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/shenandoahBrooksPointer.inline.hpp" 30 #include "gc_implementation/shenandoah/shenandoahAsserts.hpp" 31 #include "gc_implementation/shenandoah/shenandoahBarrierSet.inline.hpp" 32 #include "gc_implementation/shenandoah/shenandoahCollectionSet.hpp" 33 #include "gc_implementation/shenandoah/shenandoahCollectionSet.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 template <class T> 47 void ShenandoahUpdateRefsClosure::do_oop_work(T* p) { 48 T o = oopDesc::load_heap_oop(p); 49 if (! oopDesc::is_null(o)) { 50 oop obj = oopDesc::decode_heap_oop_not_null(o); 51 _heap->update_with_forwarded_not_null(p, obj); 52 } 53 } 54 55 void ShenandoahUpdateRefsClosure::do_oop(oop* p) { do_oop_work(p); } 56 void ShenandoahUpdateRefsClosure::do_oop(narrowOop* p) { do_oop_work(p); } 57 58 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() { 59 size_t new_index = Atomic::add((size_t) 1, &_index); 60 // get_region() provides the bounds-check and returns NULL on OOB. 61 return _heap->get_region(new_index - 1); 62 } 63 64 inline bool ShenandoahHeap::has_forwarded_objects() const { 65 return _gc_state.is_set(HAS_FORWARDED); 66 } 67 68 inline ShenandoahWorkGang* ShenandoahHeap::workers() const { 69 return _workers; 70 } 71 72 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const { 73 uintptr_t region_start = ((uintptr_t) addr); 74 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift(); 75 assert(index < num_regions(), err_msg("Region index is in bounds: " PTR_FORMAT, p2i(addr))); 76 return index; 77 } 78 79 inline ShenandoahHeapRegion* const ShenandoahHeap::heap_region_containing(const void* addr) const { 80 size_t index = heap_region_index_containing(addr); 81 ShenandoahHeapRegion* const result = get_region(index); 82 assert(addr >= result->bottom() && addr < result->end(), err_msg("Heap region contains the address: " PTR_FORMAT, p2i(addr))); 83 return result; 84 } 85 86 template <class T> 87 inline oop ShenandoahHeap::update_with_forwarded_not_null(T* p, oop obj) { 88 if (in_collection_set(obj)) { 89 shenandoah_assert_forwarded_except(p, obj, is_full_gc_in_progress() || cancelled_gc() || is_degenerated_gc_in_progress()); 90 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj); 91 oopDesc::encode_store_heap_oop(p, obj); 92 } 93 #ifdef ASSERT 94 else { 95 shenandoah_assert_not_forwarded(p, obj); 96 } 97 #endif 98 return obj; 99 } 100 101 template <class T> 102 inline oop ShenandoahHeap::maybe_update_with_forwarded(T* p) { 103 T o = oopDesc::load_heap_oop(p); 104 if (! oopDesc::is_null(o)) { 105 oop obj = oopDesc::decode_heap_oop_not_null(o); 106 return maybe_update_with_forwarded_not_null(p, obj); 107 } else { 108 return NULL; 109 } 110 } 111 112 inline oop ShenandoahHeap::atomic_compare_exchange_oop(oop n, oop* addr, oop c) { 113 return (oop) Atomic::cmpxchg_ptr(n, addr, c); 114 } 115 116 inline oop ShenandoahHeap::atomic_compare_exchange_oop(oop n, narrowOop* addr, oop c) { 117 narrowOop cmp = oopDesc::encode_heap_oop(c); 118 narrowOop val = oopDesc::encode_heap_oop(n); 119 return oopDesc::decode_heap_oop((narrowOop) Atomic::cmpxchg(val, addr, cmp)); 120 } 121 122 template <class T> 123 inline oop ShenandoahHeap::maybe_update_with_forwarded_not_null(T* p, oop heap_oop) { 124 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || is_full_gc_in_progress() || is_degenerated_gc_in_progress()); 125 shenandoah_assert_correct(p, heap_oop); 126 127 if (in_collection_set(heap_oop)) { 128 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop); 129 130 shenandoah_assert_forwarded_except(p, heap_oop, is_full_gc_in_progress() || is_degenerated_gc_in_progress()); 131 shenandoah_assert_not_in_cset_except(p, forwarded_oop, cancelled_gc()); 132 133 // If this fails, another thread wrote to p before us, it will be logged in SATB and the 134 // reference be updated later. 135 oop result = atomic_compare_exchange_oop(forwarded_oop, p, heap_oop); 136 137 if (oopDesc::unsafe_equals(result, heap_oop)) { // CAS successful. 138 return forwarded_oop; 139 } else { 140 // Note: we used to assert the following here. This doesn't work because sometimes, during 141 // marking/updating-refs, it can happen that a Java thread beats us with an arraycopy, 142 // which first copies the array, which potentially contains from-space refs, and only afterwards 143 // updates all from-space refs to to-space refs, which leaves a short window where the new array 144 // elements can be from-space. 145 // assert(oopDesc::is_null(result) || 146 // oopDesc::unsafe_equals(result, ShenandoahBarrierSet::resolve_oop_static_not_null(result)), 147 // "expect not forwarded"); 148 return NULL; 149 } 150 } else { 151 shenandoah_assert_not_forwarded(p, heap_oop); 152 return heap_oop; 153 } 154 } 155 156 inline bool ShenandoahHeap::cancelled_gc() const { 157 return _cancelled_gc.is_set(); 158 } 159 160 inline bool ShenandoahHeap::try_cancel_gc() { 161 return _cancelled_gc.try_set(); 162 } 163 164 inline void ShenandoahHeap::clear_cancelled_gc() { 165 _cancelled_gc.unset(); 166 _oom_evac_handler.clear(); 167 } 168 169 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) { 170 assert(UseTLAB, "TLABs should be enabled"); 171 172 if (!thread->gclab().is_initialized()) { 173 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(), 174 err_msg("Performance: thread should have GCLAB: %s", thread->name())); 175 // No GCLABs in this thread, fallback to shared allocation 176 return NULL; 177 } 178 HeapWord *obj = thread->gclab().allocate(size); 179 if (obj != NULL) { 180 return obj; 181 } 182 // Otherwise... 183 return allocate_from_gclab_slow(thread, size); 184 } 185 186 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread, bool& evacuated) { 187 evacuated = false; 188 189 if (Thread::current()->is_oom_during_evac()) { 190 // This thread went through the OOM during evac protocol and it is safe to return 191 // the forward pointer. It must not attempt to evacuate any more. 192 return ShenandoahBarrierSet::resolve_forwarded(p); 193 } 194 195 assert(thread->is_evac_allowed(), "must be enclosed in in oom-evac scope"); 196 197 size_t size_no_fwdptr = (size_t) p->size(); 198 size_t size_with_fwdptr = size_no_fwdptr + ShenandoahBrooksPointer::word_size(); 199 200 assert(!heap_region_containing(p)->is_humongous(), "never evacuate humongous objects"); 201 202 bool alloc_from_gclab = true; 203 HeapWord* filler = NULL; 204 205 #ifdef ASSERT 206 if (ShenandoahOOMDuringEvacALot && 207 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call 208 filler = NULL; 209 } else { 210 #endif 211 if (UseTLAB) { 212 filler = allocate_from_gclab(thread, size_with_fwdptr); 213 } 214 if (filler == NULL) { 215 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size_with_fwdptr); 216 filler = allocate_memory(req); 217 alloc_from_gclab = false; 218 } 219 #ifdef ASSERT 220 } 221 #endif 222 223 if (filler == NULL) { 224 control_thread()->handle_alloc_failure_evac(size_with_fwdptr); 225 226 _oom_evac_handler.handle_out_of_memory_during_evacuation(); 227 228 return ShenandoahBarrierSet::resolve_forwarded(p); 229 } 230 231 // Copy the object and initialize its forwarding ptr: 232 HeapWord* copy = filler + ShenandoahBrooksPointer::word_size(); 233 oop copy_val = oop(copy); 234 235 Copy::aligned_disjoint_words((HeapWord*) p, copy, size_no_fwdptr); 236 ShenandoahBrooksPointer::initialize(oop(copy)); 237 238 // Try to install the new forwarding pointer. 239 oop result = ShenandoahBrooksPointer::try_update_forwardee(p, copy_val); 240 241 if (oopDesc::unsafe_equals(result, p)) { 242 // Successfully evacuated. Our copy is now the public one! 243 evacuated = true; 244 shenandoah_assert_correct(NULL, copy_val); 245 return copy_val; 246 } else { 247 // Failed to evacuate. We need to deal with the object that is left behind. Since this 248 // new allocation is certainly after TAMS, it will be considered live in the next cycle. 249 // But if it happens to contain references to evacuated regions, those references would 250 // not get updated for this stale copy during this cycle, and we will crash while scanning 251 // it the next cycle. 252 // 253 // For GCLAB allocations, it is enough to rollback the allocation ptr. Either the next 254 // object will overwrite this stale copy, or the filler object on LAB retirement will 255 // do this. For non-GCLAB allocations, we have no way to retract the allocation, and 256 // have to explicitly overwrite the copy with the filler object. With that overwrite, 257 // we have to keep the fwdptr initialized and pointing to our (stale) copy. 258 if (alloc_from_gclab) { 259 thread->gclab().rollback(size_with_fwdptr); 260 } else { 261 fill_with_object(copy, size_no_fwdptr); 262 } 263 shenandoah_assert_correct(NULL, copy_val); 264 shenandoah_assert_correct(NULL, result); 265 return result; 266 } 267 } 268 269 inline bool ShenandoahHeap::requires_marking(const void* entry) const { 270 return !_marking_context->is_marked(oop(entry)); 271 } 272 273 template <class T> 274 inline bool ShenandoahHeap::in_collection_set(T p) const { 275 HeapWord* obj = (HeapWord*) p; 276 assert(collection_set() != NULL, "Sanity"); 277 assert(is_in(obj), "should be in heap"); 278 279 return collection_set()->is_in(obj); 280 } 281 282 inline bool ShenandoahHeap::is_stable() const { 283 return _gc_state.is_clear(); 284 } 285 286 inline bool ShenandoahHeap::is_idle() const { 287 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS); 288 } 289 290 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const { 291 return _gc_state.is_set(MARKING); 292 } 293 294 inline bool ShenandoahHeap::is_evacuation_in_progress() const { 295 return _gc_state.is_set(EVACUATION); 296 } 297 298 inline bool ShenandoahHeap::is_gc_in_progress_mask(uint mask) const { 299 return _gc_state.is_set(mask); 300 } 301 302 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const { 303 return _degenerated_gc_in_progress.is_set(); 304 } 305 306 inline bool ShenandoahHeap::is_full_gc_in_progress() const { 307 return _full_gc_in_progress.is_set(); 308 } 309 310 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const { 311 return _full_gc_move_in_progress.is_set(); 312 } 313 314 inline bool ShenandoahHeap::is_update_refs_in_progress() const { 315 return _gc_state.is_set(UPDATEREFS); 316 } 317 318 template<class T> 319 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) { 320 marked_object_iterate(region, cl, region->top()); 321 } 322 323 template<class T> 324 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) { 325 assert(ShenandoahBrooksPointer::word_offset() < 0, "skip_delta calculation below assumes the forwarding ptr is before obj"); 326 327 ShenandoahMarkingContext* const ctx = complete_marking_context(); 328 assert(ctx->is_complete(), "sanity"); 329 330 MarkBitMap* mark_bit_map = ctx->mark_bit_map(); 331 HeapWord* tams = ctx->top_at_mark_start(region); 332 333 size_t skip_bitmap_delta = ShenandoahBrooksPointer::word_size() + 1; 334 size_t skip_objsize_delta = ShenandoahBrooksPointer::word_size() /* + actual obj.size() below */; 335 HeapWord* start = region->bottom() + ShenandoahBrooksPointer::word_size(); 336 HeapWord* end = MIN2(tams + ShenandoahBrooksPointer::word_size(), region->end()); 337 338 // Step 1. Scan below the TAMS based on bitmap data. 339 HeapWord* limit_bitmap = MIN2(limit, tams); 340 341 // Try to scan the initial candidate. If the candidate is above the TAMS, it would 342 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2. 343 HeapWord* cb = mark_bit_map->getNextMarkedWordAddress(start, end); 344 345 intx dist = ShenandoahMarkScanPrefetch; 346 if (dist > 0) { 347 // Batched scan that prefetches the oop data, anticipating the access to 348 // either header, oop field, or forwarding pointer. Not that we cannot 349 // touch anything in oop, while it still being prefetched to get enough 350 // time for prefetch to work. This is why we try to scan the bitmap linearly, 351 // disregarding the object size. However, since we know forwarding pointer 352 // preceeds the object, we can skip over it. Once we cannot trust the bitmap, 353 // there is no point for prefetching the oop contents, as oop->size() will 354 // touch it prematurely. 355 356 // No variable-length arrays in standard C++, have enough slots to fit 357 // the prefetch distance. 358 static const int SLOT_COUNT = 256; 359 guarantee(dist <= SLOT_COUNT, "adjust slot count"); 360 HeapWord* slots[SLOT_COUNT]; 361 362 int avail; 363 do { 364 avail = 0; 365 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) { 366 Prefetch::read(cb, ShenandoahBrooksPointer::byte_offset()); 367 slots[avail++] = cb; 368 cb += skip_bitmap_delta; 369 if (cb < limit_bitmap) { 370 cb = mark_bit_map->getNextMarkedWordAddress(cb, limit_bitmap); 371 } 372 } 373 374 for (int c = 0; c < avail; c++) { 375 assert (slots[c] < tams, err_msg("only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams))); 376 assert (slots[c] < limit, err_msg("only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit))); 377 oop obj = oop(slots[c]); 378 assert(!oopDesc::is_null(obj), "sanity"); 379 assert(obj->is_oop(), "sanity"); 380 assert(_marking_context->is_marked(obj), "object expected to be marked"); 381 cl->do_object(obj); 382 } 383 } while (avail > 0); 384 } else { 385 while (cb < limit_bitmap) { 386 assert (cb < tams, err_msg("only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams))); 387 assert (cb < limit, err_msg("only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit))); 388 oop obj = oop(cb); 389 assert(!oopDesc::is_null(obj), "sanity"); 390 assert(obj->is_oop(), "sanity"); 391 assert(_marking_context->is_marked(obj), "object expected to be marked"); 392 cl->do_object(obj); 393 cb += skip_bitmap_delta; 394 if (cb < limit_bitmap) { 395 cb = mark_bit_map->getNextMarkedWordAddress(cb, limit_bitmap); 396 } 397 } 398 } 399 400 // Step 2. Accurate size-based traversal, happens past the TAMS. 401 // This restarts the scan at TAMS, which makes sure we traverse all objects, 402 // regardless of what happened at Step 1. 403 HeapWord* cs = tams + ShenandoahBrooksPointer::word_size(); 404 while (cs < limit) { 405 assert (cs > tams, err_msg("only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams))); 406 assert (cs < limit, err_msg("only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit))); 407 oop obj = oop(cs); 408 int size = obj->size(); 409 assert(!oopDesc::is_null(obj), "sanity"); 410 assert(obj->is_oop(), "sanity"); 411 assert(_marking_context->is_marked(obj), "object expected to be marked"); 412 cl->do_object(obj); 413 cs += size + skip_objsize_delta; 414 } 415 } 416 417 template <class T> 418 class ShenandoahObjectToOopClosure : public ObjectClosure { 419 T* _cl; 420 public: 421 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {} 422 423 void do_object(oop obj) { 424 obj->oop_iterate(_cl); 425 } 426 }; 427 428 template <class T> 429 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure { 430 T* _cl; 431 MemRegion _bounds; 432 public: 433 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) : 434 _cl(cl), _bounds(bottom, top) {} 435 436 void do_object(oop obj) { 437 obj->oop_iterate(_cl, _bounds); 438 } 439 }; 440 441 template<class T> 442 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) { 443 if (region->is_humongous()) { 444 HeapWord* bottom = region->bottom(); 445 if (top > bottom) { 446 region = region->humongous_start_region(); 447 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top); 448 marked_object_iterate(region, &objs); 449 } 450 } else { 451 ShenandoahObjectToOopClosure<T> objs(cl); 452 marked_object_iterate(region, &objs, top); 453 } 454 } 455 456 inline ShenandoahHeapRegion* const ShenandoahHeap::get_region(size_t region_idx) const { 457 if (region_idx < _num_regions) { 458 return _regions[region_idx]; 459 } else { 460 return NULL; 461 } 462 } 463 464 inline void ShenandoahHeap::mark_complete_marking_context() { 465 _marking_context->mark_complete(); 466 } 467 468 inline void ShenandoahHeap::mark_incomplete_marking_context() { 469 _marking_context->mark_incomplete(); 470 } 471 472 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const { 473 assert (_marking_context->is_complete()," sanity"); 474 return _marking_context; 475 } 476 477 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const { 478 return _marking_context; 479 } 480 481 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP