1 /* 2 * Copyright (c) 2015, 2020, Red Hat, Inc. All rights reserved. 3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP 27 #define SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP 28 29 #include "gc/shenandoah/shenandoahHeap.hpp" 30 31 #include "classfile/javaClasses.inline.hpp" 32 #include "gc/shared/markBitMap.inline.hpp" 33 #include "gc/shared/threadLocalAllocBuffer.inline.hpp" 34 #include "gc/shared/continuationGCSupport.inline.hpp" 35 #include "gc/shared/suspendibleThreadSet.hpp" 36 #include "gc/shared/tlab_globals.hpp" 37 #include "gc/shenandoah/shenandoahAsserts.hpp" 38 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp" 39 #include "gc/shenandoah/shenandoahCollectionSet.inline.hpp" 40 #include "gc/shenandoah/shenandoahForwarding.inline.hpp" 41 #include "gc/shenandoah/shenandoahWorkGroup.hpp" 42 #include "gc/shenandoah/shenandoahHeapRegionSet.inline.hpp" 43 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp" 44 #include "gc/shenandoah/shenandoahControlThread.hpp" 45 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp" 46 #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp" 47 #include "gc/shenandoah/shenandoahThreadLocalData.hpp" 48 #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp" 49 #include "gc/shenandoah/mode/shenandoahMode.hpp" 50 #include "oops/compressedOops.inline.hpp" 51 #include "oops/oop.inline.hpp" 52 #include "runtime/atomic.hpp" 53 #include "runtime/javaThread.hpp" 54 #include "runtime/prefetch.inline.hpp" 55 #include "utilities/copy.hpp" 56 #include "utilities/globalDefinitions.hpp" 57 58 inline ShenandoahHeap* ShenandoahHeap::heap() { 59 return named_heap<ShenandoahHeap>(CollectedHeap::Shenandoah); 60 } 61 62 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() { 63 size_t new_index = Atomic::add(&_index, (size_t) 1, memory_order_relaxed); 64 // get_region() provides the bounds-check and returns null on OOB. 65 return _heap->get_region(new_index - 1); 66 } 67 68 inline bool ShenandoahHeap::has_forwarded_objects() const { 69 return _gc_state.is_set(HAS_FORWARDED); 70 } 71 72 inline WorkerThreads* ShenandoahHeap::workers() const { 73 return _workers; 74 } 75 76 inline WorkerThreads* ShenandoahHeap::safepoint_workers() { 77 return _safepoint_workers; 78 } 79 80 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const { 81 uintptr_t region_start = ((uintptr_t) addr); 82 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift(); 83 assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr)); 84 return index; 85 } 86 87 inline ShenandoahHeapRegion* ShenandoahHeap::heap_region_containing(const void* addr) const { 88 size_t index = heap_region_index_containing(addr); 89 ShenandoahHeapRegion* const result = get_region(index); 90 assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr)); 91 return result; 92 } 93 94 inline void ShenandoahHeap::enter_evacuation(Thread* t) { 95 _oom_evac_handler.enter_evacuation(t); 96 } 97 98 inline void ShenandoahHeap::leave_evacuation(Thread* t) { 99 _oom_evac_handler.leave_evacuation(t); 100 } 101 102 template <class T> 103 inline void ShenandoahHeap::update_with_forwarded(T* p) { 104 T o = RawAccess<>::oop_load(p); 105 if (!CompressedOops::is_null(o)) { 106 oop obj = CompressedOops::decode_not_null(o); 107 if (in_collection_set(obj)) { 108 // Corner case: when evacuation fails, there are objects in collection 109 // set that are not really forwarded. We can still go and try and update them 110 // (uselessly) to simplify the common path. 111 shenandoah_assert_forwarded_except(p, obj, cancelled_gc()); 112 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 113 shenandoah_assert_not_in_cset_except(p, fwd, cancelled_gc()); 114 115 // Unconditionally store the update: no concurrent updates expected. 116 RawAccess<IS_NOT_NULL>::oop_store(p, fwd); 117 } 118 } 119 } 120 121 template <class T> 122 inline void ShenandoahHeap::conc_update_with_forwarded(T* p) { 123 T o = RawAccess<>::oop_load(p); 124 if (!CompressedOops::is_null(o)) { 125 oop obj = CompressedOops::decode_not_null(o); 126 if (in_collection_set(obj)) { 127 // Corner case: when evacuation fails, there are objects in collection 128 // set that are not really forwarded. We can still go and try CAS-update them 129 // (uselessly) to simplify the common path. 130 shenandoah_assert_forwarded_except(p, obj, cancelled_gc()); 131 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 132 shenandoah_assert_not_in_cset_except(p, fwd, cancelled_gc()); 133 134 // Sanity check: we should not be updating the cset regions themselves, 135 // unless we are recovering from the evacuation failure. 136 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || cancelled_gc()); 137 138 // Either we succeed in updating the reference, or something else gets in our way. 139 // We don't care if that is another concurrent GC update, or another mutator update. 140 atomic_update_oop(fwd, p, obj); 141 } 142 } 143 } 144 145 // Atomic updates of heap location. This is only expected to work with updating the same 146 // logical object with its forwardee. The reason why we need stronger-than-relaxed memory 147 // ordering has to do with coordination with GC barriers and mutator accesses. 148 // 149 // In essence, stronger CAS access is required to maintain the transitive chains that mutator 150 // accesses build by themselves. To illustrate this point, consider the following example. 151 // 152 // Suppose "o" is the object that has a field "x" and the reference to "o" is stored 153 // to field at "addr", which happens to be Java volatile field. Normally, the accesses to volatile 154 // field at "addr" would be matched with release/acquire barriers. This changes when GC moves 155 // the object under mutator feet. 156 // 157 // Thread 1 (Java) 158 // // --- previous access starts here 159 // ... 160 // T1.1: store(&o.x, 1, mo_relaxed) 161 // T1.2: store(&addr, o, mo_release) // volatile store 162 // 163 // // --- new access starts here 164 // // LRB: copy and install the new copy to fwdptr 165 // T1.3: var copy = copy(o) 166 // T1.4: cas(&fwd, t, copy, mo_release) // pointer-mediated publication 167 // <access continues> 168 // 169 // Thread 2 (GC updater) 170 // T2.1: var f = load(&fwd, mo_{consume|acquire}) // pointer-mediated acquisition 171 // T2.2: cas(&addr, o, f, mo_release) // this method 172 // 173 // Thread 3 (Java) 174 // T3.1: var o = load(&addr, mo_acquire) // volatile read 175 // T3.2: if (o != null) 176 // T3.3: var r = load(&o.x, mo_relaxed) 177 // 178 // r is guaranteed to contain "1". 179 // 180 // Without GC involvement, there is synchronizes-with edge from T1.2 to T3.1, 181 // which guarantees this. With GC involvement, when LRB copies the object and 182 // another thread updates the reference to it, we need to have the transitive edge 183 // from T1.4 to T2.1 (that one is guaranteed by forwarding accesses), plus the edge 184 // from T2.2 to T3.1 (which is brought by this CAS). 185 // 186 // Note that we do not need to "acquire" in these methods, because we do not read the 187 // failure witnesses contents on any path, and "release" is enough. 188 // 189 190 inline void ShenandoahHeap::atomic_update_oop(oop update, oop* addr, oop compare) { 191 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 192 Atomic::cmpxchg(addr, compare, update, memory_order_release); 193 } 194 195 inline void ShenandoahHeap::atomic_update_oop(oop update, narrowOop* addr, narrowOop compare) { 196 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 197 narrowOop u = CompressedOops::encode(update); 198 Atomic::cmpxchg(addr, compare, u, memory_order_release); 199 } 200 201 inline void ShenandoahHeap::atomic_update_oop(oop update, narrowOop* addr, oop compare) { 202 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 203 narrowOop c = CompressedOops::encode(compare); 204 narrowOop u = CompressedOops::encode(update); 205 Atomic::cmpxchg(addr, c, u, memory_order_release); 206 } 207 208 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, oop* addr, oop compare) { 209 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 210 return (oop) Atomic::cmpxchg(addr, compare, update, memory_order_release) == compare; 211 } 212 213 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, narrowOop* addr, narrowOop compare) { 214 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 215 narrowOop u = CompressedOops::encode(update); 216 return (narrowOop) Atomic::cmpxchg(addr, compare, u, memory_order_release) == compare; 217 } 218 219 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, narrowOop* addr, oop compare) { 220 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 221 narrowOop c = CompressedOops::encode(compare); 222 narrowOop u = CompressedOops::encode(update); 223 return CompressedOops::decode(Atomic::cmpxchg(addr, c, u, memory_order_release)) == compare; 224 } 225 226 // The memory ordering discussion above does not apply for methods that store nulls: 227 // then, there is no transitive reads in mutator (as we see nulls), and we can do 228 // relaxed memory ordering there. 229 230 inline void ShenandoahHeap::atomic_clear_oop(oop* addr, oop compare) { 231 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 232 Atomic::cmpxchg(addr, compare, oop(), memory_order_relaxed); 233 } 234 235 inline void ShenandoahHeap::atomic_clear_oop(narrowOop* addr, oop compare) { 236 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 237 narrowOop cmp = CompressedOops::encode(compare); 238 Atomic::cmpxchg(addr, cmp, narrowOop(), memory_order_relaxed); 239 } 240 241 inline void ShenandoahHeap::atomic_clear_oop(narrowOop* addr, narrowOop compare) { 242 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 243 Atomic::cmpxchg(addr, compare, narrowOop(), memory_order_relaxed); 244 } 245 246 inline bool ShenandoahHeap::cancelled_gc() const { 247 return _cancelled_gc.get() == CANCELLED; 248 } 249 250 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) { 251 if (sts_active && ShenandoahSuspendibleWorkers && !cancelled_gc()) { 252 if (SuspendibleThreadSet::should_yield()) { 253 SuspendibleThreadSet::yield(); 254 } 255 } 256 return cancelled_gc(); 257 } 258 259 inline void ShenandoahHeap::clear_cancelled_gc(bool clear_oom_handler) { 260 _cancelled_gc.set(CANCELLABLE); 261 if (_cancel_requested_time > 0) { 262 double cancel_time = os::elapsedTime() - _cancel_requested_time; 263 log_info(gc)("GC cancellation took %.3fs", cancel_time); 264 _cancel_requested_time = 0; 265 } 266 267 if (clear_oom_handler) { 268 _oom_evac_handler.clear(); 269 } 270 } 271 272 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) { 273 assert(UseTLAB, "TLABs should be enabled"); 274 275 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 276 if (gclab == nullptr) { 277 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(), 278 "Performance: thread should have GCLAB: %s", thread->name()); 279 // No GCLABs in this thread, fallback to shared allocation 280 return nullptr; 281 } 282 HeapWord* obj = gclab->allocate(size); 283 if (obj != nullptr) { 284 return obj; 285 } 286 return allocate_from_gclab_slow(thread, size); 287 } 288 289 inline HeapWord* ShenandoahHeap::allocate_from_plab(Thread* thread, size_t size, bool is_promotion) { 290 assert(UseTLAB, "TLABs should be enabled"); 291 292 PLAB* plab = ShenandoahThreadLocalData::plab(thread); 293 HeapWord* obj; 294 295 if (plab == nullptr) { 296 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(), "Performance: thread should have PLAB: %s", thread->name()); 297 // No PLABs in this thread, fallback to shared allocation 298 return nullptr; 299 } else if (is_promotion && !ShenandoahThreadLocalData::allow_plab_promotions(thread)) { 300 return nullptr; 301 } 302 // if plab->word_size() <= 0, thread's plab not yet initialized for this pass, so allow_plab_promotions() is not trustworthy 303 obj = plab->allocate(size); 304 if ((obj == nullptr) && (plab->words_remaining() < PLAB::min_size())) { 305 // allocate_from_plab_slow will establish allow_plab_promotions(thread) for future invocations 306 obj = allocate_from_plab_slow(thread, size, is_promotion); 307 } 308 // if plab->words_remaining() >= PLAB::min_size(), just return nullptr so we can use a shared allocation 309 if (obj == nullptr) { 310 return nullptr; 311 } 312 313 if (is_promotion) { 314 ShenandoahThreadLocalData::add_to_plab_promoted(thread, size * HeapWordSize); 315 } else { 316 ShenandoahThreadLocalData::add_to_plab_evacuated(thread, size * HeapWordSize); 317 } 318 return obj; 319 } 320 321 inline ShenandoahAgeCensus* ShenandoahHeap::age_census() const { 322 assert(mode()->is_generational(), "Only in generational mode"); 323 assert(_age_census != nullptr, "Error: not initialized"); 324 return _age_census; 325 } 326 327 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) { 328 assert(thread == Thread::current(), "Expected thread parameter to be current thread."); 329 if (ShenandoahThreadLocalData::is_oom_during_evac(thread)) { 330 // This thread went through the OOM during evac protocol and it is safe to return 331 // the forward pointer. It must not attempt to evacuate any more. 332 return ShenandoahBarrierSet::resolve_forwarded(p); 333 } 334 335 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope"); 336 337 ShenandoahHeapRegion* r = heap_region_containing(p); 338 assert(!r->is_humongous(), "never evacuate humongous objects"); 339 340 ShenandoahAffiliation target_gen = r->affiliation(); 341 if (mode()->is_generational() && ShenandoahHeap::heap()->is_gc_generation_young() && 342 target_gen == YOUNG_GENERATION) { 343 markWord mark = p->mark(); 344 if (mark.is_marked()) { 345 // Already forwarded. 346 return ShenandoahBarrierSet::resolve_forwarded(p); 347 } 348 if (mark.has_displaced_mark_helper()) { 349 // We don't want to deal with MT here just to ensure we read the right mark word. 350 // Skip the potential promotion attempt for this one. 351 } else if (r->age() + mark.age() >= age_census()->tenuring_threshold()) { 352 oop result = try_evacuate_object(p, thread, r, OLD_GENERATION); 353 if (result != nullptr) { 354 return result; 355 } 356 // If we failed to promote this aged object, we'll fall through to code below and evacuate to young-gen. 357 } 358 } 359 return try_evacuate_object(p, thread, r, target_gen); 360 } 361 362 // try_evacuate_object registers the object and dirties the associated remembered set information when evacuating 363 // to OLD_GENERATION. 364 inline oop ShenandoahHeap::try_evacuate_object(oop p, Thread* thread, ShenandoahHeapRegion* from_region, 365 ShenandoahAffiliation target_gen) { 366 bool alloc_from_lab = true; 367 bool has_plab = false; 368 HeapWord* copy = nullptr; 369 size_t size = p->size(); 370 bool is_promotion = (target_gen == OLD_GENERATION) && from_region->is_young(); 371 372 #ifdef ASSERT 373 if (ShenandoahOOMDuringEvacALot && 374 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call 375 copy = nullptr; 376 } else { 377 #endif 378 if (UseTLAB) { 379 switch (target_gen) { 380 case YOUNG_GENERATION: { 381 copy = allocate_from_gclab(thread, size); 382 if ((copy == nullptr) && (size < ShenandoahThreadLocalData::gclab_size(thread))) { 383 // GCLAB allocation failed because we are bumping up against the limit on young evacuation reserve. Try resetting 384 // the desired GCLAB size and retry GCLAB allocation to avoid cascading of shared memory allocations. 385 ShenandoahThreadLocalData::set_gclab_size(thread, PLAB::min_size()); 386 copy = allocate_from_gclab(thread, size); 387 // If we still get nullptr, we'll try a shared allocation below. 388 } 389 break; 390 } 391 case OLD_GENERATION: { 392 PLAB* plab = ShenandoahThreadLocalData::plab(thread); 393 if (plab != nullptr) { 394 has_plab = true; 395 } 396 copy = allocate_from_plab(thread, size, is_promotion); 397 if ((copy == nullptr) && (size < ShenandoahThreadLocalData::plab_size(thread)) && 398 ShenandoahThreadLocalData::plab_retries_enabled(thread)) { 399 // PLAB allocation failed because we are bumping up against the limit on old evacuation reserve or because 400 // the requested object does not fit within the current plab but the plab still has an "abundance" of memory, 401 // where abundance is defined as >= PLAB::min_size(). In the former case, we try resetting the desired 402 // PLAB size and retry PLAB allocation to avoid cascading of shared memory allocations. 403 404 // In this situation, PLAB memory is precious. We'll try to preserve our existing PLAB by forcing 405 // this particular allocation to be shared. 406 if (plab->words_remaining() < PLAB::min_size()) { 407 ShenandoahThreadLocalData::set_plab_size(thread, PLAB::min_size()); 408 copy = allocate_from_plab(thread, size, is_promotion); 409 // If we still get nullptr, we'll try a shared allocation below. 410 if (copy == nullptr) { 411 // If retry fails, don't continue to retry until we have success (probably in next GC pass) 412 ShenandoahThreadLocalData::disable_plab_retries(thread); 413 } 414 } 415 // else, copy still equals nullptr. this causes shared allocation below, preserving this plab for future needs. 416 } 417 break; 418 } 419 default: { 420 ShouldNotReachHere(); 421 break; 422 } 423 } 424 } 425 426 if (copy == nullptr) { 427 // If we failed to allocate in LAB, we'll try a shared allocation. 428 if (!is_promotion || !has_plab || (size > PLAB::min_size())) { 429 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size, target_gen); 430 copy = allocate_memory(req, is_promotion); 431 alloc_from_lab = false; 432 } 433 // else, we leave copy equal to nullptr, signaling a promotion failure below if appropriate. 434 // We choose not to promote objects smaller than PLAB::min_size() by way of shared allocations, as this is too 435 // costly. Instead, we'll simply "evacuate" to young-gen memory (using a GCLAB) and will promote in a future 436 // evacuation pass. This condition is denoted by: is_promotion && has_plab && (size <= PLAB::min_size()) 437 } 438 #ifdef ASSERT 439 } 440 #endif 441 442 if (copy == nullptr) { 443 if (target_gen == OLD_GENERATION) { 444 assert(mode()->is_generational(), "Should only be here in generational mode."); 445 if (from_region->is_young()) { 446 // Signal that promotion failed. Will evacuate this old object somewhere in young gen. 447 report_promotion_failure(thread, size); 448 return nullptr; 449 } else { 450 // Remember that evacuation to old gen failed. We'll want to trigger a full gc to recover from this 451 // after the evacuation threads have finished. 452 handle_old_evacuation_failure(); 453 } 454 } 455 456 control_thread()->handle_alloc_failure_evac(size); 457 458 _oom_evac_handler.handle_out_of_memory_during_evacuation(); 459 460 return ShenandoahBarrierSet::resolve_forwarded(p); 461 } 462 463 // Copy the object: 464 _evac_tracker->begin_evacuation(thread, size * HeapWordSize); 465 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(p), copy, size); 466 467 oop copy_val = cast_to_oop(copy); 468 469 if (mode()->is_generational() && target_gen == YOUNG_GENERATION && is_aging_cycle()) { 470 ShenandoahHeap::increase_object_age(copy_val, from_region->age() + 1); 471 } 472 473 // Try to install the new forwarding pointer. 474 ContinuationGCSupport::relativize_stack_chunk(copy_val); 475 476 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val); 477 if (result == copy_val) { 478 // Successfully evacuated. Our copy is now the public one! 479 _evac_tracker->end_evacuation(thread, size * HeapWordSize); 480 if (mode()->is_generational()) { 481 if (target_gen == OLD_GENERATION) { 482 handle_old_evacuation(copy, size, from_region->is_young()); 483 } else { 484 // When copying to the old generation above, we don't care 485 // about recording object age in the census stats. 486 assert(target_gen == YOUNG_GENERATION, "Error"); 487 // We record this census only when simulating pre-adaptive tenuring behavior, or 488 // when we have been asked to record the census at evacuation rather than at mark 489 if (ShenandoahGenerationalCensusAtEvac || !ShenandoahGenerationalAdaptiveTenuring) { 490 _evac_tracker->record_age(thread, size * HeapWordSize, ShenandoahHeap::get_object_age(copy_val)); 491 } 492 } 493 } 494 shenandoah_assert_correct(nullptr, copy_val); 495 return copy_val; 496 } else { 497 // Failed to evacuate. We need to deal with the object that is left behind. Since this 498 // new allocation is certainly after TAMS, it will be considered live in the next cycle. 499 // But if it happens to contain references to evacuated regions, those references would 500 // not get updated for this stale copy during this cycle, and we will crash while scanning 501 // it the next cycle. 502 if (alloc_from_lab) { 503 // For LAB allocations, it is enough to rollback the allocation ptr. Either the next 504 // object will overwrite this stale copy, or the filler object on LAB retirement will 505 // do this. 506 switch (target_gen) { 507 case YOUNG_GENERATION: { 508 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size); 509 break; 510 } 511 case OLD_GENERATION: { 512 ShenandoahThreadLocalData::plab(thread)->undo_allocation(copy, size); 513 if (is_promotion) { 514 ShenandoahThreadLocalData::subtract_from_plab_promoted(thread, size * HeapWordSize); 515 } else { 516 ShenandoahThreadLocalData::subtract_from_plab_evacuated(thread, size * HeapWordSize); 517 } 518 break; 519 } 520 default: { 521 ShouldNotReachHere(); 522 break; 523 } 524 } 525 } else { 526 // For non-LAB allocations, we have no way to retract the allocation, and 527 // have to explicitly overwrite the copy with the filler object. With that overwrite, 528 // we have to keep the fwdptr initialized and pointing to our (stale) copy. 529 assert(size >= ShenandoahHeap::min_fill_size(), "previously allocated object known to be larger than min_size"); 530 fill_with_object(copy, size); 531 shenandoah_assert_correct(nullptr, copy_val); 532 // For non-LAB allocations, the object has already been registered 533 } 534 shenandoah_assert_correct(nullptr, result); 535 return result; 536 } 537 } 538 539 void ShenandoahHeap::increase_object_age(oop obj, uint additional_age) { 540 markWord w = obj->has_displaced_mark() ? obj->displaced_mark() : obj->mark(); 541 w = w.set_age(MIN2(markWord::max_age, w.age() + additional_age)); 542 if (obj->has_displaced_mark()) { 543 obj->set_displaced_mark(w); 544 } else { 545 obj->set_mark(w); 546 } 547 } 548 549 // Return the object's age (at a safepoint or when object isn't 550 // mutable by the mutator) 551 uint ShenandoahHeap::get_object_age(oop obj) { 552 markWord w = obj->has_displaced_mark() ? obj->displaced_mark() : obj->mark(); 553 assert(w.age() <= markWord::max_age, "Impossible!"); 554 return w.age(); 555 } 556 557 // Return the object's age, or a sentinel value when the age can't 558 // necessarily be determined because of concurrent locking by the 559 // mutator 560 uint ShenandoahHeap::get_object_age_concurrent(oop obj) { 561 // This is impossible to do unless we "freeze" ABA-type oscillations 562 // With Lilliput, we can do this more easily. 563 markWord w = obj->mark(); 564 // We can do better for objects with inflated monitor 565 if (w.is_being_inflated() || w.has_displaced_mark_helper()) { 566 // Informs caller that we aren't able to determine the age 567 return markWord::max_age + 1; // sentinel 568 } 569 assert(w.age() <= markWord::max_age, "Impossible!"); 570 return w.age(); 571 } 572 573 inline bool ShenandoahHeap::clear_old_evacuation_failure() { 574 return _old_gen_oom_evac.try_unset(); 575 } 576 577 bool ShenandoahHeap::is_in(const void* p) const { 578 HeapWord* heap_base = (HeapWord*) base(); 579 HeapWord* last_region_end = heap_base + ShenandoahHeapRegion::region_size_words() * num_regions(); 580 return p >= heap_base && p < last_region_end; 581 } 582 583 inline bool ShenandoahHeap::is_in_active_generation(oop obj) const { 584 if (!mode()->is_generational()) { 585 // everything is the same single generation 586 return true; 587 } 588 589 if (active_generation() == nullptr) { 590 // no collection is happening, only expect this to be called 591 // when concurrent processing is active, but that could change 592 return false; 593 } 594 595 assert(is_in(obj), "only check if is in active generation for objects (" PTR_FORMAT ") in heap", p2i(obj)); 596 assert((active_generation() == (ShenandoahGeneration*) old_generation()) || 597 (active_generation() == (ShenandoahGeneration*) young_generation()) || 598 (active_generation() == global_generation()), "Active generation must be old, young, or global"); 599 600 size_t index = heap_region_containing(obj)->index(); 601 switch (_affiliations[index]) { 602 case ShenandoahAffiliation::FREE: 603 // Free regions are in Old, Young, Global 604 return true; 605 case ShenandoahAffiliation::YOUNG_GENERATION: 606 // Young regions are in young_generation and global_generation, not in old_generation 607 return (active_generation() != (ShenandoahGeneration*) old_generation()); 608 case ShenandoahAffiliation::OLD_GENERATION: 609 // Old regions are in old_generation and global_generation, not in young_generation 610 return (active_generation() != (ShenandoahGeneration*) young_generation()); 611 default: 612 assert(false, "Bad affiliation (%d) for region " SIZE_FORMAT, _affiliations[index], index); 613 return false; 614 } 615 } 616 617 inline bool ShenandoahHeap::is_in_young(const void* p) const { 618 return is_in(p) && (_affiliations[heap_region_index_containing(p)] == ShenandoahAffiliation::YOUNG_GENERATION); 619 } 620 621 inline bool ShenandoahHeap::is_in_old(const void* p) const { 622 return is_in(p) && (_affiliations[heap_region_index_containing(p)] == ShenandoahAffiliation::OLD_GENERATION); 623 } 624 625 inline bool ShenandoahHeap::is_old(oop obj) const { 626 return is_gc_generation_young() && is_in_old(obj); 627 } 628 629 inline ShenandoahAffiliation ShenandoahHeap::region_affiliation(const ShenandoahHeapRegion *r) { 630 return (ShenandoahAffiliation) _affiliations[r->index()]; 631 } 632 633 inline void ShenandoahHeap::assert_lock_for_affiliation(ShenandoahAffiliation orig_affiliation, 634 ShenandoahAffiliation new_affiliation) { 635 // A lock is required when changing from FREE to NON-FREE. Though it may be possible to elide the lock when 636 // transitioning from in-use to FREE, the current implementation uses a lock for this transition. A lock is 637 // not required to change from YOUNG to OLD (i.e. when promoting humongous region). 638 // 639 // new_affiliation is: FREE YOUNG OLD 640 // orig_affiliation is: FREE X L L 641 // YOUNG L X 642 // OLD L X X 643 // X means state transition won't happen (so don't care) 644 // L means lock should be held 645 // Blank means no lock required because affiliation visibility will not be required until subsequent safepoint 646 // 647 // Note: during full GC, all transitions between states are possible. During Full GC, we should be in a safepoint. 648 649 if ((orig_affiliation == ShenandoahAffiliation::FREE) || (new_affiliation == ShenandoahAffiliation::FREE)) { 650 shenandoah_assert_heaplocked_or_fullgc_safepoint(); 651 } 652 } 653 654 inline void ShenandoahHeap::set_affiliation(ShenandoahHeapRegion* r, ShenandoahAffiliation new_affiliation) { 655 #ifdef ASSERT 656 assert_lock_for_affiliation(region_affiliation(r), new_affiliation); 657 #endif 658 _affiliations[r->index()] = (uint8_t) new_affiliation; 659 } 660 661 inline ShenandoahAffiliation ShenandoahHeap::region_affiliation(size_t index) { 662 return (ShenandoahAffiliation) _affiliations[index]; 663 } 664 665 inline void ShenandoahHeap::set_affiliation(size_t index, ShenandoahAffiliation new_affiliation) { 666 #ifdef ASSERT 667 assert_lock_for_affiliation(region_affiliation(index), new_affiliation); 668 #endif 669 _affiliations[index] = (uint8_t) new_affiliation; 670 } 671 672 inline bool ShenandoahHeap::requires_marking(const void* entry) const { 673 oop obj = cast_to_oop(entry); 674 return !_marking_context->is_marked_strong(obj); 675 } 676 677 inline bool ShenandoahHeap::in_collection_set(oop p) const { 678 assert(collection_set() != nullptr, "Sanity"); 679 return collection_set()->is_in(p); 680 } 681 682 inline bool ShenandoahHeap::in_collection_set_loc(void* p) const { 683 assert(collection_set() != nullptr, "Sanity"); 684 return collection_set()->is_in_loc(p); 685 } 686 687 688 inline bool ShenandoahHeap::is_stable() const { 689 return _gc_state.is_clear(); 690 } 691 692 inline bool ShenandoahHeap::has_evacuation_reserve_quantities() const { 693 return _has_evacuation_reserve_quantities; 694 } 695 696 inline bool ShenandoahHeap::is_idle() const { 697 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS); 698 } 699 700 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const { 701 return _gc_state.is_set(MARKING); 702 } 703 704 inline bool ShenandoahHeap::is_concurrent_young_mark_in_progress() const { 705 return _gc_state.is_set(YOUNG_MARKING); 706 } 707 708 inline bool ShenandoahHeap::is_concurrent_old_mark_in_progress() const { 709 return _gc_state.is_set(OLD_MARKING); 710 } 711 712 inline bool ShenandoahHeap::is_evacuation_in_progress() const { 713 return _gc_state.is_set(EVACUATION); 714 } 715 716 inline bool ShenandoahHeap::is_gc_in_progress_mask(uint mask) const { 717 return _gc_state.is_set(mask); 718 } 719 720 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const { 721 return _degenerated_gc_in_progress.is_set(); 722 } 723 724 inline bool ShenandoahHeap::is_full_gc_in_progress() const { 725 return _full_gc_in_progress.is_set(); 726 } 727 728 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const { 729 return _full_gc_move_in_progress.is_set(); 730 } 731 732 inline bool ShenandoahHeap::is_update_refs_in_progress() const { 733 return _gc_state.is_set(UPDATEREFS); 734 } 735 736 inline bool ShenandoahHeap::is_stw_gc_in_progress() const { 737 return is_full_gc_in_progress() || is_degenerated_gc_in_progress(); 738 } 739 740 inline bool ShenandoahHeap::is_concurrent_strong_root_in_progress() const { 741 return _concurrent_strong_root_in_progress.is_set(); 742 } 743 744 inline bool ShenandoahHeap::is_concurrent_weak_root_in_progress() const { 745 return _gc_state.is_set(WEAK_ROOTS); 746 } 747 748 inline bool ShenandoahHeap::is_aging_cycle() const { 749 return _is_aging_cycle.is_set(); 750 } 751 752 inline bool ShenandoahHeap::is_prepare_for_old_mark_in_progress() const { 753 return _prepare_for_old_mark; 754 } 755 756 inline size_t ShenandoahHeap::set_promoted_reserve(size_t new_val) { 757 size_t orig = _promoted_reserve; 758 _promoted_reserve = new_val; 759 return orig; 760 } 761 762 inline size_t ShenandoahHeap::get_promoted_reserve() const { 763 return _promoted_reserve; 764 } 765 766 // returns previous value 767 size_t ShenandoahHeap::capture_old_usage(size_t old_usage) { 768 size_t previous_value = _captured_old_usage; 769 _captured_old_usage = old_usage; 770 return previous_value; 771 } 772 773 void ShenandoahHeap::set_previous_promotion(size_t promoted_bytes) { 774 shenandoah_assert_heaplocked(); 775 _previous_promotion = promoted_bytes; 776 } 777 778 size_t ShenandoahHeap::get_previous_promotion() const { 779 return _previous_promotion; 780 } 781 782 inline size_t ShenandoahHeap::set_old_evac_reserve(size_t new_val) { 783 size_t orig = _old_evac_reserve; 784 _old_evac_reserve = new_val; 785 return orig; 786 } 787 788 inline size_t ShenandoahHeap::get_old_evac_reserve() const { 789 return _old_evac_reserve; 790 } 791 792 inline void ShenandoahHeap::augment_old_evac_reserve(size_t increment) { 793 _old_evac_reserve += increment; 794 } 795 796 inline void ShenandoahHeap::augment_promo_reserve(size_t increment) { 797 _promoted_reserve += increment; 798 } 799 800 inline void ShenandoahHeap::reset_old_evac_expended() { 801 Atomic::store(&_old_evac_expended, (size_t) 0); 802 } 803 804 inline size_t ShenandoahHeap::expend_old_evac(size_t increment) { 805 return Atomic::add(&_old_evac_expended, increment); 806 } 807 808 inline size_t ShenandoahHeap::get_old_evac_expended() { 809 return Atomic::load(&_old_evac_expended); 810 } 811 812 inline void ShenandoahHeap::reset_promoted_expended() { 813 Atomic::store(&_promoted_expended, (size_t) 0); 814 } 815 816 inline size_t ShenandoahHeap::expend_promoted(size_t increment) { 817 return Atomic::add(&_promoted_expended, increment); 818 } 819 820 inline size_t ShenandoahHeap::unexpend_promoted(size_t decrement) { 821 return Atomic::sub(&_promoted_expended, decrement); 822 } 823 824 inline size_t ShenandoahHeap::get_promoted_expended() { 825 return Atomic::load(&_promoted_expended); 826 } 827 828 inline size_t ShenandoahHeap::set_young_evac_reserve(size_t new_val) { 829 size_t orig = _young_evac_reserve; 830 _young_evac_reserve = new_val; 831 return orig; 832 } 833 834 inline size_t ShenandoahHeap::get_young_evac_reserve() const { 835 return _young_evac_reserve; 836 } 837 838 template<class T> 839 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) { 840 marked_object_iterate(region, cl, region->top()); 841 } 842 843 template<class T> 844 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) { 845 assert(! region->is_humongous_continuation(), "no humongous continuation regions here"); 846 847 ShenandoahMarkingContext* const ctx = marking_context(); 848 849 HeapWord* tams = ctx->top_at_mark_start(region); 850 851 size_t skip_bitmap_delta = 1; 852 HeapWord* start = region->bottom(); 853 HeapWord* end = MIN2(tams, region->end()); 854 855 // Step 1. Scan below the TAMS based on bitmap data. 856 HeapWord* limit_bitmap = MIN2(limit, tams); 857 858 // Try to scan the initial candidate. If the candidate is above the TAMS, it would 859 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2. 860 HeapWord* cb = ctx->get_next_marked_addr(start, end); 861 862 intx dist = ShenandoahMarkScanPrefetch; 863 if (dist > 0) { 864 // Batched scan that prefetches the oop data, anticipating the access to 865 // either header, oop field, or forwarding pointer. Not that we cannot 866 // touch anything in oop, while it still being prefetched to get enough 867 // time for prefetch to work. This is why we try to scan the bitmap linearly, 868 // disregarding the object size. However, since we know forwarding pointer 869 // precedes the object, we can skip over it. Once we cannot trust the bitmap, 870 // there is no point for prefetching the oop contents, as oop->size() will 871 // touch it prematurely. 872 873 // No variable-length arrays in standard C++, have enough slots to fit 874 // the prefetch distance. 875 static const int SLOT_COUNT = 256; 876 guarantee(dist <= SLOT_COUNT, "adjust slot count"); 877 HeapWord* slots[SLOT_COUNT]; 878 879 int avail; 880 do { 881 avail = 0; 882 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) { 883 Prefetch::read(cb, oopDesc::mark_offset_in_bytes()); 884 slots[avail++] = cb; 885 cb += skip_bitmap_delta; 886 if (cb < limit_bitmap) { 887 cb = ctx->get_next_marked_addr(cb, limit_bitmap); 888 } 889 } 890 891 for (int c = 0; c < avail; c++) { 892 assert (slots[c] < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams)); 893 assert (slots[c] < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit)); 894 oop obj = cast_to_oop(slots[c]); 895 assert(oopDesc::is_oop(obj), "sanity"); 896 assert(ctx->is_marked(obj), "object expected to be marked"); 897 cl->do_object(obj); 898 } 899 } while (avail > 0); 900 } else { 901 while (cb < limit_bitmap) { 902 assert (cb < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams)); 903 assert (cb < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit)); 904 oop obj = cast_to_oop(cb); 905 assert(oopDesc::is_oop(obj), "sanity"); 906 assert(ctx->is_marked(obj), "object expected to be marked"); 907 cl->do_object(obj); 908 cb += skip_bitmap_delta; 909 if (cb < limit_bitmap) { 910 cb = ctx->get_next_marked_addr(cb, limit_bitmap); 911 } 912 } 913 } 914 915 // Step 2. Accurate size-based traversal, happens past the TAMS. 916 // This restarts the scan at TAMS, which makes sure we traverse all objects, 917 // regardless of what happened at Step 1. 918 HeapWord* cs = tams; 919 while (cs < limit) { 920 assert (cs >= tams, "only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams)); 921 assert (cs < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit)); 922 oop obj = cast_to_oop(cs); 923 assert(oopDesc::is_oop(obj), "sanity"); 924 assert(ctx->is_marked(obj), "object expected to be marked"); 925 size_t size = obj->size(); 926 cl->do_object(obj); 927 cs += size; 928 } 929 } 930 931 template <class T> 932 class ShenandoahObjectToOopClosure : public ObjectClosure { 933 T* _cl; 934 public: 935 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {} 936 937 void do_object(oop obj) { 938 obj->oop_iterate(_cl); 939 } 940 }; 941 942 template <class T> 943 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure { 944 T* _cl; 945 MemRegion _bounds; 946 public: 947 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) : 948 _cl(cl), _bounds(bottom, top) {} 949 950 void do_object(oop obj) { 951 obj->oop_iterate(_cl, _bounds); 952 } 953 }; 954 955 template<class T> 956 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) { 957 if (region->is_humongous()) { 958 HeapWord* bottom = region->bottom(); 959 if (top > bottom) { 960 region = region->humongous_start_region(); 961 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top); 962 marked_object_iterate(region, &objs); 963 } 964 } else { 965 ShenandoahObjectToOopClosure<T> objs(cl); 966 marked_object_iterate(region, &objs, top); 967 } 968 } 969 970 inline ShenandoahHeapRegion* ShenandoahHeap::get_region(size_t region_idx) const { 971 if (region_idx < _num_regions) { 972 return _regions[region_idx]; 973 } else { 974 return nullptr; 975 } 976 } 977 978 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const { 979 assert (_marking_context->is_complete()," sanity"); 980 return _marking_context; 981 } 982 983 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const { 984 return _marking_context; 985 } 986 987 inline void ShenandoahHeap::clear_cards_for(ShenandoahHeapRegion* region) { 988 if (mode()->is_generational()) { 989 _card_scan->mark_range_as_empty(region->bottom(), pointer_delta(region->end(), region->bottom())); 990 } 991 } 992 993 inline void ShenandoahHeap::dirty_cards(HeapWord* start, HeapWord* end) { 994 assert(mode()->is_generational(), "Should only be used for generational mode"); 995 size_t words = pointer_delta(end, start); 996 _card_scan->mark_range_as_dirty(start, words); 997 } 998 999 inline void ShenandoahHeap::clear_cards(HeapWord* start, HeapWord* end) { 1000 assert(mode()->is_generational(), "Should only be used for generational mode"); 1001 size_t words = pointer_delta(end, start); 1002 _card_scan->mark_range_as_clean(start, words); 1003 } 1004 1005 inline void ShenandoahHeap::mark_card_as_dirty(void* location) { 1006 if (mode()->is_generational()) { 1007 _card_scan->mark_card_as_dirty((HeapWord*)location); 1008 } 1009 } 1010 1011 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP