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