } 240 241 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) { 242 assert(UseTLAB, "TLABs should be enabled"); 243 244 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 245 if (gclab == NULL) { 246 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(), 247 "Performance: thread should have GCLAB: %s", thread->name()); 248 // No GCLABs in this thread, fallback to shared allocation 249 return NULL; 250 } 251 HeapWord* obj = gclab->allocate(size); 252 if (obj != NULL) { 253 return obj; 254 } 255 // Otherwise... 256 return allocate_from_gclab_slow(thread, size); 257 } 258 259 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) { 260 if (ShenandoahThreadLocalData::is_oom_during_evac(Thread::current())) { 261 // This thread went through the OOM during evac protocol and it is safe to return 262 // the forward pointer. It must not attempt to evacuate any more. 263 return ShenandoahBarrierSet::resolve_forwarded(p); 264 } 265 266 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope"); 267 268 size_t size = p->size(); 269 270 assert(!heap_region_containing(p)->is_humongous(), "never evacuate humongous objects"); 271 272 bool alloc_from_gclab = true; 273 HeapWord* copy = NULL; 274 275 #ifdef ASSERT 276 if (ShenandoahOOMDuringEvacALot && 277 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call 278 copy = NULL; 279 } else { 280 #endif 281 if (UseTLAB) { 282 copy = allocate_from_gclab(thread, size); 283 } 284 if (copy == NULL) { 285 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size); 286 copy = allocate_memory(req); 287 alloc_from_gclab = false; 288 } 289 #ifdef ASSERT 290 } 291 #endif 292 293 if (copy == NULL) { 294 control_thread()->handle_alloc_failure_evac(size); 295 296 _oom_evac_handler.handle_out_of_memory_during_evacuation(); 297 298 return ShenandoahBarrierSet::resolve_forwarded(p); 299 } 300 301 // Copy the object: 302 Copy::aligned_disjoint_words((HeapWord*) p, copy, size); 303 304 // Try to install the new forwarding pointer. 305 oop copy_val = oop(copy); 306 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val); 307 if (result == copy_val) { 308 // Successfully evacuated. Our copy is now the public one! 309 shenandoah_assert_correct(NULL, copy_val); 310 return copy_val; 311 } else { 312 // Failed to evacuate. We need to deal with the object that is left behind. Since this 313 // new allocation is certainly after TAMS, it will be considered live in the next cycle. 314 // But if it happens to contain references to evacuated regions, those references would 315 // not get updated for this stale copy during this cycle, and we will crash while scanning 316 // it the next cycle. 317 // 318 // For GCLAB allocations, it is enough to rollback the allocation ptr. Either the next 319 // object will overwrite this stale copy, or the filler object on LAB retirement will 320 // do this. For non-GCLAB allocations, we have no way to retract the allocation, and 321 // have to explicitly overwrite the copy with the filler object. With that overwrite, 322 // we have to keep the fwdptr initialized and pointing to our (stale) copy. 323 if (alloc_from_gclab) { 324 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size); 325 } else { 326 fill_with_object(copy, size); 327 shenandoah_assert_correct(NULL, copy_val); 328 } 329 shenandoah_assert_correct(NULL, result); 330 return result; 331 } 332 } 333 334 inline bool ShenandoahHeap::requires_marking(const void* entry) const { 335 return !_marking_context->is_marked(oop(entry)); 336 } 337 338 template <class T> 339 inline bool ShenandoahHeap::in_collection_set(T p) const { 340 HeapWord* obj = (HeapWord*) p; 341 assert(collection_set() != NULL, "Sanity"); 342 assert(is_in(obj), "should be in heap"); 343 344 return collection_set()->is_in(obj); 345 } 346 347 inline bool ShenandoahHeap::is_stable() const { 348 return _gc_state.is_clear(); 349 } 350 351 inline bool ShenandoahHeap::is_idle() const { 352 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS | TRAVERSAL); 353 } 354 355 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const { 356 return _gc_state.is_set(MARKING); 357 } 358 359 inline bool ShenandoahHeap::is_concurrent_traversal_in_progress() const { 360 return _gc_state.is_set(TRAVERSAL); 361 } 362 363 inline bool ShenandoahHeap::is_evacuation_in_progress() const { 364 return _gc_state.is_set(EVACUATION); 365 } 366 367 inline bool ShenandoahHeap::is_gc_in_progress_mask(uint mask) const { 368 return _gc_state.is_set(mask); 369 } 370 371 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const { 372 return _degenerated_gc_in_progress.is_set(); 373 } 374 375 inline bool ShenandoahHeap::is_full_gc_in_progress() const { 376 return _full_gc_in_progress.is_set(); 377 } 378 379 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const { 380 return _full_gc_move_in_progress.is_set(); 381 } 382 383 inline bool ShenandoahHeap::is_update_refs_in_progress() const { 384 return _gc_state.is_set(UPDATEREFS); 385 } 386 387 template<class T> 388 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) { 389 marked_object_iterate(region, cl, region->top()); 390 } 391 392 template<class T> 393 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) { 394 assert(! region->is_humongous_continuation(), "no humongous continuation regions here"); 395 396 ShenandoahMarkingContext* const ctx = complete_marking_context(); 397 assert(ctx->is_complete(), "sanity"); 398 399 MarkBitMap* mark_bit_map = ctx->mark_bit_map(); 400 HeapWord* tams = ctx->top_at_mark_start(region); 401 402 size_t skip_bitmap_delta = 1; 403 HeapWord* start = region->bottom(); 404 HeapWord* end = MIN2(tams, region->end()); 405 406 // Step 1. Scan below the TAMS based on bitmap data. 407 HeapWord* limit_bitmap = MIN2(limit, tams); 408 409 // Try to scan the initial candidate. If the candidate is above the TAMS, it would 410 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2. 411 HeapWord* cb = mark_bit_map->getNextMarkedWordAddress(start, end); 412 413 intx dist = ShenandoahMarkScanPrefetch; 414 if (dist > 0) { 415 // Batched scan that prefetches the oop data, anticipating the access to 416 // either header, oop field, or forwarding pointer. Not that we cannot 417 // touch anything in oop, while it still being prefetched to get enough 418 // time for prefetch to work. This is why we try to scan the bitmap linearly, 419 // disregarding the object size. However, since we know forwarding pointer 420 // preceeds the object, we can skip over it. Once we cannot trust the bitmap, 421 // there is no point for prefetching the oop contents, as oop->size() will 422 // touch it prematurely. 423 424 // No variable-length arrays in standard C++, have enough slots to fit 425 // the prefetch distance. 426 static const int SLOT_COUNT = 256; 427 guarantee(dist <= SLOT_COUNT, "adjust slot count"); 428 HeapWord* slots[SLOT_COUNT]; 429 430 int avail; 431 do { 432 avail = 0; 433 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) { 434 Prefetch::read(cb, oopDesc::mark_offset_in_bytes()); 435 slots[avail++] = cb; 436 cb += skip_bitmap_delta; 437 if (cb < limit_bitmap) { 438 cb = mark_bit_map->getNextMarkedWordAddress(cb, limit_bitmap); 439 } 440 } 441 442 for (int c = 0; c < avail; c++) { 443 assert (slots[c] < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams)); 444 assert (slots[c] < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit)); 445 oop obj = oop(slots[c]); 446 assert(oopDesc::is_oop(obj), "sanity"); 447 assert(ctx->is_marked(obj), "object expected to be marked"); 448 cl->do_object(obj); 449 } 450 } while (avail > 0); 451 } else { 452 while (cb < limit_bitmap) { 453 assert (cb < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams)); 454 assert (cb < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit)); 455 oop obj = oop(cb); 456 assert(oopDesc::is_oop(obj), "sanity"); 457 assert(ctx->is_marked(obj), "object expected to be marked"); 458 cl->do_object(obj); 459 cb += skip_bitmap_delta; 460 if (cb < limit_bitmap) { 461 cb = mark_bit_map->getNextMarkedWordAddress(cb, limit_bitmap); 462 } 463 } 464 } 465 466 // Step 2. Accurate size-based traversal, happens past the TAMS. 467 // This restarts the scan at TAMS, which makes sure we traverse all objects, 468 // regardless of what happened at Step 1. 469 HeapWord* cs = tams; 470 while (cs < limit) { 471 assert (cs >= tams, "only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams)); 472 assert (cs < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit)); 473 oop obj = oop(cs); 474 assert(oopDesc::is_oop(obj), "sanity"); 475 assert(ctx->is_marked(obj), "object expected to be marked"); 476 int size = obj->size(); 477 cl->do_object(obj); 478 cs += size; 479 } 480 } 481 482 template <class T> 483 class ShenandoahObjectToOopClosure : public ObjectClosure { 484 T* _cl; 485 public: 486 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {} 487 488 void do_object(oop obj) { 489 obj->oop_iterate(_cl); 490 } 491 }; 492 493 template <class T> 494 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure { 495 T* _cl; 496 MemRegion _bounds; 497 public: 498 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) : 499 _cl(cl), _bounds(bottom, top) {} 500 501 void do_object(oop obj) { 502 obj->oop_iterate(_cl, _bounds); 503 } 504 }; 505 506 template<class T> 507 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) { 508 if (region->is_humongous()) { 509 HeapWord* bottom = region->bottom(); 510 if (top > bottom) { 511 region = region->humongous_start_region(); 512 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top); 513 marked_object_iterate(region, &objs); 514 } 515 } else { 516 ShenandoahObjectToOopClosure<T> objs(cl); 517 marked_object_iterate(region, &objs, top); 518 } 519 } 520 521 inline ShenandoahHeapRegion* const ShenandoahHeap::get_region(size_t region_idx) const { 522 if (region_idx < _num_regions) { 523 return _regions[region_idx]; 524 } else { 525 return NULL; 526 } 527 } 528 529 inline void ShenandoahHeap::mark_complete_marking_context() { 530 _marking_context->mark_complete(); 531 } 532 533 inline void ShenandoahHeap::mark_incomplete_marking_context() { 534 _marking_context->mark_incomplete(); 535 } 536 537 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const { 538 assert (_marking_context->is_complete()," sanity"); 539 return _marking_context; 540 } 541 542 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const { 543 return _marking_context; 544 } 545 546 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP