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/g1YoungGCEvacFailureInjector.inline.hpp"
36 #include "gc/shared/continuationGCSupport.inline.hpp"
37 #include "gc/shared/partialArrayTaskStepper.inline.hpp"
38 #include "gc/shared/preservedMarks.inline.hpp"
39 #include "gc/shared/stringdedup/stringDedup.hpp"
40 #include "gc/shared/taskqueue.inline.hpp"
41 #include "memory/allocation.inline.hpp"
42 #include "oops/access.inline.hpp"
43 #include "oops/oop.inline.hpp"
44 #include "runtime/atomic.hpp"
45 #include "runtime/prefetch.inline.hpp"
46 #include "utilities/globalDefinitions.hpp"
47 #include "utilities/macros.hpp"
48
49 // In fastdebug builds the code size can get out of hand, potentially
50 // tripping over compiler limits (which may be bugs, but nevertheless
51 // need to be taken into consideration). A side benefit of limiting
52 // inlining is that we get more call frames that might aid debugging.
53 // And the fastdebug compile time for this file is much reduced.
54 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
55 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
56
57 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
58 G1RedirtyCardsQueueSet* rdcqs,
59 PreservedMarks* preserved_marks,
60 uint worker_id,
61 uint num_workers,
62 G1CollectionSet* collection_set,
63 G1EvacFailureRegions* evac_failure_regions)
64 : _g1h(g1h),
65 _task_queue(g1h->task_queue(worker_id)),
66 _rdc_local_qset(rdcqs),
67 _ct(g1h->card_table()),
68 _closures(nullptr),
69 _plab_allocator(nullptr),
70 _age_table(false),
71 _tenuring_threshold(g1h->policy()->tenuring_threshold()),
72 _scanner(g1h, this),
73 _worker_id(worker_id),
74 _last_enqueued_card(SIZE_MAX),
75 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
76 _stack_trim_lower_threshold(GCDrainStackTargetSize),
77 _trim_ticks(),
78 _surviving_young_words_base(nullptr),
79 _surviving_young_words(nullptr),
80 _surviving_words_length(collection_set->young_region_length() + 1),
81 _old_gen_is_full(false),
82 _partial_objarray_chunk_size(ParGCArrayScanChunk),
83 _partial_array_stepper(num_workers),
84 _string_dedup_requests(),
85 _max_num_optional_regions(collection_set->optional_region_length()),
86 _numa(g1h->numa()),
87 _obj_alloc_stat(nullptr),
88 EVAC_FAILURE_INJECTOR_ONLY(_evac_failure_inject_counter(0) COMMA)
89 _preserved_marks(preserved_marks),
90 _evacuation_failed_info(),
91 _evac_failure_regions(evac_failure_regions)
92 {
93 // We allocate number of young gen regions in the collection set plus one
94 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
95 // We also add a few elements at the beginning and at the end in
96 // an attempt to eliminate cache contention
97 const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
98 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
99
100 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
101 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
102 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
103
104 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
105
106 _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
107 this,
108 collection_set->only_contains_young_regions());
109
110 _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
111
112 initialize_numa_stats();
113 }
114
115 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
116 _rdc_local_qset.flush();
117 flush_numa_stats();
118 // Update allocation statistics.
119 _plab_allocator->flush_and_retire_stats(num_workers);
120 _g1h->policy()->record_age_table(&_age_table);
121
122 if (_evacuation_failed_info.has_failed()) {
123 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
124 }
125
126 size_t sum = 0;
127 for (uint i = 0; i < _surviving_words_length; i++) {
128 surviving_young_words[i] += _surviving_young_words[i];
129 sum += _surviving_young_words[i];
130 }
131 return sum;
132 }
133
134 G1ParScanThreadState::~G1ParScanThreadState() {
135 delete _plab_allocator;
136 delete _closures;
137 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
138 delete[] _oops_into_optional_regions;
139 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
140 }
141
142 size_t G1ParScanThreadState::lab_waste_words() const {
143 return _plab_allocator->waste();
144 }
145
146 size_t G1ParScanThreadState::lab_undo_waste_words() const {
147 return _plab_allocator->undo_waste();
148 }
149
150 #ifdef ASSERT
151 void G1ParScanThreadState::verify_task(narrowOop* task) const {
152 assert(task != nullptr, "invariant");
153 assert(UseCompressedOops, "sanity");
154 oop p = RawAccess<>::oop_load(task);
155 assert(_g1h->is_in_reserved(p),
156 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
157 }
158
159 void G1ParScanThreadState::verify_task(oop* task) const {
160 assert(task != nullptr, "invariant");
161 oop p = RawAccess<>::oop_load(task);
162 assert(_g1h->is_in_reserved(p),
163 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
164 }
165
166 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
167 // Must be in the collection set--it's already been copied.
168 oop p = task.to_source_array();
169 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
170 }
171
172 void G1ParScanThreadState::verify_task(ScannerTask task) const {
173 if (task.is_narrow_oop_ptr()) {
174 verify_task(task.to_narrow_oop_ptr());
175 } else if (task.is_oop_ptr()) {
176 verify_task(task.to_oop_ptr());
177 } else if (task.is_partial_array_task()) {
178 verify_task(task.to_partial_array_task());
179 } else {
180 ShouldNotReachHere();
181 }
182 }
183 #endif // ASSERT
184
185 template <class T>
186 MAYBE_INLINE_EVACUATION
187 void G1ParScanThreadState::do_oop_evac(T* p) {
188 // Reference should not be null here as such are never pushed to the task queue.
189 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
190
191 // Although we never intentionally push references outside of the collection
192 // set, due to (benign) races in the claim mechanism during RSet scanning more
193 // than one thread might claim the same card. So the same card may be
194 // processed multiple times, and so we might get references into old gen here.
195 // So we need to redo this check.
196 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
197 // References pushed onto the work stack should never point to a humongous region
198 // as they are not added to the collection set due to above precondition.
199 assert(!region_attr.is_humongous_candidate(),
200 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
201 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
202
203 if (!region_attr.is_in_cset()) {
204 // In this case somebody else already did all the work.
205 return;
206 }
207
208 markWord m = obj->mark();
209 if (m.is_marked()) {
210 obj = obj->forwardee(m);
211 } else {
212 obj = do_copy_to_survivor_space(region_attr, obj, m);
213 }
214 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
215
216 write_ref_field_post(p, obj);
217 }
218
219 MAYBE_INLINE_EVACUATION
220 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
221 oop from_obj = task.to_source_array();
222
223 assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
224 assert(from_obj->forward_safe_klass()->is_objArray_klass(), "must be obj array");
225 assert(from_obj->is_forwarded(), "must be forwarded");
226
227 oop to_obj = from_obj->forwardee();
228 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
229 assert(to_obj->is_objArray(), "must be obj array");
230 objArrayOop to_array = objArrayOop(to_obj);
231
232 PartialArrayTaskStepper::Step step
233 = _partial_array_stepper.next(objArrayOop(from_obj),
234 to_array,
235 _partial_objarray_chunk_size);
236 for (uint i = 0; i < step._ncreate; ++i) {
237 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
238 }
239
240 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
241 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
242 // Process claimed task. The length of to_array is not correct, but
243 // fortunately the iteration ignores the length field and just relies
244 // on start/end.
245 to_array->oop_iterate_range(&_scanner,
246 step._index,
247 step._index + _partial_objarray_chunk_size);
248 }
249
250 MAYBE_INLINE_EVACUATION
251 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
252 oop from_obj,
253 oop to_obj) {
254 assert(from_obj->forward_safe_klass()->is_objArray_klass(), "precondition");
255 assert(from_obj->is_forwarded(), "precondition");
256 assert(from_obj->forwardee() == to_obj, "precondition");
257 assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
258 assert(to_obj->is_objArray(), "precondition");
259
260 objArrayOop to_array = objArrayOop(to_obj);
261
262 PartialArrayTaskStepper::Step step
263 = _partial_array_stepper.start(objArrayOop(from_obj),
264 to_array,
265 _partial_objarray_chunk_size);
266
267 // Push any needed partial scan tasks. Pushed before processing the
268 // initial chunk to allow other workers to steal while we're processing.
269 for (uint i = 0; i < step._ncreate; ++i) {
270 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
271 }
272
273 // Skip the card enqueue iff the object (to_array) is in survivor region.
274 // However, HeapRegion::is_survivor() is too expensive here.
275 // Instead, we use dest_attr.is_young() because the two values are always
276 // equal: successfully allocated young regions must be survivor regions.
277 assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
278 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
279 // Process the initial chunk. No need to process the type in the
280 // klass, as it will already be handled by processing the built-in
281 // module. The length of to_array is not correct, but fortunately
282 // the iteration ignores that length field and relies on start/end.
283 to_array->oop_iterate_range(&_scanner, 0, step._index);
284 }
285
286 MAYBE_INLINE_EVACUATION
287 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
288 verify_task(task);
289 if (task.is_narrow_oop_ptr()) {
290 do_oop_evac(task.to_narrow_oop_ptr());
291 } else if (task.is_oop_ptr()) {
292 do_oop_evac(task.to_oop_ptr());
293 } else {
294 do_partial_array(task.to_partial_array_task());
295 }
296 }
297
298 // Process tasks until overflow queue is empty and local queue
299 // contains no more than threshold entries. NOINLINE to prevent
300 // inlining into steal_and_trim_queue.
301 ATTRIBUTE_FLATTEN NOINLINE
302 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
303 ScannerTask task;
304 do {
305 while (_task_queue->pop_overflow(task)) {
306 if (!_task_queue->try_push_to_taskqueue(task)) {
307 dispatch_task(task);
308 }
309 }
310 while (_task_queue->pop_local(task, threshold)) {
311 dispatch_task(task);
312 }
313 } while (!_task_queue->overflow_empty());
314 }
315
316 ATTRIBUTE_FLATTEN
317 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
318 ScannerTask stolen_task;
319 while (task_queues->steal(_worker_id, stolen_task)) {
320 dispatch_task(stolen_task);
321 // Processing stolen task may have added tasks to our queue.
322 trim_queue();
323 }
324 }
325
326 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
327 size_t word_sz,
328 bool previous_plab_refill_failed,
329 uint node_index) {
330
331 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
332
333 // Right now we only have two types of regions (young / old) so
334 // let's keep the logic here simple. We can generalize it when necessary.
335 if (dest->is_young()) {
336 bool plab_refill_in_old_failed = false;
337 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
338 word_sz,
339 &plab_refill_in_old_failed,
340 node_index);
341 // Make sure that we won't attempt to copy any other objects out
342 // of a survivor region (given that apparently we cannot allocate
343 // any new ones) to avoid coming into this slow path again and again.
344 // Only consider failed PLAB refill here: failed inline allocations are
345 // typically large, so not indicative of remaining space.
346 if (previous_plab_refill_failed) {
347 _tenuring_threshold = 0;
348 }
349
350 if (obj_ptr != nullptr) {
351 dest->set_old();
352 } else {
353 // We just failed to allocate in old gen. The same idea as explained above
354 // for making survivor gen unavailable for allocation applies for old gen.
355 _old_gen_is_full = plab_refill_in_old_failed;
356 }
357 return obj_ptr;
358 } else {
359 _old_gen_is_full = previous_plab_refill_failed;
360 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
361 // no other space to try.
362 return nullptr;
363 }
364 }
365
366 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
367 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
368
369 if (region_attr.is_young()) {
370 age = !m.has_displaced_mark_helper() ? m.age()
371 : m.displaced_mark_helper().age();
372 if (age < _tenuring_threshold) {
373 return region_attr;
374 }
375 }
376 // young-to-old (promotion) or old-to-old; destination is old in both cases.
377 return G1HeapRegionAttr::Old;
378 }
379
380 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
381 Klass* klass, size_t word_sz, uint age,
382 HeapWord * const obj_ptr, uint node_index) const {
383 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
384 if (alloc_buf->contains(obj_ptr)) {
385 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
386 dest_attr.type() == G1HeapRegionAttr::Old,
387 alloc_buf->word_sz() * HeapWordSize);
388 } else {
389 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
390 dest_attr.type() == G1HeapRegionAttr::Old);
391 }
392 }
393
394 NOINLINE
395 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
396 Klass* klass,
397 size_t word_sz,
398 uint age,
399 uint node_index) {
400 HeapWord* obj_ptr = nullptr;
401 // Try slow-path allocation unless we're allocating old and old is already full.
402 if (!(dest_attr->is_old() && _old_gen_is_full)) {
403 bool plab_refill_failed = false;
404 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
405 word_sz,
406 &plab_refill_failed,
407 node_index);
408 if (obj_ptr == nullptr) {
409 obj_ptr = allocate_in_next_plab(dest_attr,
410 word_sz,
411 plab_refill_failed,
412 node_index);
413 }
414 }
415 if (obj_ptr != nullptr) {
416 update_numa_stats(node_index);
417 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
418 // The events are checked individually as part of the actual commit
419 report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
420 }
421 }
422 return obj_ptr;
423 }
424
425 #if EVAC_FAILURE_INJECTOR
426 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
427 return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
428 }
429 #endif
430
431 NOINLINE
432 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
433 HeapWord* obj_ptr,
434 size_t word_sz,
435 uint node_index) {
436 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
437 }
438
439 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
440 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
441 HeapRegion* region = _g1h->heap_region_containing(obj_start);
442 region->update_bot_for_obj(obj_start, word_sz);
443 }
444
445 // Private inline function, for direct internal use and providing the
446 // implementation of the public not-inline function.
447 MAYBE_INLINE_EVACUATION
448 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
449 oop const old,
450 markWord const old_mark) {
451 assert(region_attr.is_in_cset(),
452 "Unexpected region attr type: %s", region_attr.get_type_str());
453
454 // Get the klass once. We'll need it again later, and this avoids
455 // re-decoding when it's compressed.
456 // NOTE: With compact headers, it is not safe to load the Klass* from o, because
457 // that would access the mark-word, and the mark-word might change at any time by
458 // concurrent promotion. The promoted mark-word would point to the forwardee, which
459 // may not yet have completed copying. Therefore we must load the Klass* from
460 // the mark-word that we have already loaded. This is safe, because we have checked
461 // that this is not yet forwarded in the caller.
462 Klass* klass = old->forward_safe_klass(old_mark);
463 const size_t word_sz = old->size_given_klass(klass);
464
465 uint age = 0;
466 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
467 HeapRegion* const from_region = _g1h->heap_region_containing(old);
468 uint node_index = from_region->node_index();
469
470 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
471
472 // PLAB allocations should succeed most of the time, so we'll
473 // normally check against null once and that's it.
474 if (obj_ptr == nullptr) {
475 obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
476 if (obj_ptr == nullptr) {
477 // This will either forward-to-self, or detect that someone else has
478 // installed a forwarding pointer.
479 return handle_evacuation_failure_par(old, old_mark, word_sz);
480 }
481 }
482
483 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
484 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
485
486 // Should this evacuation fail?
487 if (inject_evacuation_failure(from_region->hrm_index())) {
488 // Doing this after all the allocation attempts also tests the
489 // undo_allocation() method too.
490 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
491 return handle_evacuation_failure_par(old, old_mark, word_sz);
492 }
493
494 // We're going to allocate linearly, so might as well prefetch ahead.
495 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
496 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
497
498 const oop obj = cast_to_oop(obj_ptr);
499 // Because the forwarding is done with memory_order_relaxed there is no
500 // ordering with the above copy. Clients that get the forwardee must not
501 // examine its contents without other synchronization, since the contents
502 // may not be up to date for them.
503 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
504 if (forward_ptr == nullptr) {
505
506 {
507 const uint young_index = from_region->young_index_in_cset();
508 assert((from_region->is_young() && young_index > 0) ||
509 (!from_region->is_young() && young_index == 0), "invariant" );
510 _surviving_young_words[young_index] += word_sz;
511 }
512
513 if (dest_attr.is_young()) {
514 if (age < markWord::max_age) {
515 age++;
516 obj->incr_age();
517 }
518 _age_table.add(age, word_sz);
519 } else {
520 update_bot_after_copying(obj, word_sz);
521 }
522
523 // Most objects are not arrays, so do one array check rather than
524 // checking for each array category for each object.
525 if (klass->is_array_klass()) {
526 if (klass->is_objArray_klass()) {
527 start_partial_objarray(dest_attr, old, obj);
528 } else {
529 // Nothing needs to be done for typeArrays. Body doesn't contain
530 // any oops to scan, and the type in the klass will already be handled
531 // by processing the built-in module.
532 assert(klass->is_typeArray_klass(), "invariant");
533 }
534 return obj;
535 }
536
537 ContinuationGCSupport::transform_stack_chunk(obj);
538
539 // Check for deduplicating young Strings.
540 if (G1StringDedup::is_candidate_from_evacuation(klass,
541 region_attr,
542 dest_attr,
543 age)) {
544 // Record old; request adds a new weak reference, which reference
545 // processing expects to refer to a from-space object.
546 _string_dedup_requests.add(old);
547 }
548
549 // Skip the card enqueue iff the object (obj) is in survivor region.
550 // However, HeapRegion::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 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
555 obj->oop_iterate_backwards(&_scanner, klass);
556 return obj;
557 } else {
558 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
559 return forward_ptr;
560 }
561 }
562
563 // Public not-inline entry point.
564 ATTRIBUTE_FLATTEN
565 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
566 oop old,
567 markWord old_mark) {
568 return do_copy_to_survivor_space(region_attr, old, old_mark);
569 }
570
571 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
572 assert(worker_id < _num_workers, "out of bounds access");
573 if (_states[worker_id] == nullptr) {
574 _states[worker_id] =
575 new G1ParScanThreadState(_g1h, rdcqs(),
576 _preserved_marks_set.get(worker_id),
577 worker_id,
578 _num_workers,
579 _collection_set,
580 _evac_failure_regions);
581 }
582 return _states[worker_id];
583 }
584
585 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
586 assert(_flushed, "thread local state from the per thread states should have been flushed");
587 return _surviving_young_words_total;
588 }
589
590 void G1ParScanThreadStateSet::flush_stats() {
591 assert(!_flushed, "thread local state from the per thread states should be flushed once");
592
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
605 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
606 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
607 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
608
609 delete pss;
610 _states[worker_id] = nullptr;
611 }
612 _flushed = true;
613 }
614
615 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
616 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
617 G1ParScanThreadState* pss = _states[worker_index];
618 assert(pss != nullptr, "must be initialized");
619
620 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
621 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
622 }
623 }
624
625 NOINLINE
626 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
627 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
628
629 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
630 if (forward_ptr == nullptr) {
631 // Forward-to-self succeeded. We are the "owner" of the object.
632 HeapRegion* r = _g1h->heap_region_containing(old);
633
634 if (_evac_failure_regions->record(r->hrm_index())) {
635 _g1h->hr_printer()->evac_failure(r);
636 }
637
638 // Mark the failing object in the marking bitmap and later use the bitmap to handle
639 // evacuation failure recovery.
640 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
641
642 _preserved_marks->push_if_necessary(old, m);
643
644 ContinuationGCSupport::transform_stack_chunk(old);
645
646 _evacuation_failed_info.register_copy_failure(word_sz);
647
648 // For iterating objects that failed evacuation currently we can reuse the
649 // existing closure to scan evacuated objects because:
650 // - for objects referring into the collection set we do not need to gather
651 // cards at this time. The regions they are in will be unconditionally turned
652 // to old regions without remembered sets.
653 // - since we are iterating from a collection set region (i.e. never a Survivor
654 // region), we always need to gather cards for this case.
655 G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
656 old->oop_iterate_backwards(&_scanner);
657
658 return old;
659 } else {
660 // Forward-to-self failed. Either someone else managed to allocate
661 // space for this object (old != forward_ptr) or they beat us in
662 // self-forwarding it (old == forward_ptr).
663 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
664 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
665 "should not be in the CSet",
666 p2i(old), p2i(forward_ptr));
667 return forward_ptr;
668 }
669 }
670
671 void G1ParScanThreadState::initialize_numa_stats() {
672 if (_numa->is_enabled()) {
673 LogTarget(Info, gc, heap, numa) lt;
674
675 if (lt.is_enabled()) {
676 uint num_nodes = _numa->num_active_nodes();
677 // Record only if there are multiple active nodes.
678 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
679 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
680 }
681 }
682 }
683
684 void G1ParScanThreadState::flush_numa_stats() {
685 if (_obj_alloc_stat != nullptr) {
686 uint node_index = _numa->index_of_current_thread();
687 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
688 }
689 }
690
691 void G1ParScanThreadState::update_numa_stats(uint node_index) {
692 if (_obj_alloc_stat != nullptr) {
693 _obj_alloc_stat[node_index]++;
694 }
695 }
696
697 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
698 uint num_workers,
699 G1CollectionSet* collection_set,
700 G1EvacFailureRegions* evac_failure_regions) :
701 _g1h(g1h),
702 _collection_set(collection_set),
703 _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
704 _preserved_marks_set(true /* in_c_heap */),
705 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
706 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
707 _num_workers(num_workers),
708 _flushed(false),
709 _evac_failure_regions(evac_failure_regions) {
710 _preserved_marks_set.init(num_workers);
711 for (uint i = 0; i < num_workers; ++i) {
712 _states[i] = nullptr;
713 }
714 memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
715 }
716
717 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
718 assert(_flushed, "thread local state from the per thread states should have been flushed");
719 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
720 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
721 _preserved_marks_set.assert_empty();
722 _preserved_marks_set.reclaim();
723 }