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