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