1 /* 2 * Copyright (c) 2023, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2013, 2022, Red Hat, Inc. All rights reserved. 4 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved. 5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 6 * 7 * This code is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License version 2 only, as 9 * published by the Free Software Foundation. 10 * 11 * This code is distributed in the hope that it will be useful, but WITHOUT 12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 * version 2 for more details (a copy is included in the LICENSE file that 15 * accompanied this code). 16 * 17 * You should have received a copy of the GNU General Public License version 18 * 2 along with this work; if not, write to the Free Software Foundation, 19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 20 * 21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 22 * or visit www.oracle.com if you need additional information or have any 23 * questions. 24 * 25 */ 26 27 #include "precompiled.hpp" 28 #include "memory/allocation.hpp" 29 #include "memory/universe.hpp" 30 31 #include "gc/shared/gcArguments.hpp" 32 #include "gc/shared/gcTimer.hpp" 33 #include "gc/shared/gcTraceTime.inline.hpp" 34 #include "gc/shared/locationPrinter.inline.hpp" 35 #include "gc/shared/memAllocator.hpp" 36 #include "gc/shared/plab.hpp" 37 #include "gc/shared/tlab_globals.hpp" 38 39 #include "gc/shenandoah/shenandoahAgeCensus.hpp" 40 #include "gc/shenandoah/heuristics/shenandoahOldHeuristics.hpp" 41 #include "gc/shenandoah/heuristics/shenandoahYoungHeuristics.hpp" 42 #include "gc/shenandoah/shenandoahAllocRequest.hpp" 43 #include "gc/shenandoah/shenandoahBarrierSet.hpp" 44 #include "gc/shenandoah/shenandoahCardTable.hpp" 45 #include "gc/shenandoah/shenandoahClosures.inline.hpp" 46 #include "gc/shenandoah/shenandoahCollectionSet.hpp" 47 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp" 48 #include "gc/shenandoah/shenandoahConcurrentMark.hpp" 49 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp" 50 #include "gc/shenandoah/shenandoahControlThread.hpp" 51 #include "gc/shenandoah/shenandoahRegulatorThread.hpp" 52 #include "gc/shenandoah/shenandoahFreeSet.hpp" 53 #include "gc/shenandoah/shenandoahGlobalGeneration.hpp" 54 #include "gc/shenandoah/shenandoahPhaseTimings.hpp" 55 #include "gc/shenandoah/shenandoahHeap.inline.hpp" 56 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp" 57 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp" 58 #include "gc/shenandoah/shenandoahInitLogger.hpp" 59 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp" 60 #include "gc/shenandoah/shenandoahMemoryPool.hpp" 61 #include "gc/shenandoah/shenandoahMetrics.hpp" 62 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp" 63 #include "gc/shenandoah/shenandoahOldGeneration.hpp" 64 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp" 65 #include "gc/shenandoah/shenandoahPacer.inline.hpp" 66 #include "gc/shenandoah/shenandoahPadding.hpp" 67 #include "gc/shenandoah/shenandoahParallelCleaning.inline.hpp" 68 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp" 69 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp" 70 #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp" 71 #include "gc/shenandoah/shenandoahSTWMark.hpp" 72 #include "gc/shenandoah/shenandoahUtils.hpp" 73 #include "gc/shenandoah/shenandoahVerifier.hpp" 74 #include "gc/shenandoah/shenandoahCodeRoots.hpp" 75 #include "gc/shenandoah/shenandoahVMOperations.hpp" 76 #include "gc/shenandoah/shenandoahWorkGroup.hpp" 77 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp" 78 #include "gc/shenandoah/shenandoahYoungGeneration.hpp" 79 #include "gc/shenandoah/mode/shenandoahGenerationalMode.hpp" 80 #include "gc/shenandoah/mode/shenandoahIUMode.hpp" 81 #include "gc/shenandoah/mode/shenandoahPassiveMode.hpp" 82 #include "gc/shenandoah/mode/shenandoahSATBMode.hpp" 83 #include "utilities/globalDefinitions.hpp" 84 85 #if INCLUDE_JFR 86 #include "gc/shenandoah/shenandoahJfrSupport.hpp" 87 #endif 88 89 #include "classfile/systemDictionary.hpp" 90 #include "code/codeCache.hpp" 91 #include "memory/classLoaderMetaspace.hpp" 92 #include "memory/metaspaceUtils.hpp" 93 #include "oops/compressedOops.inline.hpp" 94 #include "prims/jvmtiTagMap.hpp" 95 #include "runtime/atomic.hpp" 96 #include "runtime/globals.hpp" 97 #include "runtime/interfaceSupport.inline.hpp" 98 #include "runtime/java.hpp" 99 #include "runtime/orderAccess.hpp" 100 #include "runtime/safepointMechanism.hpp" 101 #include "runtime/vmThread.hpp" 102 #include "services/mallocTracker.hpp" 103 #include "services/memTracker.hpp" 104 #include "utilities/events.hpp" 105 #include "utilities/powerOfTwo.hpp" 106 107 class ShenandoahPretouchHeapTask : public WorkerTask { 108 private: 109 ShenandoahRegionIterator _regions; 110 const size_t _page_size; 111 public: 112 ShenandoahPretouchHeapTask(size_t page_size) : 113 WorkerTask("Shenandoah Pretouch Heap"), 114 _page_size(page_size) {} 115 116 virtual void work(uint worker_id) { 117 ShenandoahHeapRegion* r = _regions.next(); 118 while (r != nullptr) { 119 if (r->is_committed()) { 120 os::pretouch_memory(r->bottom(), r->end(), _page_size); 121 } 122 r = _regions.next(); 123 } 124 } 125 }; 126 127 class ShenandoahPretouchBitmapTask : public WorkerTask { 128 private: 129 ShenandoahRegionIterator _regions; 130 char* _bitmap_base; 131 const size_t _bitmap_size; 132 const size_t _page_size; 133 public: 134 ShenandoahPretouchBitmapTask(char* bitmap_base, size_t bitmap_size, size_t page_size) : 135 WorkerTask("Shenandoah Pretouch Bitmap"), 136 _bitmap_base(bitmap_base), 137 _bitmap_size(bitmap_size), 138 _page_size(page_size) {} 139 140 virtual void work(uint worker_id) { 141 ShenandoahHeapRegion* r = _regions.next(); 142 while (r != nullptr) { 143 size_t start = r->index() * ShenandoahHeapRegion::region_size_bytes() / MarkBitMap::heap_map_factor(); 144 size_t end = (r->index() + 1) * ShenandoahHeapRegion::region_size_bytes() / MarkBitMap::heap_map_factor(); 145 assert (end <= _bitmap_size, "end is sane: " SIZE_FORMAT " < " SIZE_FORMAT, end, _bitmap_size); 146 147 if (r->is_committed()) { 148 os::pretouch_memory(_bitmap_base + start, _bitmap_base + end, _page_size); 149 } 150 151 r = _regions.next(); 152 } 153 } 154 }; 155 156 jint ShenandoahHeap::initialize() { 157 // 158 // Figure out heap sizing 159 // 160 161 size_t init_byte_size = InitialHeapSize; 162 size_t min_byte_size = MinHeapSize; 163 size_t max_byte_size = MaxHeapSize; 164 size_t heap_alignment = HeapAlignment; 165 166 size_t reg_size_bytes = ShenandoahHeapRegion::region_size_bytes(); 167 168 Universe::check_alignment(max_byte_size, reg_size_bytes, "Shenandoah heap"); 169 Universe::check_alignment(init_byte_size, reg_size_bytes, "Shenandoah heap"); 170 171 _num_regions = ShenandoahHeapRegion::region_count(); 172 assert(_num_regions == (max_byte_size / reg_size_bytes), 173 "Regions should cover entire heap exactly: " SIZE_FORMAT " != " SIZE_FORMAT "/" SIZE_FORMAT, 174 _num_regions, max_byte_size, reg_size_bytes); 175 176 size_t num_committed_regions = init_byte_size / reg_size_bytes; 177 num_committed_regions = MIN2(num_committed_regions, _num_regions); 178 assert(num_committed_regions <= _num_regions, "sanity"); 179 _initial_size = num_committed_regions * reg_size_bytes; 180 181 size_t num_min_regions = min_byte_size / reg_size_bytes; 182 num_min_regions = MIN2(num_min_regions, _num_regions); 183 assert(num_min_regions <= _num_regions, "sanity"); 184 _minimum_size = num_min_regions * reg_size_bytes; 185 186 // Default to max heap size. 187 _soft_max_size = _num_regions * reg_size_bytes; 188 189 _committed = _initial_size; 190 191 // Now we know the number of regions and heap sizes, initialize the heuristics. 192 initialize_heuristics_generations(); 193 194 size_t heap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 195 size_t bitmap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 196 size_t region_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); 197 198 // 199 // Reserve and commit memory for heap 200 // 201 202 ReservedHeapSpace heap_rs = Universe::reserve_heap(max_byte_size, heap_alignment); 203 initialize_reserved_region(heap_rs); 204 _heap_region = MemRegion((HeapWord*)heap_rs.base(), heap_rs.size() / HeapWordSize); 205 _heap_region_special = heap_rs.special(); 206 207 assert((((size_t) base()) & ShenandoahHeapRegion::region_size_bytes_mask()) == 0, 208 "Misaligned heap: " PTR_FORMAT, p2i(base())); 209 os::trace_page_sizes_for_requested_size("Heap", 210 max_byte_size, heap_alignment, 211 heap_rs.base(), 212 heap_rs.size(), heap_rs.page_size()); 213 214 #if SHENANDOAH_OPTIMIZED_MARKTASK 215 // The optimized ShenandoahMarkTask takes some bits away from the full object bits. 216 // Fail if we ever attempt to address more than we can. 217 if ((uintptr_t)heap_rs.end() >= ShenandoahMarkTask::max_addressable()) { 218 FormatBuffer<512> buf("Shenandoah reserved [" PTR_FORMAT ", " PTR_FORMAT") for the heap, \n" 219 "but max object address is " PTR_FORMAT ". Try to reduce heap size, or try other \n" 220 "VM options that allocate heap at lower addresses (HeapBaseMinAddress, AllocateHeapAt, etc).", 221 p2i(heap_rs.base()), p2i(heap_rs.end()), ShenandoahMarkTask::max_addressable()); 222 vm_exit_during_initialization("Fatal Error", buf); 223 } 224 #endif 225 226 ReservedSpace sh_rs = heap_rs.first_part(max_byte_size); 227 if (!_heap_region_special) { 228 os::commit_memory_or_exit(sh_rs.base(), _initial_size, heap_alignment, false, 229 "Cannot commit heap memory"); 230 } 231 232 BarrierSet::set_barrier_set(new ShenandoahBarrierSet(this, _heap_region)); 233 234 // 235 // After reserving the Java heap, create the card table, barriers, and workers, in dependency order 236 // 237 if (mode()->is_generational()) { 238 ShenandoahDirectCardMarkRememberedSet *rs; 239 ShenandoahCardTable* card_table = ShenandoahBarrierSet::barrier_set()->card_table(); 240 size_t card_count = card_table->cards_required(heap_rs.size() / HeapWordSize); 241 rs = new ShenandoahDirectCardMarkRememberedSet(ShenandoahBarrierSet::barrier_set()->card_table(), card_count); 242 _card_scan = new ShenandoahScanRemembered<ShenandoahDirectCardMarkRememberedSet>(rs); 243 244 // Age census structure 245 _age_census = new ShenandoahAgeCensus(); 246 } 247 248 _workers = new ShenandoahWorkerThreads("Shenandoah GC Threads", _max_workers); 249 if (_workers == nullptr) { 250 vm_exit_during_initialization("Failed necessary allocation."); 251 } else { 252 _workers->initialize_workers(); 253 } 254 255 if (ParallelGCThreads > 1) { 256 _safepoint_workers = new ShenandoahWorkerThreads("Safepoint Cleanup Thread", ParallelGCThreads); 257 _safepoint_workers->initialize_workers(); 258 } 259 260 // 261 // Reserve and commit memory for bitmap(s) 262 // 263 264 size_t bitmap_size_orig = ShenandoahMarkBitMap::compute_size(heap_rs.size()); 265 _bitmap_size = align_up(bitmap_size_orig, bitmap_page_size); 266 267 size_t bitmap_bytes_per_region = reg_size_bytes / ShenandoahMarkBitMap::heap_map_factor(); 268 269 guarantee(bitmap_bytes_per_region != 0, 270 "Bitmap bytes per region should not be zero"); 271 guarantee(is_power_of_2(bitmap_bytes_per_region), 272 "Bitmap bytes per region should be power of two: " SIZE_FORMAT, bitmap_bytes_per_region); 273 274 if (bitmap_page_size > bitmap_bytes_per_region) { 275 _bitmap_regions_per_slice = bitmap_page_size / bitmap_bytes_per_region; 276 _bitmap_bytes_per_slice = bitmap_page_size; 277 } else { 278 _bitmap_regions_per_slice = 1; 279 _bitmap_bytes_per_slice = bitmap_bytes_per_region; 280 } 281 282 guarantee(_bitmap_regions_per_slice >= 1, 283 "Should have at least one region per slice: " SIZE_FORMAT, 284 _bitmap_regions_per_slice); 285 286 guarantee(((_bitmap_bytes_per_slice) % bitmap_page_size) == 0, 287 "Bitmap slices should be page-granular: bps = " SIZE_FORMAT ", page size = " SIZE_FORMAT, 288 _bitmap_bytes_per_slice, bitmap_page_size); 289 290 ReservedSpace bitmap(_bitmap_size, bitmap_page_size); 291 os::trace_page_sizes_for_requested_size("Mark Bitmap", 292 bitmap_size_orig, bitmap_page_size, 293 bitmap.base(), 294 bitmap.size(), bitmap.page_size()); 295 MemTracker::record_virtual_memory_type(bitmap.base(), mtGC); 296 _bitmap_region = MemRegion((HeapWord*) bitmap.base(), bitmap.size() / HeapWordSize); 297 _bitmap_region_special = bitmap.special(); 298 299 size_t bitmap_init_commit = _bitmap_bytes_per_slice * 300 align_up(num_committed_regions, _bitmap_regions_per_slice) / _bitmap_regions_per_slice; 301 bitmap_init_commit = MIN2(_bitmap_size, bitmap_init_commit); 302 if (!_bitmap_region_special) { 303 os::commit_memory_or_exit((char *) _bitmap_region.start(), bitmap_init_commit, bitmap_page_size, false, 304 "Cannot commit bitmap memory"); 305 } 306 307 _marking_context = new ShenandoahMarkingContext(_heap_region, _bitmap_region, _num_regions); 308 309 if (ShenandoahVerify) { 310 ReservedSpace verify_bitmap(_bitmap_size, bitmap_page_size); 311 os::trace_page_sizes_for_requested_size("Verify Bitmap", 312 bitmap_size_orig, bitmap_page_size, 313 verify_bitmap.base(), 314 verify_bitmap.size(), verify_bitmap.page_size()); 315 if (!verify_bitmap.special()) { 316 os::commit_memory_or_exit(verify_bitmap.base(), verify_bitmap.size(), bitmap_page_size, false, 317 "Cannot commit verification bitmap memory"); 318 } 319 MemTracker::record_virtual_memory_type(verify_bitmap.base(), mtGC); 320 MemRegion verify_bitmap_region = MemRegion((HeapWord *) verify_bitmap.base(), verify_bitmap.size() / HeapWordSize); 321 _verification_bit_map.initialize(_heap_region, verify_bitmap_region); 322 _verifier = new ShenandoahVerifier(this, &_verification_bit_map); 323 } 324 325 // Reserve aux bitmap for use in object_iterate(). We don't commit it here. 326 size_t aux_bitmap_page_size = bitmap_page_size; 327 #ifdef LINUX 328 // In THP "advise" mode, we refrain from advising the system to use large pages 329 // since we know these commits will be short lived, and there is no reason to trash 330 // the THP area with this bitmap. 331 if (UseTransparentHugePages) { 332 aux_bitmap_page_size = os::vm_page_size(); 333 } 334 #endif 335 ReservedSpace aux_bitmap(_bitmap_size, aux_bitmap_page_size); 336 os::trace_page_sizes_for_requested_size("Aux Bitmap", 337 bitmap_size_orig, aux_bitmap_page_size, 338 aux_bitmap.base(), 339 aux_bitmap.size(), aux_bitmap.page_size()); 340 MemTracker::record_virtual_memory_type(aux_bitmap.base(), mtGC); 341 _aux_bitmap_region = MemRegion((HeapWord*) aux_bitmap.base(), aux_bitmap.size() / HeapWordSize); 342 _aux_bitmap_region_special = aux_bitmap.special(); 343 _aux_bit_map.initialize(_heap_region, _aux_bitmap_region); 344 345 // 346 // Create regions and region sets 347 // 348 size_t region_align = align_up(sizeof(ShenandoahHeapRegion), SHENANDOAH_CACHE_LINE_SIZE); 349 size_t region_storage_size_orig = region_align * _num_regions; 350 size_t region_storage_size = align_up(region_storage_size_orig, 351 MAX2(region_page_size, os::vm_allocation_granularity())); 352 353 ReservedSpace region_storage(region_storage_size, region_page_size); 354 os::trace_page_sizes_for_requested_size("Region Storage", 355 region_storage_size_orig, region_page_size, 356 region_storage.base(), 357 region_storage.size(), region_storage.page_size()); 358 MemTracker::record_virtual_memory_type(region_storage.base(), mtGC); 359 if (!region_storage.special()) { 360 os::commit_memory_or_exit(region_storage.base(), region_storage_size, region_page_size, false, 361 "Cannot commit region memory"); 362 } 363 364 // Try to fit the collection set bitmap at lower addresses. This optimizes code generation for cset checks. 365 // Go up until a sensible limit (subject to encoding constraints) and try to reserve the space there. 366 // If not successful, bite a bullet and allocate at whatever address. 367 { 368 const size_t cset_align = MAX2<size_t>(os::vm_page_size(), os::vm_allocation_granularity()); 369 const size_t cset_size = align_up(((size_t) sh_rs.base() + sh_rs.size()) >> ShenandoahHeapRegion::region_size_bytes_shift(), cset_align); 370 const size_t cset_page_size = os::vm_page_size(); 371 372 uintptr_t min = round_up_power_of_2(cset_align); 373 uintptr_t max = (1u << 30u); 374 ReservedSpace cset_rs; 375 376 for (uintptr_t addr = min; addr <= max; addr <<= 1u) { 377 char* req_addr = (char*)addr; 378 assert(is_aligned(req_addr, cset_align), "Should be aligned"); 379 cset_rs = ReservedSpace(cset_size, cset_align, cset_page_size, req_addr); 380 if (cset_rs.is_reserved()) { 381 assert(cset_rs.base() == req_addr, "Allocated where requested: " PTR_FORMAT ", " PTR_FORMAT, p2i(cset_rs.base()), addr); 382 _collection_set = new ShenandoahCollectionSet(this, cset_rs, sh_rs.base()); 383 break; 384 } 385 } 386 387 if (_collection_set == nullptr) { 388 cset_rs = ReservedSpace(cset_size, cset_align, os::vm_page_size()); 389 _collection_set = new ShenandoahCollectionSet(this, cset_rs, sh_rs.base()); 390 } 391 os::trace_page_sizes_for_requested_size("Collection Set", 392 cset_size, cset_page_size, 393 cset_rs.base(), 394 cset_rs.size(), cset_rs.page_size()); 395 } 396 397 _regions = NEW_C_HEAP_ARRAY(ShenandoahHeapRegion*, _num_regions, mtGC); 398 _affiliations = NEW_C_HEAP_ARRAY(uint8_t, _num_regions, mtGC); 399 _free_set = new ShenandoahFreeSet(this, _num_regions); 400 401 { 402 ShenandoahHeapLocker locker(lock()); 403 404 405 for (size_t i = 0; i < _num_regions; i++) { 406 HeapWord* start = (HeapWord*)sh_rs.base() + ShenandoahHeapRegion::region_size_words() * i; 407 bool is_committed = i < num_committed_regions; 408 void* loc = region_storage.base() + i * region_align; 409 410 ShenandoahHeapRegion* r = new (loc) ShenandoahHeapRegion(start, i, is_committed); 411 assert(is_aligned(r, SHENANDOAH_CACHE_LINE_SIZE), "Sanity"); 412 413 _marking_context->initialize_top_at_mark_start(r); 414 _regions[i] = r; 415 assert(!collection_set()->is_in(i), "New region should not be in collection set"); 416 417 _affiliations[i] = ShenandoahAffiliation::FREE; 418 } 419 420 // Initialize to complete 421 _marking_context->mark_complete(); 422 size_t young_cset_regions, old_cset_regions; 423 424 // We are initializing free set. We ignore cset region tallies. 425 _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions); 426 _free_set->rebuild(young_cset_regions, old_cset_regions); 427 } 428 429 if (AlwaysPreTouch) { 430 // For NUMA, it is important to pre-touch the storage under bitmaps with worker threads, 431 // before initialize() below zeroes it with initializing thread. For any given region, 432 // we touch the region and the corresponding bitmaps from the same thread. 433 ShenandoahPushWorkerScope scope(workers(), _max_workers, false); 434 435 _pretouch_heap_page_size = heap_page_size; 436 _pretouch_bitmap_page_size = bitmap_page_size; 437 438 #ifdef LINUX 439 // UseTransparentHugePages would madvise that backing memory can be coalesced into huge 440 // pages. But, the kernel needs to know that every small page is used, in order to coalesce 441 // them into huge one. Therefore, we need to pretouch with smaller pages. 442 if (UseTransparentHugePages) { 443 _pretouch_heap_page_size = (size_t)os::vm_page_size(); 444 _pretouch_bitmap_page_size = (size_t)os::vm_page_size(); 445 } 446 #endif 447 448 // OS memory managers may want to coalesce back-to-back pages. Make their jobs 449 // simpler by pre-touching continuous spaces (heap and bitmap) separately. 450 451 ShenandoahPretouchBitmapTask bcl(bitmap.base(), _bitmap_size, _pretouch_bitmap_page_size); 452 _workers->run_task(&bcl); 453 454 ShenandoahPretouchHeapTask hcl(_pretouch_heap_page_size); 455 _workers->run_task(&hcl); 456 } 457 458 // 459 // Initialize the rest of GC subsystems 460 // 461 462 _liveness_cache = NEW_C_HEAP_ARRAY(ShenandoahLiveData*, _max_workers, mtGC); 463 for (uint worker = 0; worker < _max_workers; worker++) { 464 _liveness_cache[worker] = NEW_C_HEAP_ARRAY(ShenandoahLiveData, _num_regions, mtGC); 465 Copy::fill_to_bytes(_liveness_cache[worker], _num_regions * sizeof(ShenandoahLiveData)); 466 } 467 468 // There should probably be Shenandoah-specific options for these, 469 // just as there are G1-specific options. 470 { 471 ShenandoahSATBMarkQueueSet& satbqs = ShenandoahBarrierSet::satb_mark_queue_set(); 472 satbqs.set_process_completed_buffers_threshold(20); // G1SATBProcessCompletedThreshold 473 satbqs.set_buffer_enqueue_threshold_percentage(60); // G1SATBBufferEnqueueingThresholdPercent 474 } 475 476 _monitoring_support = new ShenandoahMonitoringSupport(this); 477 _phase_timings = new ShenandoahPhaseTimings(max_workers()); 478 ShenandoahCodeRoots::initialize(); 479 480 if (ShenandoahPacing) { 481 _pacer = new ShenandoahPacer(this); 482 _pacer->setup_for_idle(); 483 } else { 484 _pacer = nullptr; 485 } 486 487 _control_thread = new ShenandoahControlThread(); 488 _regulator_thread = new ShenandoahRegulatorThread(_control_thread); 489 490 print_init_logger(); 491 492 return JNI_OK; 493 } 494 495 void ShenandoahHeap::print_init_logger() const { 496 ShenandoahInitLogger::print(); 497 } 498 499 size_t ShenandoahHeap::max_size_for(ShenandoahGeneration* generation) const { 500 switch (generation->type()) { 501 case YOUNG: 502 return _generation_sizer.max_young_size(); 503 case OLD: 504 return max_capacity() - _generation_sizer.min_young_size(); 505 case GLOBAL_GEN: 506 case GLOBAL_NON_GEN: 507 return max_capacity(); 508 default: 509 ShouldNotReachHere(); 510 return 0; 511 } 512 } 513 514 size_t ShenandoahHeap::min_size_for(ShenandoahGeneration* generation) const { 515 switch (generation->type()) { 516 case YOUNG: 517 return _generation_sizer.min_young_size(); 518 case OLD: 519 return max_capacity() - _generation_sizer.max_young_size(); 520 case GLOBAL_GEN: 521 case GLOBAL_NON_GEN: 522 return min_capacity(); 523 default: 524 ShouldNotReachHere(); 525 return 0; 526 } 527 } 528 529 void ShenandoahHeap::initialize_heuristics_generations() { 530 if (ShenandoahGCMode != nullptr) { 531 if (strcmp(ShenandoahGCMode, "satb") == 0) { 532 _gc_mode = new ShenandoahSATBMode(); 533 } else if (strcmp(ShenandoahGCMode, "iu") == 0) { 534 _gc_mode = new ShenandoahIUMode(); 535 } else if (strcmp(ShenandoahGCMode, "passive") == 0) { 536 _gc_mode = new ShenandoahPassiveMode(); 537 } else if (strcmp(ShenandoahGCMode, "generational") == 0) { 538 _gc_mode = new ShenandoahGenerationalMode(); 539 } else { 540 vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option"); 541 } 542 } else { 543 vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option (null)"); 544 } 545 _gc_mode->initialize_flags(); 546 if (_gc_mode->is_diagnostic() && !UnlockDiagnosticVMOptions) { 547 vm_exit_during_initialization( 548 err_msg("GC mode \"%s\" is diagnostic, and must be enabled via -XX:+UnlockDiagnosticVMOptions.", 549 _gc_mode->name())); 550 } 551 if (_gc_mode->is_experimental() && !UnlockExperimentalVMOptions) { 552 vm_exit_during_initialization( 553 err_msg("GC mode \"%s\" is experimental, and must be enabled via -XX:+UnlockExperimentalVMOptions.", 554 _gc_mode->name())); 555 } 556 557 // Max capacity is the maximum _allowed_ capacity. That is, the maximum allowed capacity 558 // for old would be total heap - minimum capacity of young. This means the sum of the maximum 559 // allowed for old and young could exceed the total heap size. It remains the case that the 560 // _actual_ capacity of young + old = total. 561 _generation_sizer.heap_size_changed(max_capacity()); 562 size_t initial_capacity_young = _generation_sizer.max_young_size(); 563 size_t max_capacity_young = _generation_sizer.max_young_size(); 564 size_t initial_capacity_old = max_capacity() - max_capacity_young; 565 size_t max_capacity_old = max_capacity() - initial_capacity_young; 566 567 _young_generation = new ShenandoahYoungGeneration(_max_workers, max_capacity_young, initial_capacity_young); 568 _old_generation = new ShenandoahOldGeneration(_max_workers, max_capacity_old, initial_capacity_old); 569 _global_generation = new ShenandoahGlobalGeneration(_gc_mode->is_generational(), _max_workers, max_capacity(), max_capacity()); 570 _global_generation->initialize_heuristics(_gc_mode); 571 if (mode()->is_generational()) { 572 _young_generation->initialize_heuristics(_gc_mode); 573 _old_generation->initialize_heuristics(_gc_mode); 574 } 575 _evac_tracker = new ShenandoahEvacuationTracker(mode()->is_generational()); 576 } 577 578 #ifdef _MSC_VER 579 #pragma warning( push ) 580 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 581 #endif 582 583 ShenandoahHeap::ShenandoahHeap(ShenandoahCollectorPolicy* policy) : 584 CollectedHeap(), 585 _gc_generation(nullptr), 586 _prepare_for_old_mark(false), 587 _initial_size(0), 588 _promotion_potential(0), 589 _committed(0), 590 _max_workers(MAX3(ConcGCThreads, ParallelGCThreads, 1U)), 591 _workers(nullptr), 592 _safepoint_workers(nullptr), 593 _heap_region_special(false), 594 _num_regions(0), 595 _regions(nullptr), 596 _affiliations(nullptr), 597 _update_refs_iterator(this), 598 _promoted_reserve(0), 599 _old_evac_reserve(0), 600 _old_evac_expended(0), 601 _young_evac_reserve(0), 602 _captured_old_usage(0), 603 _previous_promotion(0), 604 _upgraded_to_full(false), 605 _age_census(nullptr), 606 _has_evacuation_reserve_quantities(false), 607 _cancel_requested_time(0), 608 _young_generation(nullptr), 609 _global_generation(nullptr), 610 _old_generation(nullptr), 611 _control_thread(nullptr), 612 _regulator_thread(nullptr), 613 _shenandoah_policy(policy), 614 _free_set(nullptr), 615 _pacer(nullptr), 616 _verifier(nullptr), 617 _phase_timings(nullptr), 618 _evac_tracker(nullptr), 619 _mmu_tracker(), 620 _generation_sizer(&_mmu_tracker), 621 _monitoring_support(nullptr), 622 _memory_pool(nullptr), 623 _young_gen_memory_pool(nullptr), 624 _old_gen_memory_pool(nullptr), 625 _stw_memory_manager("Shenandoah Pauses"), 626 _cycle_memory_manager("Shenandoah Cycles"), 627 _gc_timer(new ConcurrentGCTimer()), 628 _soft_ref_policy(), 629 _log_min_obj_alignment_in_bytes(LogMinObjAlignmentInBytes), 630 _old_regions_surplus(0), 631 _old_regions_deficit(0), 632 _marking_context(nullptr), 633 _bitmap_size(0), 634 _bitmap_regions_per_slice(0), 635 _bitmap_bytes_per_slice(0), 636 _bitmap_region_special(false), 637 _aux_bitmap_region_special(false), 638 _liveness_cache(nullptr), 639 _collection_set(nullptr), 640 _card_scan(nullptr) 641 { 642 } 643 644 #ifdef _MSC_VER 645 #pragma warning( pop ) 646 #endif 647 648 void ShenandoahHeap::print_on(outputStream* st) const { 649 st->print_cr("Shenandoah Heap"); 650 st->print_cr(" " SIZE_FORMAT "%s max, " SIZE_FORMAT "%s soft max, " SIZE_FORMAT "%s committed, " SIZE_FORMAT "%s used", 651 byte_size_in_proper_unit(max_capacity()), proper_unit_for_byte_size(max_capacity()), 652 byte_size_in_proper_unit(soft_max_capacity()), proper_unit_for_byte_size(soft_max_capacity()), 653 byte_size_in_proper_unit(committed()), proper_unit_for_byte_size(committed()), 654 byte_size_in_proper_unit(used()), proper_unit_for_byte_size(used())); 655 st->print_cr(" " SIZE_FORMAT " x " SIZE_FORMAT"%s regions", 656 num_regions(), 657 byte_size_in_proper_unit(ShenandoahHeapRegion::region_size_bytes()), 658 proper_unit_for_byte_size(ShenandoahHeapRegion::region_size_bytes())); 659 660 st->print("Status: "); 661 if (has_forwarded_objects()) st->print("has forwarded objects, "); 662 if (is_concurrent_old_mark_in_progress()) st->print("old marking, "); 663 if (is_concurrent_young_mark_in_progress()) st->print("young marking, "); 664 if (is_evacuation_in_progress()) st->print("evacuating, "); 665 if (is_update_refs_in_progress()) st->print("updating refs, "); 666 if (is_degenerated_gc_in_progress()) st->print("degenerated gc, "); 667 if (is_full_gc_in_progress()) st->print("full gc, "); 668 if (is_full_gc_move_in_progress()) st->print("full gc move, "); 669 if (is_concurrent_weak_root_in_progress()) st->print("concurrent weak roots, "); 670 if (is_concurrent_strong_root_in_progress() && 671 !is_concurrent_weak_root_in_progress()) st->print("concurrent strong roots, "); 672 673 if (cancelled_gc()) { 674 st->print("cancelled"); 675 } else { 676 st->print("not cancelled"); 677 } 678 st->cr(); 679 680 st->print_cr("Reserved region:"); 681 st->print_cr(" - [" PTR_FORMAT ", " PTR_FORMAT ") ", 682 p2i(reserved_region().start()), 683 p2i(reserved_region().end())); 684 685 ShenandoahCollectionSet* cset = collection_set(); 686 st->print_cr("Collection set:"); 687 if (cset != nullptr) { 688 st->print_cr(" - map (vanilla): " PTR_FORMAT, p2i(cset->map_address())); 689 st->print_cr(" - map (biased): " PTR_FORMAT, p2i(cset->biased_map_address())); 690 } else { 691 st->print_cr(" (null)"); 692 } 693 694 st->cr(); 695 MetaspaceUtils::print_on(st); 696 697 if (Verbose) { 698 st->cr(); 699 print_heap_regions_on(st); 700 } 701 } 702 703 class ShenandoahInitWorkerGCLABClosure : public ThreadClosure { 704 public: 705 void do_thread(Thread* thread) { 706 assert(thread != nullptr, "Sanity"); 707 assert(thread->is_Worker_thread(), "Only worker thread expected"); 708 ShenandoahThreadLocalData::initialize_gclab(thread); 709 } 710 }; 711 712 void ShenandoahHeap::post_initialize() { 713 CollectedHeap::post_initialize(); 714 _mmu_tracker.initialize(); 715 716 MutexLocker ml(Threads_lock); 717 718 ShenandoahInitWorkerGCLABClosure init_gclabs; 719 _workers->threads_do(&init_gclabs); 720 721 // gclab can not be initialized early during VM startup, as it can not determinate its max_size. 722 // Now, we will let WorkerThreads to initialize gclab when new worker is created. 723 _workers->set_initialize_gclab(); 724 if (_safepoint_workers != nullptr) { 725 _safepoint_workers->threads_do(&init_gclabs); 726 _safepoint_workers->set_initialize_gclab(); 727 } 728 729 JFR_ONLY(ShenandoahJFRSupport::register_jfr_type_serializers()); 730 } 731 732 ShenandoahHeuristics* ShenandoahHeap::heuristics() { 733 return _global_generation->heuristics(); 734 } 735 736 ShenandoahOldHeuristics* ShenandoahHeap::old_heuristics() { 737 return (ShenandoahOldHeuristics*) _old_generation->heuristics(); 738 } 739 740 ShenandoahYoungHeuristics* ShenandoahHeap::young_heuristics() { 741 return (ShenandoahYoungHeuristics*) _young_generation->heuristics(); 742 } 743 744 bool ShenandoahHeap::doing_mixed_evacuations() { 745 return _old_generation->state() == ShenandoahOldGeneration::WAITING_FOR_EVAC; 746 } 747 748 bool ShenandoahHeap::is_old_bitmap_stable() const { 749 return _old_generation->is_mark_complete(); 750 } 751 752 bool ShenandoahHeap::is_gc_generation_young() const { 753 return _gc_generation != nullptr && _gc_generation->is_young(); 754 } 755 756 size_t ShenandoahHeap::used() const { 757 return global_generation()->used(); 758 } 759 760 size_t ShenandoahHeap::committed() const { 761 return Atomic::load(&_committed); 762 } 763 764 void ShenandoahHeap::increase_committed(size_t bytes) { 765 shenandoah_assert_heaplocked_or_safepoint(); 766 _committed += bytes; 767 } 768 769 void ShenandoahHeap::decrease_committed(size_t bytes) { 770 shenandoah_assert_heaplocked_or_safepoint(); 771 _committed -= bytes; 772 } 773 774 // For tracking usage based on allocations, it should be the case that: 775 // * The sum of regions::used == heap::used 776 // * The sum of a generation's regions::used == generation::used 777 // * The sum of a generation's humongous regions::free == generation::humongous_waste 778 // These invariants are checked by the verifier on GC safepoints. 779 // 780 // Additional notes: 781 // * When a mutator's allocation request causes a region to be retired, the 782 // free memory left in that region is considered waste. It does not contribute 783 // to the usage, but it _does_ contribute to allocation rate. 784 // * The bottom of a PLAB must be aligned on card size. In some cases this will 785 // require padding in front of the PLAB (a filler object). Because this padding 786 // is included in the region's used memory we include the padding in the usage 787 // accounting as waste. 788 // * Mutator allocations are used to compute an allocation rate. They are also 789 // sent to the Pacer for those purposes. 790 // * There are three sources of waste: 791 // 1. The padding used to align a PLAB on card size 792 // 2. Region's free is less than minimum TLAB size and is retired 793 // 3. The unused portion of memory in the last region of a humongous object 794 void ShenandoahHeap::increase_used(const ShenandoahAllocRequest& req) { 795 size_t actual_bytes = req.actual_size() * HeapWordSize; 796 size_t wasted_bytes = req.waste() * HeapWordSize; 797 ShenandoahGeneration* generation = generation_for(req.affiliation()); 798 799 if (req.is_gc_alloc()) { 800 assert(wasted_bytes == 0 || req.type() == ShenandoahAllocRequest::_alloc_plab, "Only PLABs have waste"); 801 increase_used(generation, actual_bytes + wasted_bytes); 802 } else { 803 assert(req.is_mutator_alloc(), "Expected mutator alloc here"); 804 // padding and actual size both count towards allocation counter 805 generation->increase_allocated(actual_bytes + wasted_bytes); 806 807 // only actual size counts toward usage for mutator allocations 808 increase_used(generation, actual_bytes); 809 810 // notify pacer of both actual size and waste 811 notify_mutator_alloc_words(req.actual_size(), req.waste()); 812 813 if (wasted_bytes > 0 && req.actual_size() > ShenandoahHeapRegion::humongous_threshold_words()) { 814 increase_humongous_waste(generation,wasted_bytes); 815 } 816 } 817 } 818 819 void ShenandoahHeap::increase_humongous_waste(ShenandoahGeneration* generation, size_t bytes) { 820 generation->increase_humongous_waste(bytes); 821 if (!generation->is_global()) { 822 global_generation()->increase_humongous_waste(bytes); 823 } 824 } 825 826 void ShenandoahHeap::decrease_humongous_waste(ShenandoahGeneration* generation, size_t bytes) { 827 generation->decrease_humongous_waste(bytes); 828 if (!generation->is_global()) { 829 global_generation()->decrease_humongous_waste(bytes); 830 } 831 } 832 833 void ShenandoahHeap::increase_used(ShenandoahGeneration* generation, size_t bytes) { 834 generation->increase_used(bytes); 835 if (!generation->is_global()) { 836 global_generation()->increase_used(bytes); 837 } 838 } 839 840 void ShenandoahHeap::decrease_used(ShenandoahGeneration* generation, size_t bytes) { 841 generation->decrease_used(bytes); 842 if (!generation->is_global()) { 843 global_generation()->decrease_used(bytes); 844 } 845 } 846 847 void ShenandoahHeap::notify_mutator_alloc_words(size_t words, size_t waste) { 848 if (ShenandoahPacing) { 849 control_thread()->pacing_notify_alloc(words); 850 if (waste > 0) { 851 pacer()->claim_for_alloc(waste, true); 852 } 853 } 854 } 855 856 size_t ShenandoahHeap::capacity() const { 857 return committed(); 858 } 859 860 size_t ShenandoahHeap::max_capacity() const { 861 return _num_regions * ShenandoahHeapRegion::region_size_bytes(); 862 } 863 864 size_t ShenandoahHeap::soft_max_capacity() const { 865 size_t v = Atomic::load(&_soft_max_size); 866 assert(min_capacity() <= v && v <= max_capacity(), 867 "Should be in bounds: " SIZE_FORMAT " <= " SIZE_FORMAT " <= " SIZE_FORMAT, 868 min_capacity(), v, max_capacity()); 869 return v; 870 } 871 872 void ShenandoahHeap::set_soft_max_capacity(size_t v) { 873 assert(min_capacity() <= v && v <= max_capacity(), 874 "Should be in bounds: " SIZE_FORMAT " <= " SIZE_FORMAT " <= " SIZE_FORMAT, 875 min_capacity(), v, max_capacity()); 876 Atomic::store(&_soft_max_size, v); 877 } 878 879 size_t ShenandoahHeap::min_capacity() const { 880 return _minimum_size; 881 } 882 883 size_t ShenandoahHeap::initial_capacity() const { 884 return _initial_size; 885 } 886 887 void ShenandoahHeap::op_uncommit(double shrink_before, size_t shrink_until) { 888 assert (ShenandoahUncommit, "should be enabled"); 889 890 // Application allocates from the beginning of the heap, and GC allocates at 891 // the end of it. It is more efficient to uncommit from the end, so that applications 892 // could enjoy the near committed regions. GC allocations are much less frequent, 893 // and therefore can accept the committing costs. 894 895 size_t count = 0; 896 for (size_t i = num_regions(); i > 0; i--) { // care about size_t underflow 897 ShenandoahHeapRegion* r = get_region(i - 1); 898 if (r->is_empty_committed() && (r->empty_time() < shrink_before)) { 899 ShenandoahHeapLocker locker(lock()); 900 if (r->is_empty_committed()) { 901 if (committed() < shrink_until + ShenandoahHeapRegion::region_size_bytes()) { 902 break; 903 } 904 905 r->make_uncommitted(); 906 count++; 907 } 908 } 909 SpinPause(); // allow allocators to take the lock 910 } 911 912 if (count > 0) { 913 control_thread()->notify_heap_changed(); 914 regulator_thread()->notify_heap_changed(); 915 } 916 } 917 918 void ShenandoahHeap::handle_old_evacuation(HeapWord* obj, size_t words, bool promotion) { 919 // Only register the copy of the object that won the evacuation race. 920 card_scan()->register_object_without_lock(obj); 921 922 // Mark the entire range of the evacuated object as dirty. At next remembered set scan, 923 // we will clear dirty bits that do not hold interesting pointers. It's more efficient to 924 // do this in batch, in a background GC thread than to try to carefully dirty only cards 925 // that hold interesting pointers right now. 926 card_scan()->mark_range_as_dirty(obj, words); 927 928 if (promotion) { 929 // This evacuation was a promotion, track this as allocation against old gen 930 old_generation()->increase_allocated(words * HeapWordSize); 931 } 932 } 933 934 void ShenandoahHeap::handle_old_evacuation_failure() { 935 if (_old_gen_oom_evac.try_set()) { 936 log_info(gc)("Old gen evac failure."); 937 } 938 } 939 940 void ShenandoahHeap::report_promotion_failure(Thread* thread, size_t size) { 941 // We squelch excessive reports to reduce noise in logs. 942 const size_t MaxReportsPerEpoch = 4; 943 static size_t last_report_epoch = 0; 944 static size_t epoch_report_count = 0; 945 946 size_t promotion_reserve; 947 size_t promotion_expended; 948 949 size_t gc_id = control_thread()->get_gc_id(); 950 951 if ((gc_id != last_report_epoch) || (epoch_report_count++ < MaxReportsPerEpoch)) { 952 { 953 // Promotion failures should be very rare. Invest in providing useful diagnostic info. 954 ShenandoahHeapLocker locker(lock()); 955 promotion_reserve = get_promoted_reserve(); 956 promotion_expended = get_promoted_expended(); 957 } 958 PLAB* plab = ShenandoahThreadLocalData::plab(thread); 959 size_t words_remaining = (plab == nullptr)? 0: plab->words_remaining(); 960 const char* promote_enabled = ShenandoahThreadLocalData::allow_plab_promotions(thread)? "enabled": "disabled"; 961 ShenandoahGeneration* old_gen = old_generation(); 962 size_t old_capacity = old_gen->max_capacity(); 963 size_t old_usage = old_gen->used(); 964 size_t old_free_regions = old_gen->free_unaffiliated_regions(); 965 966 log_info(gc, ergo)("Promotion failed, size " SIZE_FORMAT ", has plab? %s, PLAB remaining: " SIZE_FORMAT 967 ", plab promotions %s, promotion reserve: " SIZE_FORMAT ", promotion expended: " SIZE_FORMAT 968 ", old capacity: " SIZE_FORMAT ", old_used: " SIZE_FORMAT ", old unaffiliated regions: " SIZE_FORMAT, 969 size * HeapWordSize, plab == nullptr? "no": "yes", 970 words_remaining * HeapWordSize, promote_enabled, promotion_reserve, promotion_expended, 971 old_capacity, old_usage, old_free_regions); 972 973 if ((gc_id == last_report_epoch) && (epoch_report_count >= MaxReportsPerEpoch)) { 974 log_info(gc, ergo)("Squelching additional promotion failure reports for current epoch"); 975 } else if (gc_id != last_report_epoch) { 976 last_report_epoch = gc_id;; 977 epoch_report_count = 1; 978 } 979 } 980 } 981 982 HeapWord* ShenandoahHeap::allocate_from_gclab_slow(Thread* thread, size_t size) { 983 // New object should fit the GCLAB size 984 size_t min_size = MAX2(size, PLAB::min_size()); 985 986 // Figure out size of new GCLAB, looking back at heuristics. Expand aggressively. 987 size_t new_size = ShenandoahThreadLocalData::gclab_size(thread) * 2; 988 989 // Limit growth of GCLABs to ShenandoahMaxEvacLABRatio * the minimum size. This enables more equitable distribution of 990 // available evacuation buidget between the many threads that are coordinating in the evacuation effort. 991 if (ShenandoahMaxEvacLABRatio > 0) { 992 log_debug(gc, free)("Allocate new gclab: " SIZE_FORMAT ", " SIZE_FORMAT, new_size, PLAB::min_size() * ShenandoahMaxEvacLABRatio); 993 new_size = MIN2(new_size, PLAB::min_size() * ShenandoahMaxEvacLABRatio); 994 } 995 996 new_size = MIN2(new_size, PLAB::max_size()); 997 new_size = MAX2(new_size, PLAB::min_size()); 998 999 // Record new heuristic value even if we take any shortcut. This captures 1000 // the case when moderately-sized objects always take a shortcut. At some point, 1001 // heuristics should catch up with them. 1002 ShenandoahThreadLocalData::set_gclab_size(thread, new_size); 1003 1004 if (new_size < size) { 1005 // New size still does not fit the object. Fall back to shared allocation. 1006 // This avoids retiring perfectly good GCLABs, when we encounter a large object. 1007 log_debug(gc, free)("New gclab size (" SIZE_FORMAT ") is too small for " SIZE_FORMAT, new_size, size); 1008 return nullptr; 1009 } 1010 1011 // Retire current GCLAB, and allocate a new one. 1012 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 1013 gclab->retire(); 1014 1015 size_t actual_size = 0; 1016 HeapWord* gclab_buf = allocate_new_gclab(min_size, new_size, &actual_size); 1017 if (gclab_buf == nullptr) { 1018 return nullptr; 1019 } 1020 1021 assert (size <= actual_size, "allocation should fit"); 1022 1023 if (ZeroTLAB) { 1024 // ..and clear it. 1025 Copy::zero_to_words(gclab_buf, actual_size); 1026 } else { 1027 // ...and zap just allocated object. 1028 #ifdef ASSERT 1029 // Skip mangling the space corresponding to the object header to 1030 // ensure that the returned space is not considered parsable by 1031 // any concurrent GC thread. 1032 size_t hdr_size = oopDesc::header_size(); 1033 Copy::fill_to_words(gclab_buf + hdr_size, actual_size - hdr_size, badHeapWordVal); 1034 #endif // ASSERT 1035 } 1036 gclab->set_buf(gclab_buf, actual_size); 1037 return gclab->allocate(size); 1038 } 1039 1040 // Establish a new PLAB and allocate size HeapWords within it. 1041 HeapWord* ShenandoahHeap::allocate_from_plab_slow(Thread* thread, size_t size, bool is_promotion) { 1042 // New object should fit the PLAB size 1043 size_t min_size = MAX2(size, PLAB::min_size()); 1044 1045 // Figure out size of new PLAB, looking back at heuristics. Expand aggressively. 1046 size_t cur_size = ShenandoahThreadLocalData::plab_size(thread); 1047 if (cur_size == 0) { 1048 cur_size = PLAB::min_size(); 1049 } 1050 size_t future_size = cur_size * 2; 1051 // Limit growth of PLABs to ShenandoahMaxEvacLABRatio * the minimum size. This enables more equitable distribution of 1052 // available evacuation buidget between the many threads that are coordinating in the evacuation effort. 1053 if (ShenandoahMaxEvacLABRatio > 0) { 1054 future_size = MIN2(future_size, PLAB::min_size() * ShenandoahMaxEvacLABRatio); 1055 } 1056 future_size = MIN2(future_size, PLAB::max_size()); 1057 future_size = MAX2(future_size, PLAB::min_size()); 1058 1059 size_t unalignment = future_size % CardTable::card_size_in_words(); 1060 if (unalignment != 0) { 1061 future_size = future_size - unalignment + CardTable::card_size_in_words(); 1062 } 1063 1064 // Record new heuristic value even if we take any shortcut. This captures 1065 // the case when moderately-sized objects always take a shortcut. At some point, 1066 // heuristics should catch up with them. Note that the requested cur_size may 1067 // not be honored, but we remember that this is the preferred size. 1068 ShenandoahThreadLocalData::set_plab_size(thread, future_size); 1069 if (cur_size < size) { 1070 // The PLAB to be allocated is still not large enough to hold the object. Fall back to shared allocation. 1071 // This avoids retiring perfectly good PLABs in order to represent a single large object allocation. 1072 return nullptr; 1073 } 1074 1075 // Retire current PLAB, and allocate a new one. 1076 PLAB* plab = ShenandoahThreadLocalData::plab(thread); 1077 if (plab->words_remaining() < PLAB::min_size()) { 1078 // Retire current PLAB, and allocate a new one. 1079 // CAUTION: retire_plab may register the remnant filler object with the remembered set scanner without a lock. This 1080 // is safe iff it is assured that each PLAB is a whole-number multiple of card-mark memory size and each PLAB is 1081 // aligned with the start of a card's memory range. 1082 retire_plab(plab, thread); 1083 1084 size_t actual_size = 0; 1085 // allocate_new_plab resets plab_evacuated and plab_promoted and disables promotions if old-gen available is 1086 // less than the remaining evacuation need. It also adjusts plab_preallocated and expend_promoted if appropriate. 1087 HeapWord* plab_buf = allocate_new_plab(min_size, cur_size, &actual_size); 1088 if (plab_buf == nullptr) { 1089 if (min_size == PLAB::min_size()) { 1090 // Disable plab promotions for this thread because we cannot even allocate a plab of minimal size. This allows us 1091 // to fail faster on subsequent promotion attempts. 1092 ShenandoahThreadLocalData::disable_plab_promotions(thread); 1093 } 1094 return NULL; 1095 } else { 1096 ShenandoahThreadLocalData::enable_plab_retries(thread); 1097 } 1098 assert (size <= actual_size, "allocation should fit"); 1099 if (ZeroTLAB) { 1100 // ..and clear it. 1101 Copy::zero_to_words(plab_buf, actual_size); 1102 } else { 1103 // ...and zap just allocated object. 1104 #ifdef ASSERT 1105 // Skip mangling the space corresponding to the object header to 1106 // ensure that the returned space is not considered parsable by 1107 // any concurrent GC thread. 1108 size_t hdr_size = oopDesc::header_size(); 1109 Copy::fill_to_words(plab_buf + hdr_size, actual_size - hdr_size, badHeapWordVal); 1110 #endif // ASSERT 1111 } 1112 plab->set_buf(plab_buf, actual_size); 1113 if (is_promotion && !ShenandoahThreadLocalData::allow_plab_promotions(thread)) { 1114 return nullptr; 1115 } 1116 return plab->allocate(size); 1117 } else { 1118 // If there's still at least min_size() words available within the current plab, don't retire it. Let's gnaw 1119 // away on this plab as long as we can. Meanwhile, return nullptr to force this particular allocation request 1120 // to be satisfied with a shared allocation. By packing more promotions into the previously allocated PLAB, we 1121 // reduce the likelihood of evacuation failures, and we we reduce the need for downsizing our PLABs. 1122 return nullptr; 1123 } 1124 } 1125 1126 // TODO: It is probably most efficient to register all objects (both promotions and evacuations) that were allocated within 1127 // this plab at the time we retire the plab. A tight registration loop will run within both code and data caches. This change 1128 // would allow smaller and faster in-line implementation of alloc_from_plab(). Since plabs are aligned on card-table boundaries, 1129 // this object registration loop can be performed without acquiring a lock. 1130 void ShenandoahHeap::retire_plab(PLAB* plab, Thread* thread) { 1131 // We don't enforce limits on plab_evacuated. We let it consume all available old-gen memory in order to reduce 1132 // probability of an evacuation failure. We do enforce limits on promotion, to make sure that excessive promotion 1133 // does not result in an old-gen evacuation failure. Note that a failed promotion is relatively harmless. Any 1134 // object that fails to promote in the current cycle will be eligible for promotion in a subsequent cycle. 1135 1136 // When the plab was instantiated, its entirety was treated as if the entire buffer was going to be dedicated to 1137 // promotions. Now that we are retiring the buffer, we adjust for the reality that the plab is not entirely promotions. 1138 // 1. Some of the plab may have been dedicated to evacuations. 1139 // 2. Some of the plab may have been abandoned due to waste (at the end of the plab). 1140 size_t not_promoted = 1141 ShenandoahThreadLocalData::get_plab_preallocated_promoted(thread) - ShenandoahThreadLocalData::get_plab_promoted(thread); 1142 ShenandoahThreadLocalData::reset_plab_promoted(thread); 1143 ShenandoahThreadLocalData::reset_plab_evacuated(thread); 1144 ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, 0); 1145 if (not_promoted > 0) { 1146 unexpend_promoted(not_promoted); 1147 } 1148 size_t waste = plab->waste(); 1149 HeapWord* top = plab->top(); 1150 plab->retire(); 1151 if (top != nullptr && plab->waste() > waste && is_in_old(top)) { 1152 // If retiring the plab created a filler object, then we 1153 // need to register it with our card scanner so it can 1154 // safely walk the region backing the plab. 1155 log_debug(gc)("retire_plab() is registering remnant of size " SIZE_FORMAT " at " PTR_FORMAT, 1156 plab->waste() - waste, p2i(top)); 1157 card_scan()->register_object_without_lock(top); 1158 } 1159 } 1160 1161 void ShenandoahHeap::retire_plab(PLAB* plab) { 1162 Thread* thread = Thread::current(); 1163 retire_plab(plab, thread); 1164 } 1165 1166 void ShenandoahHeap::cancel_old_gc() { 1167 shenandoah_assert_safepoint(); 1168 assert(_old_generation != nullptr, "Should only have mixed collections in generation mode."); 1169 log_info(gc)("Terminating old gc cycle."); 1170 1171 // Stop marking 1172 old_generation()->cancel_marking(); 1173 // Stop coalescing undead objects 1174 set_prepare_for_old_mark_in_progress(false); 1175 // Stop tracking old regions 1176 old_heuristics()->abandon_collection_candidates(); 1177 // Remove old generation access to young generation mark queues 1178 young_generation()->set_old_gen_task_queues(nullptr); 1179 // Transition to IDLE now. 1180 _old_generation->transition_to(ShenandoahOldGeneration::IDLE); 1181 } 1182 1183 bool ShenandoahHeap::is_old_gc_active() { 1184 return _old_generation->state() != ShenandoahOldGeneration::IDLE; 1185 } 1186 1187 // xfer_limit is the maximum we're able to transfer from young to old 1188 void ShenandoahHeap::adjust_generation_sizes_for_next_cycle( 1189 size_t xfer_limit, size_t young_cset_regions, size_t old_cset_regions) { 1190 1191 // Make sure old-generation is large enough, but no larger, than is necessary to hold mixed evacuations 1192 // and promotions if we anticipate either. 1193 size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes(); 1194 size_t promo_load = get_promotion_potential(); 1195 // The free set will reserve this amount of memory to hold young evacuations 1196 size_t young_reserve = (young_generation()->max_capacity() * ShenandoahEvacReserve) / 100; 1197 size_t old_reserve = 0; 1198 size_t mixed_candidates = old_heuristics()->unprocessed_old_collection_candidates(); 1199 bool doing_mixed = (mixed_candidates > 0); 1200 bool doing_promotions = promo_load > 0; 1201 1202 // round down 1203 size_t max_old_region_xfer = xfer_limit / region_size_bytes; 1204 1205 // We can limit the reserve to the size of anticipated promotions 1206 size_t max_old_reserve = young_reserve * ShenandoahOldEvacRatioPercent / (100 - ShenandoahOldEvacRatioPercent); 1207 // Here's the algebra: 1208 // TotalEvacuation = OldEvacuation + YoungEvacuation 1209 // OldEvacuation = TotalEvacuation*(ShenandoahOldEvacRatioPercent/100) 1210 // OldEvacuation = YoungEvacuation * (ShenandoahOldEvacRatioPercent/100)/(1 - ShenandoahOldEvacRatioPercent/100) 1211 // OldEvacuation = YoungEvacuation * ShenandoahOldEvacRatioPercent/(100 - ShenandoahOldEvacRatioPercent) 1212 1213 size_t reserve_for_mixed, reserve_for_promo; 1214 if (doing_mixed) { 1215 assert(old_generation()->available() >= old_generation()->free_unaffiliated_regions() * region_size_bytes, 1216 "Unaffiliated available must be less than total available"); 1217 1218 // We want this much memory to be unfragmented in order to reliably evacuate old. This is conservative because we 1219 // may not evacuate the entirety of unprocessed candidates in a single mixed evacuation. 1220 size_t max_evac_need = (size_t) 1221 (old_heuristics()->unprocessed_old_collection_candidates_live_memory() * ShenandoahOldEvacWaste); 1222 size_t old_fragmented_available = 1223 old_generation()->available() - old_generation()->free_unaffiliated_regions() * region_size_bytes; 1224 reserve_for_mixed = max_evac_need + old_fragmented_available; 1225 if (reserve_for_mixed > max_old_reserve) { 1226 reserve_for_mixed = max_old_reserve; 1227 } 1228 } else { 1229 reserve_for_mixed = 0; 1230 } 1231 1232 size_t available_for_promotions = max_old_reserve - reserve_for_mixed; 1233 if (doing_promotions) { 1234 // We're only promoting and we have a maximum bound on the amount to be promoted 1235 reserve_for_promo = (size_t) (promo_load * ShenandoahPromoEvacWaste); 1236 if (reserve_for_promo > available_for_promotions) { 1237 reserve_for_promo = available_for_promotions; 1238 } 1239 } else { 1240 reserve_for_promo = 0; 1241 } 1242 old_reserve = reserve_for_mixed + reserve_for_promo; 1243 assert(old_reserve <= max_old_reserve, "cannot reserve more than max for old evacuations"); 1244 size_t old_available = old_generation()->available() + old_cset_regions * region_size_bytes; 1245 size_t young_available = young_generation()->available() + young_cset_regions * region_size_bytes; 1246 size_t old_region_deficit = 0; 1247 size_t old_region_surplus = 0; 1248 if (old_available >= old_reserve) { 1249 size_t old_excess = old_available - old_reserve; 1250 size_t excess_regions = old_excess / region_size_bytes; 1251 size_t unaffiliated_old_regions = old_generation()->free_unaffiliated_regions() + old_cset_regions; 1252 size_t unaffiliated_old = unaffiliated_old_regions * region_size_bytes; 1253 if (unaffiliated_old_regions < excess_regions) { 1254 // We'll give only unaffiliated old to young, which is known to be less than the excess. 1255 old_region_surplus = unaffiliated_old_regions; 1256 } else { 1257 // unaffiliated_old_regions > excess_regions, so we only give away the excess. 1258 old_region_surplus = excess_regions; 1259 } 1260 } else { 1261 // We need to request transfer from YOUNG. Ignore that this will directly impact young_generation()->max_capacity(), 1262 // indirectly impacting young_reserve and old_reserve. These computations are conservative. 1263 size_t old_need = old_reserve - old_available; 1264 // Round up the number of regions needed from YOUNG 1265 old_region_deficit = (old_need + region_size_bytes - 1) / region_size_bytes; 1266 } 1267 if (old_region_deficit > max_old_region_xfer) { 1268 // If we're running short on young-gen memory, limit the xfer. Old-gen collection activities will be curtailed 1269 // if the budget is smaller than desired. 1270 old_region_deficit = max_old_region_xfer; 1271 } 1272 set_old_region_surplus(old_region_surplus); 1273 set_old_region_deficit(old_region_deficit); 1274 } 1275 1276 // Called from stubs in JIT code or interpreter 1277 HeapWord* ShenandoahHeap::allocate_new_tlab(size_t min_size, 1278 size_t requested_size, 1279 size_t* actual_size) { 1280 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_tlab(min_size, requested_size); 1281 HeapWord* res = allocate_memory(req, false); 1282 if (res != nullptr) { 1283 *actual_size = req.actual_size(); 1284 } else { 1285 *actual_size = 0; 1286 } 1287 return res; 1288 } 1289 1290 HeapWord* ShenandoahHeap::allocate_new_gclab(size_t min_size, 1291 size_t word_size, 1292 size_t* actual_size) { 1293 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_gclab(min_size, word_size); 1294 HeapWord* res = allocate_memory(req, false); 1295 if (res != nullptr) { 1296 *actual_size = req.actual_size(); 1297 } else { 1298 *actual_size = 0; 1299 } 1300 return res; 1301 } 1302 1303 HeapWord* ShenandoahHeap::allocate_new_plab(size_t min_size, 1304 size_t word_size, 1305 size_t* actual_size) { 1306 // Align requested sizes to card sized multiples 1307 size_t words_in_card = CardTable::card_size_in_words(); 1308 size_t align_mask = ~(words_in_card - 1); 1309 min_size = (min_size + words_in_card - 1) & align_mask; 1310 word_size = (word_size + words_in_card - 1) & align_mask; 1311 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_plab(min_size, word_size); 1312 // Note that allocate_memory() sets a thread-local flag to prohibit further promotions by this thread 1313 // if we are at risk of infringing on the old-gen evacuation budget. 1314 HeapWord* res = allocate_memory(req, false); 1315 if (res != nullptr) { 1316 *actual_size = req.actual_size(); 1317 } else { 1318 *actual_size = 0; 1319 } 1320 return res; 1321 } 1322 1323 // is_promotion is true iff this allocation is known for sure to hold the result of young-gen evacuation 1324 // to old-gen. plab allocates are not known as such, since they may hold old-gen evacuations. 1325 HeapWord* ShenandoahHeap::allocate_memory(ShenandoahAllocRequest& req, bool is_promotion) { 1326 intptr_t pacer_epoch = 0; 1327 bool in_new_region = false; 1328 HeapWord* result = nullptr; 1329 1330 if (req.is_mutator_alloc()) { 1331 if (ShenandoahPacing) { 1332 pacer()->pace_for_alloc(req.size()); 1333 pacer_epoch = pacer()->epoch(); 1334 } 1335 1336 if (!ShenandoahAllocFailureALot || !should_inject_alloc_failure()) { 1337 result = allocate_memory_under_lock(req, in_new_region, is_promotion); 1338 } 1339 1340 // Allocation failed, block until control thread reacted, then retry allocation. 1341 // 1342 // It might happen that one of the threads requesting allocation would unblock 1343 // way later after GC happened, only to fail the second allocation, because 1344 // other threads have already depleted the free storage. In this case, a better 1345 // strategy is to try again, as long as GC makes progress (or until at least 1346 // one full GC has completed). 1347 size_t original_count = shenandoah_policy()->full_gc_count(); 1348 while (result == nullptr 1349 && (_progress_last_gc.is_set() || original_count == shenandoah_policy()->full_gc_count())) { 1350 control_thread()->handle_alloc_failure(req); 1351 result = allocate_memory_under_lock(req, in_new_region, is_promotion); 1352 } 1353 1354 } else { 1355 assert(req.is_gc_alloc(), "Can only accept GC allocs here"); 1356 result = allocate_memory_under_lock(req, in_new_region, is_promotion); 1357 // Do not call handle_alloc_failure() here, because we cannot block. 1358 // The allocation failure would be handled by the LRB slowpath with handle_alloc_failure_evac(). 1359 } 1360 1361 if (in_new_region) { 1362 control_thread()->notify_heap_changed(); 1363 regulator_thread()->notify_heap_changed(); 1364 } 1365 1366 if (result == nullptr) { 1367 req.set_actual_size(0); 1368 } 1369 1370 // This is called regardless of the outcome of the allocation to account 1371 // for any waste created by retiring regions with this request. 1372 increase_used(req); 1373 1374 if (result != nullptr) { 1375 size_t requested = req.size(); 1376 size_t actual = req.actual_size(); 1377 1378 assert (req.is_lab_alloc() || (requested == actual), 1379 "Only LAB allocations are elastic: %s, requested = " SIZE_FORMAT ", actual = " SIZE_FORMAT, 1380 ShenandoahAllocRequest::alloc_type_to_string(req.type()), requested, actual); 1381 1382 if (req.is_mutator_alloc()) { 1383 // If we requested more than we were granted, give the rest back to pacer. 1384 // This only matters if we are in the same pacing epoch: do not try to unpace 1385 // over the budget for the other phase. 1386 if (ShenandoahPacing && (pacer_epoch > 0) && (requested > actual)) { 1387 pacer()->unpace_for_alloc(pacer_epoch, requested - actual); 1388 } 1389 } 1390 } 1391 1392 return result; 1393 } 1394 1395 HeapWord* ShenandoahHeap::allocate_memory_under_lock(ShenandoahAllocRequest& req, bool& in_new_region, bool is_promotion) { 1396 bool try_smaller_lab_size = false; 1397 size_t smaller_lab_size; 1398 { 1399 // promotion_eligible pertains only to PLAB allocations, denoting that the PLAB is allowed to allocate for promotions. 1400 bool promotion_eligible = false; 1401 bool allow_allocation = true; 1402 bool plab_alloc = false; 1403 size_t requested_bytes = req.size() * HeapWordSize; 1404 HeapWord* result = nullptr; 1405 ShenandoahHeapLocker locker(lock()); 1406 Thread* thread = Thread::current(); 1407 1408 if (mode()->is_generational()) { 1409 if (req.affiliation() == YOUNG_GENERATION) { 1410 if (req.is_mutator_alloc()) { 1411 size_t young_words_available = young_generation()->available() / HeapWordSize; 1412 if (ShenandoahElasticTLAB && req.is_lab_alloc() && (req.min_size() < young_words_available)) { 1413 // Allow ourselves to try a smaller lab size even if requested_bytes <= young_available. We may need a smaller 1414 // lab size because young memory has become too fragmented. 1415 try_smaller_lab_size = true; 1416 smaller_lab_size = (young_words_available < req.size())? young_words_available: req.size(); 1417 } else if (req.size() > young_words_available) { 1418 // Can't allocate because even min_size() is larger than remaining young_available 1419 log_info(gc, ergo)("Unable to shrink %s alloc request of minimum size: " SIZE_FORMAT 1420 ", young words available: " SIZE_FORMAT, req.type_string(), 1421 HeapWordSize * (req.is_lab_alloc()? req.min_size(): req.size()), young_words_available); 1422 return nullptr; 1423 } 1424 } 1425 } else { // reg.affiliation() == OLD_GENERATION 1426 assert(req.type() != ShenandoahAllocRequest::_alloc_gclab, "GCLAB pertains only to young-gen memory"); 1427 if (req.type() == ShenandoahAllocRequest::_alloc_plab) { 1428 plab_alloc = true; 1429 size_t promotion_avail = get_promoted_reserve(); 1430 size_t promotion_expended = get_promoted_expended(); 1431 if (promotion_expended + requested_bytes > promotion_avail) { 1432 promotion_avail = 0; 1433 if (get_old_evac_reserve() == 0) { 1434 // There are no old-gen evacuations in this pass. There's no value in creating a plab that cannot 1435 // be used for promotions. 1436 allow_allocation = false; 1437 } 1438 } else { 1439 promotion_avail = promotion_avail - (promotion_expended + requested_bytes); 1440 promotion_eligible = true; 1441 } 1442 } else if (is_promotion) { 1443 // This is a shared alloc for promotion 1444 size_t promotion_avail = get_promoted_reserve(); 1445 size_t promotion_expended = get_promoted_expended(); 1446 if (promotion_expended + requested_bytes > promotion_avail) { 1447 promotion_avail = 0; 1448 } else { 1449 promotion_avail = promotion_avail - (promotion_expended + requested_bytes); 1450 } 1451 if (promotion_avail == 0) { 1452 // We need to reserve the remaining memory for evacuation. Reject this allocation. The object will be 1453 // evacuated to young-gen memory and promoted during a future GC pass. 1454 return nullptr; 1455 } 1456 // Else, we'll allow the allocation to proceed. (Since we hold heap lock, the tested condition remains true.) 1457 } else { 1458 // This is a shared allocation for evacuation. Memory has already been reserved for this purpose. 1459 } 1460 } 1461 } // This ends the is_generational() block 1462 1463 // First try the original request. If TLAB request size is greater than available, allocate() will attempt to downsize 1464 // request to fit within available memory. 1465 result = (allow_allocation)? _free_set->allocate(req, in_new_region): nullptr; 1466 if (result != nullptr) { 1467 if (req.is_old()) { 1468 ShenandoahThreadLocalData::reset_plab_promoted(thread); 1469 if (req.is_gc_alloc()) { 1470 bool disable_plab_promotions = false; 1471 if (req.type() == ShenandoahAllocRequest::_alloc_plab) { 1472 if (promotion_eligible) { 1473 size_t actual_size = req.actual_size() * HeapWordSize; 1474 // The actual size of the allocation may be larger than the requested bytes (due to alignment on card boundaries). 1475 // If this puts us over our promotion budget, we need to disable future PLAB promotions for this thread. 1476 if (get_promoted_expended() + actual_size <= get_promoted_reserve()) { 1477 // Assume the entirety of this PLAB will be used for promotion. This prevents promotion from overreach. 1478 // When we retire this plab, we'll unexpend what we don't really use. 1479 ShenandoahThreadLocalData::enable_plab_promotions(thread); 1480 expend_promoted(actual_size); 1481 assert(get_promoted_expended() <= get_promoted_reserve(), "Do not expend more promotion than budgeted"); 1482 ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, actual_size); 1483 } else { 1484 disable_plab_promotions = true; 1485 } 1486 } else { 1487 disable_plab_promotions = true; 1488 } 1489 if (disable_plab_promotions) { 1490 // Disable promotions in this thread because entirety of this PLAB must be available to hold old-gen evacuations. 1491 ShenandoahThreadLocalData::disable_plab_promotions(thread); 1492 ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, 0); 1493 } 1494 } else if (is_promotion) { 1495 // Shared promotion. Assume size is requested_bytes. 1496 expend_promoted(requested_bytes); 1497 assert(get_promoted_expended() <= get_promoted_reserve(), "Do not expend more promotion than budgeted"); 1498 } 1499 } 1500 1501 // Register the newly allocated object while we're holding the global lock since there's no synchronization 1502 // built in to the implementation of register_object(). There are potential races when multiple independent 1503 // threads are allocating objects, some of which might span the same card region. For example, consider 1504 // a card table's memory region within which three objects are being allocated by three different threads: 1505 // 1506 // objects being "concurrently" allocated: 1507 // [-----a------][-----b-----][--------------c------------------] 1508 // [---- card table memory range --------------] 1509 // 1510 // Before any objects are allocated, this card's memory range holds no objects. Note that allocation of object a 1511 // wants to set the starts-object, first-start, and last-start attributes of the preceding card region. 1512 // allocation of object b wants to set the starts-object, first-start, and last-start attributes of this card region. 1513 // allocation of object c also wants to set the starts-object, first-start, and last-start attributes of this 1514 // card region. 1515 // 1516 // The thread allocating b and the thread allocating c can "race" in various ways, resulting in confusion, such as 1517 // last-start representing object b while first-start represents object c. This is why we need to require all 1518 // register_object() invocations to be "mutually exclusive" with respect to each card's memory range. 1519 ShenandoahHeap::heap()->card_scan()->register_object(result); 1520 } 1521 } else { 1522 // The allocation failed. If this was a plab allocation, We've already retired it and no longer have a plab. 1523 if (req.is_old() && req.is_gc_alloc() && (req.type() == ShenandoahAllocRequest::_alloc_plab)) { 1524 // We don't need to disable PLAB promotions because there is no PLAB. We leave promotions enabled because 1525 // this allows the surrounding infrastructure to retry alloc_plab_slow() with a smaller PLAB size. 1526 ShenandoahThreadLocalData::set_plab_preallocated_promoted(thread, 0); 1527 } 1528 } 1529 if ((result != nullptr) || !try_smaller_lab_size) { 1530 return result; 1531 } 1532 // else, fall through to try_smaller_lab_size 1533 } // This closes the block that holds the heap lock, releasing the lock. 1534 1535 // We failed to allocate the originally requested lab size. Let's see if we can allocate a smaller lab size. 1536 if (req.size() == smaller_lab_size) { 1537 // If we were already trying to allocate min size, no value in attempting to repeat the same. End the recursion. 1538 return nullptr; 1539 } 1540 1541 // We arrive here if the tlab allocation request can be resized to fit within young_available 1542 assert((req.affiliation() == YOUNG_GENERATION) && req.is_lab_alloc() && req.is_mutator_alloc() && 1543 (smaller_lab_size < req.size()), "Only shrink allocation request size for TLAB allocations"); 1544 1545 // By convention, ShenandoahAllocationRequest is primarily read-only. The only mutable instance data is represented by 1546 // actual_size(), which is overwritten with the size of the allocaion when the allocation request is satisfied. We use a 1547 // recursive call here rather than introducing new methods to mutate the existing ShenandoahAllocationRequest argument. 1548 // Mutation of the existing object might result in astonishing results if calling contexts assume the content of immutable 1549 // fields remain constant. The original TLAB allocation request was for memory that exceeded the current capacity. We'll 1550 // attempt to allocate a smaller TLAB. If this is successful, we'll update actual_size() of our incoming 1551 // ShenandoahAllocRequest. If the recursive request fails, we'll simply return nullptr. 1552 1553 // Note that we've relinquished the HeapLock and some other thread may perform additional allocation before our recursive 1554 // call reacquires the lock. If that happens, we will need another recursive call to further reduce the size of our request 1555 // for each time another thread allocates young memory during the brief intervals that the heap lock is available to 1556 // interfering threads. We expect this interference to be rare. The recursion bottoms out when young_available is 1557 // smaller than req.min_size(). The inner-nested call to allocate_memory_under_lock() uses the same min_size() value 1558 // as this call, but it uses a preferred size() that is smaller than our preferred size, and is no larger than what we most 1559 // recently saw as the memory currently available within the young generation. 1560 1561 // TODO: At the expense of code clarity, we could rewrite this recursive solution to use iteration. We need at most one 1562 // extra instance of the ShenandoahAllocRequest, which we can re-initialize multiple times inside a loop, with one iteration 1563 // of the loop required for each time the existing solution would recurse. An iterative solution would be more efficient 1564 // in CPU time and stack memory utilization. The expectation is that it is very rare that we would recurse more than once 1565 // so making this change is not currently seen as a high priority. 1566 1567 ShenandoahAllocRequest smaller_req = ShenandoahAllocRequest::for_tlab(req.min_size(), smaller_lab_size); 1568 1569 // Note that shrinking the preferred size gets us past the gatekeeper that checks whether there's available memory to 1570 // satisfy the allocation request. The reality is the actual TLAB size is likely to be even smaller, because it will 1571 // depend on how much memory is available within mutator regions that are not yet fully used. 1572 HeapWord* result = allocate_memory_under_lock(smaller_req, in_new_region, is_promotion); 1573 if (result != nullptr) { 1574 req.set_actual_size(smaller_req.actual_size()); 1575 } 1576 return result; 1577 } 1578 1579 HeapWord* ShenandoahHeap::mem_allocate(size_t size, 1580 bool* gc_overhead_limit_was_exceeded) { 1581 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared(size); 1582 return allocate_memory(req, false); 1583 } 1584 1585 MetaWord* ShenandoahHeap::satisfy_failed_metadata_allocation(ClassLoaderData* loader_data, 1586 size_t size, 1587 Metaspace::MetadataType mdtype) { 1588 MetaWord* result; 1589 1590 // Inform metaspace OOM to GC heuristics if class unloading is possible. 1591 ShenandoahHeuristics* h = global_generation()->heuristics(); 1592 if (h->can_unload_classes()) { 1593 h->record_metaspace_oom(); 1594 } 1595 1596 // Expand and retry allocation 1597 result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype); 1598 if (result != nullptr) { 1599 return result; 1600 } 1601 1602 // Start full GC 1603 collect(GCCause::_metadata_GC_clear_soft_refs); 1604 1605 // Retry allocation 1606 result = loader_data->metaspace_non_null()->allocate(size, mdtype); 1607 if (result != nullptr) { 1608 return result; 1609 } 1610 1611 // Expand and retry allocation 1612 result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype); 1613 if (result != nullptr) { 1614 return result; 1615 } 1616 1617 // Out of memory 1618 return nullptr; 1619 } 1620 1621 class ShenandoahConcurrentEvacuateRegionObjectClosure : public ObjectClosure { 1622 private: 1623 ShenandoahHeap* const _heap; 1624 Thread* const _thread; 1625 public: 1626 ShenandoahConcurrentEvacuateRegionObjectClosure(ShenandoahHeap* heap) : 1627 _heap(heap), _thread(Thread::current()) {} 1628 1629 void do_object(oop p) { 1630 shenandoah_assert_marked(nullptr, p); 1631 if (!p->is_forwarded()) { 1632 _heap->evacuate_object(p, _thread); 1633 } 1634 } 1635 }; 1636 1637 class ShenandoahEvacuationTask : public WorkerTask { 1638 private: 1639 ShenandoahHeap* const _sh; 1640 ShenandoahCollectionSet* const _cs; 1641 bool _concurrent; 1642 public: 1643 ShenandoahEvacuationTask(ShenandoahHeap* sh, 1644 ShenandoahCollectionSet* cs, 1645 bool concurrent) : 1646 WorkerTask("Shenandoah Evacuation"), 1647 _sh(sh), 1648 _cs(cs), 1649 _concurrent(concurrent) 1650 {} 1651 1652 void work(uint worker_id) { 1653 if (_concurrent) { 1654 ShenandoahConcurrentWorkerSession worker_session(worker_id); 1655 ShenandoahSuspendibleThreadSetJoiner stsj(ShenandoahSuspendibleWorkers); 1656 ShenandoahEvacOOMScope oom_evac_scope; 1657 do_work(); 1658 } else { 1659 ShenandoahParallelWorkerSession worker_session(worker_id); 1660 ShenandoahEvacOOMScope oom_evac_scope; 1661 do_work(); 1662 } 1663 } 1664 1665 private: 1666 void do_work() { 1667 ShenandoahConcurrentEvacuateRegionObjectClosure cl(_sh); 1668 ShenandoahHeapRegion* r; 1669 while ((r =_cs->claim_next()) != nullptr) { 1670 assert(r->has_live(), "Region " SIZE_FORMAT " should have been reclaimed early", r->index()); 1671 1672 _sh->marked_object_iterate(r, &cl); 1673 1674 if (ShenandoahPacing) { 1675 _sh->pacer()->report_evac(r->used() >> LogHeapWordSize); 1676 } 1677 if (_sh->check_cancelled_gc_and_yield(_concurrent)) { 1678 break; 1679 } 1680 } 1681 } 1682 }; 1683 1684 // Unlike ShenandoahEvacuationTask, this iterates over all regions rather than just the collection set. 1685 // This is needed in order to promote humongous start regions if age() >= tenure threshold. 1686 class ShenandoahGenerationalEvacuationTask : public WorkerTask { 1687 private: 1688 ShenandoahHeap* const _sh; 1689 ShenandoahRegionIterator *_regions; 1690 bool _concurrent; 1691 uint _tenuring_threshold; 1692 1693 public: 1694 ShenandoahGenerationalEvacuationTask(ShenandoahHeap* sh, 1695 ShenandoahRegionIterator* iterator, 1696 bool concurrent) : 1697 WorkerTask("Shenandoah Evacuation"), 1698 _sh(sh), 1699 _regions(iterator), 1700 _concurrent(concurrent), 1701 _tenuring_threshold(0) 1702 { 1703 if (_sh->mode()->is_generational()) { 1704 _tenuring_threshold = _sh->age_census()->tenuring_threshold(); 1705 } 1706 } 1707 1708 void work(uint worker_id) { 1709 if (_concurrent) { 1710 ShenandoahConcurrentWorkerSession worker_session(worker_id); 1711 ShenandoahSuspendibleThreadSetJoiner stsj(ShenandoahSuspendibleWorkers); 1712 ShenandoahEvacOOMScope oom_evac_scope; 1713 do_work(); 1714 } else { 1715 ShenandoahParallelWorkerSession worker_session(worker_id); 1716 ShenandoahEvacOOMScope oom_evac_scope; 1717 do_work(); 1718 } 1719 } 1720 1721 private: 1722 void do_work() { 1723 ShenandoahConcurrentEvacuateRegionObjectClosure cl(_sh); 1724 ShenandoahHeapRegion* r; 1725 ShenandoahMarkingContext* const ctx = ShenandoahHeap::heap()->marking_context(); 1726 size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes(); 1727 size_t old_garbage_threshold = (region_size_bytes * ShenandoahOldGarbageThreshold) / 100; 1728 while ((r = _regions->next()) != nullptr) { 1729 log_debug(gc)("GenerationalEvacuationTask do_work(), looking at %s region " SIZE_FORMAT ", (age: %d) [%s, %s, %s]", 1730 r->is_old()? "old": r->is_young()? "young": "free", r->index(), r->age(), 1731 r->is_active()? "active": "inactive", 1732 r->is_humongous()? (r->is_humongous_start()? "humongous_start": "humongous_continuation"): "regular", 1733 r->is_cset()? "cset": "not-cset"); 1734 1735 if (r->is_cset()) { 1736 assert(r->has_live(), "Region " SIZE_FORMAT " should have been reclaimed early", r->index()); 1737 _sh->marked_object_iterate(r, &cl); 1738 if (ShenandoahPacing) { 1739 _sh->pacer()->report_evac(r->used() >> LogHeapWordSize); 1740 } 1741 } else if (r->is_young() && r->is_active() && (r->age() >= _tenuring_threshold)) { 1742 HeapWord* tams = ctx->top_at_mark_start(r); 1743 if (r->is_humongous_start()) { 1744 // We promote humongous_start regions along with their affiliated continuations during evacuation rather than 1745 // doing this work during a safepoint. We cannot put humongous regions into the collection set because that 1746 // triggers the load-reference barrier (LRB) to copy on reference fetch. 1747 r->promote_humongous(); 1748 } else if (r->is_regular() && (r->get_top_before_promote() != nullptr)) { 1749 assert(r->garbage_before_padded_for_promote() < old_garbage_threshold, 1750 "Region " SIZE_FORMAT " has too much garbage for promotion", r->index()); 1751 assert(r->get_top_before_promote() == tams, 1752 "Region " SIZE_FORMAT " has been used for allocations before promotion", r->index()); 1753 // Likewise, we cannot put promote-in-place regions into the collection set because that would also trigger 1754 // the LRB to copy on reference fetch. 1755 r->promote_in_place(); 1756 } 1757 // Aged humongous continuation regions are handled with their start region. If an aged regular region has 1758 // more garbage than ShenandoahOldGarbageTrheshold, we'll promote by evacuation. If there is room for evacuation 1759 // in this cycle, the region will be in the collection set. If there is not room, the region will be promoted 1760 // by evacuation in some future GC cycle. 1761 1762 // If an aged regular region has received allocations during the current cycle, we do not promote because the 1763 // newly allocated objects do not have appropriate age; this region's age will be reset to zero at end of cycle. 1764 } 1765 // else, region is free, or OLD, or not in collection set, or humongous_continuation, 1766 // or is young humongous_start that is too young to be promoted 1767 1768 if (_sh->check_cancelled_gc_and_yield(_concurrent)) { 1769 break; 1770 } 1771 } 1772 } 1773 }; 1774 1775 void ShenandoahHeap::evacuate_collection_set(bool concurrent) { 1776 if (ShenandoahHeap::heap()->mode()->is_generational()) { 1777 ShenandoahRegionIterator regions; 1778 ShenandoahGenerationalEvacuationTask task(this, ®ions, concurrent); 1779 workers()->run_task(&task); 1780 } else { 1781 ShenandoahEvacuationTask task(this, _collection_set, concurrent); 1782 workers()->run_task(&task); 1783 } 1784 } 1785 1786 void ShenandoahHeap::trash_cset_regions() { 1787 ShenandoahHeapLocker locker(lock()); 1788 1789 ShenandoahCollectionSet* set = collection_set(); 1790 ShenandoahHeapRegion* r; 1791 set->clear_current_index(); 1792 while ((r = set->next()) != nullptr) { 1793 r->make_trash(); 1794 } 1795 collection_set()->clear(); 1796 } 1797 1798 void ShenandoahHeap::print_heap_regions_on(outputStream* st) const { 1799 st->print_cr("Heap Regions:"); 1800 st->print_cr("Region state: EU=empty-uncommitted, EC=empty-committed, R=regular, H=humongous start, HP=pinned humongous start"); 1801 st->print_cr(" HC=humongous continuation, CS=collection set, TR=trash, P=pinned, CSP=pinned collection set"); 1802 st->print_cr("BTE=bottom/top/end, TAMS=top-at-mark-start"); 1803 st->print_cr("UWM=update watermark, U=used"); 1804 st->print_cr("T=TLAB allocs, G=GCLAB allocs"); 1805 st->print_cr("S=shared allocs, L=live data"); 1806 st->print_cr("CP=critical pins"); 1807 1808 for (size_t i = 0; i < num_regions(); i++) { 1809 get_region(i)->print_on(st); 1810 } 1811 } 1812 1813 size_t ShenandoahHeap::trash_humongous_region_at(ShenandoahHeapRegion* start) { 1814 assert(start->is_humongous_start(), "reclaim regions starting with the first one"); 1815 1816 oop humongous_obj = cast_to_oop(start->bottom()); 1817 size_t size = humongous_obj->size(); 1818 size_t required_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize); 1819 size_t index = start->index() + required_regions - 1; 1820 1821 assert(!start->has_live(), "liveness must be zero"); 1822 1823 for(size_t i = 0; i < required_regions; i++) { 1824 // Reclaim from tail. Otherwise, assertion fails when printing region to trace log, 1825 // as it expects that every region belongs to a humongous region starting with a humongous start region. 1826 ShenandoahHeapRegion* region = get_region(index --); 1827 1828 assert(region->is_humongous(), "expect correct humongous start or continuation"); 1829 assert(!region->is_cset(), "Humongous region should not be in collection set"); 1830 1831 region->make_trash_immediate(); 1832 } 1833 return required_regions; 1834 } 1835 1836 class ShenandoahCheckCleanGCLABClosure : public ThreadClosure { 1837 public: 1838 ShenandoahCheckCleanGCLABClosure() {} 1839 void do_thread(Thread* thread) { 1840 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 1841 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name()); 1842 assert(gclab->words_remaining() == 0, "GCLAB should not need retirement"); 1843 1844 PLAB* plab = ShenandoahThreadLocalData::plab(thread); 1845 assert(plab != nullptr, "PLAB should be initialized for %s", thread->name()); 1846 assert(plab->words_remaining() == 0, "PLAB should not need retirement"); 1847 } 1848 }; 1849 1850 class ShenandoahRetireGCLABClosure : public ThreadClosure { 1851 private: 1852 bool const _resize; 1853 public: 1854 ShenandoahRetireGCLABClosure(bool resize) : _resize(resize) {} 1855 void do_thread(Thread* thread) { 1856 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 1857 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name()); 1858 gclab->retire(); 1859 if (_resize && ShenandoahThreadLocalData::gclab_size(thread) > 0) { 1860 ShenandoahThreadLocalData::set_gclab_size(thread, 0); 1861 } 1862 1863 PLAB* plab = ShenandoahThreadLocalData::plab(thread); 1864 assert(plab != nullptr, "PLAB should be initialized for %s", thread->name()); 1865 1866 // There are two reasons to retire all plabs between old-gen evacuation passes. 1867 // 1. We need to make the plab memory parseable by remembered-set scanning. 1868 // 2. We need to establish a trustworthy UpdateWaterMark value within each old-gen heap region 1869 ShenandoahHeap::heap()->retire_plab(plab, thread); 1870 if (_resize && ShenandoahThreadLocalData::plab_size(thread) > 0) { 1871 ShenandoahThreadLocalData::set_plab_size(thread, 0); 1872 } 1873 } 1874 }; 1875 1876 void ShenandoahHeap::labs_make_parsable() { 1877 assert(UseTLAB, "Only call with UseTLAB"); 1878 1879 ShenandoahRetireGCLABClosure cl(false); 1880 1881 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) { 1882 ThreadLocalAllocBuffer& tlab = t->tlab(); 1883 tlab.make_parsable(); 1884 cl.do_thread(t); 1885 } 1886 1887 workers()->threads_do(&cl); 1888 } 1889 1890 void ShenandoahHeap::tlabs_retire(bool resize) { 1891 assert(UseTLAB, "Only call with UseTLAB"); 1892 assert(!resize || ResizeTLAB, "Only call for resize when ResizeTLAB is enabled"); 1893 1894 ThreadLocalAllocStats stats; 1895 1896 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) { 1897 ThreadLocalAllocBuffer& tlab = t->tlab(); 1898 tlab.retire(&stats); 1899 if (resize) { 1900 tlab.resize(); 1901 } 1902 } 1903 1904 stats.publish(); 1905 1906 #ifdef ASSERT 1907 ShenandoahCheckCleanGCLABClosure cl; 1908 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) { 1909 cl.do_thread(t); 1910 } 1911 workers()->threads_do(&cl); 1912 #endif 1913 } 1914 1915 void ShenandoahHeap::gclabs_retire(bool resize) { 1916 assert(UseTLAB, "Only call with UseTLAB"); 1917 assert(!resize || ResizeTLAB, "Only call for resize when ResizeTLAB is enabled"); 1918 1919 ShenandoahRetireGCLABClosure cl(resize); 1920 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) { 1921 cl.do_thread(t); 1922 } 1923 workers()->threads_do(&cl); 1924 1925 if (safepoint_workers() != nullptr) { 1926 safepoint_workers()->threads_do(&cl); 1927 } 1928 } 1929 1930 class ShenandoahTagGCLABClosure : public ThreadClosure { 1931 public: 1932 void do_thread(Thread* thread) { 1933 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread); 1934 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name()); 1935 if (gclab->words_remaining() > 0) { 1936 ShenandoahHeapRegion* r = ShenandoahHeap::heap()->heap_region_containing(gclab->allocate(0)); 1937 r->set_young_lab_flag(); 1938 } 1939 } 1940 }; 1941 1942 void ShenandoahHeap::set_young_lab_region_flags() { 1943 if (!UseTLAB) { 1944 return; 1945 } 1946 for (size_t i = 0; i < _num_regions; i++) { 1947 _regions[i]->clear_young_lab_flags(); 1948 } 1949 ShenandoahTagGCLABClosure cl; 1950 workers()->threads_do(&cl); 1951 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) { 1952 cl.do_thread(t); 1953 ThreadLocalAllocBuffer& tlab = t->tlab(); 1954 if (tlab.end() != nullptr) { 1955 ShenandoahHeapRegion* r = heap_region_containing(tlab.start()); 1956 r->set_young_lab_flag(); 1957 } 1958 } 1959 } 1960 1961 // Returns size in bytes 1962 size_t ShenandoahHeap::unsafe_max_tlab_alloc(Thread *thread) const { 1963 if (ShenandoahElasticTLAB) { 1964 if (mode()->is_generational()) { 1965 return MIN2(ShenandoahHeapRegion::max_tlab_size_bytes(), young_generation()->available()); 1966 } else { 1967 // With Elastic TLABs, return the max allowed size, and let the allocation path 1968 // figure out the safe size for current allocation. 1969 return ShenandoahHeapRegion::max_tlab_size_bytes(); 1970 } 1971 } else { 1972 return MIN2(_free_set->unsafe_peek_free(), ShenandoahHeapRegion::max_tlab_size_bytes()); 1973 } 1974 } 1975 1976 size_t ShenandoahHeap::max_tlab_size() const { 1977 // Returns size in words 1978 return ShenandoahHeapRegion::max_tlab_size_words(); 1979 } 1980 1981 void ShenandoahHeap::collect(GCCause::Cause cause) { 1982 control_thread()->request_gc(cause); 1983 } 1984 1985 void ShenandoahHeap::do_full_collection(bool clear_all_soft_refs) { 1986 //assert(false, "Shouldn't need to do full collections"); 1987 } 1988 1989 HeapWord* ShenandoahHeap::block_start(const void* addr) const { 1990 ShenandoahHeapRegion* r = heap_region_containing(addr); 1991 if (r != nullptr) { 1992 return r->block_start(addr); 1993 } 1994 return nullptr; 1995 } 1996 1997 bool ShenandoahHeap::block_is_obj(const HeapWord* addr) const { 1998 ShenandoahHeapRegion* r = heap_region_containing(addr); 1999 return r->block_is_obj(addr); 2000 } 2001 2002 bool ShenandoahHeap::print_location(outputStream* st, void* addr) const { 2003 return BlockLocationPrinter<ShenandoahHeap>::print_location(st, addr); 2004 } 2005 2006 void ShenandoahHeap::prepare_for_verify() { 2007 if (SafepointSynchronize::is_at_safepoint() && UseTLAB) { 2008 labs_make_parsable(); 2009 } 2010 } 2011 2012 void ShenandoahHeap::gc_threads_do(ThreadClosure* tcl) const { 2013 if (_shenandoah_policy->is_at_shutdown()) { 2014 return; 2015 } 2016 2017 tcl->do_thread(_control_thread); 2018 tcl->do_thread(_regulator_thread); 2019 workers()->threads_do(tcl); 2020 if (_safepoint_workers != nullptr) { 2021 _safepoint_workers->threads_do(tcl); 2022 } 2023 } 2024 2025 void ShenandoahHeap::print_tracing_info() const { 2026 LogTarget(Info, gc, stats) lt; 2027 if (lt.is_enabled()) { 2028 ResourceMark rm; 2029 LogStream ls(lt); 2030 2031 phase_timings()->print_global_on(&ls); 2032 2033 ls.cr(); 2034 ls.cr(); 2035 2036 shenandoah_policy()->print_gc_stats(&ls); 2037 2038 ls.cr(); 2039 2040 evac_tracker()->print_global_on(&ls); 2041 2042 ls.cr(); 2043 ls.cr(); 2044 } 2045 } 2046 2047 void ShenandoahHeap::on_cycle_start(GCCause::Cause cause, ShenandoahGeneration* generation) { 2048 set_gc_cause(cause); 2049 set_gc_generation(generation); 2050 2051 shenandoah_policy()->record_cycle_start(); 2052 generation->heuristics()->record_cycle_start(); 2053 } 2054 2055 void ShenandoahHeap::on_cycle_end(ShenandoahGeneration* generation) { 2056 generation->heuristics()->record_cycle_end(); 2057 if (mode()->is_generational() && (generation->is_global() || upgraded_to_full())) { 2058 // If we just completed a GLOBAL GC, claim credit for completion of young-gen and old-gen GC as well 2059 young_generation()->heuristics()->record_cycle_end(); 2060 old_generation()->heuristics()->record_cycle_end(); 2061 } 2062 set_gc_cause(GCCause::_no_gc); 2063 } 2064 2065 void ShenandoahHeap::verify(VerifyOption vo) { 2066 if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) { 2067 if (ShenandoahVerify) { 2068 verifier()->verify_generic(vo); 2069 } else { 2070 // TODO: Consider allocating verification bitmaps on demand, 2071 // and turn this on unconditionally. 2072 } 2073 } 2074 } 2075 size_t ShenandoahHeap::tlab_capacity(Thread *thr) const { 2076 return _free_set->capacity(); 2077 } 2078 2079 class ObjectIterateScanRootClosure : public BasicOopIterateClosure { 2080 private: 2081 MarkBitMap* _bitmap; 2082 ShenandoahScanObjectStack* _oop_stack; 2083 ShenandoahHeap* const _heap; 2084 ShenandoahMarkingContext* const _marking_context; 2085 2086 template <class T> 2087 void do_oop_work(T* p) { 2088 T o = RawAccess<>::oop_load(p); 2089 if (!CompressedOops::is_null(o)) { 2090 oop obj = CompressedOops::decode_not_null(o); 2091 if (_heap->is_concurrent_weak_root_in_progress() && !_marking_context->is_marked(obj)) { 2092 // There may be dead oops in weak roots in concurrent root phase, do not touch them. 2093 return; 2094 } 2095 obj = ShenandoahBarrierSet::barrier_set()->load_reference_barrier(obj); 2096 2097 assert(oopDesc::is_oop(obj), "must be a valid oop"); 2098 if (!_bitmap->is_marked(obj)) { 2099 _bitmap->mark(obj); 2100 _oop_stack->push(obj); 2101 } 2102 } 2103 } 2104 public: 2105 ObjectIterateScanRootClosure(MarkBitMap* bitmap, ShenandoahScanObjectStack* oop_stack) : 2106 _bitmap(bitmap), _oop_stack(oop_stack), _heap(ShenandoahHeap::heap()), 2107 _marking_context(_heap->marking_context()) {} 2108 void do_oop(oop* p) { do_oop_work(p); } 2109 void do_oop(narrowOop* p) { do_oop_work(p); } 2110 }; 2111 2112 /* 2113 * This is public API, used in preparation of object_iterate(). 2114 * Since we don't do linear scan of heap in object_iterate() (see comment below), we don't 2115 * need to make the heap parsable. For Shenandoah-internal linear heap scans that we can 2116 * control, we call SH::tlabs_retire, SH::gclabs_retire. 2117 */ 2118 void ShenandoahHeap::ensure_parsability(bool retire_tlabs) { 2119 // No-op. 2120 } 2121 2122 /* 2123 * Iterates objects in the heap. This is public API, used for, e.g., heap dumping. 2124 * 2125 * We cannot safely iterate objects by doing a linear scan at random points in time. Linear 2126 * scanning needs to deal with dead objects, which may have dead Klass* pointers (e.g. 2127 * calling oopDesc::size() would crash) or dangling reference fields (crashes) etc. Linear 2128 * scanning therefore depends on having a valid marking bitmap to support it. However, we only 2129 * have a valid marking bitmap after successful marking. In particular, we *don't* have a valid 2130 * marking bitmap during marking, after aborted marking or during/after cleanup (when we just 2131 * wiped the bitmap in preparation for next marking). 2132 * 2133 * For all those reasons, we implement object iteration as a single marking traversal, reporting 2134 * objects as we mark+traverse through the heap, starting from GC roots. JVMTI IterateThroughHeap 2135 * is allowed to report dead objects, but is not required to do so. 2136 */ 2137 void ShenandoahHeap::object_iterate(ObjectClosure* cl) { 2138 // Reset bitmap 2139 if (!prepare_aux_bitmap_for_iteration()) 2140 return; 2141 2142 ShenandoahScanObjectStack oop_stack; 2143 ObjectIterateScanRootClosure oops(&_aux_bit_map, &oop_stack); 2144 // Seed the stack with root scan 2145 scan_roots_for_iteration(&oop_stack, &oops); 2146 2147 // Work through the oop stack to traverse heap 2148 while (! oop_stack.is_empty()) { 2149 oop obj = oop_stack.pop(); 2150 assert(oopDesc::is_oop(obj), "must be a valid oop"); 2151 cl->do_object(obj); 2152 obj->oop_iterate(&oops); 2153 } 2154 2155 assert(oop_stack.is_empty(), "should be empty"); 2156 // Reclaim bitmap 2157 reclaim_aux_bitmap_for_iteration(); 2158 } 2159 2160 bool ShenandoahHeap::prepare_aux_bitmap_for_iteration() { 2161 assert(SafepointSynchronize::is_at_safepoint(), "safe iteration is only available during safepoints"); 2162 2163 if (!_aux_bitmap_region_special && !os::commit_memory((char*)_aux_bitmap_region.start(), _aux_bitmap_region.byte_size(), false)) { 2164 log_warning(gc)("Could not commit native memory for auxiliary marking bitmap for heap iteration"); 2165 return false; 2166 } 2167 // Reset bitmap 2168 _aux_bit_map.clear(); 2169 return true; 2170 } 2171 2172 void ShenandoahHeap::scan_roots_for_iteration(ShenandoahScanObjectStack* oop_stack, ObjectIterateScanRootClosure* oops) { 2173 // Process GC roots according to current GC cycle 2174 // This populates the work stack with initial objects 2175 // It is important to relinquish the associated locks before diving 2176 // into heap dumper 2177 uint n_workers = safepoint_workers() != nullptr ? safepoint_workers()->active_workers() : 1; 2178 ShenandoahHeapIterationRootScanner rp(n_workers); 2179 rp.roots_do(oops); 2180 } 2181 2182 void ShenandoahHeap::reclaim_aux_bitmap_for_iteration() { 2183 if (!_aux_bitmap_region_special && !os::uncommit_memory((char*)_aux_bitmap_region.start(), _aux_bitmap_region.byte_size())) { 2184 log_warning(gc)("Could not uncommit native memory for auxiliary marking bitmap for heap iteration"); 2185 } 2186 } 2187 2188 // Closure for parallelly iterate objects 2189 class ShenandoahObjectIterateParScanClosure : public BasicOopIterateClosure { 2190 private: 2191 MarkBitMap* _bitmap; 2192 ShenandoahObjToScanQueue* _queue; 2193 ShenandoahHeap* const _heap; 2194 ShenandoahMarkingContext* const _marking_context; 2195 2196 template <class T> 2197 void do_oop_work(T* p) { 2198 T o = RawAccess<>::oop_load(p); 2199 if (!CompressedOops::is_null(o)) { 2200 oop obj = CompressedOops::decode_not_null(o); 2201 if (_heap->is_concurrent_weak_root_in_progress() && !_marking_context->is_marked(obj)) { 2202 // There may be dead oops in weak roots in concurrent root phase, do not touch them. 2203 return; 2204 } 2205 obj = ShenandoahBarrierSet::barrier_set()->load_reference_barrier(obj); 2206 2207 assert(oopDesc::is_oop(obj), "Must be a valid oop"); 2208 if (_bitmap->par_mark(obj)) { 2209 _queue->push(ShenandoahMarkTask(obj)); 2210 } 2211 } 2212 } 2213 public: 2214 ShenandoahObjectIterateParScanClosure(MarkBitMap* bitmap, ShenandoahObjToScanQueue* q) : 2215 _bitmap(bitmap), _queue(q), _heap(ShenandoahHeap::heap()), 2216 _marking_context(_heap->marking_context()) {} 2217 void do_oop(oop* p) { do_oop_work(p); } 2218 void do_oop(narrowOop* p) { do_oop_work(p); } 2219 }; 2220 2221 // Object iterator for parallel heap iteraion. 2222 // The root scanning phase happenes in construction as a preparation of 2223 // parallel marking queues. 2224 // Every worker processes it's own marking queue. work-stealing is used 2225 // to balance workload. 2226 class ShenandoahParallelObjectIterator : public ParallelObjectIteratorImpl { 2227 private: 2228 uint _num_workers; 2229 bool _init_ready; 2230 MarkBitMap* _aux_bit_map; 2231 ShenandoahHeap* _heap; 2232 ShenandoahScanObjectStack _roots_stack; // global roots stack 2233 ShenandoahObjToScanQueueSet* _task_queues; 2234 public: 2235 ShenandoahParallelObjectIterator(uint num_workers, MarkBitMap* bitmap) : 2236 _num_workers(num_workers), 2237 _init_ready(false), 2238 _aux_bit_map(bitmap), 2239 _heap(ShenandoahHeap::heap()) { 2240 // Initialize bitmap 2241 _init_ready = _heap->prepare_aux_bitmap_for_iteration(); 2242 if (!_init_ready) { 2243 return; 2244 } 2245 2246 ObjectIterateScanRootClosure oops(_aux_bit_map, &_roots_stack); 2247 _heap->scan_roots_for_iteration(&_roots_stack, &oops); 2248 2249 _init_ready = prepare_worker_queues(); 2250 } 2251 2252 ~ShenandoahParallelObjectIterator() { 2253 // Reclaim bitmap 2254 _heap->reclaim_aux_bitmap_for_iteration(); 2255 // Reclaim queue for workers 2256 if (_task_queues!= nullptr) { 2257 for (uint i = 0; i < _num_workers; ++i) { 2258 ShenandoahObjToScanQueue* q = _task_queues->queue(i); 2259 if (q != nullptr) { 2260 delete q; 2261 _task_queues->register_queue(i, nullptr); 2262 } 2263 } 2264 delete _task_queues; 2265 _task_queues = nullptr; 2266 } 2267 } 2268 2269 virtual void object_iterate(ObjectClosure* cl, uint worker_id) { 2270 if (_init_ready) { 2271 object_iterate_parallel(cl, worker_id, _task_queues); 2272 } 2273 } 2274 2275 private: 2276 // Divide global root_stack into worker queues 2277 bool prepare_worker_queues() { 2278 _task_queues = new ShenandoahObjToScanQueueSet((int) _num_workers); 2279 // Initialize queues for every workers 2280 for (uint i = 0; i < _num_workers; ++i) { 2281 ShenandoahObjToScanQueue* task_queue = new ShenandoahObjToScanQueue(); 2282 _task_queues->register_queue(i, task_queue); 2283 } 2284 // Divide roots among the workers. Assume that object referencing distribution 2285 // is related with root kind, use round-robin to make every worker have same chance 2286 // to process every kind of roots 2287 size_t roots_num = _roots_stack.size(); 2288 if (roots_num == 0) { 2289 // No work to do 2290 return false; 2291 } 2292 2293 for (uint j = 0; j < roots_num; j++) { 2294 uint stack_id = j % _num_workers; 2295 oop obj = _roots_stack.pop(); 2296 _task_queues->queue(stack_id)->push(ShenandoahMarkTask(obj)); 2297 } 2298 return true; 2299 } 2300 2301 void object_iterate_parallel(ObjectClosure* cl, 2302 uint worker_id, 2303 ShenandoahObjToScanQueueSet* queue_set) { 2304 assert(SafepointSynchronize::is_at_safepoint(), "safe iteration is only available during safepoints"); 2305 assert(queue_set != nullptr, "task queue must not be null"); 2306 2307 ShenandoahObjToScanQueue* q = queue_set->queue(worker_id); 2308 assert(q != nullptr, "object iterate queue must not be null"); 2309 2310 ShenandoahMarkTask t; 2311 ShenandoahObjectIterateParScanClosure oops(_aux_bit_map, q); 2312 2313 // Work through the queue to traverse heap. 2314 // Steal when there is no task in queue. 2315 while (q->pop(t) || queue_set->steal(worker_id, t)) { 2316 oop obj = t.obj(); 2317 assert(oopDesc::is_oop(obj), "must be a valid oop"); 2318 cl->do_object(obj); 2319 obj->oop_iterate(&oops); 2320 } 2321 assert(q->is_empty(), "should be empty"); 2322 } 2323 }; 2324 2325 ParallelObjectIteratorImpl* ShenandoahHeap::parallel_object_iterator(uint workers) { 2326 return new ShenandoahParallelObjectIterator(workers, &_aux_bit_map); 2327 } 2328 2329 // Keep alive an object that was loaded with AS_NO_KEEPALIVE. 2330 void ShenandoahHeap::keep_alive(oop obj) { 2331 if (is_concurrent_mark_in_progress() && (obj != nullptr)) { 2332 ShenandoahBarrierSet::barrier_set()->enqueue(obj); 2333 } 2334 } 2335 2336 void ShenandoahHeap::heap_region_iterate(ShenandoahHeapRegionClosure* blk) const { 2337 for (size_t i = 0; i < num_regions(); i++) { 2338 ShenandoahHeapRegion* current = get_region(i); 2339 blk->heap_region_do(current); 2340 } 2341 } 2342 2343 class ShenandoahParallelHeapRegionTask : public WorkerTask { 2344 private: 2345 ShenandoahHeap* const _heap; 2346 ShenandoahHeapRegionClosure* const _blk; 2347 2348 shenandoah_padding(0); 2349 volatile size_t _index; 2350 shenandoah_padding(1); 2351 2352 public: 2353 ShenandoahParallelHeapRegionTask(ShenandoahHeapRegionClosure* blk) : 2354 WorkerTask("Shenandoah Parallel Region Operation"), 2355 _heap(ShenandoahHeap::heap()), _blk(blk), _index(0) {} 2356 2357 void work(uint worker_id) { 2358 ShenandoahParallelWorkerSession worker_session(worker_id); 2359 size_t stride = ShenandoahParallelRegionStride; 2360 2361 size_t max = _heap->num_regions(); 2362 while (Atomic::load(&_index) < max) { 2363 size_t cur = Atomic::fetch_then_add(&_index, stride, memory_order_relaxed); 2364 size_t start = cur; 2365 size_t end = MIN2(cur + stride, max); 2366 if (start >= max) break; 2367 2368 for (size_t i = cur; i < end; i++) { 2369 ShenandoahHeapRegion* current = _heap->get_region(i); 2370 _blk->heap_region_do(current); 2371 } 2372 } 2373 } 2374 }; 2375 2376 void ShenandoahHeap::parallel_heap_region_iterate(ShenandoahHeapRegionClosure* blk) const { 2377 assert(blk->is_thread_safe(), "Only thread-safe closures here"); 2378 if (num_regions() > ShenandoahParallelRegionStride) { 2379 ShenandoahParallelHeapRegionTask task(blk); 2380 workers()->run_task(&task); 2381 } else { 2382 heap_region_iterate(blk); 2383 } 2384 } 2385 2386 class ShenandoahRendezvousClosure : public HandshakeClosure { 2387 public: 2388 inline ShenandoahRendezvousClosure() : HandshakeClosure("ShenandoahRendezvous") {} 2389 inline void do_thread(Thread* thread) {} 2390 }; 2391 2392 void ShenandoahHeap::rendezvous_threads() { 2393 ShenandoahRendezvousClosure cl; 2394 Handshake::execute(&cl); 2395 } 2396 2397 void ShenandoahHeap::recycle_trash() { 2398 free_set()->recycle_trash(); 2399 } 2400 2401 void ShenandoahHeap::do_class_unloading() { 2402 _unloader.unload(); 2403 } 2404 2405 void ShenandoahHeap::stw_weak_refs(bool full_gc) { 2406 // Weak refs processing 2407 ShenandoahPhaseTimings::Phase phase = full_gc ? ShenandoahPhaseTimings::full_gc_weakrefs 2408 : ShenandoahPhaseTimings::degen_gc_weakrefs; 2409 ShenandoahTimingsTracker t(phase); 2410 ShenandoahGCWorkerPhase worker_phase(phase); 2411 active_generation()->ref_processor()->process_references(phase, workers(), false /* concurrent */); 2412 } 2413 2414 void ShenandoahHeap::prepare_update_heap_references(bool concurrent) { 2415 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at safepoint"); 2416 2417 // Evacuation is over, no GCLABs are needed anymore. GCLABs are under URWM, so we need to 2418 // make them parsable for update code to work correctly. Plus, we can compute new sizes 2419 // for future GCLABs here. 2420 if (UseTLAB) { 2421 ShenandoahGCPhase phase(concurrent ? 2422 ShenandoahPhaseTimings::init_update_refs_manage_gclabs : 2423 ShenandoahPhaseTimings::degen_gc_init_update_refs_manage_gclabs); 2424 gclabs_retire(ResizeTLAB); 2425 } 2426 2427 _update_refs_iterator.reset(); 2428 } 2429 2430 void ShenandoahHeap::set_gc_state_all_threads(char state) { 2431 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) { 2432 ShenandoahThreadLocalData::set_gc_state(t, state); 2433 } 2434 } 2435 2436 void ShenandoahHeap::set_gc_state_mask(uint mask, bool value) { 2437 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Should really be Shenandoah safepoint"); 2438 _gc_state.set_cond(mask, value); 2439 set_gc_state_all_threads(_gc_state.raw_value()); 2440 } 2441 2442 void ShenandoahHeap::set_evacuation_reserve_quantities(bool is_valid) { 2443 _has_evacuation_reserve_quantities = is_valid; 2444 } 2445 2446 void ShenandoahHeap::set_concurrent_young_mark_in_progress(bool in_progress) { 2447 uint mask; 2448 assert(!has_forwarded_objects(), "Young marking is not concurrent with evacuation"); 2449 if (!in_progress && is_concurrent_old_mark_in_progress()) { 2450 assert(mode()->is_generational(), "Only generational GC has old marking"); 2451 assert(_gc_state.is_set(MARKING), "concurrent_old_marking_in_progress implies MARKING"); 2452 // If old-marking is in progress when we turn off YOUNG_MARKING, leave MARKING (and OLD_MARKING) on 2453 mask = YOUNG_MARKING; 2454 } else { 2455 mask = MARKING | YOUNG_MARKING; 2456 } 2457 set_gc_state_mask(mask, in_progress); 2458 manage_satb_barrier(in_progress); 2459 } 2460 2461 void ShenandoahHeap::set_concurrent_old_mark_in_progress(bool in_progress) { 2462 #ifdef ASSERT 2463 // has_forwarded_objects() iff UPDATEREFS or EVACUATION 2464 bool has_forwarded = has_forwarded_objects()? 1: 0; 2465 bool updating_or_evacuating = _gc_state.is_set(UPDATEREFS | EVACUATION)? 1: 0; 2466 assert (has_forwarded == updating_or_evacuating, "Has forwarded objects iff updating or evacuating"); 2467 #endif 2468 if (!in_progress && is_concurrent_young_mark_in_progress()) { 2469 // If young-marking is in progress when we turn off OLD_MARKING, leave MARKING (and YOUNG_MARKING) on 2470 assert(_gc_state.is_set(MARKING), "concurrent_young_marking_in_progress implies MARKING"); 2471 set_gc_state_mask(OLD_MARKING, in_progress); 2472 } else { 2473 set_gc_state_mask(MARKING | OLD_MARKING, in_progress); 2474 } 2475 manage_satb_barrier(in_progress); 2476 } 2477 2478 void ShenandoahHeap::set_prepare_for_old_mark_in_progress(bool in_progress) { 2479 // Unlike other set-gc-state functions, this may happen outside safepoint. 2480 // Is only set and queried by control thread, so no coherence issues. 2481 _prepare_for_old_mark = in_progress; 2482 } 2483 2484 void ShenandoahHeap::set_aging_cycle(bool in_progress) { 2485 _is_aging_cycle.set_cond(in_progress); 2486 } 2487 2488 void ShenandoahHeap::manage_satb_barrier(bool active) { 2489 if (is_concurrent_mark_in_progress()) { 2490 // Ignore request to deactivate barrier while concurrent mark is in progress. 2491 // Do not attempt to re-activate the barrier if it is already active. 2492 if (active && !ShenandoahBarrierSet::satb_mark_queue_set().is_active()) { 2493 ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active); 2494 } 2495 } else { 2496 // No concurrent marking is in progress so honor request to deactivate, 2497 // but only if the barrier is already active. 2498 if (!active && ShenandoahBarrierSet::satb_mark_queue_set().is_active()) { 2499 ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active); 2500 } 2501 } 2502 } 2503 2504 void ShenandoahHeap::set_evacuation_in_progress(bool in_progress) { 2505 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Only call this at safepoint"); 2506 set_gc_state_mask(EVACUATION, in_progress); 2507 } 2508 2509 void ShenandoahHeap::set_concurrent_strong_root_in_progress(bool in_progress) { 2510 if (in_progress) { 2511 _concurrent_strong_root_in_progress.set(); 2512 } else { 2513 _concurrent_strong_root_in_progress.unset(); 2514 } 2515 } 2516 2517 void ShenandoahHeap::set_concurrent_weak_root_in_progress(bool cond) { 2518 set_gc_state_mask(WEAK_ROOTS, cond); 2519 } 2520 2521 GCTracer* ShenandoahHeap::tracer() { 2522 return shenandoah_policy()->tracer(); 2523 } 2524 2525 size_t ShenandoahHeap::tlab_used(Thread* thread) const { 2526 return _free_set->used(); 2527 } 2528 2529 bool ShenandoahHeap::try_cancel_gc() { 2530 jbyte prev = _cancelled_gc.cmpxchg(CANCELLED, CANCELLABLE); 2531 return prev == CANCELLABLE; 2532 } 2533 2534 void ShenandoahHeap::cancel_concurrent_mark() { 2535 _young_generation->cancel_marking(); 2536 _old_generation->cancel_marking(); 2537 _global_generation->cancel_marking(); 2538 2539 ShenandoahBarrierSet::satb_mark_queue_set().abandon_partial_marking(); 2540 } 2541 2542 void ShenandoahHeap::cancel_gc(GCCause::Cause cause) { 2543 if (try_cancel_gc()) { 2544 FormatBuffer<> msg("Cancelling GC: %s", GCCause::to_string(cause)); 2545 log_info(gc)("%s", msg.buffer()); 2546 Events::log(Thread::current(), "%s", msg.buffer()); 2547 _cancel_requested_time = os::elapsedTime(); 2548 if (cause == GCCause::_shenandoah_upgrade_to_full_gc) { 2549 _upgraded_to_full = true; 2550 } 2551 } 2552 } 2553 2554 uint ShenandoahHeap::max_workers() { 2555 return _max_workers; 2556 } 2557 2558 void ShenandoahHeap::stop() { 2559 // The shutdown sequence should be able to terminate when GC is running. 2560 2561 // Step 1. Notify policy to disable event recording and prevent visiting gc threads during shutdown 2562 _shenandoah_policy->record_shutdown(); 2563 2564 // Step 2. Stop requesting collections. 2565 regulator_thread()->stop(); 2566 2567 // Step 3. Notify control thread that we are in shutdown. 2568 // Note that we cannot do that with stop(), because stop() is blocking and waits for the actual shutdown. 2569 // Doing stop() here would wait for the normal GC cycle to complete, never falling through to cancel below. 2570 control_thread()->prepare_for_graceful_shutdown(); 2571 2572 // Step 4. Notify GC workers that we are cancelling GC. 2573 cancel_gc(GCCause::_shenandoah_stop_vm); 2574 2575 // Step 5. Wait until GC worker exits normally. 2576 control_thread()->stop(); 2577 } 2578 2579 void ShenandoahHeap::stw_unload_classes(bool full_gc) { 2580 if (!unload_classes()) return; 2581 // Unload classes and purge SystemDictionary. 2582 { 2583 ShenandoahPhaseTimings::Phase phase = full_gc ? 2584 ShenandoahPhaseTimings::full_gc_purge_class_unload : 2585 ShenandoahPhaseTimings::degen_gc_purge_class_unload; 2586 ShenandoahIsAliveSelector is_alive; 2587 CodeCache::UnloadingScope scope(is_alive.is_alive_closure()); 2588 ShenandoahGCPhase gc_phase(phase); 2589 ShenandoahGCWorkerPhase worker_phase(phase); 2590 bool purged_class = SystemDictionary::do_unloading(gc_timer()); 2591 2592 uint num_workers = _workers->active_workers(); 2593 ShenandoahClassUnloadingTask unlink_task(phase, num_workers, purged_class); 2594 _workers->run_task(&unlink_task); 2595 } 2596 2597 { 2598 ShenandoahGCPhase phase(full_gc ? 2599 ShenandoahPhaseTimings::full_gc_purge_cldg : 2600 ShenandoahPhaseTimings::degen_gc_purge_cldg); 2601 ClassLoaderDataGraph::purge(/*at_safepoint*/true); 2602 } 2603 // Resize and verify metaspace 2604 MetaspaceGC::compute_new_size(); 2605 DEBUG_ONLY(MetaspaceUtils::verify();) 2606 } 2607 2608 // Weak roots are either pre-evacuated (final mark) or updated (final updaterefs), 2609 // so they should not have forwarded oops. 2610 // However, we do need to "null" dead oops in the roots, if can not be done 2611 // in concurrent cycles. 2612 void ShenandoahHeap::stw_process_weak_roots(bool full_gc) { 2613 uint num_workers = _workers->active_workers(); 2614 ShenandoahPhaseTimings::Phase timing_phase = full_gc ? 2615 ShenandoahPhaseTimings::full_gc_purge_weak_par : 2616 ShenandoahPhaseTimings::degen_gc_purge_weak_par; 2617 ShenandoahGCPhase phase(timing_phase); 2618 ShenandoahGCWorkerPhase worker_phase(timing_phase); 2619 // Cleanup weak roots 2620 if (has_forwarded_objects()) { 2621 ShenandoahForwardedIsAliveClosure is_alive; 2622 ShenandoahUpdateRefsClosure keep_alive; 2623 ShenandoahParallelWeakRootsCleaningTask<ShenandoahForwardedIsAliveClosure, ShenandoahUpdateRefsClosure> 2624 cleaning_task(timing_phase, &is_alive, &keep_alive, num_workers); 2625 _workers->run_task(&cleaning_task); 2626 } else { 2627 ShenandoahIsAliveClosure is_alive; 2628 #ifdef ASSERT 2629 ShenandoahAssertNotForwardedClosure verify_cl; 2630 ShenandoahParallelWeakRootsCleaningTask<ShenandoahIsAliveClosure, ShenandoahAssertNotForwardedClosure> 2631 cleaning_task(timing_phase, &is_alive, &verify_cl, num_workers); 2632 #else 2633 ShenandoahParallelWeakRootsCleaningTask<ShenandoahIsAliveClosure, DoNothingClosure> 2634 cleaning_task(timing_phase, &is_alive, &do_nothing_cl, num_workers); 2635 #endif 2636 _workers->run_task(&cleaning_task); 2637 } 2638 } 2639 2640 void ShenandoahHeap::parallel_cleaning(bool full_gc) { 2641 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint"); 2642 assert(is_stw_gc_in_progress(), "Only for Degenerated and Full GC"); 2643 ShenandoahGCPhase phase(full_gc ? 2644 ShenandoahPhaseTimings::full_gc_purge : 2645 ShenandoahPhaseTimings::degen_gc_purge); 2646 stw_weak_refs(full_gc); 2647 stw_process_weak_roots(full_gc); 2648 stw_unload_classes(full_gc); 2649 } 2650 2651 void ShenandoahHeap::set_has_forwarded_objects(bool cond) { 2652 set_gc_state_mask(HAS_FORWARDED, cond); 2653 } 2654 2655 void ShenandoahHeap::set_unload_classes(bool uc) { 2656 _unload_classes.set_cond(uc); 2657 } 2658 2659 bool ShenandoahHeap::unload_classes() const { 2660 return _unload_classes.is_set(); 2661 } 2662 2663 address ShenandoahHeap::in_cset_fast_test_addr() { 2664 ShenandoahHeap* heap = ShenandoahHeap::heap(); 2665 assert(heap->collection_set() != nullptr, "Sanity"); 2666 return (address) heap->collection_set()->biased_map_address(); 2667 } 2668 2669 address ShenandoahHeap::gc_state_addr() { 2670 return (address) ShenandoahHeap::heap()->_gc_state.addr_of(); 2671 } 2672 2673 void ShenandoahHeap::reset_bytes_allocated_since_gc_start() { 2674 if (mode()->is_generational()) { 2675 young_generation()->reset_bytes_allocated_since_gc_start(); 2676 old_generation()->reset_bytes_allocated_since_gc_start(); 2677 } 2678 2679 global_generation()->reset_bytes_allocated_since_gc_start(); 2680 } 2681 2682 void ShenandoahHeap::set_degenerated_gc_in_progress(bool in_progress) { 2683 _degenerated_gc_in_progress.set_cond(in_progress); 2684 } 2685 2686 void ShenandoahHeap::set_full_gc_in_progress(bool in_progress) { 2687 _full_gc_in_progress.set_cond(in_progress); 2688 } 2689 2690 void ShenandoahHeap::set_full_gc_move_in_progress(bool in_progress) { 2691 assert (is_full_gc_in_progress(), "should be"); 2692 _full_gc_move_in_progress.set_cond(in_progress); 2693 } 2694 2695 void ShenandoahHeap::set_update_refs_in_progress(bool in_progress) { 2696 set_gc_state_mask(UPDATEREFS, in_progress); 2697 } 2698 2699 void ShenandoahHeap::register_nmethod(nmethod* nm) { 2700 ShenandoahCodeRoots::register_nmethod(nm); 2701 } 2702 2703 void ShenandoahHeap::unregister_nmethod(nmethod* nm) { 2704 ShenandoahCodeRoots::unregister_nmethod(nm); 2705 } 2706 2707 void ShenandoahHeap::pin_object(JavaThread* thr, oop o) { 2708 heap_region_containing(o)->record_pin(); 2709 } 2710 2711 void ShenandoahHeap::unpin_object(JavaThread* thr, oop o) { 2712 ShenandoahHeapRegion* r = heap_region_containing(o); 2713 assert(r != nullptr, "Sanity"); 2714 assert(r->pin_count() > 0, "Region " SIZE_FORMAT " should have non-zero pins", r->index()); 2715 r->record_unpin(); 2716 } 2717 2718 void ShenandoahHeap::sync_pinned_region_status() { 2719 ShenandoahHeapLocker locker(lock()); 2720 2721 for (size_t i = 0; i < num_regions(); i++) { 2722 ShenandoahHeapRegion *r = get_region(i); 2723 if (r->is_active()) { 2724 if (r->is_pinned()) { 2725 if (r->pin_count() == 0) { 2726 r->make_unpinned(); 2727 } 2728 } else { 2729 if (r->pin_count() > 0) { 2730 r->make_pinned(); 2731 } 2732 } 2733 } 2734 } 2735 2736 assert_pinned_region_status(); 2737 } 2738 2739 #ifdef ASSERT 2740 void ShenandoahHeap::assert_pinned_region_status() { 2741 for (size_t i = 0; i < num_regions(); i++) { 2742 ShenandoahHeapRegion* r = get_region(i); 2743 if (active_generation()->contains(r)) { 2744 assert((r->is_pinned() && r->pin_count() > 0) || (!r->is_pinned() && r->pin_count() == 0), 2745 "Region " SIZE_FORMAT " pinning status is inconsistent", i); 2746 } 2747 } 2748 } 2749 #endif 2750 2751 ConcurrentGCTimer* ShenandoahHeap::gc_timer() const { 2752 return _gc_timer; 2753 } 2754 2755 void ShenandoahHeap::prepare_concurrent_roots() { 2756 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint"); 2757 assert(!is_stw_gc_in_progress(), "Only concurrent GC"); 2758 set_concurrent_strong_root_in_progress(!collection_set()->is_empty()); 2759 set_concurrent_weak_root_in_progress(true); 2760 if (unload_classes()) { 2761 _unloader.prepare(); 2762 } 2763 } 2764 2765 void ShenandoahHeap::finish_concurrent_roots() { 2766 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint"); 2767 assert(!is_stw_gc_in_progress(), "Only concurrent GC"); 2768 if (unload_classes()) { 2769 _unloader.finish(); 2770 } 2771 } 2772 2773 #ifdef ASSERT 2774 void ShenandoahHeap::assert_gc_workers(uint nworkers) { 2775 assert(nworkers > 0 && nworkers <= max_workers(), "Sanity"); 2776 2777 if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) { 2778 if (UseDynamicNumberOfGCThreads) { 2779 assert(nworkers <= ParallelGCThreads, "Cannot use more than it has"); 2780 } else { 2781 // Use ParallelGCThreads inside safepoints 2782 assert(nworkers == ParallelGCThreads, "Use ParallelGCThreads within safepoints"); 2783 } 2784 } else { 2785 if (UseDynamicNumberOfGCThreads) { 2786 assert(nworkers <= ConcGCThreads, "Cannot use more than it has"); 2787 } else { 2788 // Use ConcGCThreads outside safepoints 2789 assert(nworkers == ConcGCThreads, "Use ConcGCThreads outside safepoints"); 2790 } 2791 } 2792 } 2793 #endif 2794 2795 ShenandoahVerifier* ShenandoahHeap::verifier() { 2796 guarantee(ShenandoahVerify, "Should be enabled"); 2797 assert (_verifier != nullptr, "sanity"); 2798 return _verifier; 2799 } 2800 2801 template<bool CONCURRENT> 2802 class ShenandoahUpdateHeapRefsTask : public WorkerTask { 2803 private: 2804 ShenandoahHeap* _heap; 2805 ShenandoahRegionIterator* _regions; 2806 ShenandoahRegionChunkIterator* _work_chunks; 2807 2808 public: 2809 explicit ShenandoahUpdateHeapRefsTask(ShenandoahRegionIterator* regions, 2810 ShenandoahRegionChunkIterator* work_chunks) : 2811 WorkerTask("Shenandoah Update References"), 2812 _heap(ShenandoahHeap::heap()), 2813 _regions(regions), 2814 _work_chunks(work_chunks) 2815 { 2816 } 2817 2818 void work(uint worker_id) { 2819 if (CONCURRENT) { 2820 ShenandoahConcurrentWorkerSession worker_session(worker_id); 2821 ShenandoahSuspendibleThreadSetJoiner stsj(ShenandoahSuspendibleWorkers); 2822 do_work<ShenandoahConcUpdateRefsClosure>(worker_id); 2823 } else { 2824 ShenandoahParallelWorkerSession worker_session(worker_id); 2825 do_work<ShenandoahSTWUpdateRefsClosure>(worker_id); 2826 } 2827 } 2828 2829 private: 2830 template<class T> 2831 void do_work(uint worker_id) { 2832 T cl; 2833 if (CONCURRENT && (worker_id == 0)) { 2834 // We ask the first worker to replenish the Mutator free set by moving regions previously reserved to hold the 2835 // results of evacuation. These reserves are no longer necessary because evacuation has completed. 2836 size_t cset_regions = _heap->collection_set()->count(); 2837 // We cannot transfer any more regions than will be reclaimed when the existing collection set is recycled, because 2838 // we need the reclaimed collection set regions to replenish the collector reserves 2839 _heap->free_set()->move_collector_sets_to_mutator(cset_regions); 2840 } 2841 // If !CONCURRENT, there's no value in expanding Mutator free set 2842 2843 ShenandoahHeapRegion* r = _regions->next(); 2844 // We update references for global, old, and young collections. 2845 assert(_heap->active_generation()->is_mark_complete(), "Expected complete marking"); 2846 ShenandoahMarkingContext* const ctx = _heap->marking_context(); 2847 bool is_mixed = _heap->collection_set()->has_old_regions(); 2848 while (r != nullptr) { 2849 HeapWord* update_watermark = r->get_update_watermark(); 2850 assert (update_watermark >= r->bottom(), "sanity"); 2851 2852 log_debug(gc)("ShenandoahUpdateHeapRefsTask::do_work(%u) looking at region " SIZE_FORMAT, worker_id, r->index()); 2853 bool region_progress = false; 2854 if (r->is_active() && !r->is_cset()) { 2855 if (!_heap->mode()->is_generational() || r->is_young()) { 2856 _heap->marked_object_oop_iterate(r, &cl, update_watermark); 2857 region_progress = true; 2858 } else if (r->is_old()) { 2859 if (_heap->active_generation()->is_global()) { 2860 // Note that GLOBAL collection is not as effectively balanced as young and mixed cycles. This is because 2861 // concurrent GC threads are parceled out entire heap regions of work at a time and there 2862 // is no "catchup phase" consisting of remembered set scanning, during which parcels of work are smaller 2863 // and more easily distributed more fairly across threads. 2864 2865 // TODO: Consider an improvement to load balance GLOBAL GC. 2866 _heap->marked_object_oop_iterate(r, &cl, update_watermark); 2867 region_progress = true; 2868 } 2869 // Otherwise, this is an old region in a young or mixed cycle. Process it during a second phase, below. 2870 // Don't bother to report pacing progress in this case. 2871 } else { 2872 // Because updating of references runs concurrently, it is possible that a FREE inactive region transitions 2873 // to a non-free active region while this loop is executing. Whenever this happens, the changing of a region's 2874 // active status may propagate at a different speed than the changing of the region's affiliation. 2875 2876 // When we reach this control point, it is because a race has allowed a region's is_active() status to be seen 2877 // by this thread before the region's affiliation() is seen by this thread. 2878 2879 // It's ok for this race to occur because the newly transformed region does not have any references to be 2880 // updated. 2881 2882 assert(r->get_update_watermark() == r->bottom(), 2883 "%s Region " SIZE_FORMAT " is_active but not recognized as YOUNG or OLD so must be newly transitioned from FREE", 2884 r->affiliation_name(), r->index()); 2885 } 2886 } 2887 if (region_progress && ShenandoahPacing) { 2888 _heap->pacer()->report_updaterefs(pointer_delta(update_watermark, r->bottom())); 2889 } 2890 if (_heap->check_cancelled_gc_and_yield(CONCURRENT)) { 2891 return; 2892 } 2893 r = _regions->next(); 2894 } 2895 2896 if (_heap->mode()->is_generational() && !_heap->active_generation()->is_global()) { 2897 // Since this is generational and not GLOBAL, we have to process the remembered set. There's no remembered 2898 // set processing if not in generational mode or if GLOBAL mode. 2899 2900 // After this thread has exhausted its traditional update-refs work, it continues with updating refs within remembered set. 2901 // The remembered set workload is better balanced between threads, so threads that are "behind" can catch up with other 2902 // threads during this phase, allowing all threads to work more effectively in parallel. 2903 struct ShenandoahRegionChunk assignment; 2904 RememberedScanner* scanner = _heap->card_scan(); 2905 2906 while (!_heap->check_cancelled_gc_and_yield(CONCURRENT) && _work_chunks->next(&assignment)) { 2907 // Keep grabbing next work chunk to process until finished, or asked to yield 2908 ShenandoahHeapRegion* r = assignment._r; 2909 if (r->is_active() && !r->is_cset() && r->is_old()) { 2910 HeapWord* start_of_range = r->bottom() + assignment._chunk_offset; 2911 HeapWord* end_of_range = r->get_update_watermark(); 2912 if (end_of_range > start_of_range + assignment._chunk_size) { 2913 end_of_range = start_of_range + assignment._chunk_size; 2914 } 2915 2916 // Old region in a young cycle or mixed cycle. 2917 if (is_mixed) { 2918 // TODO: For mixed evac, consider building an old-gen remembered set that allows restricted updating 2919 // within old-gen HeapRegions. This remembered set can be constructed by old-gen concurrent marking 2920 // and augmented by card marking. For example, old-gen concurrent marking can remember for each old-gen 2921 // card which other old-gen regions it refers to: none, one-other specifically, multiple-other non-specific. 2922 // Update-references when _mixed_evac processess each old-gen memory range that has a traditional DIRTY 2923 // card or if the "old-gen remembered set" indicates that this card holds pointers specifically to an 2924 // old-gen region in the most recent collection set, or if this card holds pointers to other non-specific 2925 // old-gen heap regions. 2926 2927 if (r->is_humongous()) { 2928 if (start_of_range < end_of_range) { 2929 // Need to examine both dirty and clean cards during mixed evac. 2930 r->oop_iterate_humongous_slice(&cl, false, start_of_range, assignment._chunk_size, true); 2931 } 2932 } else { 2933 // Since this is mixed evacuation, old regions that are candidates for collection have not been coalesced 2934 // and filled. Use mark bits to find objects that need to be updated. 2935 // 2936 // Future TODO: establish a second remembered set to identify which old-gen regions point to other old-gen 2937 // regions which are in the collection set for a particular mixed evacuation. 2938 if (start_of_range < end_of_range) { 2939 HeapWord* p = nullptr; 2940 size_t card_index = scanner->card_index_for_addr(start_of_range); 2941 // In case last object in my range spans boundary of my chunk, I may need to scan all the way to top() 2942 ShenandoahObjectToOopBoundedClosure<T> objs(&cl, start_of_range, r->top()); 2943 2944 // Any object that begins in a previous range is part of a different scanning assignment. Any object that 2945 // starts after end_of_range is also not my responsibility. (Either allocated during evacuation, so does 2946 // not hold pointers to from-space, or is beyond the range of my assigned work chunk.) 2947 2948 // Find the first object that begins in my range, if there is one. 2949 p = start_of_range; 2950 oop obj = cast_to_oop(p); 2951 HeapWord* tams = ctx->top_at_mark_start(r); 2952 if (p >= tams) { 2953 // We cannot use ctx->is_marked(obj) to test whether an object begins at this address. Instead, 2954 // we need to use the remembered set crossing map to advance p to the first object that starts 2955 // within the enclosing card. 2956 2957 while (true) { 2958 HeapWord* first_object = scanner->first_object_in_card(card_index); 2959 if (first_object != nullptr) { 2960 p = first_object; 2961 break; 2962 } else if (scanner->addr_for_card_index(card_index + 1) < end_of_range) { 2963 card_index++; 2964 } else { 2965 // Force the loop that follows to immediately terminate. 2966 p = end_of_range; 2967 break; 2968 } 2969 } 2970 obj = cast_to_oop(p); 2971 // Note: p may be >= end_of_range 2972 } else if (!ctx->is_marked(obj)) { 2973 p = ctx->get_next_marked_addr(p, tams); 2974 obj = cast_to_oop(p); 2975 // If there are no more marked objects before tams, this returns tams. 2976 // Note that tams is either >= end_of_range, or tams is the start of an object that is marked. 2977 } 2978 while (p < end_of_range) { 2979 // p is known to point to the beginning of marked object obj 2980 objs.do_object(obj); 2981 HeapWord* prev_p = p; 2982 p += obj->size(); 2983 if (p < tams) { 2984 p = ctx->get_next_marked_addr(p, tams); 2985 // If there are no more marked objects before tams, this returns tams. Note that tams is 2986 // either >= end_of_range, or tams is the start of an object that is marked. 2987 } 2988 assert(p != prev_p, "Lack of forward progress"); 2989 obj = cast_to_oop(p); 2990 } 2991 } 2992 } 2993 } else { 2994 // This is a young evac.. 2995 if (start_of_range < end_of_range) { 2996 size_t cluster_size = 2997 CardTable::card_size_in_words() * ShenandoahCardCluster<ShenandoahDirectCardMarkRememberedSet>::CardsPerCluster; 2998 size_t clusters = assignment._chunk_size / cluster_size; 2999 assert(clusters * cluster_size == assignment._chunk_size, "Chunk assignment must align on cluster boundaries"); 3000 scanner->process_region_slice(r, assignment._chunk_offset, clusters, end_of_range, &cl, true, worker_id); 3001 } 3002 } 3003 if (ShenandoahPacing && (start_of_range < end_of_range)) { 3004 _heap->pacer()->report_updaterefs(pointer_delta(end_of_range, start_of_range)); 3005 } 3006 } 3007 } 3008 } 3009 } 3010 }; 3011 3012 void ShenandoahHeap::update_heap_references(bool concurrent) { 3013 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC"); 3014 uint nworkers = workers()->active_workers(); 3015 ShenandoahRegionChunkIterator work_list(nworkers); 3016 3017 if (concurrent) { 3018 ShenandoahUpdateHeapRefsTask<true> task(&_update_refs_iterator, &work_list); 3019 workers()->run_task(&task); 3020 } else { 3021 ShenandoahUpdateHeapRefsTask<false> task(&_update_refs_iterator, &work_list); 3022 workers()->run_task(&task); 3023 } 3024 if (ShenandoahEnableCardStats && card_scan()!=nullptr) { // generational check proxy 3025 card_scan()->log_card_stats(nworkers, CARD_STAT_UPDATE_REFS); 3026 } 3027 } 3028 3029 class ShenandoahFinalUpdateRefsUpdateRegionStateClosure : public ShenandoahHeapRegionClosure { 3030 private: 3031 ShenandoahMarkingContext* _ctx; 3032 ShenandoahHeapLock* const _lock; 3033 bool _is_generational; 3034 3035 public: 3036 ShenandoahFinalUpdateRefsUpdateRegionStateClosure( 3037 ShenandoahMarkingContext* ctx) : _ctx(ctx), _lock(ShenandoahHeap::heap()->lock()), 3038 _is_generational(ShenandoahHeap::heap()->mode()->is_generational()) { } 3039 3040 void heap_region_do(ShenandoahHeapRegion* r) { 3041 3042 // Maintenance of region age must follow evacuation in order to account for evacuation allocations within survivor 3043 // regions. We consult region age during the subsequent evacuation to determine whether certain objects need to 3044 // be promoted. 3045 if (_is_generational && r->is_young() && r->is_active()) { 3046 HeapWord *tams = _ctx->top_at_mark_start(r); 3047 HeapWord *top = r->top(); 3048 3049 // Allocations move the watermark when top moves. However compacting 3050 // objects will sometimes lower top beneath the watermark, after which, 3051 // attempts to read the watermark will assert out (watermark should not be 3052 // higher than top). 3053 if (top > tams) { 3054 // There have been allocations in this region since the start of the cycle. 3055 // Any objects new to this region must not assimilate elevated age. 3056 r->reset_age(); 3057 } else if (ShenandoahHeap::heap()->is_aging_cycle()) { 3058 r->increment_age(); 3059 } 3060 } 3061 3062 // Drop unnecessary "pinned" state from regions that does not have CP marks 3063 // anymore, as this would allow trashing them. 3064 if (r->is_active()) { 3065 if (r->is_pinned()) { 3066 if (r->pin_count() == 0) { 3067 ShenandoahHeapLocker locker(_lock); 3068 r->make_unpinned(); 3069 } 3070 } else { 3071 if (r->pin_count() > 0) { 3072 ShenandoahHeapLocker locker(_lock); 3073 r->make_pinned(); 3074 } 3075 } 3076 } 3077 } 3078 3079 bool is_thread_safe() { return true; } 3080 }; 3081 3082 void ShenandoahHeap::update_heap_region_states(bool concurrent) { 3083 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint"); 3084 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC"); 3085 3086 { 3087 ShenandoahGCPhase phase(concurrent ? 3088 ShenandoahPhaseTimings::final_update_refs_update_region_states : 3089 ShenandoahPhaseTimings::degen_gc_final_update_refs_update_region_states); 3090 ShenandoahFinalUpdateRefsUpdateRegionStateClosure cl (active_generation()->complete_marking_context()); 3091 parallel_heap_region_iterate(&cl); 3092 3093 assert_pinned_region_status(); 3094 } 3095 3096 { 3097 ShenandoahGCPhase phase(concurrent ? 3098 ShenandoahPhaseTimings::final_update_refs_trash_cset : 3099 ShenandoahPhaseTimings::degen_gc_final_update_refs_trash_cset); 3100 trash_cset_regions(); 3101 } 3102 } 3103 3104 void ShenandoahHeap::rebuild_free_set(bool concurrent) { 3105 ShenandoahGCPhase phase(concurrent ? 3106 ShenandoahPhaseTimings::final_update_refs_rebuild_freeset : 3107 ShenandoahPhaseTimings::degen_gc_final_update_refs_rebuild_freeset); 3108 size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes(); 3109 ShenandoahHeapLocker locker(lock()); 3110 size_t young_cset_regions, old_cset_regions; 3111 _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions); 3112 3113 if (mode()->is_generational()) { 3114 assert(verify_generation_usage(true, old_generation()->used_regions(), 3115 old_generation()->used(), old_generation()->get_humongous_waste(), 3116 true, young_generation()->used_regions(), 3117 young_generation()->used(), young_generation()->get_humongous_waste()), 3118 "Generation accounts are inaccurate"); 3119 3120 // The computation of bytes_of_allocation_runway_before_gc_trigger is quite conservative so consider all of this 3121 // available for transfer to old. Note that transfer of humongous regions does not impact available. 3122 size_t allocation_runway = young_heuristics()->bytes_of_allocation_runway_before_gc_trigger(young_cset_regions); 3123 adjust_generation_sizes_for_next_cycle(allocation_runway, young_cset_regions, old_cset_regions); 3124 3125 // Total old_available may have been expanded to hold anticipated promotions. We trigger if the fragmented available 3126 // memory represents more than 16 regions worth of data. Note that fragmentation may increase when we promote regular 3127 // regions in place when many of these regular regions have an abundant amount of available memory within them. Fragmentation 3128 // will decrease as promote-by-copy consumes the available memory within these partially consumed regions. 3129 // 3130 // We consider old-gen to have excessive fragmentation if more than 12.5% of old-gen is free memory that resides 3131 // within partially consumed regions of memory. 3132 } 3133 // Rebuild free set based on adjusted generation sizes. 3134 _free_set->rebuild(young_cset_regions, old_cset_regions); 3135 3136 if (mode()->is_generational()) { 3137 size_t old_available = old_generation()->available(); 3138 size_t old_unaffiliated_available = old_generation()->free_unaffiliated_regions() * region_size_bytes; 3139 size_t old_fragmented_available; 3140 assert(old_available >= old_unaffiliated_available, "unaffiliated available is a subset of total available"); 3141 old_fragmented_available = old_available - old_unaffiliated_available; 3142 3143 size_t old_capacity = old_generation()->max_capacity(); 3144 size_t heap_capacity = capacity(); 3145 if ((old_capacity > heap_capacity / 8) && (old_fragmented_available > old_capacity / 8)) { 3146 old_heuristics()->trigger_old_is_fragmented(); 3147 } 3148 3149 size_t old_used = old_generation()->used() + old_generation()->get_humongous_waste(); 3150 size_t trigger_threshold = old_generation()->usage_trigger_threshold(); 3151 // Detects unsigned arithmetic underflow 3152 assert(old_used <= capacity(), 3153 "Old used (" SIZE_FORMAT ", " SIZE_FORMAT") must not be more than heap capacity (" SIZE_FORMAT ")", 3154 old_generation()->used(), old_generation()->get_humongous_waste(), capacity()); 3155 3156 if (old_used > trigger_threshold) { 3157 old_heuristics()->trigger_old_has_grown(); 3158 } 3159 } 3160 } 3161 3162 void ShenandoahHeap::print_extended_on(outputStream *st) const { 3163 print_on(st); 3164 st->cr(); 3165 print_heap_regions_on(st); 3166 } 3167 3168 bool ShenandoahHeap::is_bitmap_slice_committed(ShenandoahHeapRegion* r, bool skip_self) { 3169 size_t slice = r->index() / _bitmap_regions_per_slice; 3170 3171 size_t regions_from = _bitmap_regions_per_slice * slice; 3172 size_t regions_to = MIN2(num_regions(), _bitmap_regions_per_slice * (slice + 1)); 3173 for (size_t g = regions_from; g < regions_to; g++) { 3174 assert (g / _bitmap_regions_per_slice == slice, "same slice"); 3175 if (skip_self && g == r->index()) continue; 3176 if (get_region(g)->is_committed()) { 3177 return true; 3178 } 3179 } 3180 return false; 3181 } 3182 3183 bool ShenandoahHeap::commit_bitmap_slice(ShenandoahHeapRegion* r) { 3184 shenandoah_assert_heaplocked(); 3185 3186 // Bitmaps in special regions do not need commits 3187 if (_bitmap_region_special) { 3188 return true; 3189 } 3190 3191 if (is_bitmap_slice_committed(r, true)) { 3192 // Some other region from the group is already committed, meaning the bitmap 3193 // slice is already committed, we exit right away. 3194 return true; 3195 } 3196 3197 // Commit the bitmap slice: 3198 size_t slice = r->index() / _bitmap_regions_per_slice; 3199 size_t off = _bitmap_bytes_per_slice * slice; 3200 size_t len = _bitmap_bytes_per_slice; 3201 char* start = (char*) _bitmap_region.start() + off; 3202 3203 if (!os::commit_memory(start, len, false)) { 3204 return false; 3205 } 3206 3207 if (AlwaysPreTouch) { 3208 os::pretouch_memory(start, start + len, _pretouch_bitmap_page_size); 3209 } 3210 3211 return true; 3212 } 3213 3214 bool ShenandoahHeap::uncommit_bitmap_slice(ShenandoahHeapRegion *r) { 3215 shenandoah_assert_heaplocked(); 3216 3217 // Bitmaps in special regions do not need uncommits 3218 if (_bitmap_region_special) { 3219 return true; 3220 } 3221 3222 if (is_bitmap_slice_committed(r, true)) { 3223 // Some other region from the group is still committed, meaning the bitmap 3224 // slice is should stay committed, exit right away. 3225 return true; 3226 } 3227 3228 // Uncommit the bitmap slice: 3229 size_t slice = r->index() / _bitmap_regions_per_slice; 3230 size_t off = _bitmap_bytes_per_slice * slice; 3231 size_t len = _bitmap_bytes_per_slice; 3232 if (!os::uncommit_memory((char*)_bitmap_region.start() + off, len)) { 3233 return false; 3234 } 3235 return true; 3236 } 3237 3238 void ShenandoahHeap::safepoint_synchronize_begin() { 3239 if (ShenandoahSuspendibleWorkers) { 3240 SuspendibleThreadSet::synchronize(); 3241 } 3242 } 3243 3244 void ShenandoahHeap::safepoint_synchronize_end() { 3245 if (ShenandoahSuspendibleWorkers) { 3246 SuspendibleThreadSet::desynchronize(); 3247 } 3248 } 3249 3250 void ShenandoahHeap::entry_uncommit(double shrink_before, size_t shrink_until) { 3251 static const char *msg = "Concurrent uncommit"; 3252 ShenandoahConcurrentPhase gc_phase(msg, ShenandoahPhaseTimings::conc_uncommit, true /* log_heap_usage */); 3253 EventMark em("%s", msg); 3254 3255 op_uncommit(shrink_before, shrink_until); 3256 } 3257 3258 void ShenandoahHeap::try_inject_alloc_failure() { 3259 if (ShenandoahAllocFailureALot && !cancelled_gc() && ((os::random() % 1000) > 950)) { 3260 _inject_alloc_failure.set(); 3261 os::naked_short_sleep(1); 3262 if (cancelled_gc()) { 3263 log_info(gc)("Allocation failure was successfully injected"); 3264 } 3265 } 3266 } 3267 3268 bool ShenandoahHeap::should_inject_alloc_failure() { 3269 return _inject_alloc_failure.is_set() && _inject_alloc_failure.try_unset(); 3270 } 3271 3272 void ShenandoahHeap::initialize_serviceability() { 3273 if (mode()->is_generational()) { 3274 _young_gen_memory_pool = new ShenandoahYoungGenMemoryPool(this); 3275 _old_gen_memory_pool = new ShenandoahOldGenMemoryPool(this); 3276 _cycle_memory_manager.add_pool(_young_gen_memory_pool); 3277 _cycle_memory_manager.add_pool(_old_gen_memory_pool); 3278 _stw_memory_manager.add_pool(_young_gen_memory_pool); 3279 _stw_memory_manager.add_pool(_old_gen_memory_pool); 3280 } else { 3281 _memory_pool = new ShenandoahMemoryPool(this); 3282 _cycle_memory_manager.add_pool(_memory_pool); 3283 _stw_memory_manager.add_pool(_memory_pool); 3284 } 3285 } 3286 3287 GrowableArray<GCMemoryManager*> ShenandoahHeap::memory_managers() { 3288 GrowableArray<GCMemoryManager*> memory_managers(2); 3289 memory_managers.append(&_cycle_memory_manager); 3290 memory_managers.append(&_stw_memory_manager); 3291 return memory_managers; 3292 } 3293 3294 GrowableArray<MemoryPool*> ShenandoahHeap::memory_pools() { 3295 GrowableArray<MemoryPool*> memory_pools(1); 3296 if (mode()->is_generational()) { 3297 memory_pools.append(_young_gen_memory_pool); 3298 memory_pools.append(_old_gen_memory_pool); 3299 } else { 3300 memory_pools.append(_memory_pool); 3301 } 3302 return memory_pools; 3303 } 3304 3305 MemoryUsage ShenandoahHeap::memory_usage() { 3306 return MemoryUsage(_initial_size, used(), committed(), max_capacity()); 3307 } 3308 3309 ShenandoahRegionIterator::ShenandoahRegionIterator() : 3310 _heap(ShenandoahHeap::heap()), 3311 _index(0) {} 3312 3313 ShenandoahRegionIterator::ShenandoahRegionIterator(ShenandoahHeap* heap) : 3314 _heap(heap), 3315 _index(0) {} 3316 3317 void ShenandoahRegionIterator::reset() { 3318 _index = 0; 3319 } 3320 3321 bool ShenandoahRegionIterator::has_next() const { 3322 return _index < _heap->num_regions(); 3323 } 3324 3325 char ShenandoahHeap::gc_state() const { 3326 return _gc_state.raw_value(); 3327 } 3328 3329 ShenandoahLiveData* ShenandoahHeap::get_liveness_cache(uint worker_id) { 3330 #ifdef ASSERT 3331 assert(_liveness_cache != nullptr, "sanity"); 3332 assert(worker_id < _max_workers, "sanity"); 3333 for (uint i = 0; i < num_regions(); i++) { 3334 assert(_liveness_cache[worker_id][i] == 0, "liveness cache should be empty"); 3335 } 3336 #endif 3337 return _liveness_cache[worker_id]; 3338 } 3339 3340 void ShenandoahHeap::flush_liveness_cache(uint worker_id) { 3341 assert(worker_id < _max_workers, "sanity"); 3342 assert(_liveness_cache != nullptr, "sanity"); 3343 ShenandoahLiveData* ld = _liveness_cache[worker_id]; 3344 3345 for (uint i = 0; i < num_regions(); i++) { 3346 ShenandoahLiveData live = ld[i]; 3347 if (live > 0) { 3348 ShenandoahHeapRegion* r = get_region(i); 3349 r->increase_live_data_gc_words(live); 3350 ld[i] = 0; 3351 } 3352 } 3353 } 3354 3355 bool ShenandoahHeap::requires_barriers(stackChunkOop obj) const { 3356 if (is_idle()) return false; 3357 3358 // Objects allocated after marking start are implicitly alive, don't need any barriers during 3359 // marking phase. 3360 if (is_concurrent_mark_in_progress() && 3361 !marking_context()->allocated_after_mark_start(obj)) { 3362 return true; 3363 } 3364 3365 // Can not guarantee obj is deeply good. 3366 if (has_forwarded_objects()) { 3367 return true; 3368 } 3369 3370 return false; 3371 } 3372 3373 void ShenandoahHeap::transfer_old_pointers_from_satb() { 3374 _old_generation->transfer_pointers_from_satb(); 3375 } 3376 3377 template<> 3378 void ShenandoahGenerationRegionClosure<YOUNG>::heap_region_do(ShenandoahHeapRegion* region) { 3379 // Visit young and free regions 3380 if (!region->is_old()) { 3381 _cl->heap_region_do(region); 3382 } 3383 } 3384 3385 template<> 3386 void ShenandoahGenerationRegionClosure<OLD>::heap_region_do(ShenandoahHeapRegion* region) { 3387 // Visit old and free regions 3388 if (!region->is_young()) { 3389 _cl->heap_region_do(region); 3390 } 3391 } 3392 3393 template<> 3394 void ShenandoahGenerationRegionClosure<GLOBAL_GEN>::heap_region_do(ShenandoahHeapRegion* region) { 3395 _cl->heap_region_do(region); 3396 } 3397 3398 template<> 3399 void ShenandoahGenerationRegionClosure<GLOBAL_NON_GEN>::heap_region_do(ShenandoahHeapRegion* region) { 3400 _cl->heap_region_do(region); 3401 } 3402 3403 bool ShenandoahHeap::verify_generation_usage(bool verify_old, size_t old_regions, size_t old_bytes, size_t old_waste, 3404 bool verify_young, size_t young_regions, size_t young_bytes, size_t young_waste) { 3405 size_t tally_old_regions = 0; 3406 size_t tally_old_bytes = 0; 3407 size_t tally_old_waste = 0; 3408 size_t tally_young_regions = 0; 3409 size_t tally_young_bytes = 0; 3410 size_t tally_young_waste = 0; 3411 3412 shenandoah_assert_heaplocked_or_safepoint(); 3413 for (size_t i = 0; i < num_regions(); i++) { 3414 ShenandoahHeapRegion* r = get_region(i); 3415 if (r->is_old()) { 3416 tally_old_regions++; 3417 tally_old_bytes += r->used(); 3418 if (r->is_humongous()) { 3419 ShenandoahHeapRegion* start = r->humongous_start_region(); 3420 HeapWord* obj_addr = start->bottom(); 3421 oop obj = cast_to_oop(obj_addr); 3422 size_t word_size = obj->size(); 3423 HeapWord* end_addr = obj_addr + word_size; 3424 if (end_addr <= r->end()) { 3425 tally_old_waste += (r->end() - end_addr) * HeapWordSize; 3426 } 3427 } 3428 } else if (r->is_young()) { 3429 tally_young_regions++; 3430 tally_young_bytes += r->used(); 3431 if (r->is_humongous()) { 3432 ShenandoahHeapRegion* start = r->humongous_start_region(); 3433 HeapWord* obj_addr = start->bottom(); 3434 oop obj = cast_to_oop(obj_addr); 3435 size_t word_size = obj->size(); 3436 HeapWord* end_addr = obj_addr + word_size; 3437 if (end_addr <= r->end()) { 3438 tally_young_waste += (r->end() - end_addr) * HeapWordSize; 3439 } 3440 } 3441 } 3442 } 3443 if (verify_young && 3444 ((young_regions != tally_young_regions) || (young_bytes != tally_young_bytes) || (young_waste != tally_young_waste))) { 3445 return false; 3446 } else if (verify_old && 3447 ((old_regions != tally_old_regions) || (old_bytes != tally_old_bytes) || (old_waste != tally_old_waste))) { 3448 return false; 3449 } else { 3450 return true; 3451 } 3452 } 3453 3454 ShenandoahGeneration* ShenandoahHeap::generation_for(ShenandoahAffiliation affiliation) const { 3455 if (!mode()->is_generational()) { 3456 return global_generation(); 3457 } else if (affiliation == YOUNG_GENERATION) { 3458 return young_generation(); 3459 } else if (affiliation == OLD_GENERATION) { 3460 return old_generation(); 3461 } 3462 3463 ShouldNotReachHere(); 3464 return nullptr; 3465 } 3466 3467 void ShenandoahHeap::log_heap_status(const char* msg) const { 3468 if (mode()->is_generational()) { 3469 young_generation()->log_status(msg); 3470 old_generation()->log_status(msg); 3471 } else { 3472 global_generation()->log_status(msg); 3473 } 3474 }