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