1 /*
2 * Copyright (c) 2014, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/g1/g1Allocator.inline.hpp"
26 #include "gc/g1/g1CollectedHeap.inline.hpp"
27 #include "gc/g1/g1CollectionSet.hpp"
28 #include "gc/g1/g1EvacFailureRegions.inline.hpp"
29 #include "gc/g1/g1HeapRegionPrinter.hpp"
30 #include "gc/g1/g1OopClosures.inline.hpp"
31 #include "gc/g1/g1ParScanThreadState.inline.hpp"
32 #include "gc/g1/g1RootClosures.hpp"
33 #include "gc/g1/g1StringDedup.hpp"
34 #include "gc/g1/g1Trace.hpp"
35 #include "gc/g1/g1YoungGCAllocationFailureInjector.inline.hpp"
36 #include "gc/shared/continuationGCSupport.inline.hpp"
37 #include "gc/shared/partialArraySplitter.inline.hpp"
38 #include "gc/shared/partialArrayState.hpp"
39 #include "gc/shared/partialArrayTaskStats.hpp"
40 #include "gc/shared/stringdedup/stringDedup.hpp"
41 #include "gc/shared/taskqueue.inline.hpp"
42 #include "memory/allocation.inline.hpp"
43 #include "oops/access.inline.hpp"
44 #include "oops/oop.inline.hpp"
45 #include "runtime/mutexLocker.hpp"
46 #include "runtime/prefetch.inline.hpp"
47 #include "utilities/globalDefinitions.hpp"
48 #include "utilities/macros.hpp"
49
50 // In fastdebug builds the code size can get out of hand, potentially
51 // tripping over compiler limits (which may be bugs, but nevertheless
52 // need to be taken into consideration). A side benefit of limiting
53 // inlining is that we get more call frames that might aid debugging.
54 // And the fastdebug compile time for this file is much reduced.
55 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
56 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
57
58 // Good estimate for the initial table size.
59 static uint initial_nmethod_table_size(G1CollectedHeap* g1h) {
60 // The +1 is both to consider the retained old region likely to be added, and avoid zero-sized initial tables.
61 return MIN3(g1h->collection_set()->num_regions(), g1h->max_num_regions() / 2, g1h->num_available_regions()) + 1;
62 }
63
64 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
65 G1ParScanThreadStateSet* per_thread_states,
66 uint worker_id,
67 uint num_workers,
68 G1CollectionSet* collection_set,
69 G1EvacFailureRegions* evac_failure_regions)
70 : _g1h(g1h),
71 _per_thread_states(per_thread_states),
72 _task_queue(g1h->task_queue(worker_id)),
73 _ct(g1h->refinement_table()),
74 _closures(nullptr),
75 _plab_allocator(nullptr),
76 _age_table(false),
77 _tenuring_threshold(g1h->policy()->tenuring_threshold()),
78 _scanner(g1h, this),
79 _worker_id(worker_id),
80 _num_cards_marked_dirty(0),
81 _num_cards_marked_to_cset(0),
82 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
83 _stack_trim_lower_threshold(GCDrainStackTargetSize),
84 _trim_ticks(),
85 _surviving_young_words_base(nullptr),
86 _surviving_young_words(nullptr),
87 _surviving_words_length(collection_set->num_young_regions() + 1),
88 _old_gen_is_full(false),
89 _partial_array_splitter(g1h->partial_array_state_manager(), num_workers),
90 _string_dedup_requests(),
91 _max_num_optional_regions(collection_set->num_optional_regions()),
92 _numa(g1h->numa()),
93 _obj_alloc_stat(nullptr),
94 // The initial size estimate is relatively conservative, assuming that all regions
95 // in the collection set get evacuated into the same amount of new regions.
96 _nmethods_to_add(initial_nmethod_table_size(g1h),
97 MAX2(initial_nmethod_table_size(g1h), _g1h->max_num_regions() / 2)),
98 ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
99 _evacuation_failed_info(),
100 _evac_failure_regions(evac_failure_regions),
101 _num_cards_from_evac_failure(0)
102 {
103 // We allocate number of young gen regions in the collection set plus one
104 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
105 // We also add a few elements at the beginning and at the end in
106 // an attempt to eliminate cache contention
107 const size_t padding_elem_num = (DEFAULT_PADDING_SIZE / sizeof(size_t));
108 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
109
110 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
111 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
112 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
113
114 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
115
116 _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
117 this,
118 collection_set->only_contains_young_regions());
119
120 _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
121
122 initialize_numa_stats();
123 }
124
125 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
126 flush_numa_stats();
127 // Update allocation statistics.
128 _plab_allocator->flush_and_retire_stats(num_workers);
129 _g1h->policy()->record_age_table(&_age_table);
130
131 if (_evacuation_failed_info.has_failed()) {
132 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
133 }
134
135 size_t sum = 0;
136 for (uint i = 0; i < _surviving_words_length; i++) {
137 surviving_young_words[i] += _surviving_young_words[i];
138 sum += _surviving_young_words[i];
139 }
140 return sum;
141 }
142
143 G1ParScanThreadState::~G1ParScanThreadState() {
144 auto delete_all = [&] (uint region, G1NmethodSet* nmethods) -> bool {
145 delete nmethods;
146 return true;
147 };
148 _nmethods_to_add.iterate(delete_all);
149
150 delete _plab_allocator;
151 delete _closures;
152 FREE_C_HEAP_ARRAY(_surviving_young_words_base);
153 delete[] _oops_into_optional_regions;
154 FREE_C_HEAP_ARRAY(_obj_alloc_stat);
155 }
156
157 size_t G1ParScanThreadState::lab_waste_words() const {
158 return _plab_allocator->waste();
159 }
160
161 size_t G1ParScanThreadState::lab_undo_waste_words() const {
162 return _plab_allocator->undo_waste();
163 }
164
165 size_t G1ParScanThreadState::num_cards_pending() const {
166 return _num_cards_marked_dirty + _num_cards_from_evac_failure;
167 }
168
169 size_t G1ParScanThreadState::num_cards_marked() const {
170 return num_cards_pending() + _num_cards_marked_to_cset;
171 }
172
173 size_t G1ParScanThreadState::num_cards_from_evac_failure() const {
174 return _num_cards_from_evac_failure;
175 }
176
177 #ifdef ASSERT
178 void G1ParScanThreadState::verify_task(narrowOop* task) const {
179 assert(task != nullptr, "invariant");
180 assert(UseCompressedOops, "sanity");
181 oop p = RawAccess<>::oop_load(task);
182 assert(_g1h->is_in_reserved(p),
183 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
184 }
185
186 void G1ParScanThreadState::verify_task(oop* task) const {
187 assert(task != nullptr, "invariant");
188 oop p = RawAccess<>::oop_load(task);
189 assert(_g1h->is_in_reserved(p),
190 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
191 }
192
193 void G1ParScanThreadState::verify_task(PartialArrayState* task) const {
194 assert(task != nullptr, "invariant");
195 // Source isn't used for processing, so not recorded in task.
196 assert(task->source() == nullptr, "invariant");
197 oop p = task->destination();
198 assert(_g1h->is_in_reserved(p),
199 "task=" PTR_FORMAT " dest=" PTR_FORMAT, p2i(task), p2i(p));
200 }
201
202 void G1ParScanThreadState::verify_task(ScannerTask task) const {
203 if (task.is_narrow_oop_ptr()) {
204 verify_task(task.to_narrow_oop_ptr());
205 } else if (task.is_oop_ptr()) {
206 verify_task(task.to_oop_ptr());
207 } else if (task.is_partial_array_state()) {
208 verify_task(task.to_partial_array_state());
209 } else {
210 ShouldNotReachHere();
211 }
212 }
213 #endif // ASSERT
214
215 template <class T>
216 MAYBE_INLINE_EVACUATION
217 void G1ParScanThreadState::do_oop_evac(T* p) {
218 // Reference should not be null here as such are never pushed to the task queue.
219 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
220
221 // Although we never intentionally push references outside of the collection
222 // set, due to (benign) races in the claim mechanism during RSet scanning more
223 // than one thread might claim the same card. So the same card may be
224 // processed multiple times, and so we might get references into old gen here.
225 // So we need to redo this check.
226 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
227 // References pushed onto the work stack should never point to a humongous region
228 // as they are not added to the collection set due to above precondition.
229 assert(!region_attr.is_humongous_candidate(),
230 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
231 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
232
233 if (!region_attr.is_in_cset()) {
234 // In this case somebody else already did all the work.
235 return;
236 }
237
238 markWord m = obj->mark();
239 if (m.is_forwarded()) {
240 obj = obj->forwardee(m);
241 } else {
242 obj = do_copy_to_survivor_space(region_attr, obj, m);
243 }
244 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
245
246 write_ref_field_post(p, obj);
247 }
248
249 ALWAYSINLINE
250 void G1ParScanThreadState::process_array_chunk(objArrayOop obj, size_t start, size_t end) {
251 obj->oop_iterate_elements_range(&_scanner,
252 checked_cast<int>(start),
253 checked_cast<int>(end));
254 }
255
256 MAYBE_INLINE_EVACUATION
257 void G1ParScanThreadState::do_partial_array(PartialArrayState* state, bool stolen) {
258 // Access state before release by claim().
259 objArrayOop to_array = objArrayOop(state->destination());
260 PartialArraySplitter::Claim claim =
261 _partial_array_splitter.claim(state, _task_queue, stolen);
262 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
263 G1SkipCardMarkSetter x(&_scanner, dest_attr.is_new_survivor());
264 // Process claimed task.
265 assert(to_array->is_objArray(), "Must be");
266 process_array_chunk(to_array, claim._start, claim._end);
267 }
268
269 MAYBE_INLINE_EVACUATION
270 void G1ParScanThreadState::start_partial_objarray(oop from_obj,
271 oop to_obj) {
272 assert(from_obj->is_forwarded(), "precondition");
273 assert(from_obj->forwardee() == to_obj, "precondition");
274 assert(to_obj->is_objArray(), "precondition");
275 assert(!_scanner.do_metadata(), "precondition");
276 assert(_scanner.skip_card_mark_set(), "precondition");
277
278 objArrayOop to_array = objArrayOop(to_obj);
279 size_t array_length = to_array->length();
280 size_t initial_chunk_size =
281 // The source array is unused when processing states.
282 _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length, ParGCArrayScanChunk);
283
284 process_array_chunk(to_array, 0, initial_chunk_size);
285 }
286
287 MAYBE_INLINE_EVACUATION
288 void G1ParScanThreadState::dispatch_task(ScannerTask task, bool stolen) {
289 verify_task(task);
290 if (task.is_narrow_oop_ptr()) {
291 do_oop_evac(task.to_narrow_oop_ptr());
292 } else if (task.is_oop_ptr()) {
293 do_oop_evac(task.to_oop_ptr());
294 } else {
295 do_partial_array(task.to_partial_array_state(), stolen);
296 }
297 }
298
299 // Process tasks until overflow queue is empty and local queue
300 // contains no more than threshold entries. NOINLINE to prevent
301 // inlining into steal_and_trim_queue.
302 ATTRIBUTE_FLATTEN NOINLINE
303 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
304 ScannerTask task;
305 do {
306 while (_task_queue->pop_overflow(task)) {
307 if (!_task_queue->try_push_to_taskqueue(task)) {
308 dispatch_task(task, false);
309 }
310 }
311 while (_task_queue->pop_local(task, threshold)) {
312 dispatch_task(task, false);
313 }
314 } while (!_task_queue->overflow_empty());
315 }
316
317 ATTRIBUTE_FLATTEN
318 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
319 ScannerTask stolen_task;
320 while (task_queues->steal(_worker_id, stolen_task)) {
321 dispatch_task(stolen_task, true);
322 // Processing stolen task may have added tasks to our queue.
323 trim_queue();
324 }
325 }
326
327 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
328 size_t word_sz,
329 bool previous_plab_refill_failed,
330 uint node_index) {
331
332 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
333
334 // Right now we only have two types of regions (young / old) so
335 // let's keep the logic here simple. We can generalize it when necessary.
336 if (dest->is_young()) {
337 bool plab_refill_in_old_failed = false;
338 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
339 word_sz,
340 &plab_refill_in_old_failed,
341 node_index);
342 // Make sure that we won't attempt to copy any other objects out
343 // of a survivor region (given that apparently we cannot allocate
344 // any new ones) to avoid coming into this slow path again and again.
345 // Only consider failed PLAB refill here: failed inline allocations are
346 // typically large, so not indicative of remaining space.
347 if (previous_plab_refill_failed) {
348 _tenuring_threshold = 0;
349 }
350
351 if (obj_ptr != nullptr) {
352 dest->set_old();
353 } else {
354 // We just failed to allocate in old gen. The same idea as explained above
355 // for making survivor gen unavailable for allocation applies for old gen.
356 _old_gen_is_full = plab_refill_in_old_failed;
357 }
358 return obj_ptr;
359 } else {
360 _old_gen_is_full = previous_plab_refill_failed;
361 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
362 // no other space to try.
363 return nullptr;
364 }
365 }
366
367 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
368 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
369
370 if (region_attr.is_young()) {
371 age = !m.has_displaced_mark_helper() ? m.age()
372 : m.displaced_mark_helper().age();
373 if (age < _tenuring_threshold) {
374 return region_attr;
375 }
376 }
377 // young-to-old (promotion) or old-to-old; destination is old in both cases.
378 return G1HeapRegionAttr::Old;
379 }
380
381 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
382 Klass* klass, size_t word_sz, uint age,
383 HeapWord * const obj_ptr, uint node_index) const {
384 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
385 if (alloc_buf->contains(obj_ptr)) {
386 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
387 dest_attr.type() == G1HeapRegionAttr::Old,
388 alloc_buf->word_sz() * HeapWordSize);
389 } else {
390 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
391 dest_attr.type() == G1HeapRegionAttr::Old);
392 }
393 }
394
395 NOINLINE
396 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
397 Klass* klass,
398 size_t word_sz,
399 uint age,
400 uint node_index) {
401 HeapWord* obj_ptr = nullptr;
402 // Try slow-path allocation unless we're allocating old and old is already full.
403 if (!(dest_attr->is_old() && _old_gen_is_full)) {
404 bool plab_refill_failed = false;
405 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
406 word_sz,
407 &plab_refill_failed,
408 node_index);
409 if (obj_ptr == nullptr) {
410 obj_ptr = allocate_in_next_plab(dest_attr,
411 word_sz,
412 plab_refill_failed,
413 node_index);
414 }
415 }
416 if (obj_ptr != nullptr) {
417 update_numa_stats(node_index);
418 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
419 // The events are checked individually as part of the actual commit
420 report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
421 }
422 }
423 return obj_ptr;
424 }
425
426 #if ALLOCATION_FAILURE_INJECTOR
427 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
428 return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
429 }
430 #endif
431
432 NOINLINE
433 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
434 HeapWord* obj_ptr,
435 size_t word_sz,
436 uint node_index) {
437 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
438 }
439
440 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
441 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
442 G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
443 region->update_bot_for_block(obj_start, obj_start + word_sz);
444 }
445
446 ALWAYSINLINE
447 void G1ParScanThreadState::do_iterate_object(oop const obj,
448 oop const old,
449 Klass* const klass,
450 G1HeapRegionAttr const region_attr,
451 G1HeapRegionAttr const dest_attr,
452 uint age) {
453 // Most objects are not arrays, so do one array check rather than
454 // checking for each array category for each object.
455 if (klass->is_array_klass()) {
456 assert(!klass->is_stack_chunk_instance_klass(), "must be");
457
458 if (klass->is_objArray_klass()) {
459 start_partial_objarray(old, obj);
460 } else {
461 // Nothing needs to be done for typeArrays. Body doesn't contain
462 // any oops to scan, and the type in the klass will already be handled
463 // by processing the built-in module.
464 assert(klass->is_typeArray_klass(), "invariant");
465 }
466 return;
467 }
468
469 ContinuationGCSupport::transform_stack_chunk(obj, klass);
470
471 // Check for deduplicating young Strings.
472 if (G1StringDedup::is_candidate_from_evacuation(klass,
473 region_attr,
474 dest_attr,
475 age)) {
476 // Record old; request adds a new weak reference, which reference
477 // processing expects to refer to a from-space object.
478 _string_dedup_requests.add(old);
479 }
480
481 assert(_scanner.skip_card_mark_set(), "must be");
482 obj->oop_iterate_backwards(&_scanner, klass);
483 }
484
485 // Private inline function, for direct internal use and providing the
486 // implementation of the public not-inline function.
487 MAYBE_INLINE_EVACUATION
488 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
489 oop const old,
490 markWord const old_mark) {
491 assert(region_attr.is_in_cset(),
492 "Unexpected region attr type: %s", region_attr.get_type_str());
493
494 // NOTE: With compact headers, it is not safe to load the Klass* from old, because
495 // that would access the mark-word, that might change at any time by concurrent
496 // workers.
497 // This mark word would refer to a forwardee, which may not yet have completed
498 // copying. Therefore we must load the Klass* from the mark-word that we already
499 // loaded. This is safe, because we only enter here if not yet forwarded.
500 assert(!old_mark.is_forwarded(), "precondition");
501 Klass* klass = UseCompactObjectHeaders
502 ? old_mark.klass()
503 : old->klass();
504
505 const size_t word_sz = old->size_given_klass(klass);
506
507 // JNI only allows pinning of typeArrays, so we only need to keep those in place.
508 if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
509 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, true /* cause_pinned */);
510 }
511
512 uint age = 0;
513 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
514 G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
515 uint node_index = from_region->node_index();
516
517 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
518
519 // PLAB allocations should succeed most of the time, so we'll
520 // normally check against null once and that's it.
521 if (obj_ptr == nullptr) {
522 obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
523 if (obj_ptr == nullptr) {
524 // This will either forward-to-self, or detect that someone else has
525 // installed a forwarding pointer.
526 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
527 }
528 }
529
530 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
531 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
532
533 // Should this evacuation fail?
534 if (inject_allocation_failure(from_region->hrm_index())) {
535 // Doing this after all the allocation attempts also tests the
536 // undo_allocation() method too.
537 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
538 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
539 }
540
541 // We're going to allocate linearly, so might as well prefetch ahead.
542 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
543 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
544
545 const oop obj = cast_to_oop(obj_ptr);
546 // Because the forwarding is done with memory_order_relaxed there is no
547 // ordering with the above copy. Clients that get the forwardee must not
548 // examine its contents without other synchronization, since the contents
549 // may not be up to date for them.
550 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
551 if (forward_ptr == nullptr) {
552
553 {
554 const uint young_index = from_region->young_index_in_cset();
555 assert((from_region->is_young() && young_index > 0) ||
556 (!from_region->is_young() && young_index == 0), "invariant" );
557 _surviving_young_words[young_index] += word_sz;
558 }
559
560 if (dest_attr.is_young()) {
561 if (age < markWord::max_age) {
562 age++;
563 obj->incr_age();
564 }
565 _age_table.add(age, word_sz);
566 } else {
567 update_bot_after_copying(obj, word_sz);
568 }
569
570 {
571 // Skip the card enqueue iff the object (obj) is in survivor region.
572 // However, G1HeapRegion::is_survivor() is too expensive here.
573 // Instead, we use dest_attr.is_young() because the two values are always
574 // equal: successfully allocated young regions must be survivor regions.
575 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
576 G1SkipCardMarkSetter x(&_scanner, dest_attr.is_young());
577 do_iterate_object(obj, old, klass, region_attr, dest_attr, age);
578 }
579
580 return obj;
581 } else {
582 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
583 return forward_ptr;
584 }
585 }
586
587 // Public not-inline entry point.
588 ATTRIBUTE_FLATTEN
589 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
590 oop old,
591 markWord old_mark) {
592 return do_copy_to_survivor_space(region_attr, old, old_mark);
593 }
594
595 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
596 assert(worker_id < _num_workers, "out of bounds access");
597 if (_states[worker_id] == nullptr) {
598 _states[worker_id] =
599 new G1ParScanThreadState(_g1h,
600 this,
601 worker_id,
602 _num_workers,
603 _collection_set,
604 _evac_failure_regions);
605 }
606 return _states[worker_id];
607 }
608
609 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
610 assert(_flushed, "thread local state from the per thread states should have been flushed");
611 return _surviving_young_words_total;
612 }
613
614 void G1ParScanThreadStateSet::flush_stats() {
615 assert(!_flushed, "thread local state from the per thread states should be flushed once");
616 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
617 G1ParScanThreadState* pss = _states[worker_id];
618 assert(pss != nullptr, "must be initialized");
619
620 G1GCPhaseTimes* p = _g1h->phase_times();
621
622 // Need to get the following two before the call to G1ParThreadScanState::flush()
623 // because it resets the PLAB allocator where we get this info from.
624 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
625 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
626 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers) * HeapWordSize;
627 size_t pending_cards = pss->num_cards_pending();
628 size_t to_young_gen_cards = pss->num_cards_marked() - pss->num_cards_pending();
629 size_t evac_failure_cards = pss->num_cards_from_evac_failure();
630 size_t marked_cards = pss->num_cards_marked();
631
632 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, copied_bytes, G1GCPhaseTimes::FlushPSSCopiedBytes);
633 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::FlushPSSLABWasteBytes);
634 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::FlushPSSLABUndoWasteBytes);
635 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, pending_cards, G1GCPhaseTimes::FlushPSSPendingCards);
636 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, to_young_gen_cards, G1GCPhaseTimes::FlushPSSToYoungGenCards);
637 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, evac_failure_cards, G1GCPhaseTimes::FlushPSSEvacFail);
638 p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, marked_cards, G1GCPhaseTimes::FlushPSSMarked);
639 }
640
641 _flushed = true;
642 }
643
644 void G1ParScanThreadStateSet::destroy_worker_states() {
645 assert(_flushed, "statistics must already be flushed");
646 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
647 delete _states[worker_id];
648 _states[worker_id] = nullptr;
649 }
650 }
651
652 void G1ParScanThreadStateSet::update_nmethod_regions_to_add(G1NmethodsToAdd* nmethods) {
653 if (nmethods->number_of_entries() == 0) {
654 return;
655 }
656
657 // Take the key set, look which are not yet in the global set, and update the necessary ones.
658 ResourceMark rm;
659 GrowableArray<uint> regions_to_add = GrowableArray<uint>(nmethods->table_size());
660
661 nmethods->iterate_all([&] (uint& region, void*) {
662 if (_has_nmethods_to_add.par_set_bit(region, memory_order_relaxed)) {
663 regions_to_add.push(region);
664 }
665 });
666
667 uint num_regions_to_add = (uint)regions_to_add.length();
668
669 if (num_regions_to_add == 0) {
670 return;
671 }
672
673 uint first_index = _num_nmethod_regions_to_add.fetch_then_add(num_regions_to_add, memory_order_relaxed);
674 guarantee(first_index + num_regions_to_add <= _g1h->max_num_regions(), "must be");
675
676 memcpy(&_nmethod_regions_to_add[first_index], regions_to_add.adr_at(0), num_regions_to_add * sizeof(uint));
677 }
678
679 void G1ParScanThreadStateSet::par_iterate_nmethod_regions_to_add(G1HeapRegionClosure* cl,
680 G1HeapRegionClaimer* claimer,
681 uint worker_id) {
682 _g1h->par_iterate_regions_array(cl, claimer, _nmethod_regions_to_add, num_nmethod_regions_to_add(), worker_id);
683 }
684
685 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
686 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
687 G1ParScanThreadState* pss = _states[worker_index];
688 assert(pss != nullptr, "must be initialized");
689
690 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
691 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
692 }
693 }
694
695 void G1ParScanThreadState::record_evacuation_failed_region(G1HeapRegion* r, uint worker_id, bool cause_pinned) {
696 if (_evac_failure_regions->record(worker_id, r->hrm_index(), cause_pinned)) {
697 G1HeapRegionPrinter::evac_failure(r);
698 }
699 }
700
701 NOINLINE
702 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, Klass* klass, G1HeapRegionAttr attr, size_t word_sz, bool cause_pinned) {
703 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
704
705 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
706 if (forward_ptr == nullptr) {
707 // Forward-to-self succeeded. We are the "owner" of the object.
708 G1HeapRegion* r = _g1h->heap_region_containing(old);
709
710 record_evacuation_failed_region(r, _worker_id, cause_pinned);
711
712 // Mark the failing object in the marking bitmap and later use the bitmap to handle
713 // evacuation failure recovery.
714 _g1h->mark_evac_failure_object(old);
715
716 _evacuation_failed_info.register_copy_failure(word_sz);
717
718 {
719 // For iterating objects that failed evacuation currently we can reuse the
720 // existing closure to scan evacuated objects; since we are iterating from a
721 // collection set region (i.e. never a Survivor region), we always need to
722 // gather cards for this case.
723 G1SkipCardMarkSetter x(&_scanner, false /* skip_card_mark */);
724 do_iterate_object(old, old, klass, attr, attr, m.age());
725 }
726
727 return old;
728 } else {
729 // Forward-to-self failed. Either someone else managed to allocate
730 // space for this object (old != forward_ptr) or they beat us in
731 // self-forwarding it (old == forward_ptr).
732 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
733 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
734 "should not be in the CSet",
735 p2i(old), p2i(forward_ptr));
736 return forward_ptr;
737 }
738 }
739
740 void G1ParScanThreadState::update_nmethod_regions_to_add() {
741 _per_thread_states->update_nmethod_regions_to_add(&_nmethods_to_add);
742 }
743
744 void G1ParScanThreadState::initialize_numa_stats() {
745 if (_numa->is_enabled()) {
746 LogTarget(Info, gc, heap, numa) lt;
747
748 if (lt.is_enabled()) {
749 uint num_nodes = _numa->num_active_nodes();
750 // Record only if there are multiple active nodes.
751 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
752 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
753 }
754 }
755 }
756
757 void G1ParScanThreadState::flush_numa_stats() {
758 if (_obj_alloc_stat != nullptr) {
759 uint node_index = _numa->index_of_current_thread();
760 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
761 }
762 }
763
764 void G1ParScanThreadState::update_numa_stats(uint node_index) {
765 if (_obj_alloc_stat != nullptr) {
766 _obj_alloc_stat[node_index]++;
767 }
768 }
769
770 #if TASKQUEUE_STATS
771
772 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
773 return _partial_array_splitter.stats();
774 }
775
776 #endif // TASKQUEUE_STATS
777
778 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
779 uint num_workers,
780 G1CollectionSet* collection_set,
781 G1EvacFailureRegions* evac_failure_regions) :
782 _g1h(g1h),
783 _collection_set(collection_set),
784 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
785 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->num_young_regions() + 1, mtGC)),
786 _num_workers(num_workers),
787 _flushed(false),
788 _evac_failure_regions(evac_failure_regions),
789 _has_nmethods_to_add(g1h->max_num_regions(), mtGC),
790 _num_nmethod_regions_to_add(0),
791 _nmethod_regions_to_add(NEW_C_HEAP_ARRAY(uint, g1h->max_num_regions(), mtGC)) // Conservative length estimation.
792 {
793 for (uint i = 0; i < num_workers; ++i) {
794 _states[i] = nullptr;
795 }
796 memset(_surviving_young_words_total, 0, (collection_set->num_young_regions() + 1) * sizeof(size_t));
797 }
798
799 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
800 for (uint i = 0; i < _num_workers; i++) {
801 assert(_states[i] == nullptr, "must be");
802 }
803 FREE_C_HEAP_ARRAY(_nmethod_regions_to_add);
804 FREE_C_HEAP_ARRAY(_states);
805 FREE_C_HEAP_ARRAY(_surviving_young_words_total);
806 }
807
808 #if TASKQUEUE_STATS
809
810 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
811 auto get_stats = [&](uint i) {
812 return state_for_worker(i)->partial_array_task_stats();
813 };
814 PartialArrayTaskStats::log_set(_num_workers, get_stats,
815 "Young GC Partial Array");
816 }
817
818 #endif // TASKQUEUE_STATS