1 /*
2 * Copyright (c) 2014, 2024, 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 "precompiled.hpp"
26 #include "gc/g1/g1Allocator.inline.hpp"
27 #include "gc/g1/g1CollectedHeap.inline.hpp"
28 #include "gc/g1/g1CollectionSet.hpp"
29 #include "gc/g1/g1EvacFailureRegions.inline.hpp"
30 #include "gc/g1/g1HeapRegionPrinter.hpp"
31 #include "gc/g1/g1OopClosures.inline.hpp"
32 #include "gc/g1/g1ParScanThreadState.inline.hpp"
33 #include "gc/g1/g1RootClosures.hpp"
34 #include "gc/g1/g1StringDedup.hpp"
35 #include "gc/g1/g1Trace.hpp"
36 #include "gc/g1/g1YoungGCAllocationFailureInjector.inline.hpp"
37 #include "gc/shared/continuationGCSupport.inline.hpp"
38 #include "gc/shared/partialArrayState.hpp"
39 #include "gc/shared/partialArrayTaskStepper.inline.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/atomic.hpp"
46 #include "runtime/mutexLocker.hpp"
47 #include "runtime/prefetch.inline.hpp"
48 #include "utilities/globalDefinitions.hpp"
49 #include "utilities/macros.hpp"
50
51 // In fastdebug builds the code size can get out of hand, potentially
52 // tripping over compiler limits (which may be bugs, but nevertheless
53 // need to be taken into consideration). A side benefit of limiting
54 // inlining is that we get more call frames that might aid debugging.
55 // And the fastdebug compile time for this file is much reduced.
56 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
57 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
58
59 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
60 G1RedirtyCardsQueueSet* rdcqs,
61 uint worker_id,
62 uint num_workers,
63 G1CollectionSet* collection_set,
64 G1EvacFailureRegions* evac_failure_regions)
65 : _g1h(g1h),
66 _task_queue(g1h->task_queue(worker_id)),
67 _rdc_local_qset(rdcqs),
68 _ct(g1h->card_table()),
69 _closures(nullptr),
70 _plab_allocator(nullptr),
71 _age_table(false),
72 _tenuring_threshold(g1h->policy()->tenuring_threshold()),
73 _scanner(g1h, this),
74 _worker_id(worker_id),
75 _last_enqueued_card(SIZE_MAX),
76 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
77 _stack_trim_lower_threshold(GCDrainStackTargetSize),
78 _trim_ticks(),
79 _surviving_young_words_base(nullptr),
80 _surviving_young_words(nullptr),
81 _surviving_words_length(collection_set->young_region_length() + 1),
82 _old_gen_is_full(false),
83 _partial_array_state_allocator(g1h->partial_array_state_manager()),
84 _partial_array_stepper(num_workers, ParGCArrayScanChunk),
85 _string_dedup_requests(),
86 _max_num_optional_regions(collection_set->optional_region_length()),
87 _numa(g1h->numa()),
88 _obj_alloc_stat(nullptr),
89 ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
90 _evacuation_failed_info(),
91 _evac_failure_regions(evac_failure_regions),
92 _evac_failure_enqueued_cards(0)
93 {
94 // We allocate number of young gen regions in the collection set plus one
95 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
96 // We also add a few elements at the beginning and at the end in
97 // an attempt to eliminate cache contention
98 const size_t padding_elem_num = (DEFAULT_PADDING_SIZE / sizeof(size_t));
99 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
100
101 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
102 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
103 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
104
105 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
106
107 _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
108 this,
109 collection_set->only_contains_young_regions());
110
111 _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
112
113 initialize_numa_stats();
114 }
115
116 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers, BufferNodeList* rdc_buffers) {
117 *rdc_buffers = _rdc_local_qset.flush();
118 flush_numa_stats();
119 // Update allocation statistics.
120 _plab_allocator->flush_and_retire_stats(num_workers);
121 _g1h->policy()->record_age_table(&_age_table);
122
123 if (_evacuation_failed_info.has_failed()) {
124 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
125 }
126
127 size_t sum = 0;
128 for (uint i = 0; i < _surviving_words_length; i++) {
129 surviving_young_words[i] += _surviving_young_words[i];
130 sum += _surviving_young_words[i];
131 }
132 return sum;
133 }
134
135 G1ParScanThreadState::~G1ParScanThreadState() {
136 delete _plab_allocator;
137 delete _closures;
138 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
139 delete[] _oops_into_optional_regions;
140 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
141 }
142
143 size_t G1ParScanThreadState::lab_waste_words() const {
144 return _plab_allocator->waste();
145 }
146
147 size_t G1ParScanThreadState::lab_undo_waste_words() const {
148 return _plab_allocator->undo_waste();
149 }
150
151 size_t G1ParScanThreadState::evac_failure_enqueued_cards() const {
152 return _evac_failure_enqueued_cards;
153 }
154
155 #ifdef ASSERT
156 void G1ParScanThreadState::verify_task(narrowOop* task) const {
157 assert(task != nullptr, "invariant");
158 assert(UseCompressedOops, "sanity");
159 oop p = RawAccess<>::oop_load(task);
160 assert(_g1h->is_in_reserved(p),
161 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
162 }
163
164 void G1ParScanThreadState::verify_task(oop* task) const {
165 assert(task != nullptr, "invariant");
166 oop p = RawAccess<>::oop_load(task);
167 assert(_g1h->is_in_reserved(p),
168 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
169 }
170
171 void G1ParScanThreadState::verify_task(PartialArrayState* task) const {
172 // Must be in the collection set--it's already been copied.
173 oop p = task->source();
174 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
175 }
176
177 void G1ParScanThreadState::verify_task(ScannerTask task) const {
178 if (task.is_narrow_oop_ptr()) {
179 verify_task(task.to_narrow_oop_ptr());
180 } else if (task.is_oop_ptr()) {
181 verify_task(task.to_oop_ptr());
182 } else if (task.is_partial_array_state()) {
183 verify_task(task.to_partial_array_state());
184 } else {
185 ShouldNotReachHere();
186 }
187 }
188 #endif // ASSERT
189
190 template <class T>
191 MAYBE_INLINE_EVACUATION
192 void G1ParScanThreadState::do_oop_evac(T* p) {
193 // Reference should not be null here as such are never pushed to the task queue.
194 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
195
196 // Although we never intentionally push references outside of the collection
197 // set, due to (benign) races in the claim mechanism during RSet scanning more
198 // than one thread might claim the same card. So the same card may be
199 // processed multiple times, and so we might get references into old gen here.
200 // So we need to redo this check.
201 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
202 // References pushed onto the work stack should never point to a humongous region
203 // as they are not added to the collection set due to above precondition.
204 assert(!region_attr.is_humongous_candidate(),
205 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
206 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
207
208 if (!region_attr.is_in_cset()) {
209 // In this case somebody else already did all the work.
210 return;
211 }
212
213 markWord m = obj->mark();
214 if (m.is_forwarded()) {
215 obj = obj->forwardee(m);
216 } else {
217 obj = do_copy_to_survivor_space(region_attr, obj, m);
218 }
219 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
220
221 write_ref_field_post(p, obj);
222 }
223
224 MAYBE_INLINE_EVACUATION
225 void G1ParScanThreadState::do_partial_array(PartialArrayState* state) {
226 oop to_obj = state->destination();
227
228 #ifdef ASSERT
229 oop from_obj = state->source();
230 assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
231 assert(from_obj->is_forwarded(), "must be forwarded");
232 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
233 assert(to_obj->is_objArray(), "must be obj array");
234 #endif // ASSERT
235
236 objArrayOop to_array = objArrayOop(to_obj);
237
238 // Claim a chunk and get number of additional tasks to enqueue.
239 PartialArrayTaskStepper::Step step = _partial_array_stepper.next(state);
240 // Push any additional partial scan tasks needed. Pushed before processing
241 // the claimed chunk to allow other workers to steal while we're processing.
242 if (step._ncreate > 0) {
243 state->add_references(step._ncreate);
244 for (uint i = 0; i < step._ncreate; ++i) {
245 push_on_queue(ScannerTask(state));
246 }
247 }
248
249 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
250 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
251 // Process claimed task.
252 to_array->oop_iterate_range(&_scanner,
253 checked_cast<int>(step._index),
254 checked_cast<int>(step._index + _partial_array_stepper.chunk_size()));
255 // Release reference to the state, now that we're done with it.
256 _partial_array_state_allocator.release(state);
257 }
258
259 MAYBE_INLINE_EVACUATION
260 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
261 oop from_obj,
262 oop to_obj) {
263 assert(from_obj->is_forwarded(), "precondition");
264 assert(from_obj->forwardee() == to_obj, "precondition");
265 assert(to_obj->is_objArray(), "precondition");
266
267 objArrayOop to_array = objArrayOop(to_obj);
268
269 size_t array_length = to_array->length();
270 PartialArrayTaskStepper::Step step = _partial_array_stepper.start(array_length);
271
272 // Push any needed partial scan tasks. Pushed before processing the
273 // initial chunk to allow other workers to steal while we're processing.
274 if (step._ncreate > 0) {
275 assert(step._index < array_length, "invariant");
276 assert(((array_length - step._index) % _partial_array_stepper.chunk_size()) == 0,
277 "invariant");
278 PartialArrayState* state =
279 _partial_array_state_allocator.allocate(from_obj, to_obj,
280 step._index,
281 array_length,
282 step._ncreate);
283 for (uint i = 0; i < step._ncreate; ++i) {
284 push_on_queue(ScannerTask(state));
285 }
286 } else {
287 assert(step._index == array_length, "invariant");
288 }
289
290 // Skip the card enqueue iff the object (to_array) is in survivor region.
291 // However, G1HeapRegion::is_survivor() is too expensive here.
292 // Instead, we use dest_attr.is_young() because the two values are always
293 // equal: successfully allocated young regions must be survivor regions.
294 assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
295 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
296 // Process the initial chunk. No need to process the type in the
297 // klass, as it will already be handled by processing the built-in
298 // module.
299 to_array->oop_iterate_range(&_scanner, 0, checked_cast<int>(step._index));
300 }
301
302 MAYBE_INLINE_EVACUATION
303 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
304 verify_task(task);
305 if (task.is_narrow_oop_ptr()) {
306 do_oop_evac(task.to_narrow_oop_ptr());
307 } else if (task.is_oop_ptr()) {
308 do_oop_evac(task.to_oop_ptr());
309 } else {
310 do_partial_array(task.to_partial_array_state());
311 }
312 }
313
314 // Process tasks until overflow queue is empty and local queue
315 // contains no more than threshold entries. NOINLINE to prevent
316 // inlining into steal_and_trim_queue.
317 ATTRIBUTE_FLATTEN NOINLINE
318 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
319 ScannerTask task;
320 do {
321 while (_task_queue->pop_overflow(task)) {
322 if (!_task_queue->try_push_to_taskqueue(task)) {
323 dispatch_task(task);
324 }
325 }
326 while (_task_queue->pop_local(task, threshold)) {
327 dispatch_task(task);
328 }
329 } while (!_task_queue->overflow_empty());
330 }
331
332 ATTRIBUTE_FLATTEN
333 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
334 ScannerTask stolen_task;
335 while (task_queues->steal(_worker_id, stolen_task)) {
336 dispatch_task(stolen_task);
337 // Processing stolen task may have added tasks to our queue.
338 trim_queue();
339 }
340 }
341
342 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
343 size_t word_sz,
344 bool previous_plab_refill_failed,
345 uint node_index) {
346
347 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
348
349 // Right now we only have two types of regions (young / old) so
350 // let's keep the logic here simple. We can generalize it when necessary.
351 if (dest->is_young()) {
352 bool plab_refill_in_old_failed = false;
353 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
354 word_sz,
355 &plab_refill_in_old_failed,
356 node_index);
357 // Make sure that we won't attempt to copy any other objects out
358 // of a survivor region (given that apparently we cannot allocate
359 // any new ones) to avoid coming into this slow path again and again.
360 // Only consider failed PLAB refill here: failed inline allocations are
361 // typically large, so not indicative of remaining space.
362 if (previous_plab_refill_failed) {
363 _tenuring_threshold = 0;
364 }
365
366 if (obj_ptr != nullptr) {
367 dest->set_old();
368 } else {
369 // We just failed to allocate in old gen. The same idea as explained above
370 // for making survivor gen unavailable for allocation applies for old gen.
371 _old_gen_is_full = plab_refill_in_old_failed;
372 }
373 return obj_ptr;
374 } else {
375 _old_gen_is_full = previous_plab_refill_failed;
376 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
377 // no other space to try.
378 return nullptr;
379 }
380 }
381
382 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
383 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
384
385 if (region_attr.is_young()) {
386 age = !m.has_displaced_mark_helper() ? m.age()
387 : m.displaced_mark_helper().age();
388 if (age < _tenuring_threshold) {
389 return region_attr;
390 }
391 }
392 // young-to-old (promotion) or old-to-old; destination is old in both cases.
393 return G1HeapRegionAttr::Old;
394 }
395
396 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
397 Klass* klass, size_t word_sz, uint age,
398 HeapWord * const obj_ptr, uint node_index) const {
399 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
400 if (alloc_buf->contains(obj_ptr)) {
401 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
402 dest_attr.type() == G1HeapRegionAttr::Old,
403 alloc_buf->word_sz() * HeapWordSize);
404 } else {
405 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
406 dest_attr.type() == G1HeapRegionAttr::Old);
407 }
408 }
409
410 NOINLINE
411 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
412 Klass* klass,
413 size_t word_sz,
414 uint age,
415 uint node_index) {
416 HeapWord* obj_ptr = nullptr;
417 // Try slow-path allocation unless we're allocating old and old is already full.
418 if (!(dest_attr->is_old() && _old_gen_is_full)) {
419 bool plab_refill_failed = false;
420 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
421 word_sz,
422 &plab_refill_failed,
423 node_index);
424 if (obj_ptr == nullptr) {
425 obj_ptr = allocate_in_next_plab(dest_attr,
426 word_sz,
427 plab_refill_failed,
428 node_index);
429 }
430 }
431 if (obj_ptr != nullptr) {
432 update_numa_stats(node_index);
433 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
434 // The events are checked individually as part of the actual commit
435 report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
436 }
437 }
438 return obj_ptr;
439 }
440
441 #if ALLOCATION_FAILURE_INJECTOR
442 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
443 return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
444 }
445 #endif
446
447 NOINLINE
448 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
449 HeapWord* obj_ptr,
450 size_t word_sz,
451 uint node_index) {
452 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
453 }
454
455 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
456 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
457 G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
458 region->update_bot_for_block(obj_start, obj_start + word_sz);
459 }
460
461 // Private inline function, for direct internal use and providing the
462 // implementation of the public not-inline function.
463 MAYBE_INLINE_EVACUATION
464 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
465 oop const old,
466 markWord const old_mark) {
467 assert(region_attr.is_in_cset(),
468 "Unexpected region attr type: %s", region_attr.get_type_str());
469
470 // NOTE: With compact headers, it is not safe to load the Klass* from old, because
471 // that would access the mark-word, that might change at any time by concurrent
472 // workers.
473 // This mark word would refer to a forwardee, which may not yet have completed
474 // copying. Therefore we must load the Klass* from the mark-word that we already
475 // loaded. This is safe, because we only enter here if not yet forwarded.
476 assert(!old_mark.is_forwarded(), "precondition");
477 Klass* klass = UseCompactObjectHeaders
478 ? old_mark.klass()
479 : old->klass();
480
481 const size_t old_size = old->size_given_mark_and_klass(old_mark, klass);
482 const size_t word_sz = old->copy_size(old_size, old_mark);
483
484 // JNI only allows pinning of typeArrays, so we only need to keep those in place.
485 if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
486 return handle_evacuation_failure_par(old, old_mark, word_sz, true /* cause_pinned */);
487 }
488
489 uint age = 0;
490 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
491 G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
492 uint node_index = from_region->node_index();
493
494 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
495
496 // PLAB allocations should succeed most of the time, so we'll
497 // normally check against null once and that's it.
498 if (obj_ptr == nullptr) {
499 obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
500 if (obj_ptr == nullptr) {
501 // This will either forward-to-self, or detect that someone else has
502 // installed a forwarding pointer.
503 return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
504 }
505 }
506
507 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
508 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
509
510 // Should this evacuation fail?
511 if (inject_allocation_failure(from_region->hrm_index())) {
512 // Doing this after all the allocation attempts also tests the
513 // undo_allocation() method too.
514 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
515 return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
516 }
517
518 // We're going to allocate linearly, so might as well prefetch ahead.
519 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
520 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, old_size);
521
522 const oop obj = cast_to_oop(obj_ptr);
523 // Because the forwarding is done with memory_order_relaxed there is no
524 // ordering with the above copy. Clients that get the forwardee must not
525 // examine its contents without other synchronization, since the contents
526 // may not be up to date for them.
527 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
528 if (forward_ptr == nullptr) {
529
530 {
531 const uint young_index = from_region->young_index_in_cset();
532 assert((from_region->is_young() && young_index > 0) ||
533 (!from_region->is_young() && young_index == 0), "invariant" );
534 _surviving_young_words[young_index] += word_sz;
535 }
536
537 obj->initialize_hash_if_necessary(old);
538
539 if (dest_attr.is_young()) {
540 if (age < markWord::max_age) {
541 age++;
542 obj->incr_age();
543 }
544 _age_table.add(age, word_sz);
545 } else {
546 update_bot_after_copying(obj, word_sz);
547 }
548
549 // Most objects are not arrays, so do one array check rather than
550 // checking for each array category for each object.
551 if (klass->is_array_klass()) {
552 if (klass->is_objArray_klass()) {
553 start_partial_objarray(dest_attr, old, obj);
554 } else {
555 // Nothing needs to be done for typeArrays. Body doesn't contain
556 // any oops to scan, and the type in the klass will already be handled
557 // by processing the built-in module.
558 assert(klass->is_typeArray_klass(), "invariant");
559 }
560 return obj;
561 }
562
563 ContinuationGCSupport::transform_stack_chunk(obj);
564
565 // Check for deduplicating young Strings.
566 if (G1StringDedup::is_candidate_from_evacuation(klass,
567 region_attr,
568 dest_attr,
569 age)) {
570 // Record old; request adds a new weak reference, which reference
571 // processing expects to refer to a from-space object.
572 _string_dedup_requests.add(old);
573 }
574
575 // Skip the card enqueue iff the object (obj) is in survivor region.
576 // However, G1HeapRegion::is_survivor() is too expensive here.
577 // Instead, we use dest_attr.is_young() because the two values are always
578 // equal: successfully allocated young regions must be survivor regions.
579 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
580 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
581 obj->oop_iterate_backwards(&_scanner, klass);
582 return obj;
583 } else {
584 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
585 return forward_ptr;
586 }
587 }
588
589 // Public not-inline entry point.
590 ATTRIBUTE_FLATTEN
591 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
592 oop old,
593 markWord old_mark) {
594 return do_copy_to_survivor_space(region_attr, old, old_mark);
595 }
596
597 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
598 assert(worker_id < _num_workers, "out of bounds access");
599 if (_states[worker_id] == nullptr) {
600 _states[worker_id] =
601 new G1ParScanThreadState(_g1h, rdcqs(),
602 worker_id,
603 _num_workers,
604 _collection_set,
605 _evac_failure_regions);
606 }
607 return _states[worker_id];
608 }
609
610 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
611 assert(_flushed, "thread local state from the per thread states should have been flushed");
612 return _surviving_young_words_total;
613 }
614
615 void G1ParScanThreadStateSet::flush_stats() {
616 assert(!_flushed, "thread local state from the per thread states should be flushed once");
617 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
618 G1ParScanThreadState* pss = _states[worker_id];
619 assert(pss != nullptr, "must be initialized");
620
621 G1GCPhaseTimes* p = _g1h->phase_times();
622
623 // Need to get the following two before the call to G1ParThreadScanState::flush()
624 // because it resets the PLAB allocator where we get this info from.
625 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
626 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
627 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers, &_rdc_buffers[worker_id]) * HeapWordSize;
628 size_t evac_fail_enqueued_cards = pss->evac_failure_enqueued_cards();
629
630 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
631 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
632 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
633 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_fail_enqueued_cards, G1GCPhaseTimes::MergePSSEvacFailExtra);
634
635 delete pss;
636 _states[worker_id] = nullptr;
637 }
638
639 G1DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
640 dcq.merge_bufferlists(rdcqs());
641 rdcqs()->verify_empty();
642
643 _flushed = true;
644 }
645
646 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
647 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
648 G1ParScanThreadState* pss = _states[worker_index];
649 assert(pss != nullptr, "must be initialized");
650
651 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
652 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
653 }
654 }
655
656 NOINLINE
657 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz, bool cause_pinned) {
658 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
659
660 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
661 if (forward_ptr == nullptr) {
662 // Forward-to-self succeeded. We are the "owner" of the object.
663 G1HeapRegion* r = _g1h->heap_region_containing(old);
664
665 if (_evac_failure_regions->record(_worker_id, r->hrm_index(), cause_pinned)) {
666 G1HeapRegionPrinter::evac_failure(r);
667 }
668
669 // Mark the failing object in the marking bitmap and later use the bitmap to handle
670 // evacuation failure recovery.
671 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
672
673 ContinuationGCSupport::transform_stack_chunk(old);
674
675 _evacuation_failed_info.register_copy_failure(word_sz);
676
677 // For iterating objects that failed evacuation currently we can reuse the
678 // existing closure to scan evacuated objects; since we are iterating from a
679 // collection set region (i.e. never a Survivor region), we always need to
680 // gather cards for this case.
681 G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
682 old->oop_iterate_backwards(&_scanner);
683
684 return old;
685 } else {
686 // Forward-to-self failed. Either someone else managed to allocate
687 // space for this object (old != forward_ptr) or they beat us in
688 // self-forwarding it (old == forward_ptr).
689 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
690 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
691 "should not be in the CSet",
692 p2i(old), p2i(forward_ptr));
693 return forward_ptr;
694 }
695 }
696
697 void G1ParScanThreadState::initialize_numa_stats() {
698 if (_numa->is_enabled()) {
699 LogTarget(Info, gc, heap, numa) lt;
700
701 if (lt.is_enabled()) {
702 uint num_nodes = _numa->num_active_nodes();
703 // Record only if there are multiple active nodes.
704 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
705 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
706 }
707 }
708 }
709
710 void G1ParScanThreadState::flush_numa_stats() {
711 if (_obj_alloc_stat != nullptr) {
712 uint node_index = _numa->index_of_current_thread();
713 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
714 }
715 }
716
717 void G1ParScanThreadState::update_numa_stats(uint node_index) {
718 if (_obj_alloc_stat != nullptr) {
719 _obj_alloc_stat[node_index]++;
720 }
721 }
722
723 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
724 uint num_workers,
725 G1CollectionSet* collection_set,
726 G1EvacFailureRegions* evac_failure_regions) :
727 _g1h(g1h),
728 _collection_set(collection_set),
729 _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
730 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
731 _rdc_buffers(NEW_C_HEAP_ARRAY(BufferNodeList, num_workers, mtGC)),
732 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
733 _num_workers(num_workers),
734 _flushed(false),
735 _evac_failure_regions(evac_failure_regions)
736 {
737 for (uint i = 0; i < num_workers; ++i) {
738 _states[i] = nullptr;
739 _rdc_buffers[i] = BufferNodeList();
740 }
741 memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
742 }
743
744 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
745 assert(_flushed, "thread local state from the per thread states should have been flushed");
746 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
747 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
748 FREE_C_HEAP_ARRAY(BufferNodeList, _rdc_buffers);
749 }