1 /*
2 * Copyright (c) 2014, 2023, 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/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/partialArrayTaskStepper.inline.hpp"
38 #include "gc/shared/stringdedup/stringDedup.hpp"
39 #include "gc/shared/taskqueue.inline.hpp"
40 #include "memory/allocation.inline.hpp"
41 #include "oops/access.inline.hpp"
42 #include "oops/oop.inline.hpp"
43 #include "runtime/atomic.hpp"
44 #include "runtime/prefetch.inline.hpp"
45 #include "utilities/globalDefinitions.hpp"
46 #include "utilities/macros.hpp"
47
48 // In fastdebug builds the code size can get out of hand, potentially
49 // tripping over compiler limits (which may be bugs, but nevertheless
50 // need to be taken into consideration). A side benefit of limiting
51 // inlining is that we get more call frames that might aid debugging.
52 // And the fastdebug compile time for this file is much reduced.
53 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
54 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
55
56 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
57 G1RedirtyCardsQueueSet* rdcqs,
58 uint worker_id,
59 uint num_workers,
60 G1CollectionSet* collection_set,
61 G1EvacFailureRegions* evac_failure_regions)
62 : _g1h(g1h),
63 _task_queue(g1h->task_queue(worker_id)),
64 _rdc_local_qset(rdcqs),
65 _ct(g1h->card_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 _last_enqueued_card(SIZE_MAX),
73 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
74 _stack_trim_lower_threshold(GCDrainStackTargetSize),
75 _trim_ticks(),
76 _surviving_young_words_base(nullptr),
77 _surviving_young_words(nullptr),
78 _surviving_words_length(collection_set->young_region_length() + 1),
79 _old_gen_is_full(false),
80 _partial_objarray_chunk_size(ParGCArrayScanChunk),
81 _partial_array_stepper(num_workers),
82 _string_dedup_requests(),
83 _max_num_optional_regions(collection_set->optional_region_length()),
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 _evac_failure_enqueued_cards(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, BufferNodeList* rdc_buffers) {
114 *rdc_buffers = _rdc_local_qset.flush();
115 flush_numa_stats();
116 // Update allocation statistics.
117 _plab_allocator->flush_and_retire_stats(num_workers);
118 _g1h->policy()->record_age_table(&_age_table);
119
120 if (_evacuation_failed_info.has_failed()) {
121 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
122 }
123
124 size_t sum = 0;
125 for (uint i = 0; i < _surviving_words_length; i++) {
126 surviving_young_words[i] += _surviving_young_words[i];
127 sum += _surviving_young_words[i];
128 }
129 return sum;
130 }
131
132 G1ParScanThreadState::~G1ParScanThreadState() {
133 delete _plab_allocator;
134 delete _closures;
135 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
136 delete[] _oops_into_optional_regions;
137 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
138 }
139
140 size_t G1ParScanThreadState::lab_waste_words() const {
141 return _plab_allocator->waste();
142 }
143
144 size_t G1ParScanThreadState::lab_undo_waste_words() const {
145 return _plab_allocator->undo_waste();
146 }
147
148 size_t G1ParScanThreadState::evac_failure_enqueued_cards() const {
149 return _evac_failure_enqueued_cards;
150 }
151
152 #ifdef ASSERT
153 void G1ParScanThreadState::verify_task(narrowOop* task) const {
154 assert(task != nullptr, "invariant");
155 assert(UseCompressedOops, "sanity");
156 oop p = RawAccess<>::oop_load(task);
157 assert(_g1h->is_in_reserved(p),
158 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
159 }
160
161 void G1ParScanThreadState::verify_task(oop* task) const {
162 assert(task != nullptr, "invariant");
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(PartialArrayScanTask task) const {
169 // Must be in the collection set--it's already been copied.
170 oop p = task.to_source_array();
171 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
172 }
173
174 void G1ParScanThreadState::verify_task(ScannerTask task) const {
175 if (task.is_narrow_oop_ptr()) {
176 verify_task(task.to_narrow_oop_ptr());
177 } else if (task.is_oop_ptr()) {
178 verify_task(task.to_oop_ptr());
179 } else if (task.is_partial_array_task()) {
180 verify_task(task.to_partial_array_task());
181 } else {
182 ShouldNotReachHere();
183 }
184 }
185 #endif // ASSERT
186
187 template <class T>
188 MAYBE_INLINE_EVACUATION
189 void G1ParScanThreadState::do_oop_evac(T* p) {
190 // Reference should not be null here as such are never pushed to the task queue.
191 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
192
193 // Although we never intentionally push references outside of the collection
194 // set, due to (benign) races in the claim mechanism during RSet scanning more
195 // than one thread might claim the same card. So the same card may be
196 // processed multiple times, and so we might get references into old gen here.
197 // So we need to redo this check.
198 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
199 // References pushed onto the work stack should never point to a humongous region
200 // as they are not added to the collection set due to above precondition.
201 assert(!region_attr.is_humongous_candidate(),
202 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
203 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
204
205 if (!region_attr.is_in_cset()) {
206 // In this case somebody else already did all the work.
207 return;
208 }
209
210 markWord m = obj->mark();
211 if (m.is_forwarded()) {
212 obj = obj->forwardee(m);
213 } else {
214 obj = do_copy_to_survivor_space(region_attr, obj, m);
215 }
216 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
217
218 write_ref_field_post(p, obj);
219 }
220
221 MAYBE_INLINE_EVACUATION
222 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
223 oop from_obj = task.to_source_array();
224
225 assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
226 assert(from_obj->forward_safe_klass()->is_objArray_klass(), "must be obj array");
227 assert(from_obj->is_forwarded(), "must be forwarded");
228
229 oop to_obj = from_obj->forwardee();
230 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
231 assert(to_obj->is_objArray(), "must be obj array");
232 objArrayOop to_array = objArrayOop(to_obj);
233
234 PartialArrayTaskStepper::Step step
235 = _partial_array_stepper.next(objArrayOop(from_obj),
236 to_array,
237 _partial_objarray_chunk_size);
238 for (uint i = 0; i < step._ncreate; ++i) {
239 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
240 }
241
242 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
243 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
244 // Process claimed task. The length of to_array is not correct, but
245 // fortunately the iteration ignores the length field and just relies
246 // on start/end.
247 to_array->oop_iterate_range(&_scanner,
248 step._index,
249 step._index + _partial_objarray_chunk_size);
250 }
251
252 MAYBE_INLINE_EVACUATION
253 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
254 oop from_obj,
255 oop to_obj) {
256 assert(from_obj->forward_safe_klass()->is_objArray_klass(), "precondition");
257 assert(from_obj->is_forwarded(), "precondition");
258 assert(from_obj->forwardee() == to_obj, "precondition");
259 assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
260 assert(to_obj->is_objArray(), "precondition");
261
262 objArrayOop to_array = objArrayOop(to_obj);
263
264 PartialArrayTaskStepper::Step step
265 = _partial_array_stepper.start(objArrayOop(from_obj),
266 to_array,
267 _partial_objarray_chunk_size);
268
269 // Push any needed partial scan tasks. Pushed before processing the
270 // initial chunk to allow other workers to steal while we're processing.
271 for (uint i = 0; i < step._ncreate; ++i) {
272 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
273 }
274
275 // Skip the card enqueue iff the object (to_array) is in survivor region.
276 // However, HeapRegion::is_survivor() is too expensive here.
277 // Instead, we use dest_attr.is_young() because the two values are always
278 // equal: successfully allocated young regions must be survivor regions.
279 assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
280 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
281 // Process the initial chunk. No need to process the type in the
282 // klass, as it will already be handled by processing the built-in
283 // module. The length of to_array is not correct, but fortunately
284 // the iteration ignores that length field and relies on start/end.
285 to_array->oop_iterate_range(&_scanner, 0, step._index);
286 }
287
288 MAYBE_INLINE_EVACUATION
289 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
290 verify_task(task);
291 if (task.is_narrow_oop_ptr()) {
292 do_oop_evac(task.to_narrow_oop_ptr());
293 } else if (task.is_oop_ptr()) {
294 do_oop_evac(task.to_oop_ptr());
295 } else {
296 do_partial_array(task.to_partial_array_task());
297 }
298 }
299
300 // Process tasks until overflow queue is empty and local queue
301 // contains no more than threshold entries. NOINLINE to prevent
302 // inlining into steal_and_trim_queue.
303 ATTRIBUTE_FLATTEN NOINLINE
304 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
305 ScannerTask task;
306 do {
307 while (_task_queue->pop_overflow(task)) {
308 if (!_task_queue->try_push_to_taskqueue(task)) {
309 dispatch_task(task);
310 }
311 }
312 while (_task_queue->pop_local(task, threshold)) {
313 dispatch_task(task);
314 }
315 } while (!_task_queue->overflow_empty());
316 }
317
318 ATTRIBUTE_FLATTEN
319 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
320 ScannerTask stolen_task;
321 while (task_queues->steal(_worker_id, stolen_task)) {
322 dispatch_task(stolen_task);
323 // Processing stolen task may have added tasks to our queue.
324 trim_queue();
325 }
326 }
327
328 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
329 size_t word_sz,
330 bool previous_plab_refill_failed,
331 uint node_index) {
332
333 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
334
335 // Right now we only have two types of regions (young / old) so
336 // let's keep the logic here simple. We can generalize it when necessary.
337 if (dest->is_young()) {
338 bool plab_refill_in_old_failed = false;
339 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
340 word_sz,
341 &plab_refill_in_old_failed,
342 node_index);
343 // Make sure that we won't attempt to copy any other objects out
344 // of a survivor region (given that apparently we cannot allocate
345 // any new ones) to avoid coming into this slow path again and again.
346 // Only consider failed PLAB refill here: failed inline allocations are
347 // typically large, so not indicative of remaining space.
348 if (previous_plab_refill_failed) {
349 _tenuring_threshold = 0;
350 }
351
352 if (obj_ptr != nullptr) {
353 dest->set_old();
354 } else {
355 // We just failed to allocate in old gen. The same idea as explained above
356 // for making survivor gen unavailable for allocation applies for old gen.
357 _old_gen_is_full = plab_refill_in_old_failed;
358 }
359 return obj_ptr;
360 } else {
361 _old_gen_is_full = previous_plab_refill_failed;
362 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
363 // no other space to try.
364 return nullptr;
365 }
366 }
367
368 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
369 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
370
371 if (region_attr.is_young()) {
372 age = !m.has_displaced_mark_helper() ? m.age()
373 : m.displaced_mark_helper().age();
374 if (age < _tenuring_threshold) {
375 return region_attr;
376 }
377 }
378 // young-to-old (promotion) or old-to-old; destination is old in both cases.
379 return G1HeapRegionAttr::Old;
380 }
381
382 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
383 Klass* klass, size_t word_sz, uint age,
384 HeapWord * const obj_ptr, uint node_index) const {
385 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
386 if (alloc_buf->contains(obj_ptr)) {
387 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
388 dest_attr.type() == G1HeapRegionAttr::Old,
389 alloc_buf->word_sz() * HeapWordSize);
390 } else {
391 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
392 dest_attr.type() == G1HeapRegionAttr::Old);
393 }
394 }
395
396 NOINLINE
397 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
398 Klass* klass,
399 size_t word_sz,
400 uint age,
401 uint node_index) {
402 HeapWord* obj_ptr = nullptr;
403 // Try slow-path allocation unless we're allocating old and old is already full.
404 if (!(dest_attr->is_old() && _old_gen_is_full)) {
405 bool plab_refill_failed = false;
406 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
407 word_sz,
408 &plab_refill_failed,
409 node_index);
410 if (obj_ptr == nullptr) {
411 obj_ptr = allocate_in_next_plab(dest_attr,
412 word_sz,
413 plab_refill_failed,
414 node_index);
415 }
416 }
417 if (obj_ptr != nullptr) {
418 update_numa_stats(node_index);
419 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
420 // The events are checked individually as part of the actual commit
421 report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
422 }
423 }
424 return obj_ptr;
425 }
426
427 #if ALLOCATION_FAILURE_INJECTOR
428 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
429 return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
430 }
431 #endif
432
433 NOINLINE
434 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
435 HeapWord* obj_ptr,
436 size_t word_sz,
437 uint node_index) {
438 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
439 }
440
441 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
442 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
443 HeapRegion* region = _g1h->heap_region_containing(obj_start);
444 region->update_bot_for_obj(obj_start, word_sz);
445 }
446
447 // Private inline function, for direct internal use and providing the
448 // implementation of the public not-inline function.
449 MAYBE_INLINE_EVACUATION
450 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
451 oop const old,
452 markWord const old_mark) {
453 assert(region_attr.is_in_cset(),
454 "Unexpected region attr type: %s", region_attr.get_type_str());
455
456 // Get the klass once. We'll need it again later, and this avoids
457 // re-decoding when it's compressed.
458 // NOTE: With compact headers, it is not safe to load the Klass* from o, because
459 // that would access the mark-word, and the mark-word might change at any time by
460 // concurrent promotion. The promoted mark-word would point to the forwardee, which
461 // may not yet have completed copying. Therefore we must load the Klass* from
462 // the mark-word that we have already loaded. This is safe, because we have checked
463 // that this is not yet forwarded in the caller.
464 Klass* klass = old->forward_safe_klass(old_mark);
465 const size_t word_sz = old->size_given_klass(klass);
466
467 // JNI only allows pinning of typeArrays, so we only need to keep those in place.
468 if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
469 return handle_evacuation_failure_par(old, old_mark, word_sz, true /* cause_pinned */);
470 }
471
472 uint age = 0;
473 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
474 HeapRegion* const from_region = _g1h->heap_region_containing(old);
475 uint node_index = from_region->node_index();
476
477 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
478
479 // PLAB allocations should succeed most of the time, so we'll
480 // normally check against null once and that's it.
481 if (obj_ptr == nullptr) {
482 obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
483 if (obj_ptr == nullptr) {
484 // This will either forward-to-self, or detect that someone else has
485 // installed a forwarding pointer.
486 return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
487 }
488 }
489
490 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
491 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
492
493 // Should this evacuation fail?
494 if (inject_allocation_failure(from_region->hrm_index())) {
495 // Doing this after all the allocation attempts also tests the
496 // undo_allocation() method too.
497 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
498 return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
499 }
500
501 // We're going to allocate linearly, so might as well prefetch ahead.
502 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
503 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
504
505 const oop obj = cast_to_oop(obj_ptr);
506 // Because the forwarding is done with memory_order_relaxed there is no
507 // ordering with the above copy. Clients that get the forwardee must not
508 // examine its contents without other synchronization, since the contents
509 // may not be up to date for them.
510 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
511 if (forward_ptr == nullptr) {
512
513 {
514 const uint young_index = from_region->young_index_in_cset();
515 assert((from_region->is_young() && young_index > 0) ||
516 (!from_region->is_young() && young_index == 0), "invariant" );
517 _surviving_young_words[young_index] += word_sz;
518 }
519
520 if (dest_attr.is_young()) {
521 if (age < markWord::max_age) {
522 age++;
523 obj->incr_age();
524 }
525 _age_table.add(age, word_sz);
526 } else {
527 update_bot_after_copying(obj, word_sz);
528 }
529
530 // Most objects are not arrays, so do one array check rather than
531 // checking for each array category for each object.
532 if (klass->is_array_klass()) {
533 if (klass->is_objArray_klass()) {
534 start_partial_objarray(dest_attr, old, obj);
535 } else {
536 // Nothing needs to be done for typeArrays. Body doesn't contain
537 // any oops to scan, and the type in the klass will already be handled
538 // by processing the built-in module.
539 assert(klass->is_typeArray_klass(), "invariant");
540 }
541 return obj;
542 }
543
544 ContinuationGCSupport::transform_stack_chunk(obj);
545
546 // Check for deduplicating young Strings.
547 if (G1StringDedup::is_candidate_from_evacuation(klass,
548 region_attr,
549 dest_attr,
550 age)) {
551 // Record old; request adds a new weak reference, which reference
552 // processing expects to refer to a from-space object.
553 _string_dedup_requests.add(old);
554 }
555
556 // Skip the card enqueue iff the object (obj) is in survivor region.
557 // However, HeapRegion::is_survivor() is too expensive here.
558 // Instead, we use dest_attr.is_young() because the two values are always
559 // equal: successfully allocated young regions must be survivor regions.
560 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
561 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
562 obj->oop_iterate_backwards(&_scanner, klass);
563 return obj;
564 } else {
565 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
566 return forward_ptr;
567 }
568 }
569
570 // Public not-inline entry point.
571 ATTRIBUTE_FLATTEN
572 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
573 oop old,
574 markWord old_mark) {
575 return do_copy_to_survivor_space(region_attr, old, old_mark);
576 }
577
578 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
579 assert(worker_id < _num_workers, "out of bounds access");
580 if (_states[worker_id] == nullptr) {
581 _states[worker_id] =
582 new G1ParScanThreadState(_g1h, rdcqs(),
583 worker_id,
584 _num_workers,
585 _collection_set,
586 _evac_failure_regions);
587 }
588 return _states[worker_id];
589 }
590
591 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
592 assert(_flushed, "thread local state from the per thread states should have been flushed");
593 return _surviving_young_words_total;
594 }
595
596 void G1ParScanThreadStateSet::flush_stats() {
597 assert(!_flushed, "thread local state from the per thread states should be flushed once");
598 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
599 G1ParScanThreadState* pss = _states[worker_id];
600 assert(pss != nullptr, "must be initialized");
601
602 G1GCPhaseTimes* p = _g1h->phase_times();
603
604 // Need to get the following two before the call to G1ParThreadScanState::flush()
605 // because it resets the PLAB allocator where we get this info from.
606 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
607 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
608 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers, &_rdc_buffers[worker_id]) * HeapWordSize;
609 size_t evac_fail_enqueued_cards = pss->evac_failure_enqueued_cards();
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, evac_fail_enqueued_cards, G1GCPhaseTimes::MergePSSEvacFailExtra);
615
616 delete pss;
617 _states[worker_id] = nullptr;
618 }
619
620 G1DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
621 dcq.merge_bufferlists(rdcqs());
622 rdcqs()->verify_empty();
623
624 _flushed = true;
625 }
626
627 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
628 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
629 G1ParScanThreadState* pss = _states[worker_index];
630 assert(pss != nullptr, "must be initialized");
631
632 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
633 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
634 }
635 }
636
637 NOINLINE
638 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz, bool cause_pinned) {
639 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
640
641 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
642 if (forward_ptr == nullptr) {
643 // Forward-to-self succeeded. We are the "owner" of the object.
644 HeapRegion* r = _g1h->heap_region_containing(old);
645
646 if (_evac_failure_regions->record(_worker_id, r->hrm_index(), cause_pinned)) {
647 _g1h->hr_printer()->evac_failure(r);
648 }
649
650 // Mark the failing object in the marking bitmap and later use the bitmap to handle
651 // evacuation failure recovery.
652 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
653
654 ContinuationGCSupport::transform_stack_chunk(old);
655
656 _evacuation_failed_info.register_copy_failure(word_sz);
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 G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
663 old->oop_iterate_backwards(&_scanner);
664
665 return old;
666 } else {
667 // Forward-to-self failed. Either someone else managed to allocate
668 // space for this object (old != forward_ptr) or they beat us in
669 // self-forwarding it (old == forward_ptr).
670 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
671 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
672 "should not be in the CSet",
673 p2i(old), p2i(forward_ptr));
674 return forward_ptr;
675 }
676 }
677
678 void G1ParScanThreadState::initialize_numa_stats() {
679 if (_numa->is_enabled()) {
680 LogTarget(Info, gc, heap, numa) lt;
681
682 if (lt.is_enabled()) {
683 uint num_nodes = _numa->num_active_nodes();
684 // Record only if there are multiple active nodes.
685 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
686 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
687 }
688 }
689 }
690
691 void G1ParScanThreadState::flush_numa_stats() {
692 if (_obj_alloc_stat != nullptr) {
693 uint node_index = _numa->index_of_current_thread();
694 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
695 }
696 }
697
698 void G1ParScanThreadState::update_numa_stats(uint node_index) {
699 if (_obj_alloc_stat != nullptr) {
700 _obj_alloc_stat[node_index]++;
701 }
702 }
703
704 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
705 uint num_workers,
706 G1CollectionSet* collection_set,
707 G1EvacFailureRegions* evac_failure_regions) :
708 _g1h(g1h),
709 _collection_set(collection_set),
710 _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
711 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
712 _rdc_buffers(NEW_C_HEAP_ARRAY(BufferNodeList, num_workers, mtGC)),
713 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
714 _num_workers(num_workers),
715 _flushed(false),
716 _evac_failure_regions(evac_failure_regions) {
717 for (uint i = 0; i < num_workers; ++i) {
718 _states[i] = nullptr;
719 _rdc_buffers[i] = BufferNodeList();
720 }
721 memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
722 }
723
724 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
725 assert(_flushed, "thread local state from the per thread states should have been flushed");
726 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
727 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
728 FREE_C_HEAP_ARRAY(BufferNodeList, _rdc_buffers);
729 }