1 /* 2 * Copyright (c) 2014, 2021, Red Hat, Inc. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 27 #include "compiler/oopMap.hpp" 28 #include "gc/shared/continuationGCSupport.hpp" 29 #include "gc/shared/gcTraceTime.inline.hpp" 30 #include "gc/shared/preservedMarks.inline.hpp" 31 #include "gc/shared/tlab_globals.hpp" 32 #include "gc/shared/workerThread.hpp" 33 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp" 34 #include "gc/shenandoah/shenandoahConcurrentGC.hpp" 35 #include "gc/shenandoah/shenandoahCollectionSet.hpp" 36 #include "gc/shenandoah/shenandoahFreeSet.hpp" 37 #include "gc/shenandoah/shenandoahFullGC.hpp" 38 #include "gc/shenandoah/shenandoahPhaseTimings.hpp" 39 #include "gc/shenandoah/shenandoahMark.inline.hpp" 40 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp" 41 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp" 42 #include "gc/shenandoah/shenandoahHeap.inline.hpp" 43 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp" 44 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp" 45 #include "gc/shenandoah/shenandoahMetrics.hpp" 46 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp" 47 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp" 48 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp" 49 #include "gc/shenandoah/shenandoahSTWMark.hpp" 50 #include "gc/shenandoah/shenandoahUtils.hpp" 51 #include "gc/shenandoah/shenandoahVerifier.hpp" 52 #include "gc/shenandoah/shenandoahVMOperations.hpp" 53 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp" 54 #include "memory/metaspaceUtils.hpp" 55 #include "memory/universe.hpp" 56 #include "oops/compressedOops.inline.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "runtime/javaThread.hpp" 59 #include "runtime/orderAccess.hpp" 60 #include "runtime/vmThread.hpp" 61 #include "utilities/copy.hpp" 62 #include "utilities/events.hpp" 63 #include "utilities/growableArray.hpp" 64 65 ShenandoahFullGC::ShenandoahFullGC() : 66 _gc_timer(ShenandoahHeap::heap()->gc_timer()), 67 _preserved_marks(new PreservedMarksSet(true)) {} 68 69 ShenandoahFullGC::~ShenandoahFullGC() { 70 delete _preserved_marks; 71 } 72 73 bool ShenandoahFullGC::collect(GCCause::Cause cause) { 74 vmop_entry_full(cause); 75 // Always success 76 return true; 77 } 78 79 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) { 80 ShenandoahHeap* const heap = ShenandoahHeap::heap(); 81 TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters()); 82 ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross); 83 84 heap->try_inject_alloc_failure(); 85 VM_ShenandoahFullGC op(cause, this); 86 VMThread::execute(&op); 87 } 88 89 void ShenandoahFullGC::entry_full(GCCause::Cause cause) { 90 static const char* msg = "Pause Full"; 91 ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */); 92 EventMark em("%s", msg); 93 94 ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(), 95 ShenandoahWorkerPolicy::calc_workers_for_fullgc(), 96 "full gc"); 97 98 op_full(cause); 99 } 100 101 void ShenandoahFullGC::op_full(GCCause::Cause cause) { 102 ShenandoahMetricsSnapshot metrics; 103 metrics.snap_before(); 104 105 // Perform full GC 106 do_it(cause); 107 108 metrics.snap_after(); 109 110 if (metrics.is_good_progress()) { 111 ShenandoahHeap::heap()->notify_gc_progress(); 112 } else { 113 // Nothing to do. Tell the allocation path that we have failed to make 114 // progress, and it can finally fail. 115 ShenandoahHeap::heap()->notify_gc_no_progress(); 116 } 117 } 118 119 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) { 120 ShenandoahHeap* heap = ShenandoahHeap::heap(); 121 122 if (ShenandoahVerify) { 123 heap->verifier()->verify_before_fullgc(); 124 } 125 126 if (VerifyBeforeGC) { 127 Universe::verify(); 128 } 129 130 // Degenerated GC may carry concurrent root flags when upgrading to 131 // full GC. We need to reset it before mutators resume. 132 heap->set_concurrent_strong_root_in_progress(false); 133 heap->set_concurrent_weak_root_in_progress(false); 134 135 heap->set_full_gc_in_progress(true); 136 137 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint"); 138 assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped"); 139 140 { 141 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre); 142 heap->pre_full_gc_dump(_gc_timer); 143 } 144 145 { 146 ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare); 147 // Full GC is supposed to recover from any GC state: 148 149 // a0. Remember if we have forwarded objects 150 bool has_forwarded_objects = heap->has_forwarded_objects(); 151 152 // a1. Cancel evacuation, if in progress 153 if (heap->is_evacuation_in_progress()) { 154 heap->set_evacuation_in_progress(false); 155 } 156 assert(!heap->is_evacuation_in_progress(), "sanity"); 157 158 // a2. Cancel update-refs, if in progress 159 if (heap->is_update_refs_in_progress()) { 160 heap->set_update_refs_in_progress(false); 161 } 162 assert(!heap->is_update_refs_in_progress(), "sanity"); 163 164 // b. Cancel concurrent mark, if in progress 165 if (heap->is_concurrent_mark_in_progress()) { 166 ShenandoahConcurrentGC::cancel(); 167 heap->set_concurrent_mark_in_progress(false); 168 } 169 assert(!heap->is_concurrent_mark_in_progress(), "sanity"); 170 171 // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots. 172 if (has_forwarded_objects) { 173 update_roots(true /*full_gc*/); 174 } 175 176 // d. Reset the bitmaps for new marking 177 heap->reset_mark_bitmap(); 178 assert(heap->marking_context()->is_bitmap_clear(), "sanity"); 179 assert(!heap->marking_context()->is_complete(), "sanity"); 180 181 // e. Abandon reference discovery and clear all discovered references. 182 ShenandoahReferenceProcessor* rp = heap->ref_processor(); 183 rp->abandon_partial_discovery(); 184 185 // f. Sync pinned region status from the CP marks 186 heap->sync_pinned_region_status(); 187 188 // The rest of prologue: 189 _preserved_marks->init(heap->workers()->active_workers()); 190 191 assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change"); 192 } 193 194 if (UseTLAB) { 195 heap->gclabs_retire(ResizeTLAB); 196 heap->tlabs_retire(ResizeTLAB); 197 } 198 199 OrderAccess::fence(); 200 201 phase1_mark_heap(); 202 203 // Once marking is done, which may have fixed up forwarded objects, we can drop it. 204 // Coming out of Full GC, we would not have any forwarded objects. 205 // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3. 206 heap->set_has_forwarded_objects(false); 207 208 heap->set_full_gc_move_in_progress(true); 209 210 // Setup workers for the rest 211 OrderAccess::fence(); 212 213 // Initialize worker slices 214 ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC); 215 for (uint i = 0; i < heap->max_workers(); i++) { 216 worker_slices[i] = new ShenandoahHeapRegionSet(); 217 } 218 219 { 220 // The rest of code performs region moves, where region status is undefined 221 // until all phases run together. 222 ShenandoahHeapLocker lock(heap->lock()); 223 224 phase2_calculate_target_addresses(worker_slices); 225 226 OrderAccess::fence(); 227 228 phase3_update_references(); 229 230 phase4_compact_objects(worker_slices); 231 } 232 233 { 234 // Epilogue 235 _preserved_marks->restore(heap->workers()); 236 _preserved_marks->reclaim(); 237 } 238 239 // Resize metaspace 240 MetaspaceGC::compute_new_size(); 241 242 // Free worker slices 243 for (uint i = 0; i < heap->max_workers(); i++) { 244 delete worker_slices[i]; 245 } 246 FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices); 247 248 heap->set_full_gc_move_in_progress(false); 249 heap->set_full_gc_in_progress(false); 250 251 if (ShenandoahVerify) { 252 heap->verifier()->verify_after_fullgc(); 253 } 254 255 if (VerifyAfterGC) { 256 Universe::verify(); 257 } 258 259 { 260 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post); 261 heap->post_full_gc_dump(_gc_timer); 262 } 263 } 264 265 class ShenandoahPrepareForMarkClosure: public ShenandoahHeapRegionClosure { 266 private: 267 ShenandoahMarkingContext* const _ctx; 268 269 public: 270 ShenandoahPrepareForMarkClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {} 271 272 void heap_region_do(ShenandoahHeapRegion *r) { 273 _ctx->capture_top_at_mark_start(r); 274 r->clear_live_data(); 275 } 276 }; 277 278 void ShenandoahFullGC::phase1_mark_heap() { 279 GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer); 280 ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark); 281 282 ShenandoahHeap* heap = ShenandoahHeap::heap(); 283 284 ShenandoahPrepareForMarkClosure cl; 285 heap->heap_region_iterate(&cl); 286 287 heap->set_unload_classes(heap->heuristics()->can_unload_classes()); 288 289 ShenandoahReferenceProcessor* rp = heap->ref_processor(); 290 // enable ("weak") refs discovery 291 rp->set_soft_reference_policy(true); // forcefully purge all soft references 292 293 ShenandoahSTWMark mark(true /*full_gc*/); 294 mark.mark(); 295 heap->parallel_cleaning(true /* full_gc */); 296 } 297 298 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure { 299 private: 300 PreservedMarks* const _preserved_marks; 301 ShenandoahHeap* const _heap; 302 GrowableArray<ShenandoahHeapRegion*>& _empty_regions; 303 int _empty_regions_pos; 304 ShenandoahHeapRegion* _to_region; 305 ShenandoahHeapRegion* _from_region; 306 HeapWord* _compact_point; 307 308 public: 309 ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks, 310 GrowableArray<ShenandoahHeapRegion*>& empty_regions, 311 ShenandoahHeapRegion* to_region) : 312 _preserved_marks(preserved_marks), 313 _heap(ShenandoahHeap::heap()), 314 _empty_regions(empty_regions), 315 _empty_regions_pos(0), 316 _to_region(to_region), 317 _from_region(nullptr), 318 _compact_point(to_region->bottom()) {} 319 320 void set_from_region(ShenandoahHeapRegion* from_region) { 321 _from_region = from_region; 322 } 323 324 void finish_region() { 325 assert(_to_region != nullptr, "should not happen"); 326 _to_region->set_new_top(_compact_point); 327 } 328 329 bool is_compact_same_region() { 330 return _from_region == _to_region; 331 } 332 333 int empty_regions_pos() { 334 return _empty_regions_pos; 335 } 336 337 void do_object(oop p) { 338 assert(_from_region != nullptr, "must set before work"); 339 assert(_heap->complete_marking_context()->is_marked(p), "must be marked"); 340 assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked"); 341 342 size_t obj_size = p->size(); 343 if (_compact_point + obj_size > _to_region->end()) { 344 finish_region(); 345 346 // Object doesn't fit. Pick next empty region and start compacting there. 347 ShenandoahHeapRegion* new_to_region; 348 if (_empty_regions_pos < _empty_regions.length()) { 349 new_to_region = _empty_regions.at(_empty_regions_pos); 350 _empty_regions_pos++; 351 } else { 352 // Out of empty region? Compact within the same region. 353 new_to_region = _from_region; 354 } 355 356 assert(new_to_region != _to_region, "must not reuse same to-region"); 357 assert(new_to_region != nullptr, "must not be null"); 358 _to_region = new_to_region; 359 _compact_point = _to_region->bottom(); 360 } 361 362 // Object fits into current region, record new location: 363 assert(_compact_point + obj_size <= _to_region->end(), "must fit"); 364 shenandoah_assert_not_forwarded(nullptr, p); 365 _preserved_marks->push_if_necessary(p, p->mark()); 366 p->forward_to(cast_to_oop(_compact_point)); 367 _compact_point += obj_size; 368 } 369 }; 370 371 class ShenandoahPrepareForCompactionTask : public WorkerTask { 372 private: 373 PreservedMarksSet* const _preserved_marks; 374 ShenandoahHeap* const _heap; 375 ShenandoahHeapRegionSet** const _worker_slices; 376 377 public: 378 ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) : 379 WorkerTask("Shenandoah Prepare For Compaction"), 380 _preserved_marks(preserved_marks), 381 _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) { 382 } 383 384 static bool is_candidate_region(ShenandoahHeapRegion* r) { 385 // Empty region: get it into the slice to defragment the slice itself. 386 // We could have skipped this without violating correctness, but we really 387 // want to compact all live regions to the start of the heap, which sometimes 388 // means moving them into the fully empty regions. 389 if (r->is_empty()) return true; 390 391 // Can move the region, and this is not the humongous region. Humongous 392 // moves are special cased here, because their moves are handled separately. 393 return r->is_stw_move_allowed() && !r->is_humongous(); 394 } 395 396 void work(uint worker_id) { 397 ShenandoahParallelWorkerSession worker_session(worker_id); 398 ShenandoahHeapRegionSet* slice = _worker_slices[worker_id]; 399 ShenandoahHeapRegionSetIterator it(slice); 400 ShenandoahHeapRegion* from_region = it.next(); 401 // No work? 402 if (from_region == nullptr) { 403 return; 404 } 405 406 // Sliding compaction. Walk all regions in the slice, and compact them. 407 // Remember empty regions and reuse them as needed. 408 ResourceMark rm; 409 410 GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions()); 411 412 ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region); 413 414 while (from_region != nullptr) { 415 assert(is_candidate_region(from_region), "Sanity"); 416 417 cl.set_from_region(from_region); 418 if (from_region->has_live()) { 419 _heap->marked_object_iterate(from_region, &cl); 420 } 421 422 // Compacted the region to somewhere else? From-region is empty then. 423 if (!cl.is_compact_same_region()) { 424 empty_regions.append(from_region); 425 } 426 from_region = it.next(); 427 } 428 cl.finish_region(); 429 430 // Mark all remaining regions as empty 431 for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) { 432 ShenandoahHeapRegion* r = empty_regions.at(pos); 433 r->set_new_top(r->bottom()); 434 } 435 } 436 }; 437 438 void ShenandoahFullGC::calculate_target_humongous_objects() { 439 ShenandoahHeap* heap = ShenandoahHeap::heap(); 440 441 // Compute the new addresses for humongous objects. We need to do this after addresses 442 // for regular objects are calculated, and we know what regions in heap suffix are 443 // available for humongous moves. 444 // 445 // Scan the heap backwards, because we are compacting humongous regions towards the end. 446 // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide 447 // humongous start there. 448 // 449 // The complication is potential non-movable regions during the scan. If such region is 450 // detected, then sliding restarts towards that non-movable region. 451 452 size_t to_begin = heap->num_regions(); 453 size_t to_end = heap->num_regions(); 454 455 for (size_t c = heap->num_regions(); c > 0; c--) { 456 ShenandoahHeapRegion *r = heap->get_region(c - 1); 457 if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) { 458 // To-region candidate: record this, and continue scan 459 to_begin = r->index(); 460 continue; 461 } 462 463 if (r->is_humongous_start() && r->is_stw_move_allowed()) { 464 // From-region candidate: movable humongous region 465 oop old_obj = cast_to_oop(r->bottom()); 466 size_t words_size = old_obj->size(); 467 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize); 468 469 size_t start = to_end - num_regions; 470 471 if (start >= to_begin && start != r->index()) { 472 // Fits into current window, and the move is non-trivial. Record the move then, and continue scan. 473 _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark()); 474 old_obj->forward_to(cast_to_oop(heap->get_region(start)->bottom())); 475 to_end = start; 476 continue; 477 } 478 } 479 480 // Failed to fit. Scan starting from current region. 481 to_begin = r->index(); 482 to_end = r->index(); 483 } 484 } 485 486 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure { 487 private: 488 ShenandoahHeap* const _heap; 489 490 public: 491 ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {} 492 void heap_region_do(ShenandoahHeapRegion* r) { 493 if (r->is_trash()) { 494 r->recycle(); 495 } 496 if (r->is_cset()) { 497 r->make_regular_bypass(); 498 } 499 if (r->is_empty_uncommitted()) { 500 r->make_committed_bypass(); 501 } 502 assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index()); 503 504 // Record current region occupancy: this communicates empty regions are free 505 // to the rest of Full GC code. 506 r->set_new_top(r->top()); 507 } 508 }; 509 510 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure { 511 private: 512 ShenandoahHeap* const _heap; 513 ShenandoahMarkingContext* const _ctx; 514 515 public: 516 ShenandoahTrashImmediateGarbageClosure() : 517 _heap(ShenandoahHeap::heap()), 518 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {} 519 520 void heap_region_do(ShenandoahHeapRegion* r) { 521 if (r->is_humongous_start()) { 522 oop humongous_obj = cast_to_oop(r->bottom()); 523 if (!_ctx->is_marked(humongous_obj)) { 524 assert(!r->has_live(), 525 "Region " SIZE_FORMAT " is not marked, should not have live", r->index()); 526 _heap->trash_humongous_region_at(r); 527 } else { 528 assert(r->has_live(), 529 "Region " SIZE_FORMAT " should have live", r->index()); 530 } 531 } else if (r->is_humongous_continuation()) { 532 // If we hit continuation, the non-live humongous starts should have been trashed already 533 assert(r->humongous_start_region()->has_live(), 534 "Region " SIZE_FORMAT " should have live", r->index()); 535 } else if (r->is_regular()) { 536 if (!r->has_live()) { 537 r->make_trash_immediate(); 538 } 539 } 540 } 541 }; 542 543 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) { 544 ShenandoahHeap* heap = ShenandoahHeap::heap(); 545 546 uint n_workers = heap->workers()->active_workers(); 547 size_t n_regions = heap->num_regions(); 548 549 // What we want to accomplish: have the dense prefix of data, while still balancing 550 // out the parallel work. 551 // 552 // Assuming the amount of work is driven by the live data that needs moving, we can slice 553 // the entire heap into equal-live-sized prefix slices, and compact into them. So, each 554 // thread takes all regions in its prefix subset, and then it takes some regions from 555 // the tail. 556 // 557 // Tail region selection becomes interesting. 558 // 559 // First, we want to distribute the regions fairly between the workers, and those regions 560 // might have different amount of live data. So, until we sure no workers need live data, 561 // we need to only take what the worker needs. 562 // 563 // Second, since we slide everything to the left in each slice, the most busy regions 564 // would be the ones on the left. Which means we want to have all workers have their after-tail 565 // regions as close to the left as possible. 566 // 567 // The easiest way to do this is to distribute after-tail regions in round-robin between 568 // workers that still need live data. 569 // 570 // Consider parallel workers A, B, C, then the target slice layout would be: 571 // 572 // AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA 573 // 574 // (.....dense-prefix.....) (.....................tail...................) 575 // [all regions fully live] [left-most regions are fuller that right-most] 576 // 577 578 // Compute how much live data is there. This would approximate the size of dense prefix 579 // we target to create. 580 size_t total_live = 0; 581 for (size_t idx = 0; idx < n_regions; idx++) { 582 ShenandoahHeapRegion *r = heap->get_region(idx); 583 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) { 584 total_live += r->get_live_data_words(); 585 } 586 } 587 588 // Estimate the size for the dense prefix. Note that we specifically count only the 589 // "full" regions, so there would be some non-full regions in the slice tail. 590 size_t live_per_worker = total_live / n_workers; 591 size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words(); 592 size_t prefix_regions_total = prefix_regions_per_worker * n_workers; 593 prefix_regions_total = MIN2(prefix_regions_total, n_regions); 594 assert(prefix_regions_total <= n_regions, "Sanity"); 595 596 // There might be non-candidate regions in the prefix. To compute where the tail actually 597 // ends up being, we need to account those as well. 598 size_t prefix_end = prefix_regions_total; 599 for (size_t idx = 0; idx < prefix_regions_total; idx++) { 600 ShenandoahHeapRegion *r = heap->get_region(idx); 601 if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) { 602 prefix_end++; 603 } 604 } 605 prefix_end = MIN2(prefix_end, n_regions); 606 assert(prefix_end <= n_regions, "Sanity"); 607 608 // Distribute prefix regions per worker: each thread definitely gets its own same-sized 609 // subset of dense prefix. 610 size_t prefix_idx = 0; 611 612 size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC); 613 614 for (size_t wid = 0; wid < n_workers; wid++) { 615 ShenandoahHeapRegionSet* slice = worker_slices[wid]; 616 617 live[wid] = 0; 618 size_t regs = 0; 619 620 // Add all prefix regions for this worker 621 while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) { 622 ShenandoahHeapRegion *r = heap->get_region(prefix_idx); 623 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) { 624 slice->add_region(r); 625 live[wid] += r->get_live_data_words(); 626 regs++; 627 } 628 prefix_idx++; 629 } 630 } 631 632 // Distribute the tail among workers in round-robin fashion. 633 size_t wid = n_workers - 1; 634 635 for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) { 636 ShenandoahHeapRegion *r = heap->get_region(tail_idx); 637 if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) { 638 assert(wid < n_workers, "Sanity"); 639 640 size_t live_region = r->get_live_data_words(); 641 642 // Select next worker that still needs live data. 643 size_t old_wid = wid; 644 do { 645 wid++; 646 if (wid == n_workers) wid = 0; 647 } while (live[wid] + live_region >= live_per_worker && old_wid != wid); 648 649 if (old_wid == wid) { 650 // Circled back to the same worker? This means liveness data was 651 // miscalculated. Bump the live_per_worker limit so that 652 // everyone gets a piece of the leftover work. 653 live_per_worker += ShenandoahHeapRegion::region_size_words(); 654 } 655 656 worker_slices[wid]->add_region(r); 657 live[wid] += live_region; 658 } 659 } 660 661 FREE_C_HEAP_ARRAY(size_t, live); 662 663 #ifdef ASSERT 664 ResourceBitMap map(n_regions); 665 for (size_t wid = 0; wid < n_workers; wid++) { 666 ShenandoahHeapRegionSetIterator it(worker_slices[wid]); 667 ShenandoahHeapRegion* r = it.next(); 668 while (r != nullptr) { 669 size_t idx = r->index(); 670 assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx); 671 assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx); 672 map.at_put(idx, true); 673 r = it.next(); 674 } 675 } 676 677 for (size_t rid = 0; rid < n_regions; rid++) { 678 bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid)); 679 bool is_distributed = map.at(rid); 680 assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid); 681 } 682 #endif 683 } 684 685 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) { 686 GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer); 687 ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses); 688 689 ShenandoahHeap* heap = ShenandoahHeap::heap(); 690 691 // About to figure out which regions can be compacted, make sure pinning status 692 // had been updated in GC prologue. 693 heap->assert_pinned_region_status(); 694 695 { 696 // Trash the immediately collectible regions before computing addresses 697 ShenandoahTrashImmediateGarbageClosure tigcl; 698 heap->heap_region_iterate(&tigcl); 699 700 // Make sure regions are in good state: committed, active, clean. 701 // This is needed because we are potentially sliding the data through them. 702 ShenandoahEnsureHeapActiveClosure ecl; 703 heap->heap_region_iterate(&ecl); 704 } 705 706 // Compute the new addresses for regular objects 707 { 708 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular); 709 710 distribute_slices(worker_slices); 711 712 ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices); 713 heap->workers()->run_task(&task); 714 } 715 716 // Compute the new addresses for humongous objects 717 { 718 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong); 719 calculate_target_humongous_objects(); 720 } 721 } 722 723 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure { 724 private: 725 ShenandoahHeap* const _heap; 726 ShenandoahMarkingContext* const _ctx; 727 728 template <class T> 729 inline void do_oop_work(T* p) { 730 T o = RawAccess<>::oop_load(p); 731 if (!CompressedOops::is_null(o)) { 732 oop obj = CompressedOops::decode_not_null(o); 733 assert(_ctx->is_marked(obj), "must be marked"); 734 if (obj->is_forwarded()) { 735 oop forw = obj->forwardee(); 736 RawAccess<IS_NOT_NULL>::oop_store(p, forw); 737 } 738 } 739 } 740 741 public: 742 ShenandoahAdjustPointersClosure() : 743 _heap(ShenandoahHeap::heap()), 744 _ctx(ShenandoahHeap::heap()->complete_marking_context()) {} 745 746 void do_oop(oop* p) { do_oop_work(p); } 747 void do_oop(narrowOop* p) { do_oop_work(p); } 748 void do_method(Method* m) {} 749 void do_nmethod(nmethod* nm) {} 750 }; 751 752 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure { 753 private: 754 ShenandoahHeap* const _heap; 755 ShenandoahAdjustPointersClosure _cl; 756 757 public: 758 ShenandoahAdjustPointersObjectClosure() : 759 _heap(ShenandoahHeap::heap()) { 760 } 761 void do_object(oop p) { 762 assert(_heap->complete_marking_context()->is_marked(p), "must be marked"); 763 p->oop_iterate(&_cl); 764 } 765 }; 766 767 class ShenandoahAdjustPointersTask : public WorkerTask { 768 private: 769 ShenandoahHeap* const _heap; 770 ShenandoahRegionIterator _regions; 771 772 public: 773 ShenandoahAdjustPointersTask() : 774 WorkerTask("Shenandoah Adjust Pointers"), 775 _heap(ShenandoahHeap::heap()) { 776 } 777 778 void work(uint worker_id) { 779 ShenandoahParallelWorkerSession worker_session(worker_id); 780 ShenandoahAdjustPointersObjectClosure obj_cl; 781 ShenandoahHeapRegion* r = _regions.next(); 782 while (r != nullptr) { 783 if (!r->is_humongous_continuation() && r->has_live()) { 784 _heap->marked_object_iterate(r, &obj_cl); 785 } 786 r = _regions.next(); 787 } 788 } 789 }; 790 791 class ShenandoahAdjustRootPointersTask : public WorkerTask { 792 private: 793 ShenandoahRootAdjuster* _rp; 794 PreservedMarksSet* _preserved_marks; 795 public: 796 ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) : 797 WorkerTask("Shenandoah Adjust Root Pointers"), 798 _rp(rp), 799 _preserved_marks(preserved_marks) {} 800 801 void work(uint worker_id) { 802 ShenandoahParallelWorkerSession worker_session(worker_id); 803 ShenandoahAdjustPointersClosure cl; 804 _rp->roots_do(worker_id, &cl); 805 _preserved_marks->get(worker_id)->adjust_during_full_gc(); 806 } 807 }; 808 809 void ShenandoahFullGC::phase3_update_references() { 810 GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer); 811 ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers); 812 813 ShenandoahHeap* heap = ShenandoahHeap::heap(); 814 815 WorkerThreads* workers = heap->workers(); 816 uint nworkers = workers->active_workers(); 817 { 818 #if COMPILER2_OR_JVMCI 819 DerivedPointerTable::clear(); 820 #endif 821 ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots); 822 ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks); 823 workers->run_task(&task); 824 #if COMPILER2_OR_JVMCI 825 DerivedPointerTable::update_pointers(); 826 #endif 827 } 828 829 ShenandoahAdjustPointersTask adjust_pointers_task; 830 workers->run_task(&adjust_pointers_task); 831 } 832 833 class ShenandoahCompactObjectsClosure : public ObjectClosure { 834 private: 835 ShenandoahHeap* const _heap; 836 uint const _worker_id; 837 838 public: 839 ShenandoahCompactObjectsClosure(uint worker_id) : 840 _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {} 841 842 void do_object(oop p) { 843 assert(_heap->complete_marking_context()->is_marked(p), "must be marked"); 844 size_t size = p->size(); 845 if (p->is_forwarded()) { 846 HeapWord* compact_from = cast_from_oop<HeapWord*>(p); 847 HeapWord* compact_to = cast_from_oop<HeapWord*>(p->forwardee()); 848 Copy::aligned_conjoint_words(compact_from, compact_to, size); 849 oop new_obj = cast_to_oop(compact_to); 850 851 ContinuationGCSupport::relativize_stack_chunk(new_obj); 852 new_obj->init_mark(); 853 } 854 } 855 }; 856 857 class ShenandoahCompactObjectsTask : public WorkerTask { 858 private: 859 ShenandoahHeap* const _heap; 860 ShenandoahHeapRegionSet** const _worker_slices; 861 862 public: 863 ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) : 864 WorkerTask("Shenandoah Compact Objects"), 865 _heap(ShenandoahHeap::heap()), 866 _worker_slices(worker_slices) { 867 } 868 869 void work(uint worker_id) { 870 ShenandoahParallelWorkerSession worker_session(worker_id); 871 ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]); 872 873 ShenandoahCompactObjectsClosure cl(worker_id); 874 ShenandoahHeapRegion* r = slice.next(); 875 while (r != nullptr) { 876 assert(!r->is_humongous(), "must not get humongous regions here"); 877 if (r->has_live()) { 878 _heap->marked_object_iterate(r, &cl); 879 } 880 r->set_top(r->new_top()); 881 r = slice.next(); 882 } 883 } 884 }; 885 886 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure { 887 private: 888 ShenandoahHeap* const _heap; 889 size_t _live; 890 891 public: 892 ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) { 893 _heap->free_set()->clear(); 894 } 895 896 void heap_region_do(ShenandoahHeapRegion* r) { 897 assert (!r->is_cset(), "cset regions should have been demoted already"); 898 899 // Need to reset the complete-top-at-mark-start pointer here because 900 // the complete marking bitmap is no longer valid. This ensures 901 // size-based iteration in marked_object_iterate(). 902 // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip 903 // pinned regions. 904 if (!r->is_pinned()) { 905 _heap->complete_marking_context()->reset_top_at_mark_start(r); 906 } 907 908 size_t live = r->used(); 909 910 // Make empty regions that have been allocated into regular 911 if (r->is_empty() && live > 0) { 912 r->make_regular_bypass(); 913 } 914 915 // Reclaim regular regions that became empty 916 if (r->is_regular() && live == 0) { 917 r->make_trash(); 918 } 919 920 // Recycle all trash regions 921 if (r->is_trash()) { 922 live = 0; 923 r->recycle(); 924 } 925 926 r->set_live_data(live); 927 r->reset_alloc_metadata(); 928 _live += live; 929 } 930 931 size_t get_live() { 932 return _live; 933 } 934 }; 935 936 void ShenandoahFullGC::compact_humongous_objects() { 937 // Compact humongous regions, based on their fwdptr objects. 938 // 939 // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases, 940 // humongous regions are already compacted, and do not require further moves, which alleviates 941 // sliding costs. We may consider doing this in parallel in future. 942 943 ShenandoahHeap* heap = ShenandoahHeap::heap(); 944 945 for (size_t c = heap->num_regions(); c > 0; c--) { 946 ShenandoahHeapRegion* r = heap->get_region(c - 1); 947 if (r->is_humongous_start()) { 948 oop old_obj = cast_to_oop(r->bottom()); 949 if (!old_obj->is_forwarded()) { 950 // No need to move the object, it stays at the same slot 951 continue; 952 } 953 size_t words_size = old_obj->size(); 954 size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize); 955 956 size_t old_start = r->index(); 957 size_t old_end = old_start + num_regions - 1; 958 size_t new_start = heap->heap_region_index_containing(old_obj->forwardee()); 959 size_t new_end = new_start + num_regions - 1; 960 assert(old_start != new_start, "must be real move"); 961 assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index()); 962 963 Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size); 964 ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom())); 965 966 oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom()); 967 new_obj->init_mark(); 968 969 { 970 for (size_t c = old_start; c <= old_end; c++) { 971 ShenandoahHeapRegion* r = heap->get_region(c); 972 r->make_regular_bypass(); 973 r->set_top(r->bottom()); 974 } 975 976 for (size_t c = new_start; c <= new_end; c++) { 977 ShenandoahHeapRegion* r = heap->get_region(c); 978 if (c == new_start) { 979 r->make_humongous_start_bypass(); 980 } else { 981 r->make_humongous_cont_bypass(); 982 } 983 984 // Trailing region may be non-full, record the remainder there 985 size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask(); 986 if ((c == new_end) && (remainder != 0)) { 987 r->set_top(r->bottom() + remainder); 988 } else { 989 r->set_top(r->end()); 990 } 991 992 r->reset_alloc_metadata(); 993 } 994 } 995 } 996 } 997 } 998 999 // This is slightly different to ShHeap::reset_next_mark_bitmap: 1000 // we need to remain able to walk pinned regions. 1001 // Since pinned region do not move and don't get compacted, we will get holes with 1002 // unreachable objects in them (which may have pointers to unloaded Klasses and thus 1003 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using 1004 // a valid marking bitmap and valid TAMS pointer. This class only resets marking 1005 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions. 1006 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask { 1007 private: 1008 ShenandoahRegionIterator _regions; 1009 1010 public: 1011 ShenandoahMCResetCompleteBitmapTask() : 1012 WorkerTask("Shenandoah Reset Bitmap") { 1013 } 1014 1015 void work(uint worker_id) { 1016 ShenandoahParallelWorkerSession worker_session(worker_id); 1017 ShenandoahHeapRegion* region = _regions.next(); 1018 ShenandoahHeap* heap = ShenandoahHeap::heap(); 1019 ShenandoahMarkingContext* const ctx = heap->complete_marking_context(); 1020 while (region != nullptr) { 1021 if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) { 1022 ctx->clear_bitmap(region); 1023 } 1024 region = _regions.next(); 1025 } 1026 } 1027 }; 1028 1029 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) { 1030 GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer); 1031 ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects); 1032 1033 ShenandoahHeap* heap = ShenandoahHeap::heap(); 1034 1035 // Compact regular objects first 1036 { 1037 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular); 1038 ShenandoahCompactObjectsTask compact_task(worker_slices); 1039 heap->workers()->run_task(&compact_task); 1040 } 1041 1042 // Compact humongous objects after regular object moves 1043 { 1044 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong); 1045 compact_humongous_objects(); 1046 } 1047 1048 // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer 1049 // and must ensure the bitmap is in sync. 1050 { 1051 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete); 1052 ShenandoahMCResetCompleteBitmapTask task; 1053 heap->workers()->run_task(&task); 1054 } 1055 1056 // Bring regions in proper states after the collection, and set heap properties. 1057 { 1058 ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild); 1059 1060 ShenandoahPostCompactClosure post_compact; 1061 heap->heap_region_iterate(&post_compact); 1062 heap->set_used(post_compact.get_live()); 1063 1064 heap->collection_set()->clear(); 1065 heap->free_set()->rebuild(); 1066 } 1067 1068 heap->clear_cancelled_gc(); 1069 }