1 /* 2 * Copyright (c) 2014, 2019, Oracle and/or its affiliates. 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 #include "precompiled.hpp" 26 #include "gc/g1/g1Allocator.inline.hpp" 27 #include "gc/g1/g1CollectedHeap.inline.hpp" 28 #include "gc/g1/g1CollectionSet.hpp" 29 #include "gc/g1/g1OopClosures.inline.hpp" 30 #include "gc/g1/g1ParScanThreadState.inline.hpp" 31 #include "gc/g1/g1RootClosures.hpp" 32 #include "gc/g1/g1StringDedup.hpp" 33 #include "gc/g1/g1Trace.hpp" 34 #include "gc/shared/taskqueue.inline.hpp" 35 #include "memory/allocation.inline.hpp" 36 #include "oops/access.inline.hpp" 37 #include "oops/oop.inline.hpp" 38 #include "runtime/prefetch.inline.hpp" 39 40 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, 41 G1RedirtyCardsQueueSet* rdcqs, 42 uint worker_id, 43 size_t young_cset_length, 44 size_t optional_cset_length) 45 : _g1h(g1h), 46 _refs(g1h->task_queue(worker_id)), 47 _rdcq(rdcqs), 48 _ct(g1h->card_table()), 49 _closures(NULL), 50 _plab_allocator(NULL), 51 _age_table(false), 52 _tenuring_threshold(g1h->policy()->tenuring_threshold()), 53 _scanner(g1h, this), 54 _worker_id(worker_id), 55 _last_enqueued_card(SIZE_MAX), 56 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1), 57 _stack_trim_lower_threshold(GCDrainStackTargetSize), 58 _trim_ticks(), 59 _old_gen_is_full(false), 60 _num_optional_regions(optional_cset_length) 61 { 62 // We allocate number of young gen regions in the collection set plus one 63 // entries, since entry 0 keeps track of surviving bytes for non-young regions. 64 // We also add a few elements at the beginning and at the end in 65 // an attempt to eliminate cache contention 66 size_t real_length = young_cset_length + 1; 67 size_t array_length = PADDING_ELEM_NUM + real_length + PADDING_ELEM_NUM; 68 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 69 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 70 memset(_surviving_young_words, 0, real_length * sizeof(size_t)); 71 72 _plab_allocator = new G1PLABAllocator(_g1h->allocator()); 73 74 // The dest for Young is used when the objects are aged enough to 75 // need to be moved to the next space. 76 _dest[G1HeapRegionAttr::Young] = G1HeapRegionAttr::Old; 77 _dest[G1HeapRegionAttr::Old] = G1HeapRegionAttr::Old; 78 79 _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h); 80 81 _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions]; 82 } 83 84 // Pass locally gathered statistics to global state. 85 void G1ParScanThreadState::flush(size_t* surviving_young_words) { 86 _rdcq.flush(); 87 // Update allocation statistics. 88 _plab_allocator->flush_and_retire_stats(); 89 _g1h->policy()->record_age_table(&_age_table); 90 91 uint length = _g1h->collection_set()->young_region_length() + 1; 92 for (uint i = 0; i < length; i++) { 93 surviving_young_words[i] += _surviving_young_words[i]; 94 } 95 } 96 97 G1ParScanThreadState::~G1ParScanThreadState() { 98 delete _plab_allocator; 99 delete _closures; 100 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base); 101 delete[] _oops_into_optional_regions; 102 } 103 104 size_t G1ParScanThreadState::lab_waste_words() const { 105 return _plab_allocator->waste(); 106 } 107 108 size_t G1ParScanThreadState::lab_undo_waste_words() const { 109 return _plab_allocator->undo_waste(); 110 } 111 112 #ifdef ASSERT 113 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 114 assert(ref != NULL, "invariant"); 115 assert(UseCompressedOops, "sanity"); 116 assert(!has_partial_array_mask(ref), "ref=" PTR_FORMAT, p2i(ref)); 117 oop p = RawAccess<>::oop_load(ref); 118 assert(_g1h->is_in_g1_reserved(p), 119 "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p)); 120 return true; 121 } 122 123 bool G1ParScanThreadState::verify_ref(oop* ref) const { 124 assert(ref != NULL, "invariant"); 125 if (has_partial_array_mask(ref)) { 126 // Must be in the collection set--it's already been copied. 127 oop p = clear_partial_array_mask(ref); 128 assert(_g1h->is_in_cset(p), 129 "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p)); 130 } else { 131 oop p = RawAccess<>::oop_load(ref); 132 assert(_g1h->is_in_g1_reserved(p), 133 "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p)); 134 } 135 return true; 136 } 137 138 bool G1ParScanThreadState::verify_task(StarTask ref) const { 139 if (ref.is_narrow()) { 140 return verify_ref((narrowOop*) ref); 141 } else { 142 return verify_ref((oop*) ref); 143 } 144 } 145 #endif // ASSERT 146 147 void G1ParScanThreadState::trim_queue() { 148 StarTask ref; 149 do { 150 // Fully drain the queue. 151 trim_queue_to_threshold(0); 152 } while (!_refs->is_empty()); 153 } 154 155 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest, 156 size_t word_sz, 157 bool previous_plab_refill_failed, 158 uint node_index) { 159 160 assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str()); 161 162 // Right now we only have two types of regions (young / old) so 163 // let's keep the logic here simple. We can generalize it when necessary. 164 if (dest->is_young()) { 165 bool plab_refill_in_old_failed = false; 166 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old, 167 word_sz, 168 &plab_refill_in_old_failed, 169 node_index); 170 // Make sure that we won't attempt to copy any other objects out 171 // of a survivor region (given that apparently we cannot allocate 172 // any new ones) to avoid coming into this slow path again and again. 173 // Only consider failed PLAB refill here: failed inline allocations are 174 // typically large, so not indicative of remaining space. 175 if (previous_plab_refill_failed) { 176 _tenuring_threshold = 0; 177 } 178 179 if (obj_ptr != NULL) { 180 dest->set_old(); 181 } else { 182 // We just failed to allocate in old gen. The same idea as explained above 183 // for making survivor gen unavailable for allocation applies for old gen. 184 _old_gen_is_full = plab_refill_in_old_failed; 185 } 186 return obj_ptr; 187 } else { 188 _old_gen_is_full = previous_plab_refill_failed; 189 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str()); 190 // no other space to try. 191 return NULL; 192 } 193 } 194 195 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) { 196 if (region_attr.is_young()) { 197 age = !m.has_displaced_mark_helper() ? m.age() 198 : m.displaced_mark_helper().age(); 199 if (age < _tenuring_threshold) { 200 return region_attr; 201 } 202 } 203 return dest(region_attr); 204 } 205 206 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr, 207 oop const old, size_t word_sz, uint age, 208 HeapWord * const obj_ptr, uint node_index) const { 209 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index); 210 if (alloc_buf->contains(obj_ptr)) { 211 _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age, 212 dest_attr.type() == G1HeapRegionAttr::Old, 213 alloc_buf->word_sz() * HeapWordSize); 214 } else { 215 _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age, 216 dest_attr.type() == G1HeapRegionAttr::Old); 217 } 218 } 219 220 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr const region_attr, 221 oop const old, 222 markWord const old_mark) { 223 const size_t word_sz = old->size(); 224 225 uint age = 0; 226 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age); 227 // The second clause is to prevent premature evacuation failure in case there 228 // is still space in survivor, but old gen is full. 229 if (_old_gen_is_full && dest_attr.is_old()) { 230 return handle_evacuation_failure_par(old, old_mark); 231 } 232 HeapRegion* const from_region = _g1h->heap_region_containing(old); 233 uint node_index = from_region->node_index(); 234 235 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index); 236 237 // PLAB allocations should succeed most of the time, so we'll 238 // normally check against NULL once and that's it. 239 if (obj_ptr == NULL) { 240 bool plab_refill_failed = false; 241 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(dest_attr, word_sz, &plab_refill_failed, node_index); 242 if (obj_ptr == NULL) { 243 assert(region_attr.is_in_cset(), "Unexpected region attr type: %s", region_attr.get_type_str()); 244 obj_ptr = allocate_in_next_plab(&dest_attr, word_sz, plab_refill_failed, node_index); 245 if (obj_ptr == NULL) { 246 // This will either forward-to-self, or detect that someone else has 247 // installed a forwarding pointer. 248 return handle_evacuation_failure_par(old, old_mark); 249 } 250 } 251 if (_g1h->_gc_tracer_stw->should_report_promotion_events()) { 252 // The events are checked individually as part of the actual commit 253 report_promotion_event(dest_attr, old, word_sz, age, obj_ptr, node_index); 254 } 255 } 256 257 assert(obj_ptr != NULL, "when we get here, allocation should have succeeded"); 258 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap"); 259 260 #ifndef PRODUCT 261 // Should this evacuation fail? 262 if (_g1h->evacuation_should_fail()) { 263 // Doing this after all the allocation attempts also tests the 264 // undo_allocation() method too. 265 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index); 266 return handle_evacuation_failure_par(old, old_mark); 267 } 268 #endif // !PRODUCT 269 270 // We're going to allocate linearly, so might as well prefetch ahead. 271 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 272 273 const oop obj = oop(obj_ptr); 274 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed); 275 if (forward_ptr == NULL) { 276 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 277 278 const uint young_index = from_region->young_index_in_cset(); 279 280 assert((from_region->is_young() && young_index > 0) || 281 (!from_region->is_young() && young_index == 0), "invariant" ); 282 283 if (dest_attr.is_young()) { 284 if (age < markWord::max_age) { 285 age++; 286 } 287 if (old_mark.has_displaced_mark_helper()) { 288 // In this case, we have to install the mark word first, 289 // otherwise obj looks to be forwarded (the old mark word, 290 // which contains the forward pointer, was copied) 291 obj->set_mark_raw(old_mark); 292 markWord new_mark = old_mark.displaced_mark_helper().set_age(age); 293 old_mark.set_displaced_mark_helper(new_mark); 294 } else { 295 obj->set_mark_raw(old_mark.set_age(age)); 296 } 297 _age_table.add(age, word_sz); 298 } else { 299 obj->set_mark_raw(old_mark); 300 } 301 302 if (G1StringDedup::is_enabled()) { 303 const bool is_from_young = region_attr.is_young(); 304 const bool is_to_young = dest_attr.is_young(); 305 assert(is_from_young == from_region->is_young(), 306 "sanity"); 307 assert(is_to_young == _g1h->heap_region_containing(obj)->is_young(), 308 "sanity"); 309 G1StringDedup::enqueue_from_evacuation(is_from_young, 310 is_to_young, 311 _worker_id, 312 obj); 313 } 314 315 _surviving_young_words[young_index] += word_sz; 316 317 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 318 // We keep track of the next start index in the length field of 319 // the to-space object. The actual length can be found in the 320 // length field of the from-space object. 321 arrayOop(obj)->set_length(0); 322 oop* old_p = set_partial_array_mask(old); 323 do_oop_partial_array(old_p); 324 } else { 325 G1ScanInYoungSetter x(&_scanner, dest_attr.is_young()); 326 obj->oop_iterate_backwards(&_scanner); 327 } 328 return obj; 329 } else { 330 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index); 331 return forward_ptr; 332 } 333 } 334 335 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) { 336 assert(worker_id < _n_workers, "out of bounds access"); 337 if (_states[worker_id] == NULL) { 338 _states[worker_id] = 339 new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length); 340 } 341 return _states[worker_id]; 342 } 343 344 const size_t* G1ParScanThreadStateSet::surviving_young_words() const { 345 assert(_flushed, "thread local state from the per thread states should have been flushed"); 346 return _surviving_young_words_total; 347 } 348 349 void G1ParScanThreadStateSet::flush() { 350 assert(!_flushed, "thread local state from the per thread states should be flushed once"); 351 352 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) { 353 G1ParScanThreadState* pss = _states[worker_index]; 354 355 if (pss == NULL) { 356 continue; 357 } 358 359 pss->flush(_surviving_young_words_total); 360 delete pss; 361 _states[worker_index] = NULL; 362 } 363 _flushed = true; 364 } 365 366 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) { 367 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) { 368 G1ParScanThreadState* pss = _states[worker_index]; 369 370 if (pss == NULL) { 371 continue; 372 } 373 374 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory(); 375 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory); 376 } 377 } 378 379 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m) { 380 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old)); 381 382 oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed); 383 if (forward_ptr == NULL) { 384 // Forward-to-self succeeded. We are the "owner" of the object. 385 HeapRegion* r = _g1h->heap_region_containing(old); 386 387 if (!r->evacuation_failed()) { 388 r->set_evacuation_failed(true); 389 _g1h->hr_printer()->evac_failure(r); 390 } 391 392 _g1h->preserve_mark_during_evac_failure(_worker_id, old, m); 393 394 G1ScanInYoungSetter x(&_scanner, r->is_young()); 395 old->oop_iterate_backwards(&_scanner); 396 397 return old; 398 } else { 399 // Forward-to-self failed. Either someone else managed to allocate 400 // space for this object (old != forward_ptr) or they beat us in 401 // self-forwarding it (old == forward_ptr). 402 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr), 403 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " " 404 "should not be in the CSet", 405 p2i(old), p2i(forward_ptr)); 406 return forward_ptr; 407 } 408 } 409 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h, 410 G1RedirtyCardsQueueSet* rdcqs, 411 uint n_workers, 412 size_t young_cset_length, 413 size_t optional_cset_length) : 414 _g1h(g1h), 415 _rdcqs(rdcqs), 416 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)), 417 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)), 418 _young_cset_length(young_cset_length), 419 _optional_cset_length(optional_cset_length), 420 _n_workers(n_workers), 421 _flushed(false) { 422 for (uint i = 0; i < n_workers; ++i) { 423 _states[i] = NULL; 424 } 425 memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t)); 426 } 427 428 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() { 429 assert(_flushed, "thread local state from the per thread states should have been flushed"); 430 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states); 431 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total); 432 }