1 /*
2 * Copyright (c) 2014, 2021, 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/partialArrayTaskStepper.inline.hpp"
37 #include "gc/shared/preservedMarks.inline.hpp"
38 #include "gc/shared/stringdedup/stringDedup.hpp"
39 #include "gc/shared/taskqueue.inline.hpp"
40 #include "memory/allocation.inline.hpp"
41 #include "oops/access.inline.hpp"
42 #include "oops/oop.inline.hpp"
43 #include "runtime/atomic.hpp"
44 #include "runtime/prefetch.inline.hpp"
45 #include "utilities/globalDefinitions.hpp"
46 #include "utilities/macros.hpp"
47
48 // In fastdebug builds the code size can get out of hand, potentially
49 // tripping over compiler limits (which may be bugs, but nevertheless
50 // need to be taken into consideration). A side benefit of limiting
51 // inlining is that we get more call frames that might aid debugging.
52 // And the fastdebug compile time for this file is much reduced.
53 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
54 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
55
56 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
57 G1RedirtyCardsQueueSet* rdcqs,
58 PreservedMarks* preserved_marks,
59 uint worker_id,
60 uint n_workers,
61 size_t young_cset_length,
62 size_t optional_cset_length,
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(NULL),
69 _plab_allocator(NULL),
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(NULL),
79 _surviving_young_words(NULL),
80 _surviving_words_length(young_cset_length + 1),
81 _old_gen_is_full(false),
82 _partial_objarray_chunk_size(ParGCArrayScanChunk),
83 _partial_array_stepper(n_workers),
84 _string_dedup_requests(),
85 _num_optional_regions(optional_cset_length),
86 _numa(g1h->numa()),
87 _obj_alloc_stat(NULL),
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(this, _g1h);
107
108 _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
109
110 initialize_numa_stats();
111 }
112
113 size_t G1ParScanThreadState::flush(size_t* surviving_young_words) {
114 _rdc_local_qset.flush();
115 flush_numa_stats();
116 // Update allocation statistics.
117 _plab_allocator->flush_and_retire_stats();
118 _g1h->policy()->record_age_table(&_age_table);
119
120 if (_evacuation_failed_info.has_failed()) {
121 _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
122 }
123
124 size_t sum = 0;
125 for (uint i = 0; i < _surviving_words_length; i++) {
126 surviving_young_words[i] += _surviving_young_words[i];
127 sum += _surviving_young_words[i];
128 }
129 return sum;
130 }
131
132 G1ParScanThreadState::~G1ParScanThreadState() {
133 delete _plab_allocator;
134 delete _closures;
135 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
136 delete[] _oops_into_optional_regions;
137 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
138 }
139
140 size_t G1ParScanThreadState::lab_waste_words() const {
141 return _plab_allocator->waste();
142 }
143
144 size_t G1ParScanThreadState::lab_undo_waste_words() const {
145 return _plab_allocator->undo_waste();
146 }
147
148 #ifdef ASSERT
149 void G1ParScanThreadState::verify_task(narrowOop* task) const {
150 assert(task != NULL, "invariant");
151 assert(UseCompressedOops, "sanity");
152 oop p = RawAccess<>::oop_load(task);
153 assert(_g1h->is_in_reserved(p),
154 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
155 }
156
157 void G1ParScanThreadState::verify_task(oop* task) const {
158 assert(task != NULL, "invariant");
159 oop p = RawAccess<>::oop_load(task);
160 assert(_g1h->is_in_reserved(p),
161 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
162 }
163
164 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
165 // Must be in the collection set--it's already been copied.
166 oop p = task.to_source_array();
167 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
168 }
169
170 void G1ParScanThreadState::verify_task(ScannerTask task) const {
171 if (task.is_narrow_oop_ptr()) {
172 verify_task(task.to_narrow_oop_ptr());
173 } else if (task.is_oop_ptr()) {
174 verify_task(task.to_oop_ptr());
175 } else if (task.is_partial_array_task()) {
176 verify_task(task.to_partial_array_task());
177 } else {
178 ShouldNotReachHere();
179 }
180 }
181 #endif // ASSERT
182
183 template <class T>
184 MAYBE_INLINE_EVACUATION
185 void G1ParScanThreadState::do_oop_evac(T* p) {
186 // Reference should not be NULL here as such are never pushed to the task queue.
187 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
188
189 // Although we never intentionally push references outside of the collection
190 // set, due to (benign) races in the claim mechanism during RSet scanning more
191 // than one thread might claim the same card. So the same card may be
192 // processed multiple times, and so we might get references into old gen here.
193 // So we need to redo this check.
194 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
195 // References pushed onto the work stack should never point to a humongous region
196 // as they are not added to the collection set due to above precondition.
197 assert(!region_attr.is_humongous(),
198 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
199 p2i(obj), _g1h->addr_to_region(cast_from_oop<HeapWord*>(obj)), p2i(p));
200
201 if (!region_attr.is_in_cset()) {
202 // In this case somebody else already did all the work.
203 return;
204 }
205
206 markWord m = obj->mark();
207 if (m.is_marked()) {
208 obj = obj->forwardee(m);
209 } else {
210 obj = do_copy_to_survivor_space(region_attr, obj, m);
211 }
212 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
213
214 write_ref_field_post(p, obj);
215 }
216
217 MAYBE_INLINE_EVACUATION
218 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
219 oop from_obj = task.to_source_array();
220
221 assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
222 assert(from_obj->is_forwarded(), "must be forwarded");
223
224 oop to_obj = from_obj->forwardee();
225 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
226 assert(to_obj->is_objArray(), "must be obj array");
227 objArrayOop to_array = objArrayOop(to_obj);
228
229 PartialArrayTaskStepper::Step step
230 = _partial_array_stepper.next(objArrayOop(from_obj),
231 to_array,
232 _partial_objarray_chunk_size);
233 for (uint i = 0; i < step._ncreate; ++i) {
234 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
235 }
236
237 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
238 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
239 // Process claimed task. The length of to_array is not correct, but
240 // fortunately the iteration ignores the length field and just relies
241 // on start/end.
242 to_array->oop_iterate_range(&_scanner,
243 step._index,
244 step._index + _partial_objarray_chunk_size);
245 }
246
247 MAYBE_INLINE_EVACUATION
248 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
249 oop from_obj,
250 oop to_obj) {
251 assert(from_obj->is_forwarded(), "precondition");
252 assert(from_obj->forwardee() == to_obj, "precondition");
253 assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
254 assert(to_obj->is_objArray(), "precondition");
255
256 objArrayOop to_array = objArrayOop(to_obj);
257
258 PartialArrayTaskStepper::Step step
259 = _partial_array_stepper.start(objArrayOop(from_obj),
260 to_array,
261 _partial_objarray_chunk_size);
262
263 // Push any needed partial scan tasks. Pushed before processing the
264 // intitial chunk to allow other workers to steal while we're processing.
265 for (uint i = 0; i < step._ncreate; ++i) {
266 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
267 }
268
269 // Skip the card enqueue iff the object (to_array) is in survivor region.
270 // However, HeapRegion::is_survivor() is too expensive here.
271 // Instead, we use dest_attr.is_young() because the two values are always
272 // equal: successfully allocated young regions must be survivor regions.
273 assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
274 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
275 // Process the initial chunk. No need to process the type in the
276 // klass, as it will already be handled by processing the built-in
277 // module. The length of to_array is not correct, but fortunately
278 // the iteration ignores that length field and relies on start/end.
279 to_array->oop_iterate_range(&_scanner, 0, step._index);
280 }
281
282 MAYBE_INLINE_EVACUATION
283 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
284 verify_task(task);
285 if (task.is_narrow_oop_ptr()) {
286 do_oop_evac(task.to_narrow_oop_ptr());
287 } else if (task.is_oop_ptr()) {
288 do_oop_evac(task.to_oop_ptr());
289 } else {
290 do_partial_array(task.to_partial_array_task());
291 }
292 }
293
294 // Process tasks until overflow queue is empty and local queue
295 // contains no more than threshold entries. NOINLINE to prevent
296 // inlining into steal_and_trim_queue.
297 ATTRIBUTE_FLATTEN NOINLINE
298 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
299 ScannerTask task;
300 do {
301 while (_task_queue->pop_overflow(task)) {
302 if (!_task_queue->try_push_to_taskqueue(task)) {
303 dispatch_task(task);
304 }
305 }
306 while (_task_queue->pop_local(task, threshold)) {
307 dispatch_task(task);
308 }
309 } while (!_task_queue->overflow_empty());
310 }
311
312 ATTRIBUTE_FLATTEN
313 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
314 ScannerTask stolen_task;
315 while (task_queues->steal(_worker_id, stolen_task)) {
316 dispatch_task(stolen_task);
317 // Processing stolen task may have added tasks to our queue.
318 trim_queue();
319 }
320 }
321
322 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
323 size_t word_sz,
324 bool previous_plab_refill_failed,
325 uint node_index) {
326
327 assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
328
329 // Right now we only have two types of regions (young / old) so
330 // let's keep the logic here simple. We can generalize it when necessary.
331 if (dest->is_young()) {
332 bool plab_refill_in_old_failed = false;
333 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
334 word_sz,
335 &plab_refill_in_old_failed,
336 node_index);
337 // Make sure that we won't attempt to copy any other objects out
338 // of a survivor region (given that apparently we cannot allocate
339 // any new ones) to avoid coming into this slow path again and again.
340 // Only consider failed PLAB refill here: failed inline allocations are
341 // typically large, so not indicative of remaining space.
342 if (previous_plab_refill_failed) {
343 _tenuring_threshold = 0;
344 }
345
346 if (obj_ptr != NULL) {
347 dest->set_old();
348 } else {
349 // We just failed to allocate in old gen. The same idea as explained above
350 // for making survivor gen unavailable for allocation applies for old gen.
351 _old_gen_is_full = plab_refill_in_old_failed;
352 }
353 return obj_ptr;
354 } else {
355 _old_gen_is_full = previous_plab_refill_failed;
356 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
357 // no other space to try.
358 return NULL;
359 }
360 }
361
362 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
363 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
364
365 if (region_attr.is_young()) {
366 age = !m.has_displaced_mark_helper() ? m.age()
367 : m.displaced_mark_helper().age();
368 if (age < _tenuring_threshold) {
369 return region_attr;
370 }
371 }
372 // young-to-old (promotion) or old-to-old; destination is old in both cases.
373 return G1HeapRegionAttr::Old;
374 }
375
376 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
377 oop const old, Klass* klass, size_t word_sz, uint age,
378 HeapWord * const obj_ptr, uint node_index) const {
379 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
380 if (alloc_buf->contains(obj_ptr)) {
381 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
382 dest_attr.type() == G1HeapRegionAttr::Old,
383 alloc_buf->word_sz() * HeapWordSize);
384 } else {
385 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
386 dest_attr.type() == G1HeapRegionAttr::Old);
387 }
388 }
389
390 NOINLINE
391 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
392 oop old,
393 Klass* klass,
394 size_t word_sz,
395 uint age,
396 uint node_index) {
397 HeapWord* obj_ptr = NULL;
398 // Try slow-path allocation unless we're allocating old and old is already full.
399 if (!(dest_attr->is_old() && _old_gen_is_full)) {
400 bool plab_refill_failed = false;
401 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
402 word_sz,
403 &plab_refill_failed,
404 node_index);
405 if (obj_ptr == NULL) {
406 obj_ptr = allocate_in_next_plab(dest_attr,
407 word_sz,
408 plab_refill_failed,
409 node_index);
410 }
411 }
412 if (obj_ptr != NULL) {
413 update_numa_stats(node_index);
414 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
415 // The events are checked individually as part of the actual commit
416 report_promotion_event(*dest_attr, old, klass, word_sz, age, obj_ptr, node_index);
417 }
418 }
419 return obj_ptr;
420 }
421
422 #if EVAC_FAILURE_INJECTOR
423 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
424 return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
425 }
426 #endif
427
428 NOINLINE
429 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
430 HeapWord* obj_ptr,
431 size_t word_sz,
432 uint node_index) {
433 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
434 }
435
436 // Private inline function, for direct internal use and providing the
437 // implementation of the public not-inline function.
438 MAYBE_INLINE_EVACUATION
439 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
440 oop const old,
441 markWord const old_mark) {
442 assert(region_attr.is_in_cset(),
443 "Unexpected region attr type: %s", region_attr.get_type_str());
444
445 if (old_mark.is_marked()) {
446 // Already forwarded by somebody else, return forwardee.
447 return old->forwardee(old_mark);
448 }
449 // Get the klass once. We'll need it again later, and this avoids
450 // re-decoding when it's compressed.
451 #ifdef _LP64
452 Klass* klass = old_mark.safe_klass();
453 #else
454 Klass* klass = old->klass();
455 #endif
456 const size_t word_sz = old->size_given_klass(klass);
457
458 uint age = 0;
459 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
460 HeapRegion* const from_region = _g1h->heap_region_containing(old);
461 uint node_index = from_region->node_index();
462
463 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
464
465 // PLAB allocations should succeed most of the time, so we'll
466 // normally check against NULL once and that's it.
467 if (obj_ptr == NULL) {
468 obj_ptr = allocate_copy_slow(&dest_attr, old, klass, word_sz, age, node_index);
469 if (obj_ptr == NULL) {
470 // This will either forward-to-self, or detect that someone else has
471 // installed a forwarding pointer.
472 return handle_evacuation_failure_par(old, old_mark, word_sz);
473 }
474 }
475
476 assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
477 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
478
479 // Should this evacuation fail?
480 if (inject_evacuation_failure(from_region->hrm_index())) {
481 // Doing this after all the allocation attempts also tests the
482 // undo_allocation() method too.
483 undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
484 return handle_evacuation_failure_par(old, old_mark, word_sz);
485 }
486
487 // We're going to allocate linearly, so might as well prefetch ahead.
488 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
489 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
490
491 const oop obj = cast_to_oop(obj_ptr);
492 // Because the forwarding is done with memory_order_relaxed there is no
493 // ordering with the above copy. Clients that get the forwardee must not
494 // examine its contents without other synchronization, since the contents
495 // may not be up to date for them.
496 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
497 if (forward_ptr == NULL) {
498
499 {
500 const uint young_index = from_region->young_index_in_cset();
501 assert((from_region->is_young() && young_index > 0) ||
502 (!from_region->is_young() && young_index == 0), "invariant" );
503 _surviving_young_words[young_index] += word_sz;
504 }
505
506 if (dest_attr.is_young()) {
507 if (age < markWord::max_age) {
508 age++;
509 obj->incr_age();
510 }
511 _age_table.add(age, word_sz);
512 } else {
513 // Currently we only have two destinations and we only need BOT updates for
514 // old. If the current allocation was done outside the PLAB this call will
515 // have no effect since the _top of the PLAB has not changed.
516 _plab_allocator->update_bot_for_plab_allocation(dest_attr, word_sz, node_index);
517 }
518
519 // Most objects are not arrays, so do one array check rather than
520 // checking for each array category for each object.
521 if (klass->is_array_klass()) {
522 if (klass->is_objArray_klass()) {
523 start_partial_objarray(dest_attr, old, obj);
524 } else {
525 // Nothing needs to be done for typeArrays. Body doesn't contain
526 // any oops to scan, and the type in the klass will already be handled
527 // by processing the built-in module.
528 assert(klass->is_typeArray_klass(), "invariant");
529 }
530 return obj;
531 }
532
533 // Check for deduplicating young Strings.
534 if (G1StringDedup::is_candidate_from_evacuation(klass,
535 region_attr,
536 dest_attr,
537 age)) {
538 // Record old; request adds a new weak reference, which reference
539 // processing expects to refer to a from-space object.
540 _string_dedup_requests.add(old);
541 }
542
543 // Skip the card enqueue iff the object (obj) is in survivor region.
544 // However, HeapRegion::is_survivor() is too expensive here.
545 // Instead, we use dest_attr.is_young() because the two values are always
546 // equal: successfully allocated young regions must be survivor regions.
547 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
548 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
549 obj->oop_iterate_backwards(&_scanner, klass);
550 return obj;
551 } else {
552 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
553 return forward_ptr;
554 }
555 }
556
557 // Public not-inline entry point.
558 ATTRIBUTE_FLATTEN
559 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
560 oop old,
561 markWord old_mark) {
562 return do_copy_to_survivor_space(region_attr, old, old_mark);
563 }
564
565 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
566 assert(worker_id < _n_workers, "out of bounds access");
567 if (_states[worker_id] == NULL) {
568 _states[worker_id] =
569 new G1ParScanThreadState(_g1h, rdcqs(),
570 _preserved_marks_set->get(worker_id),
571 worker_id, _n_workers,
572 _young_cset_length, _optional_cset_length,
573 _evac_failure_regions);
574 }
575 return _states[worker_id];
576 }
577
578 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
579 assert(_flushed, "thread local state from the per thread states should have been flushed");
580 return _surviving_young_words_total;
581 }
582
583 void G1ParScanThreadStateSet::flush() {
584 assert(!_flushed, "thread local state from the per thread states should be flushed once");
585
586 for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
587 G1ParScanThreadState* pss = _states[worker_id];
588 assert(pss != nullptr, "must be initialized");
589
590 G1GCPhaseTimes* p = _g1h->phase_times();
591
592 // Need to get the following two before the call to G1ParThreadScanState::flush()
593 // because it resets the PLAB allocator where we get this info from.
594 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
595 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
596 size_t copied_bytes = pss->flush(_surviving_young_words_total) * HeapWordSize;
597
598 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
599 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
600 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
601
602 delete pss;
603 _states[worker_id] = NULL;
604 }
605 _flushed = true;
606 }
607
608 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
609 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
610 G1ParScanThreadState* pss = _states[worker_index];
611 assert(pss != nullptr, "must be initialized");
612
613 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
614 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
615 }
616 }
617
618 NOINLINE
619 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
620 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
621
622 oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
623 if (forward_ptr == NULL) {
624 // Forward-to-self succeeded. We are the "owner" of the object.
625 HeapRegion* r = _g1h->heap_region_containing(old);
626 // Records evac failure objs, this will help speed up iteration
627 // of these objs later in *remove self forward* phase of post evacuation.
628 r->record_evac_failure_obj(old);
629
630 if (_evac_failure_regions->record(r->hrm_index())) {
631 _g1h->hr_printer()->evac_failure(r);
632 }
633
634 _preserved_marks->push_if_necessary(old, m);
635 _evacuation_failed_info.register_copy_failure(word_sz);
636
637 // For iterating objects that failed evacuation currently we can reuse the
638 // existing closure to scan evacuated objects because:
639 // - for objects referring into the collection set we do not need to gather
640 // cards at this time. The regions they are in will be unconditionally turned
641 // to old regions without remembered sets.
642 // - since we are iterating from a collection set region (i.e. never a Survivor
643 // region), we always need to gather cards for this case.
644 G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
645 old->oop_iterate_backwards(&_scanner);
646
647 return old;
648 } else {
649 // Forward-to-self failed. Either someone else managed to allocate
650 // space for this object (old != forward_ptr) or they beat us in
651 // self-forwarding it (old == forward_ptr).
652 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
653 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
654 "should not be in the CSet",
655 p2i(old), p2i(forward_ptr));
656 return forward_ptr;
657 }
658 }
659
660 void G1ParScanThreadState::initialize_numa_stats() {
661 if (_numa->is_enabled()) {
662 LogTarget(Info, gc, heap, numa) lt;
663
664 if (lt.is_enabled()) {
665 uint num_nodes = _numa->num_active_nodes();
666 // Record only if there are multiple active nodes.
667 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
668 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
669 }
670 }
671 }
672
673 void G1ParScanThreadState::flush_numa_stats() {
674 if (_obj_alloc_stat != NULL) {
675 uint node_index = _numa->index_of_current_thread();
676 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
677 }
678 }
679
680 void G1ParScanThreadState::update_numa_stats(uint node_index) {
681 if (_obj_alloc_stat != NULL) {
682 _obj_alloc_stat[node_index]++;
683 }
684 }
685
686 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
687 G1RedirtyCardsQueueSet* rdcqs,
688 PreservedMarksSet* preserved_marks_set,
689 uint n_workers,
690 size_t young_cset_length,
691 size_t optional_cset_length,
692 G1EvacFailureRegions* evac_failure_regions) :
693 _g1h(g1h),
694 _rdcqs(rdcqs),
695 _preserved_marks_set(preserved_marks_set),
696 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
697 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
698 _young_cset_length(young_cset_length),
699 _optional_cset_length(optional_cset_length),
700 _n_workers(n_workers),
701 _flushed(false),
702 _evac_failure_regions(evac_failure_regions) {
703 for (uint i = 0; i < n_workers; ++i) {
704 _states[i] = NULL;
705 }
706 memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
707 }
708
709 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
710 assert(_flushed, "thread local state from the per thread states should have been flushed");
711 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
712 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
713 }