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 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
59 uint worker_id,
60 uint num_workers,
61 G1CollectionSet* collection_set,
62 G1EvacFailureRegions* evac_failure_regions)
63 : _g1h(g1h),
64 _task_queue(g1h->task_queue(worker_id)),
65 _ct(g1h->refinement_table()),
66 _closures(nullptr),
67 _plab_allocator(nullptr),
68 _age_table(false),
69 _tenuring_threshold(g1h->policy()->tenuring_threshold()),
70 _scanner(g1h, this),
71 _worker_id(worker_id),
72 _num_cards_marked_dirty(0),
73 _num_cards_marked_to_cset(0),
74 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
75 _stack_trim_lower_threshold(GCDrainStackTargetSize),
76 _trim_ticks(),
77 _surviving_young_words_base(nullptr),
78 _surviving_young_words(nullptr),
79 _surviving_words_length(collection_set->young_region_length() + 1),
80 _old_gen_is_full(false),
81 _partial_array_splitter(g1h->partial_array_state_manager(), num_workers, ParGCArrayScanChunk),
82 _string_dedup_requests(),
83 _max_num_optional_regions(collection_set->num_optional_regions()),
84 _numa(g1h->numa()),
85 _obj_alloc_stat(nullptr),
86 ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
87 _evacuation_failed_info(),
88 _evac_failure_regions(evac_failure_regions),
89 _num_cards_from_evac_failure(0)
90 {
91 // We allocate number of young gen regions in the collection set plus one
92 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
93 // We also add a few elements at the beginning and at the end in
94 // an attempt to eliminate cache contention
95 const size_t padding_elem_num = (DEFAULT_PADDING_SIZE / sizeof(size_t));
96 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
97
98 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
99 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
100 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
101
102 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
103
104 _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
105 this,
106 collection_set->only_contains_young_regions());
107
108 _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
109
110 initialize_numa_stats();
111 }
112
113 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
114 flush_numa_stats();
115 // Update allocation statistics.
116 _plab_allocator->flush_and_retire_stats(num_workers);
117 _g1h->policy()->record_age_table(&_age_table);
118
119 if (_evacuation_failed_info.has_failed()) {
120 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
121 }
122
123 size_t sum = 0;
124 for (uint i = 0; i < _surviving_words_length; i++) {
125 surviving_young_words[i] += _surviving_young_words[i];
126 sum += _surviving_young_words[i];
127 }
128 return sum;
129 }
130
131 G1ParScanThreadState::~G1ParScanThreadState() {
132 delete _plab_allocator;
133 delete _closures;
134 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
135 delete[] _oops_into_optional_regions;
136 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
137 }
138
139 size_t G1ParScanThreadState::lab_waste_words() const {
140 return _plab_allocator->waste();
141 }
142
143 size_t G1ParScanThreadState::lab_undo_waste_words() const {
144 return _plab_allocator->undo_waste();
145 }
146
147 size_t G1ParScanThreadState::num_cards_pending() const {
148 return _num_cards_marked_dirty + _num_cards_from_evac_failure;
149 }
150
151 size_t G1ParScanThreadState::num_cards_marked() const {
152 return num_cards_pending() + _num_cards_marked_to_cset;
153 }
154
155 size_t G1ParScanThreadState::num_cards_from_evac_failure() const {
156 return _num_cards_from_evac_failure;
157 }
158
159 #ifdef ASSERT
160 void G1ParScanThreadState::verify_task(narrowOop* task) const {
161 assert(task != nullptr, "invariant");
162 assert(UseCompressedOops, "sanity");
163 oop p = RawAccess<>::oop_load(task);
164 assert(_g1h->is_in_reserved(p),
165 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
166 }
167
168 void G1ParScanThreadState::verify_task(oop* task) const {
169 assert(task != nullptr, "invariant");
170 oop p = RawAccess<>::oop_load(task);
171 assert(_g1h->is_in_reserved(p),
172 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
173 }
174
175 void G1ParScanThreadState::verify_task(PartialArrayState* task) const {
176 assert(task != nullptr, "invariant");
177 // Source isn't used for processing, so not recorded in task.
178 assert(task->source() == nullptr, "invariant");
179 oop p = task->destination();
180 assert(_g1h->is_in_reserved(p),
181 "task=" PTR_FORMAT " dest=" PTR_FORMAT, p2i(task), p2i(p));
182 }
183
184 void G1ParScanThreadState::verify_task(ScannerTask task) const {
185 if (task.is_narrow_oop_ptr()) {
186 verify_task(task.to_narrow_oop_ptr());
187 } else if (task.is_oop_ptr()) {
188 verify_task(task.to_oop_ptr());
189 } else if (task.is_partial_array_state()) {
190 verify_task(task.to_partial_array_state());
191 } else {
192 ShouldNotReachHere();
193 }
194 }
195 #endif // ASSERT
196
197 template <class T>
198 MAYBE_INLINE_EVACUATION
199 void G1ParScanThreadState::do_oop_evac(T* p) {
200 // Reference should not be null here as such are never pushed to the task queue.
201 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
202
203 // Although we never intentionally push references outside of the collection
204 // set, due to (benign) races in the claim mechanism during RSet scanning more
205 // than one thread might claim the same card. So the same card may be
206 // processed multiple times, and so we might get references into old gen here.
207 // So we need to redo this check.
208 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
209 // References pushed onto the work stack should never point to a humongous region
210 // as they are not added to the collection set due to above precondition.
211 assert(!region_attr.is_humongous_candidate(),
212 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
213 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
214
215 if (!region_attr.is_in_cset()) {
216 // In this case somebody else already did all the work.
217 return;
218 }
219
220 markWord m = obj->mark();
221 if (m.is_forwarded()) {
222 obj = obj->forwardee(m);
223 } else {
224 obj = do_copy_to_survivor_space(region_attr, obj, m);
225 }
226 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
227
228 write_ref_field_post(p, obj);
229 }
230
231 MAYBE_INLINE_EVACUATION
232 void G1ParScanThreadState::do_partial_array(PartialArrayState* state, bool stolen) {
233 // Access state before release by claim().
234 objArrayOop to_array = objArrayOop(state->destination());
235 PartialArraySplitter::Claim claim =
236 _partial_array_splitter.claim(state, _task_queue, stolen);
237 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
238 G1SkipCardMarkSetter x(&_scanner, dest_attr.is_new_survivor());
239 // Process claimed task.
240 assert(to_array->is_refArray(), "Must be");
241 refArrayOop(to_array)->oop_iterate_elements_range(&_scanner,
242 checked_cast<int>(claim._start),
243 checked_cast<int>(claim._end));
244 }
245
246 MAYBE_INLINE_EVACUATION
247 void G1ParScanThreadState::start_partial_objarray(oop from_obj,
248 oop to_obj) {
249 assert(from_obj->is_forwarded(), "precondition");
250 assert(from_obj->forwardee() == to_obj, "precondition");
251 assert(to_obj->is_objArray(), "precondition");
252
253 objArrayOop to_array = objArrayOop(to_obj);
254 size_t array_length = to_array->length();
255 size_t initial_chunk_size =
256 // The source array is unused when processing states.
257 _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length);
258
259 assert(_scanner.skip_card_mark_set(), "must be");
260 // Process the initial chunk. No need to process the type in the
261 // klass, as it will already be handled by processing the built-in
262 // module.
263 assert(to_array->is_refArray(), "Must be");
264 refArrayOop(to_array)->oop_iterate_elements_range(&_scanner, 0, checked_cast<int>(initial_chunk_size));
265 }
266
267 MAYBE_INLINE_EVACUATION
268 void G1ParScanThreadState::dispatch_task(ScannerTask task, bool stolen) {
269 verify_task(task);
270 if (task.is_narrow_oop_ptr()) {
271 do_oop_evac(task.to_narrow_oop_ptr());
272 } else if (task.is_oop_ptr()) {
273 do_oop_evac(task.to_oop_ptr());
274 } else {
275 do_partial_array(task.to_partial_array_state(), stolen);
276 }
277 }
278
279 // Process tasks until overflow queue is empty and local queue
280 // contains no more than threshold entries. NOINLINE to prevent
281 // inlining into steal_and_trim_queue.
282 ATTRIBUTE_FLATTEN NOINLINE
283 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
284 ScannerTask task;
285 do {
286 while (_task_queue->pop_overflow(task)) {
287 if (!_task_queue->try_push_to_taskqueue(task)) {
288 dispatch_task(task, false);
289 }
290 }
291 while (_task_queue->pop_local(task, threshold)) {
292 dispatch_task(task, false);
293 }
294 } while (!_task_queue->overflow_empty());
295 }
296
297 ATTRIBUTE_FLATTEN
298 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
299 ScannerTask stolen_task;
300 while (task_queues->steal(_worker_id, stolen_task)) {
301 dispatch_task(stolen_task, true);
302 // Processing stolen task may have added tasks to our queue.
303 trim_queue();
304 }
305 }
306
307 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
308 size_t word_sz,
309 bool previous_plab_refill_failed,
310 uint node_index) {
311
312 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
313
314 // Right now we only have two types of regions (young / old) so
315 // let's keep the logic here simple. We can generalize it when necessary.
316 if (dest->is_young()) {
317 bool plab_refill_in_old_failed = false;
318 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
319 word_sz,
320 &plab_refill_in_old_failed,
321 node_index);
322 // Make sure that we won't attempt to copy any other objects out
323 // of a survivor region (given that apparently we cannot allocate
324 // any new ones) to avoid coming into this slow path again and again.
325 // Only consider failed PLAB refill here: failed inline allocations are
326 // typically large, so not indicative of remaining space.
327 if (previous_plab_refill_failed) {
328 _tenuring_threshold = 0;
329 }
330
331 if (obj_ptr != nullptr) {
332 dest->set_old();
333 } else {
334 // We just failed to allocate in old gen. The same idea as explained above
335 // for making survivor gen unavailable for allocation applies for old gen.
336 _old_gen_is_full = plab_refill_in_old_failed;
337 }
338 return obj_ptr;
339 } else {
340 _old_gen_is_full = previous_plab_refill_failed;
341 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
342 // no other space to try.
343 return nullptr;
344 }
345 }
346
347 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
348 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
349
350 if (region_attr.is_young()) {
351 age = !m.has_displaced_mark_helper() ? m.age()
352 : m.displaced_mark_helper().age();
353 if (age < _tenuring_threshold) {
354 return region_attr;
355 }
356 }
357 // young-to-old (promotion) or old-to-old; destination is old in both cases.
358 return G1HeapRegionAttr::Old;
359 }
360
361 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
362 Klass* klass, size_t word_sz, uint age,
363 HeapWord * const obj_ptr, uint node_index) const {
364 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
365 if (alloc_buf->contains(obj_ptr)) {
366 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
367 dest_attr.type() == G1HeapRegionAttr::Old,
368 alloc_buf->word_sz() * HeapWordSize);
369 } else {
370 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
371 dest_attr.type() == G1HeapRegionAttr::Old);
372 }
373 }
374
375 NOINLINE
376 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
377 Klass* klass,
378 size_t word_sz,
379 uint age,
380 uint node_index) {
381 HeapWord* obj_ptr = nullptr;
382 // Try slow-path allocation unless we're allocating old and old is already full.
383 if (!(dest_attr->is_old() && _old_gen_is_full)) {
384 bool plab_refill_failed = false;
385 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
386 word_sz,
387 &plab_refill_failed,
388 node_index);
389 if (obj_ptr == nullptr) {
390 obj_ptr = allocate_in_next_plab(dest_attr,
391 word_sz,
392 plab_refill_failed,
393 node_index);
394 }
395 }
396 if (obj_ptr != nullptr) {
397 update_numa_stats(node_index);
398 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
399 // The events are checked individually as part of the actual commit
400 report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
401 }
402 }
403 return obj_ptr;
404 }
405
406 #if ALLOCATION_FAILURE_INJECTOR
407 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
408 return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
409 }
410 #endif
411
412 NOINLINE
413 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
414 HeapWord* obj_ptr,
415 size_t word_sz,
416 uint node_index) {
417 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
418 }
419
420 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
421 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
422 G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
423 region->update_bot_for_block(obj_start, obj_start + word_sz);
424 }
425
426 ALWAYSINLINE
427 void G1ParScanThreadState::do_iterate_object(oop const obj,
428 oop const old,
429 Klass* const klass,
430 G1HeapRegionAttr const region_attr,
431 G1HeapRegionAttr const dest_attr,
432 uint age) {
433 // Most objects are not arrays, so do one array check rather than
434 // checking for each array category for each object.
435 if (klass->is_array_klass() && !klass->is_flatArray_klass()) {
436 assert(!klass->is_stack_chunk_instance_klass(), "must be");
437
438 if (klass->is_refArray_klass()) {
439 start_partial_objarray(old, obj);
440 } else {
441 // Nothing needs to be done for typeArrays. Body doesn't contain
442 // any oops to scan, and the type in the klass will already be handled
443 // by processing the built-in module.
444 assert(klass->is_typeArray_klass() || klass->is_objArray_klass(), "invariant");
445 }
446 return;
447 }
448
449 ContinuationGCSupport::transform_stack_chunk(obj);
450
451 // Check for deduplicating young Strings.
452 if (G1StringDedup::is_candidate_from_evacuation(klass,
453 region_attr,
454 dest_attr,
455 age)) {
456 // Record old; request adds a new weak reference, which reference
457 // processing expects to refer to a from-space object.
458 _string_dedup_requests.add(old);
459 }
460
461 assert(_scanner.skip_card_mark_set(), "must be");
462 obj->oop_iterate_backwards(&_scanner, klass);
463 }
464
465 // Private inline function, for direct internal use and providing the
466 // implementation of the public not-inline function.
467 MAYBE_INLINE_EVACUATION
468 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
469 oop const old,
470 markWord const old_mark) {
471 assert(region_attr.is_in_cset(),
472 "Unexpected region attr type: %s", region_attr.get_type_str());
473
474 // NOTE: With compact headers, it is not safe to load the Klass* from old, because
475 // that would access the mark-word, that might change at any time by concurrent
476 // workers.
477 // This mark word would refer to a forwardee, which may not yet have completed
478 // copying. Therefore we must load the Klass* from the mark-word that we already
479 // loaded. This is safe, because we only enter here if not yet forwarded.
480 assert(!old_mark.is_forwarded(), "precondition");
481 Klass* klass = UseCompactObjectHeaders
482 ? old_mark.klass()
483 : old->klass();
484
485 const size_t word_sz = old->size_given_klass(klass);
486
487 // JNI only allows pinning of typeArrays, so we only need to keep those in place.
488 if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
489 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, true /* cause_pinned */);
490 }
491
492 uint age = 0;
493 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
494 G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
495 uint node_index = from_region->node_index();
496
497 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
498
499 // PLAB allocations should succeed most of the time, so we'll
500 // normally check against null once and that's it.
501 if (obj_ptr == nullptr) {
502 obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
503 if (obj_ptr == nullptr) {
504 // This will either forward-to-self, or detect that someone else has
505 // installed a forwarding pointer.
506 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
507 }
508 }
509
510 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
511 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
512
513 // Should this evacuation fail?
514 if (inject_allocation_failure(from_region->hrm_index())) {
515 // Doing this after all the allocation attempts also tests the
516 // undo_allocation() method too.
517 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
518 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
519 }
520
521 // We're going to allocate linearly, so might as well prefetch ahead.
522 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
523 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
524
525 const oop obj = cast_to_oop(obj_ptr);
526 // Because the forwarding is done with memory_order_relaxed there is no
527 // ordering with the above copy. Clients that get the forwardee must not
528 // examine its contents without other synchronization, since the contents
529 // may not be up to date for them.
530 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
531 if (forward_ptr == nullptr) {
532
533 {
534 const uint young_index = from_region->young_index_in_cset();
535 assert((from_region->is_young() && young_index > 0) ||
536 (!from_region->is_young() && young_index == 0), "invariant" );
537 _surviving_young_words[young_index] += word_sz;
538 }
539
540 if (dest_attr.is_young()) {
541 if (age < markWord::max_age) {
542 age++;
543 obj->incr_age();
544 }
545 _age_table.add(age, word_sz);
546 } else {
547 update_bot_after_copying(obj, word_sz);
548 }
549
550 {
551 // Skip the card enqueue iff the object (obj) is in survivor region.
552 // However, G1HeapRegion::is_survivor() is too expensive here.
553 // Instead, we use dest_attr.is_young() because the two values are always
554 // equal: successfully allocated young regions must be survivor regions.
555 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
556 G1SkipCardMarkSetter x(&_scanner, dest_attr.is_young());
557 do_iterate_object(obj, old, klass, region_attr, dest_attr, age);
558 }
559
560 return obj;
561 } else {
562 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
563 return forward_ptr;
564 }
565 }
566
567 // Public not-inline entry point.
568 ATTRIBUTE_FLATTEN
569 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
570 oop old,
571 markWord old_mark) {
572 return do_copy_to_survivor_space(region_attr, old, old_mark);
573 }
574
575 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
576 assert(worker_id < _num_workers, "out of bounds access");
577 if (_states[worker_id] == nullptr) {
578 _states[worker_id] =
579 new G1ParScanThreadState(_g1h,
580 worker_id,
581 _num_workers,
582 _collection_set,
583 _evac_failure_regions);
584 }
585 return _states[worker_id];
586 }
587
588 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
589 assert(_flushed, "thread local state from the per thread states should have been flushed");
590 return _surviving_young_words_total;
591 }
592
593 void G1ParScanThreadStateSet::flush_stats() {
594 assert(!_flushed, "thread local state from the per thread states should be flushed once");
595 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
596 G1ParScanThreadState* pss = _states[worker_id];
597 assert(pss != nullptr, "must be initialized");
598
599 G1GCPhaseTimes* p = _g1h->phase_times();
600
601 // Need to get the following two before the call to G1ParThreadScanState::flush()
602 // because it resets the PLAB allocator where we get this info from.
603 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
604 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
605 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers) * HeapWordSize;
606 size_t pending_cards = pss->num_cards_pending();
607 size_t to_young_gen_cards = pss->num_cards_marked() - pss->num_cards_pending();
608 size_t evac_failure_cards = pss->num_cards_from_evac_failure();
609 size_t marked_cards = pss->num_cards_marked();
610
611 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
612 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
613 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
614 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, pending_cards, G1GCPhaseTimes::MergePSSPendingCards);
615 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, to_young_gen_cards, G1GCPhaseTimes::MergePSSToYoungGenCards);
616 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_failure_cards, G1GCPhaseTimes::MergePSSEvacFail);
617 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, marked_cards, G1GCPhaseTimes::MergePSSMarked);
618
619 delete pss;
620 _states[worker_id] = nullptr;
621 }
622
623 _flushed = true;
624 }
625
626 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
627 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
628 G1ParScanThreadState* pss = _states[worker_index];
629 assert(pss != nullptr, "must be initialized");
630
631 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
632 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
633 }
634 }
635
636 void G1ParScanThreadState::record_evacuation_failed_region(G1HeapRegion* r, uint worker_id, bool cause_pinned) {
637 if (_evac_failure_regions->record(worker_id, r->hrm_index(), cause_pinned)) {
638 G1HeapRegionPrinter::evac_failure(r);
639 }
640 }
641
642 NOINLINE
643 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, Klass* klass, G1HeapRegionAttr attr, size_t word_sz, bool cause_pinned) {
644 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
645
646 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
647 if (forward_ptr == nullptr) {
648 // Forward-to-self succeeded. We are the "owner" of the object.
649 G1HeapRegion* r = _g1h->heap_region_containing(old);
650
651 record_evacuation_failed_region(r, _worker_id, cause_pinned);
652
653 // Mark the failing object in the marking bitmap and later use the bitmap to handle
654 // evacuation failure recovery.
655 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
656
657 _evacuation_failed_info.register_copy_failure(word_sz);
658
659 {
660 // For iterating objects that failed evacuation currently we can reuse the
661 // existing closure to scan evacuated objects; since we are iterating from a
662 // collection set region (i.e. never a Survivor region), we always need to
663 // gather cards for this case.
664 G1SkipCardMarkSetter x(&_scanner, false /* skip_card_mark */);
665 do_iterate_object(old, old, klass, attr, attr, m.age());
666 }
667
668 return old;
669 } else {
670 // Forward-to-self failed. Either someone else managed to allocate
671 // space for this object (old != forward_ptr) or they beat us in
672 // self-forwarding it (old == forward_ptr).
673 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
674 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
675 "should not be in the CSet",
676 p2i(old), p2i(forward_ptr));
677 return forward_ptr;
678 }
679 }
680
681 void G1ParScanThreadState::initialize_numa_stats() {
682 if (_numa->is_enabled()) {
683 LogTarget(Info, gc, heap, numa) lt;
684
685 if (lt.is_enabled()) {
686 uint num_nodes = _numa->num_active_nodes();
687 // Record only if there are multiple active nodes.
688 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
689 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
690 }
691 }
692 }
693
694 void G1ParScanThreadState::flush_numa_stats() {
695 if (_obj_alloc_stat != nullptr) {
696 uint node_index = _numa->index_of_current_thread();
697 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
698 }
699 }
700
701 void G1ParScanThreadState::update_numa_stats(uint node_index) {
702 if (_obj_alloc_stat != nullptr) {
703 _obj_alloc_stat[node_index]++;
704 }
705 }
706
707 #if TASKQUEUE_STATS
708
709 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
710 return _partial_array_splitter.stats();
711 }
712
713 #endif // TASKQUEUE_STATS
714
715 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
716 uint num_workers,
717 G1CollectionSet* collection_set,
718 G1EvacFailureRegions* evac_failure_regions) :
719 _g1h(g1h),
720 _collection_set(collection_set),
721 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
722 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
723 _num_workers(num_workers),
724 _flushed(false),
725 _evac_failure_regions(evac_failure_regions)
726 {
727 for (uint i = 0; i < num_workers; ++i) {
728 _states[i] = nullptr;
729 }
730 memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
731 }
732
733 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
734 assert(_flushed, "thread local state from the per thread states should have been flushed");
735 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
736 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
737 }
738
739 #if TASKQUEUE_STATS
740
741 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
742 auto get_stats = [&](uint i) {
743 return state_for_worker(i)->partial_array_task_stats();
744 };
745 PartialArrayTaskStats::log_set(_num_workers, get_stats,
746 "Partial Array Task Stats");
747 }
748
749 #endif // TASKQUEUE_STATS