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/preservedMarks.inline.hpp"
41 #include "gc/shared/stringdedup/stringDedup.hpp"
42 #include "gc/shared/taskqueue.inline.hpp"
43 #include "memory/allocation.inline.hpp"
44 #include "oops/access.inline.hpp"
45 #include "oops/oop.inline.hpp"
46 #include "runtime/atomic.hpp"
47 #include "runtime/mutexLocker.hpp"
48 #include "runtime/prefetch.inline.hpp"
49 #include "utilities/globalDefinitions.hpp"
50 #include "utilities/macros.hpp"
51
52 // In fastdebug builds the code size can get out of hand, potentially
53 // tripping over compiler limits (which may be bugs, but nevertheless
54 // need to be taken into consideration). A side benefit of limiting
55 // inlining is that we get more call frames that might aid debugging.
56 // And the fastdebug compile time for this file is much reduced.
57 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
58 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
59
60 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
61 G1RedirtyCardsQueueSet* rdcqs,
62 PreservedMarks* preserved_marks,
63 uint worker_id,
64 uint num_workers,
65 G1CollectionSet* collection_set,
66 G1EvacFailureRegions* evac_failure_regions,
67 PartialArrayStateAllocator* pas_allocator)
68 : _g1h(g1h),
69 _task_queue(g1h->task_queue(worker_id)),
70 _rdc_local_qset(rdcqs),
71 _ct(g1h->card_table()),
72 _closures(nullptr),
73 _plab_allocator(nullptr),
74 _age_table(false),
75 _tenuring_threshold(g1h->policy()->tenuring_threshold()),
76 _scanner(g1h, this),
77 _worker_id(worker_id),
78 _last_enqueued_card(SIZE_MAX),
79 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
80 _stack_trim_lower_threshold(GCDrainStackTargetSize),
81 _trim_ticks(),
82 _surviving_young_words_base(nullptr),
83 _surviving_young_words(nullptr),
84 _surviving_words_length(collection_set->young_region_length() + 1),
85 _old_gen_is_full(false),
86 _partial_array_state_allocator(pas_allocator),
87 _partial_array_stepper(num_workers, ParGCArrayScanChunk),
88 _string_dedup_requests(),
89 _max_num_optional_regions(collection_set->optional_region_length()),
90 _numa(g1h->numa()),
91 _obj_alloc_stat(nullptr),
92 ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
93 _preserved_marks(preserved_marks),
94 _evacuation_failed_info(),
95 _evac_failure_regions(evac_failure_regions),
96 _evac_failure_enqueued_cards(0)
97 {
98 // We allocate number of young gen regions in the collection set plus one
99 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
100 // We also add a few elements at the beginning and at the end in
101 // an attempt to eliminate cache contention
102 const size_t padding_elem_num = (DEFAULT_PADDING_SIZE / sizeof(size_t));
103 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
104
105 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
106 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
107 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
108
109 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
110
111 _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
112 this,
113 collection_set->only_contains_young_regions());
114
115 _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
116
117 initialize_numa_stats();
118 }
119
120 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers, BufferNodeList* rdc_buffers) {
121 *rdc_buffers = _rdc_local_qset.flush();
122 flush_numa_stats();
123 // Update allocation statistics.
124 _plab_allocator->flush_and_retire_stats(num_workers);
125 _g1h->policy()->record_age_table(&_age_table);
126
127 if (_evacuation_failed_info.has_failed()) {
128 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
129 }
130
131 size_t sum = 0;
132 for (uint i = 0; i < _surviving_words_length; i++) {
133 surviving_young_words[i] += _surviving_young_words[i];
134 sum += _surviving_young_words[i];
135 }
136 return sum;
137 }
138
139 G1ParScanThreadState::~G1ParScanThreadState() {
140 delete _plab_allocator;
141 delete _closures;
142 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
143 delete[] _oops_into_optional_regions;
144 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
145 }
146
147 size_t G1ParScanThreadState::lab_waste_words() const {
148 return _plab_allocator->waste();
149 }
150
151 size_t G1ParScanThreadState::lab_undo_waste_words() const {
152 return _plab_allocator->undo_waste();
153 }
154
155 size_t G1ParScanThreadState::evac_failure_enqueued_cards() const {
156 return _evac_failure_enqueued_cards;
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 // Must be in the collection set--it's already been copied.
177 oop p = task->source();
178 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
179 }
180
181 void G1ParScanThreadState::verify_task(ScannerTask task) const {
182 if (task.is_narrow_oop_ptr()) {
183 verify_task(task.to_narrow_oop_ptr());
184 } else if (task.is_oop_ptr()) {
185 verify_task(task.to_oop_ptr());
186 } else if (task.is_partial_array_state()) {
187 verify_task(task.to_partial_array_state());
188 } else {
189 ShouldNotReachHere();
190 }
191 }
192 #endif // ASSERT
193
194 template <class T>
195 MAYBE_INLINE_EVACUATION
196 void G1ParScanThreadState::do_oop_evac(T* p) {
197 // Reference should not be null here as such are never pushed to the task queue.
198 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
199
200 // Although we never intentionally push references outside of the collection
201 // set, due to (benign) races in the claim mechanism during RSet scanning more
202 // than one thread might claim the same card. So the same card may be
203 // processed multiple times, and so we might get references into old gen here.
204 // So we need to redo this check.
205 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
206 // References pushed onto the work stack should never point to a humongous region
207 // as they are not added to the collection set due to above precondition.
208 assert(!region_attr.is_humongous_candidate(),
209 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
210 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
211
212 if (!region_attr.is_in_cset()) {
213 // In this case somebody else already did all the work.
214 return;
215 }
216
217 markWord m = obj->mark();
218 if (m.is_forwarded()) {
219 obj = m.forwardee();
220 } else {
221 obj = do_copy_to_survivor_space(region_attr, obj, m);
222 }
223 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
224
225 write_ref_field_post(p, obj);
226 }
227
228 MAYBE_INLINE_EVACUATION
229 void G1ParScanThreadState::do_partial_array(PartialArrayState* state) {
230 oop to_obj = state->destination();
231
232 #ifdef ASSERT
233 oop from_obj = state->source();
234 assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
235 assert(from_obj->is_objArray(), "must be obj array");
236 assert(from_obj->is_forwarded(), "must be forwarded");
237 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
238 assert(to_obj->is_objArray(), "must be obj array");
239 #endif // ASSERT
240
241 objArrayOop to_array = objArrayOop(to_obj);
242
243 // Claim a chunk and get number of additional tasks to enqueue.
244 PartialArrayTaskStepper::Step step = _partial_array_stepper.next(state);
245 // Push any additional partial scan tasks needed. Pushed before processing
246 // the claimed chunk to allow other workers to steal while we're processing.
247 if (step._ncreate > 0) {
248 state->add_references(step._ncreate);
249 for (uint i = 0; i < step._ncreate; ++i) {
250 push_on_queue(ScannerTask(state));
251 }
252 }
253
254 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
255 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
256 // Process claimed task.
257 to_array->oop_iterate_range(&_scanner,
258 checked_cast<int>(step._index),
259 checked_cast<int>(step._index + _partial_array_stepper.chunk_size()));
260 // Release reference to the state, now that we're done with it.
261 _partial_array_state_allocator->release(_worker_id, state);
262 }
263
264 MAYBE_INLINE_EVACUATION
265 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
266 oop from_obj,
267 oop to_obj) {
268 assert(from_obj->is_objArray(), "precondition");
269 assert(from_obj->is_forwarded(), "precondition");
270 assert(from_obj->forwardee() == to_obj, "precondition");
271 assert(to_obj->is_objArray(), "precondition");
272
273 objArrayOop to_array = objArrayOop(to_obj);
274
275 size_t array_length = to_array->length();
276 PartialArrayTaskStepper::Step step = _partial_array_stepper.start(array_length);
277
278 // Push any needed partial scan tasks. Pushed before processing the
279 // initial chunk to allow other workers to steal while we're processing.
280 if (step._ncreate > 0) {
281 assert(step._index < array_length, "invariant");
282 assert(((array_length - step._index) % _partial_array_stepper.chunk_size()) == 0,
283 "invariant");
284 PartialArrayState* state =
285 _partial_array_state_allocator->allocate(_worker_id,
286 from_obj, to_obj,
287 step._index,
288 array_length,
289 step._ncreate);
290 for (uint i = 0; i < step._ncreate; ++i) {
291 push_on_queue(ScannerTask(state));
292 }
293 } else {
294 assert(step._index == array_length, "invariant");
295 }
296
297 // Skip the card enqueue iff the object (to_array) is in survivor region.
298 // However, G1HeapRegion::is_survivor() is too expensive here.
299 // Instead, we use dest_attr.is_young() because the two values are always
300 // equal: successfully allocated young regions must be survivor regions.
301 assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
302 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
303 // Process the initial chunk. No need to process the type in the
304 // klass, as it will already be handled by processing the built-in
305 // module.
306 to_array->oop_iterate_range(&_scanner, 0, checked_cast<int>(step._index));
307 }
308
309 MAYBE_INLINE_EVACUATION
310 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
311 verify_task(task);
312 if (task.is_narrow_oop_ptr()) {
313 do_oop_evac(task.to_narrow_oop_ptr());
314 } else if (task.is_oop_ptr()) {
315 do_oop_evac(task.to_oop_ptr());
316 } else {
317 do_partial_array(task.to_partial_array_state());
318 }
319 }
320
321 // Process tasks until overflow queue is empty and local queue
322 // contains no more than threshold entries. NOINLINE to prevent
323 // inlining into steal_and_trim_queue.
324 ATTRIBUTE_FLATTEN NOINLINE
325 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
326 ScannerTask task;
327 do {
328 while (_task_queue->pop_overflow(task)) {
329 if (!_task_queue->try_push_to_taskqueue(task)) {
330 dispatch_task(task);
331 }
332 }
333 while (_task_queue->pop_local(task, threshold)) {
334 dispatch_task(task);
335 }
336 } while (!_task_queue->overflow_empty());
337 }
338
339 ATTRIBUTE_FLATTEN
340 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
341 ScannerTask stolen_task;
342 while (task_queues->steal(_worker_id, stolen_task)) {
343 dispatch_task(stolen_task);
344 // Processing stolen task may have added tasks to our queue.
345 trim_queue();
346 }
347 }
348
349 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
350 size_t word_sz,
351 bool previous_plab_refill_failed,
352 uint node_index) {
353
354 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
355
356 // Right now we only have two types of regions (young / old) so
357 // let's keep the logic here simple. We can generalize it when necessary.
358 if (dest->is_young()) {
359 bool plab_refill_in_old_failed = false;
360 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
361 word_sz,
362 &plab_refill_in_old_failed,
363 node_index);
364 // Make sure that we won't attempt to copy any other objects out
365 // of a survivor region (given that apparently we cannot allocate
366 // any new ones) to avoid coming into this slow path again and again.
367 // Only consider failed PLAB refill here: failed inline allocations are
368 // typically large, so not indicative of remaining space.
369 if (previous_plab_refill_failed) {
370 _tenuring_threshold = 0;
371 }
372
373 if (obj_ptr != nullptr) {
374 dest->set_old();
375 } else {
376 // We just failed to allocate in old gen. The same idea as explained above
377 // for making survivor gen unavailable for allocation applies for old gen.
378 _old_gen_is_full = plab_refill_in_old_failed;
379 }
380 return obj_ptr;
381 } else {
382 _old_gen_is_full = previous_plab_refill_failed;
383 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
384 // no other space to try.
385 return nullptr;
386 }
387 }
388
389 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
390 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
391
392 if (region_attr.is_young()) {
393 age = !m.has_displaced_mark_helper() ? m.age()
394 : m.displaced_mark_helper().age();
395 if (age < _tenuring_threshold) {
396 return region_attr;
397 }
398 }
399 // young-to-old (promotion) or old-to-old; destination is old in both cases.
400 return G1HeapRegionAttr::Old;
401 }
402
403 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
404 oop const old, size_t word_sz, uint age,
405 HeapWord * const obj_ptr, uint node_index) const {
406 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
407 if (alloc_buf->contains(obj_ptr)) {
408 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
409 dest_attr.type() == G1HeapRegionAttr::Old,
410 alloc_buf->word_sz() * HeapWordSize);
411 } else {
412 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
413 dest_attr.type() == G1HeapRegionAttr::Old);
414 }
415 }
416
417 NOINLINE
418 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
419 oop old,
420 size_t word_sz,
421 uint age,
422 uint node_index) {
423 HeapWord* obj_ptr = nullptr;
424 // Try slow-path allocation unless we're allocating old and old is already full.
425 if (!(dest_attr->is_old() && _old_gen_is_full)) {
426 bool plab_refill_failed = false;
427 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
428 word_sz,
429 &plab_refill_failed,
430 node_index);
431 if (obj_ptr == nullptr) {
432 obj_ptr = allocate_in_next_plab(dest_attr,
433 word_sz,
434 plab_refill_failed,
435 node_index);
436 }
437 }
438 if (obj_ptr != nullptr) {
439 update_numa_stats(node_index);
440 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
441 // The events are checked individually as part of the actual commit
442 report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
443 }
444 }
445 return obj_ptr;
446 }
447
448 #if ALLOCATION_FAILURE_INJECTOR
449 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
450 return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
451 }
452 #endif
453
454 NOINLINE
455 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
456 HeapWord* obj_ptr,
457 size_t word_sz,
458 uint node_index) {
459 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
460 }
461
462 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
463 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
464 G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
465 region->update_bot_for_block(obj_start, obj_start + word_sz);
466 }
467
468 // Private inline function, for direct internal use and providing the
469 // implementation of the public not-inline function.
470 MAYBE_INLINE_EVACUATION
471 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
472 oop const old,
473 markWord const old_mark) {
474 assert(region_attr.is_in_cset(),
475 "Unexpected region attr type: %s", region_attr.get_type_str());
476
477 // Get the klass once. We'll need it again later, and this avoids
478 // re-decoding when it's compressed.
479 Klass* klass = old->klass();
480 const size_t word_sz = old->size_given_klass(klass);
481
482 // JNI only allows pinning of typeArrays, so we only need to keep those in place.
483 if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
484 return handle_evacuation_failure_par(old, old_mark, word_sz, true /* cause_pinned */);
485 }
486
487 uint age = 0;
488 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
489 G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
490 uint node_index = from_region->node_index();
491
492 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
493
494 // PLAB allocations should succeed most of the time, so we'll
495 // normally check against null once and that's it.
496 if (obj_ptr == nullptr) {
497 obj_ptr = allocate_copy_slow(&dest_attr, old, word_sz, age, node_index);
498 if (obj_ptr == nullptr) {
499 // This will either forward-to-self, or detect that someone else has
500 // installed a forwarding pointer.
501 return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
502 }
503 }
504
505 assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
506 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
507
508 // Should this evacuation fail?
509 if (inject_allocation_failure(from_region->hrm_index())) {
510 // Doing this after all the allocation attempts also tests the
511 // undo_allocation() method too.
512 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
513 return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
514 }
515
516 // We're going to allocate linearly, so might as well prefetch ahead.
517 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
518 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
519
520 const oop obj = cast_to_oop(obj_ptr);
521 // Because the forwarding is done with memory_order_relaxed there is no
522 // ordering with the above copy. Clients that get the forwardee must not
523 // examine its contents without other synchronization, since the contents
524 // may not be up to date for them.
525 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
526 if (forward_ptr == nullptr) {
527
528 {
529 const uint young_index = from_region->young_index_in_cset();
530 assert((from_region->is_young() && young_index > 0) ||
531 (!from_region->is_young() && young_index == 0), "invariant" );
532 _surviving_young_words[young_index] += word_sz;
533 }
534
535 if (dest_attr.is_young()) {
536 if (age < markWord::max_age) {
537 age++;
538 obj->incr_age();
539 }
540 _age_table.add(age, word_sz);
541 } else {
542 update_bot_after_copying(obj, word_sz);
543 }
544
545 // Most objects are not arrays, so do one array check rather than
546 // checking for each array category for each object.
547 if (klass->is_array_klass()) {
548 if (klass->is_objArray_klass()) {
549 start_partial_objarray(dest_attr, old, obj);
550 } else {
551 // Nothing needs to be done for typeArrays. Body doesn't contain
552 // any oops to scan, and the type in the klass will already be handled
553 // by processing the built-in module.
554 assert(klass->is_typeArray_klass(), "invariant");
555 }
556 return obj;
557 }
558
559 ContinuationGCSupport::transform_stack_chunk(obj);
560
561 // Check for deduplicating young Strings.
562 if (G1StringDedup::is_candidate_from_evacuation(klass,
563 region_attr,
564 dest_attr,
565 age)) {
566 // Record old; request adds a new weak reference, which reference
567 // processing expects to refer to a from-space object.
568 _string_dedup_requests.add(old);
569 }
570
571 // Skip the card enqueue iff the object (obj) is in survivor region.
572 // However, G1HeapRegion::is_survivor() is too expensive here.
573 // Instead, we use dest_attr.is_young() because the two values are always
574 // equal: successfully allocated young regions must be survivor regions.
575 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
576 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
577 obj->oop_iterate_backwards(&_scanner, klass);
578 return obj;
579 } else {
580 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
581 return forward_ptr;
582 }
583 }
584
585 // Public not-inline entry point.
586 ATTRIBUTE_FLATTEN
587 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
588 oop old,
589 markWord old_mark) {
590 return do_copy_to_survivor_space(region_attr, old, old_mark);
591 }
592
593 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
594 assert(worker_id < _num_workers, "out of bounds access");
595 if (_states[worker_id] == nullptr) {
596 _states[worker_id] =
597 new G1ParScanThreadState(_g1h, rdcqs(),
598 _preserved_marks_set.get(worker_id),
599 worker_id,
600 _num_workers,
601 _collection_set,
602 _evac_failure_regions,
603 &_partial_array_state_allocator);
604 }
605 return _states[worker_id];
606 }
607
608 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
609 assert(_flushed, "thread local state from the per thread states should have been flushed");
610 return _surviving_young_words_total;
611 }
612
613 void G1ParScanThreadStateSet::flush_stats() {
614 assert(!_flushed, "thread local state from the per thread states should be flushed once");
615 for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
616 G1ParScanThreadState* pss = _states[worker_id];
617 assert(pss != nullptr, "must be initialized");
618
619 G1GCPhaseTimes* p = _g1h->phase_times();
620
621 // Need to get the following two before the call to G1ParThreadScanState::flush()
622 // because it resets the PLAB allocator where we get this info from.
623 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
624 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
625 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers, &_rdc_buffers[worker_id]) * HeapWordSize;
626 size_t evac_fail_enqueued_cards = pss->evac_failure_enqueued_cards();
627
628 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
629 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
630 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
631 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_fail_enqueued_cards, G1GCPhaseTimes::MergePSSEvacFailExtra);
632
633 delete pss;
634 _states[worker_id] = nullptr;
635 }
636
637 G1DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
638 dcq.merge_bufferlists(rdcqs());
639 rdcqs()->verify_empty();
640
641 _flushed = true;
642 }
643
644 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
645 for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
646 G1ParScanThreadState* pss = _states[worker_index];
647 assert(pss != nullptr, "must be initialized");
648
649 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
650 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
651 }
652 }
653
654 NOINLINE
655 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz, bool cause_pinned) {
656 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
657
658 oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
659 if (forward_ptr == nullptr) {
660 // Forward-to-self succeeded. We are the "owner" of the object.
661 G1HeapRegion* r = _g1h->heap_region_containing(old);
662
663 if (_evac_failure_regions->record(_worker_id, r->hrm_index(), cause_pinned)) {
664 G1HeapRegionPrinter::evac_failure(r);
665 }
666
667 // Mark the failing object in the marking bitmap and later use the bitmap to handle
668 // evacuation failure recovery.
669 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
670
671 _preserved_marks->push_if_necessary(old, m);
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 _preserved_marks_set(true /* in_c_heap */),
731 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
732 _rdc_buffers(NEW_C_HEAP_ARRAY(BufferNodeList, num_workers, mtGC)),
733 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
734 _num_workers(num_workers),
735 _flushed(false),
736 _evac_failure_regions(evac_failure_regions),
737 _partial_array_state_allocator(num_workers)
738 {
739 _preserved_marks_set.init(num_workers);
740 for (uint i = 0; i < num_workers; ++i) {
741 _states[i] = nullptr;
742 _rdc_buffers[i] = BufferNodeList();
743 }
744 memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
745 }
746
747 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
748 assert(_flushed, "thread local state from the per thread states should have been flushed");
749 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
750 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
751 FREE_C_HEAP_ARRAY(BufferNodeList, _rdc_buffers);
752 _preserved_marks_set.reclaim();
753 }