1 /* 2 * Copyright (c) 2015, 2020, Red Hat, Inc. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP 26 #define SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP 27 28 #include "classfile/javaClasses.inline.hpp" 29 #include "gc/shared/markBitMap.inline.hpp" 30 #include "gc/shared/threadLocalAllocBuffer.inline.hpp" 31 #include "gc/shared/suspendibleThreadSet.hpp" 32 #include "gc/shenandoah/shenandoahAsserts.hpp" 33 #include "gc/shenandoah/shenandoahBarrierSet.inline.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/compressedOops.inline.hpp" 44 #include "oops/oop.inline.hpp" 45 #include "runtime/atomic.hpp" 46 #include "runtime/prefetch.inline.hpp" 47 #include "runtime/thread.hpp" 48 #include "utilities/copy.hpp" 49 #include "utilities/globalDefinitions.hpp" 50 51 inline ShenandoahHeap* ShenandoahHeap::heap() { 52 assert(_heap != NULL, "Heap is not initialized yet"); 53 return _heap; 54 } 55 56 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() { 57 size_t new_index = Atomic::add(&_index, (size_t) 1); 58 // get_region() provides the bounds-check and returns NULL on OOB. 59 return _heap->get_region(new_index - 1); 60 } 61 62 inline bool ShenandoahHeap::has_forwarded_objects() const { 63 return _gc_state.is_set(HAS_FORWARDED); 64 } 65 66 inline WorkGang* ShenandoahHeap::workers() const { 67 return _workers; 68 } 69 70 inline WorkGang* ShenandoahHeap::get_safepoint_workers() { 71 return _safepoint_workers; 72 } 73 74 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const { 75 uintptr_t region_start = ((uintptr_t) addr); 76 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift(); 77 assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr)); 78 return index; 79 } 80 81 inline ShenandoahHeapRegion* const ShenandoahHeap::heap_region_containing(const void* addr) const { 82 size_t index = heap_region_index_containing(addr); 83 ShenandoahHeapRegion* const result = get_region(index); 84 assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr)); 85 return result; 86 } 87 88 template <class T> 89 inline oop ShenandoahHeap::update_with_forwarded_not_null(T* p, oop obj) { 90 if (in_collection_set(obj)) { 91 shenandoah_assert_forwarded_except(p, obj, is_full_gc_in_progress() || cancelled_gc() || is_degenerated_gc_in_progress()); 92 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 93 RawAccess<IS_NOT_NULL>::oop_store(p, obj); 94 } 95 #ifdef ASSERT 96 else { 97 shenandoah_assert_not_forwarded(p, obj); 98 } 99 #endif 100 return obj; 101 } 102 103 template <class T> 104 inline oop ShenandoahHeap::maybe_update_with_forwarded(T* p) { 105 T o = RawAccess<>::oop_load(p); 106 if (!CompressedOops::is_null(o)) { 107 oop obj = CompressedOops::decode_not_null(o); 108 return maybe_update_with_forwarded_not_null(p, obj); 109 } else { 110 return NULL; 111 } 112 } 113 114 template <class T> 115 inline oop ShenandoahHeap::evac_update_with_forwarded(T* p) { 116 T o = RawAccess<>::oop_load(p); 117 if (!CompressedOops::is_null(o)) { 118 oop obj = CompressedOops::decode_not_null(o); 119 if (in_collection_set(obj)) { 120 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 121 if (fwd == obj) { 122 fwd = evacuate_object(obj, Thread::current()); 123 } 124 if (fwd != obj) { // evac failure otherwise 125 oop prev = cas_oop(fwd, p, obj); 126 if (prev == obj) { 127 return fwd; 128 } 129 } 130 } 131 } 132 return NULL; 133 } 134 135 inline oop ShenandoahHeap::cas_oop(oop n, oop* addr, oop c) { 136 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 137 return (oop) Atomic::cmpxchg(addr, c, n); 138 } 139 140 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, narrowOop c) { 141 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 142 narrowOop val = CompressedOops::encode(n); 143 return CompressedOops::decode((narrowOop) Atomic::cmpxchg(addr, c, val)); 144 } 145 146 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, oop c) { 147 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr)); 148 narrowOop cmp = CompressedOops::encode(c); 149 narrowOop val = CompressedOops::encode(n); 150 return CompressedOops::decode((narrowOop) Atomic::cmpxchg(addr, cmp, val)); 151 } 152 153 template <class T> 154 inline oop ShenandoahHeap::maybe_update_with_forwarded_not_null(T* p, oop heap_oop) { 155 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || is_full_gc_in_progress() || is_degenerated_gc_in_progress()); 156 shenandoah_assert_correct(p, heap_oop); 157 158 if (in_collection_set(heap_oop)) { 159 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop); 160 if (forwarded_oop == heap_oop) { 161 // E.g. during evacuation. 162 return forwarded_oop; 163 } 164 165 shenandoah_assert_forwarded_except(p, heap_oop, is_full_gc_in_progress() || is_degenerated_gc_in_progress()); 166 shenandoah_assert_not_forwarded(p, forwarded_oop); 167 shenandoah_assert_not_in_cset_except(p, forwarded_oop, cancelled_gc()); 168 169 // If this fails, another thread wrote to p before us, it will be logged in SATB and the 170 // reference be updated later. 171 oop witness = cas_oop(forwarded_oop, p, heap_oop); 172 173 if (witness != heap_oop) { 174 // CAS failed, someone had beat us to it. Normally, we would return the failure witness, 175 // because that would be the proper write of to-space object, enforced by strong barriers. 176 // However, there is a corner case with arraycopy. It can happen that a Java thread 177 // beats us with an arraycopy, which first copies the array, which potentially contains 178 // from-space refs, and only afterwards updates all from-space refs to to-space refs, 179 // which leaves a short window where the new array elements can be from-space. 180 // In this case, we can just resolve the result again. As we resolve, we need to consider 181 // the contended write might have been NULL. 182 oop result = ShenandoahBarrierSet::resolve_forwarded(witness); 183 shenandoah_assert_not_forwarded_except(p, result, (result == NULL)); 184 shenandoah_assert_not_in_cset_except(p, result, (result == NULL) || cancelled_gc()); 185 return result; 186 } else { 187 // Success! We have updated with known to-space copy. We have already asserted it is sane. 188 return forwarded_oop; 189 } 190 } else { 191 shenandoah_assert_not_forwarded(p, heap_oop); 192 return heap_oop; 193 } 194 } 195 196 inline bool ShenandoahHeap::cancelled_gc() const { 197 return _cancelled_gc.get() == CANCELLED; 198 } 199 200 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) { 201 if (! (sts_active && ShenandoahSuspendibleWorkers)) { 202 return cancelled_gc(); 203 } 204 205 jbyte prev = _cancelled_gc.cmpxchg(NOT_CANCELLED, CANCELLABLE); 206 if (prev == CANCELLABLE || prev == NOT_CANCELLED) { 207 if (SuspendibleThreadSet::should_yield()) { 208 SuspendibleThreadSet::yield(); 209 } 210 211 // Back to CANCELLABLE. The thread that poked NOT_CANCELLED first gets 212 // to restore to CANCELLABLE. 213 if (prev == CANCELLABLE) { 214 _cancelled_gc.set(CANCELLABLE); 215 } 216 return false; 217 } else { 218 return true; 219 } 220 } 221 222 inline void ShenandoahHeap::clear_cancelled_gc() { 223 _cancelled_gc.set(CANCELLABLE); 224 _oom_evac_handler.clear(); 225 } 226 227 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) { 228 assert(UseTLAB, "TLABs should be enabled"); 229 230 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 231 if (gclab == NULL) { 232 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(), 233 "Performance: thread should have GCLAB: %s", thread->name()); 234 // No GCLABs in this thread, fallback to shared allocation 235 return NULL; 236 } 237 HeapWord* obj = gclab->allocate(size); 238 if (obj != NULL) { 239 return obj; 240 } 241 // Otherwise... 242 return allocate_from_gclab_slow(thread, size); 243 } 244 245 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) { 246 if (ShenandoahThreadLocalData::is_oom_during_evac(Thread::current())) { 247 // This thread went through the OOM during evac protocol and it is safe to return 248 // the forward pointer. It must not attempt to evacuate any more. 249 return ShenandoahBarrierSet::resolve_forwarded(p); 250 } 251 252 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope"); 253 254 size_t size = p->size(); 255 256 assert(!heap_region_containing(p)->is_humongous(), "never evacuate humongous objects"); 257 258 bool alloc_from_gclab = true; 259 HeapWord* copy = NULL; 260 261 #ifdef ASSERT 262 if (ShenandoahOOMDuringEvacALot && 263 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call 264 copy = NULL; 265 } else { 266 #endif 267 if (UseTLAB) { 268 copy = allocate_from_gclab(thread, size); 269 } 270 if (copy == NULL) { 271 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size); 272 copy = allocate_memory(req); 273 alloc_from_gclab = false; 274 } 275 #ifdef ASSERT 276 } 277 #endif 278 279 if (copy == NULL) { 280 control_thread()->handle_alloc_failure_evac(size); 281 282 _oom_evac_handler.handle_out_of_memory_during_evacuation(); 283 284 return ShenandoahBarrierSet::resolve_forwarded(p); 285 } 286 287 // Copy the object: 288 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(p), copy, size); 289 290 // Try to install the new forwarding pointer. 291 oop copy_val = oop(copy); 292 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val); 293 if (result == copy_val) { 294 // Successfully evacuated. Our copy is now the public one! 295 shenandoah_assert_correct(NULL, copy_val); 296 return copy_val; 297 } else { 298 // Failed to evacuate. We need to deal with the object that is left behind. Since this 299 // new allocation is certainly after TAMS, it will be considered live in the next cycle. 300 // But if it happens to contain references to evacuated regions, those references would 301 // not get updated for this stale copy during this cycle, and we will crash while scanning 302 // it the next cycle. 303 // 304 // For GCLAB allocations, it is enough to rollback the allocation ptr. Either the next 305 // object will overwrite this stale copy, or the filler object on LAB retirement will 306 // do this. For non-GCLAB allocations, we have no way to retract the allocation, and 307 // have to explicitly overwrite the copy with the filler object. With that overwrite, 308 // we have to keep the fwdptr initialized and pointing to our (stale) copy. 309 if (alloc_from_gclab) { 310 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size); 311 } else { 312 fill_with_object(copy, size); 313 shenandoah_assert_correct(NULL, copy_val); 314 } 315 shenandoah_assert_correct(NULL, result); 316 return result; 317 } 318 } 319 320 template<bool RESOLVE> 321 inline bool ShenandoahHeap::requires_marking(const void* entry) const { 322 oop obj = oop(entry); 323 if (RESOLVE) { 324 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 325 } 326 return !_marking_context->is_marked(obj); 327 } 328 329 inline bool ShenandoahHeap::in_collection_set(oop p) const { 330 assert(collection_set() != NULL, "Sanity"); 331 return collection_set()->is_in(p); 332 } 333 334 inline bool ShenandoahHeap::in_collection_set_loc(void* p) const { 335 assert(collection_set() != NULL, "Sanity"); 336 return collection_set()->is_in_loc(p); 337 } 338 339 inline bool ShenandoahHeap::is_stable() const { 340 return _gc_state.is_clear(); 341 } 342 343 inline bool ShenandoahHeap::is_idle() const { 344 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS); 345 } 346 347 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const { 348 return _gc_state.is_set(MARKING); 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_strong_root_in_progress() const { 380 return _concurrent_strong_root_in_progress.is_set(); 381 } 382 383 inline bool ShenandoahHeap::is_concurrent_weak_root_in_progress() const { 384 return _concurrent_weak_root_in_progress.is_set(); 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->get_next_marked_addr(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->get_next_marked_addr(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->get_next_marked_addr(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 inline ShenandoahHeapRegion* const ShenandoahHeap::get_region(size_t region_idx) const { 483 if (region_idx < _num_regions) { 484 return _regions[region_idx]; 485 } else { 486 return NULL; 487 } 488 } 489 490 inline void ShenandoahHeap::mark_complete_marking_context() { 491 _marking_context->mark_complete(); 492 } 493 494 inline void ShenandoahHeap::mark_incomplete_marking_context() { 495 _marking_context->mark_incomplete(); 496 } 497 498 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const { 499 assert (_marking_context->is_complete()," sanity"); 500 return _marking_context; 501 } 502 503 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const { 504 return _marking_context; 505 } 506 507 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP