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 to_array->oop_iterate_elements_range(&_scanner,
241 checked_cast<int>(claim._start),
242 checked_cast<int>(claim._end));
243 }
244
245 MAYBE_INLINE_EVACUATION
246 void G1ParScanThreadState::start_partial_objarray(oop from_obj,
247 oop to_obj) {
248 assert(from_obj->is_forwarded(), "precondition");
249 assert(from_obj->forwardee() == to_obj, "precondition");
250 assert(to_obj->is_objArray(), "precondition");
251
252 objArrayOop to_array = objArrayOop(to_obj);
253 size_t array_length = to_array->length();
254 size_t initial_chunk_size =
255 // The source array is unused when processing states.
256 _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length);
257
258 assert(_scanner.skip_card_mark_set(), "must be");
259 // Process the initial chunk. No need to process the type in the
260 // klass, as it will already be handled by processing the built-in
261 // module.
262 to_array->oop_iterate_elements_range(&_scanner, 0, checked_cast<int>(initial_chunk_size));
263 }
264
265 MAYBE_INLINE_EVACUATION
266 void G1ParScanThreadState::dispatch_task(ScannerTask task, bool stolen) {
267 verify_task(task);
268 if (task.is_narrow_oop_ptr()) {
269 do_oop_evac(task.to_narrow_oop_ptr());
270 } else if (task.is_oop_ptr()) {
271 do_oop_evac(task.to_oop_ptr());
272 } else {
273 do_partial_array(task.to_partial_array_state(), stolen);
274 }
275 }
276
277 // Process tasks until overflow queue is empty and local queue
278 // contains no more than threshold entries. NOINLINE to prevent
279 // inlining into steal_and_trim_queue.
280 ATTRIBUTE_FLATTEN NOINLINE
281 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
282 ScannerTask task;
283 do {
284 while (_task_queue->pop_overflow(task)) {
285 if (!_task_queue->try_push_to_taskqueue(task)) {
286 dispatch_task(task, false);
287 }
288 }
289 while (_task_queue->pop_local(task, threshold)) {
290 dispatch_task(task, false);
291 }
292 } while (!_task_queue->overflow_empty());
293 }
294
295 ATTRIBUTE_FLATTEN
296 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
297 ScannerTask stolen_task;
298 while (task_queues->steal(_worker_id, stolen_task)) {
299 dispatch_task(stolen_task, true);
300 // Processing stolen task may have added tasks to our queue.
301 trim_queue();
302 }
303 }
304
305 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
306 size_t word_sz,
307 bool previous_plab_refill_failed,
308 uint node_index) {
309
310 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
311
312 // Right now we only have two types of regions (young / old) so
313 // let's keep the logic here simple. We can generalize it when necessary.
314 if (dest->is_young()) {
315 bool plab_refill_in_old_failed = false;
316 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
317 word_sz,
318 &plab_refill_in_old_failed,
319 node_index);
320 // Make sure that we won't attempt to copy any other objects out
321 // of a survivor region (given that apparently we cannot allocate
322 // any new ones) to avoid coming into this slow path again and again.
323 // Only consider failed PLAB refill here: failed inline allocations are
324 // typically large, so not indicative of remaining space.
325 if (previous_plab_refill_failed) {
326 _tenuring_threshold = 0;
327 }
328
329 if (obj_ptr != nullptr) {
330 dest->set_old();
331 } else {
332 // We just failed to allocate in old gen. The same idea as explained above
333 // for making survivor gen unavailable for allocation applies for old gen.
334 _old_gen_is_full = plab_refill_in_old_failed;
335 }
336 return obj_ptr;
337 } else {
338 _old_gen_is_full = previous_plab_refill_failed;
339 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
340 // no other space to try.
341 return nullptr;
342 }
343 }
344
345 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
346 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
347
348 if (region_attr.is_young()) {
349 age = !m.has_displaced_mark_helper() ? m.age()
350 : m.displaced_mark_helper().age();
351 if (age < _tenuring_threshold) {
352 return region_attr;
353 }
354 }
355 // young-to-old (promotion) or old-to-old; destination is old in both cases.
356 return G1HeapRegionAttr::Old;
357 }
358
359 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
360 Klass* klass, size_t word_sz, uint age,
361 HeapWord * const obj_ptr, uint node_index) const {
362 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
363 if (alloc_buf->contains(obj_ptr)) {
364 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
365 dest_attr.type() == G1HeapRegionAttr::Old,
366 alloc_buf->word_sz() * HeapWordSize);
367 } else {
368 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
369 dest_attr.type() == G1HeapRegionAttr::Old);
370 }
371 }
372
373 NOINLINE
374 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
375 Klass* klass,
376 size_t word_sz,
377 uint age,
378 uint node_index) {
379 HeapWord* obj_ptr = nullptr;
380 // Try slow-path allocation unless we're allocating old and old is already full.
381 if (!(dest_attr->is_old() && _old_gen_is_full)) {
382 bool plab_refill_failed = false;
383 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
384 word_sz,
385 &plab_refill_failed,
386 node_index);
387 if (obj_ptr == nullptr) {
388 obj_ptr = allocate_in_next_plab(dest_attr,
389 word_sz,
390 plab_refill_failed,
391 node_index);
392 }
393 }
394 if (obj_ptr != nullptr) {
395 update_numa_stats(node_index);
396 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
397 // The events are checked individually as part of the actual commit
398 report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
399 }
400 }
401 return obj_ptr;
402 }
403
404 #if ALLOCATION_FAILURE_INJECTOR
405 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
406 return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
407 }
408 #endif
409
410 NOINLINE
411 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
412 HeapWord* obj_ptr,
413 size_t word_sz,
414 uint node_index) {
415 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
416 }
417
418 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
419 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
420 G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
421 region->update_bot_for_block(obj_start, obj_start + word_sz);
422 }
423
424 ALWAYSINLINE
425 void G1ParScanThreadState::do_iterate_object(oop const obj,
426 oop const old,
427 Klass* const klass,
428 G1HeapRegionAttr const region_attr,
429 G1HeapRegionAttr const dest_attr,
430 uint age) {
431 // Most objects are not arrays, so do one array check rather than
432 // checking for each array category for each object.
433 if (klass->is_array_klass()) {
434 assert(!klass->is_stack_chunk_instance_klass(), "must be");
435
436 if (klass->is_objArray_klass()) {
437 start_partial_objarray(old, obj);
438 } else {
439 // Nothing needs to be done for typeArrays. Body doesn't contain
440 // any oops to scan, and the type in the klass will already be handled
441 // by processing the built-in module.
442 assert(klass->is_typeArray_klass(), "invariant");
443 }
444 return;
445 }
446
447 ContinuationGCSupport::transform_stack_chunk(obj);
448
449 // Check for deduplicating young Strings.
450 if (G1StringDedup::is_candidate_from_evacuation(klass,
451 region_attr,
452 dest_attr,
453 age)) {
454 // Record old; request adds a new weak reference, which reference
455 // processing expects to refer to a from-space object.
456 _string_dedup_requests.add(old);
457 }
458
459 assert(_scanner.skip_card_mark_set(), "must be");
460 obj->oop_iterate_backwards(&_scanner, klass);
461 }
462
463 // Private inline function, for direct internal use and providing the
464 // implementation of the public not-inline function.
465 MAYBE_INLINE_EVACUATION
466 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
467 oop const old,
468 markWord const old_mark) {
469 assert(region_attr.is_in_cset(),
470 "Unexpected region attr type: %s", region_attr.get_type_str());
471
472 // NOTE: With compact headers, it is not safe to load the Klass* from old, because
473 // that would access the mark-word, that might change at any time by concurrent
474 // workers.
475 // This mark word would refer to a forwardee, which may not yet have completed
476 // copying. Therefore we must load the Klass* from the mark-word that we already
477 // loaded. This is safe, because we only enter here if not yet forwarded.
478 assert(!old_mark.is_forwarded(), "precondition");
479 Klass* klass = UseCompactObjectHeaders
480 ? old_mark.klass()
481 : old->klass();
482
483 const size_t word_sz = old->size_given_klass(klass);
484
485 // JNI only allows pinning of typeArrays, so we only need to keep those in place.
486 if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
487 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, true /* cause_pinned */);
488 }
489
490 uint age = 0;
491 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
492 G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
493 uint node_index = from_region->node_index();
494
495 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
496
497 // PLAB allocations should succeed most of the time, so we'll
498 // normally check against null once and that's it.
499 if (obj_ptr == nullptr) {
500 obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
501 if (obj_ptr == nullptr) {
502 // This will either forward-to-self, or detect that someone else has
503 // installed a forwarding pointer.
504 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
505 }
506 }
507
508 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
509 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
510
511 // Should this evacuation fail?
512 if (inject_allocation_failure(from_region->hrm_index())) {
513 // Doing this after all the allocation attempts also tests the
514 // undo_allocation() method too.
515 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
516 return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
517 }
518
519 // We're going to allocate linearly, so might as well prefetch ahead.
520 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
521 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
522
523 const oop obj = cast_to_oop(obj_ptr);
524 // Because the forwarding is done with memory_order_relaxed there is no
525 // ordering with the above copy. Clients that get the forwardee must not
526 // examine its contents without other synchronization, since the contents
527 // may not be up to date for them.
528 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
529 if (forward_ptr == nullptr) {
530
531 {
532 const uint young_index = from_region->young_index_in_cset();
533 assert((from_region->is_young() && young_index > 0) ||
534 (!from_region->is_young() && young_index == 0), "invariant" );
535 _surviving_young_words[young_index] += word_sz;
536 }
537
538 if (dest_attr.is_young()) {
539 if (age < markWord::max_age) {
540 age++;
541 obj->incr_age();
542 }
543 _age_table.add(age, word_sz);
544 } else {
545 update_bot_after_copying(obj, word_sz);
546 }
547
548 {
549 // Skip the card enqueue iff the object (obj) is in survivor region.
550 // However, G1HeapRegion::is_survivor() is too expensive here.
551 // Instead, we use dest_attr.is_young() because the two values are always
552 // equal: successfully allocated young regions must be survivor regions.
553 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
554 G1SkipCardMarkSetter x(&_scanner, dest_attr.is_young());
555 do_iterate_object(obj, old, klass, region_attr, dest_attr, age);
556 }
557
558 return obj;
559 } else {
560 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
561 return forward_ptr;
562 }
563 }
564
565 // Public not-inline entry point.
566 ATTRIBUTE_FLATTEN
567 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
568 oop old,
569 markWord old_mark) {
570 return do_copy_to_survivor_space(region_attr, old, old_mark);
571 }
572
573 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
574 assert(worker_id < _num_workers, "out of bounds access");
575 if (_states[worker_id] == nullptr) {
576 _states[worker_id] =
577 new G1ParScanThreadState(_g1h,
578 worker_id,
579 _num_workers,
580 _collection_set,
581 _evac_failure_regions);
582 }
583 return _states[worker_id];
584 }
585
586 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
587 assert(_flushed, "thread local state from the per thread states should have been flushed");
588 return _surviving_young_words_total;
589 }
590
591 void G1ParScanThreadStateSet::flush_stats() {
592 assert(!_flushed, "thread local state from the per thread states should be flushed once");
593 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
594 G1ParScanThreadState* pss = _states[worker_id];
595 assert(pss != nullptr, "must be initialized");
596
597 G1GCPhaseTimes* p = _g1h->phase_times();
598
599 // Need to get the following two before the call to G1ParThreadScanState::flush()
600 // because it resets the PLAB allocator where we get this info from.
601 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
602 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
603 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers) * HeapWordSize;
604 size_t pending_cards = pss->num_cards_pending();
605 size_t to_young_gen_cards = pss->num_cards_marked() - pss->num_cards_pending();
606 size_t evac_failure_cards = pss->num_cards_from_evac_failure();
607 size_t marked_cards = pss->num_cards_marked();
608
609 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
610 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
611 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
612 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, pending_cards, G1GCPhaseTimes::MergePSSPendingCards);
613 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, to_young_gen_cards, G1GCPhaseTimes::MergePSSToYoungGenCards);
614 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_failure_cards, G1GCPhaseTimes::MergePSSEvacFail);
615 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, marked_cards, G1GCPhaseTimes::MergePSSMarked);
616
617 delete pss;
618 _states[worker_id] = nullptr;
619 }
620
621 _flushed = true;
622 }
623
624 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
625 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
626 G1ParScanThreadState* pss = _states[worker_index];
627 assert(pss != nullptr, "must be initialized");
628
629 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
630 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
631 }
632 }
633
634 void G1ParScanThreadState::record_evacuation_failed_region(G1HeapRegion* r, uint worker_id, bool cause_pinned) {
635 if (_evac_failure_regions->record(worker_id, r->hrm_index(), cause_pinned)) {
636 G1HeapRegionPrinter::evac_failure(r);
637 }
638 }
639
640 NOINLINE
641 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, Klass* klass, G1HeapRegionAttr attr, size_t word_sz, bool cause_pinned) {
642 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
643
644 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
645 if (forward_ptr == nullptr) {
646 // Forward-to-self succeeded. We are the "owner" of the object.
647 G1HeapRegion* r = _g1h->heap_region_containing(old);
648
649 record_evacuation_failed_region(r, _worker_id, cause_pinned);
650
651 // Mark the failing object in the marking bitmap and later use the bitmap to handle
652 // evacuation failure recovery.
653 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
654
655 _evacuation_failed_info.register_copy_failure(word_sz);
656
657 {
658 // For iterating objects that failed evacuation currently we can reuse the
659 // existing closure to scan evacuated objects; since we are iterating from a
660 // collection set region (i.e. never a Survivor region), we always need to
661 // gather cards for this case.
662 G1SkipCardMarkSetter x(&_scanner, false /* skip_card_mark */);
663 do_iterate_object(old, old, klass, attr, attr, m.age());
664 }
665
666 return old;
667 } else {
668 // Forward-to-self failed. Either someone else managed to allocate
669 // space for this object (old != forward_ptr) or they beat us in
670 // self-forwarding it (old == forward_ptr).
671 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
672 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
673 "should not be in the CSet",
674 p2i(old), p2i(forward_ptr));
675 return forward_ptr;
676 }
677 }
678
679 void G1ParScanThreadState::initialize_numa_stats() {
680 if (_numa->is_enabled()) {
681 LogTarget(Info, gc, heap, numa) lt;
682
683 if (lt.is_enabled()) {
684 uint num_nodes = _numa->num_active_nodes();
685 // Record only if there are multiple active nodes.
686 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
687 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
688 }
689 }
690 }
691
692 void G1ParScanThreadState::flush_numa_stats() {
693 if (_obj_alloc_stat != nullptr) {
694 uint node_index = _numa->index_of_current_thread();
695 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
696 }
697 }
698
699 void G1ParScanThreadState::update_numa_stats(uint node_index) {
700 if (_obj_alloc_stat != nullptr) {
701 _obj_alloc_stat[node_index]++;
702 }
703 }
704
705 #if TASKQUEUE_STATS
706
707 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
708 return _partial_array_splitter.stats();
709 }
710
711 #endif // TASKQUEUE_STATS
712
713 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
714 uint num_workers,
715 G1CollectionSet* collection_set,
716 G1EvacFailureRegions* evac_failure_regions) :
717 _g1h(g1h),
718 _collection_set(collection_set),
719 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
720 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
721 _num_workers(num_workers),
722 _flushed(false),
723 _evac_failure_regions(evac_failure_regions)
724 {
725 for (uint i = 0; i < num_workers; ++i) {
726 _states[i] = nullptr;
727 }
728 memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
729 }
730
731 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
732 assert(_flushed, "thread local state from the per thread states should have been flushed");
733 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
734 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
735 }
736
737 #if TASKQUEUE_STATS
738
739 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
740 auto get_stats = [&](uint i) {
741 return state_for_worker(i)->partial_array_task_stats();
742 };
743 PartialArrayTaskStats::log_set(_num_workers, get_stats,
744 "Partial Array Task Stats");
745 }
746
747 #endif // TASKQUEUE_STATS