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