1 /*
2 * Copyright (c) 2014, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/g1Allocator.inline.hpp"
27 #include "gc/g1/g1CollectedHeap.inline.hpp"
28 #include "gc/g1/g1CollectionSet.hpp"
29 #include "gc/g1/g1EvacFailureRegions.inline.hpp"
30 #include "gc/g1/g1OopClosures.inline.hpp"
31 #include "gc/g1/g1ParScanThreadState.inline.hpp"
32 #include "gc/g1/g1RootClosures.hpp"
33 #include "gc/g1/g1StringDedup.hpp"
34 #include "gc/g1/g1Trace.hpp"
35 #include "gc/g1/g1YoungGCEvacFailureInjector.inline.hpp"
36 #include "gc/shared/continuationGCSupport.inline.hpp"
37 #include "gc/shared/partialArrayTaskStepper.inline.hpp"
38 #include "gc/shared/preservedMarks.inline.hpp"
39 #include "gc/shared/stringdedup/stringDedup.hpp"
40 #include "gc/shared/taskqueue.inline.hpp"
41 #include "memory/allocation.inline.hpp"
42 #include "oops/access.inline.hpp"
43 #include "oops/oop.inline.hpp"
44 #include "runtime/atomic.hpp"
45 #include "runtime/prefetch.inline.hpp"
46 #include "utilities/globalDefinitions.hpp"
47 #include "utilities/macros.hpp"
48
49 // In fastdebug builds the code size can get out of hand, potentially
50 // tripping over compiler limits (which may be bugs, but nevertheless
51 // need to be taken into consideration). A side benefit of limiting
52 // inlining is that we get more call frames that might aid debugging.
53 // And the fastdebug compile time for this file is much reduced.
54 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
55 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
56
57 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
58 G1RedirtyCardsQueueSet* rdcqs,
59 PreservedMarks* preserved_marks,
60 uint worker_id,
61 uint n_workers,
62 size_t young_cset_length,
63 size_t optional_cset_length,
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(NULL),
70 _plab_allocator(NULL),
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(NULL),
80 _surviving_young_words(NULL),
81 _surviving_words_length(young_cset_length + 1),
82 _old_gen_is_full(false),
83 _partial_objarray_chunk_size(ParGCArrayScanChunk),
84 _partial_array_stepper(n_workers),
85 _string_dedup_requests(),
86 _num_optional_regions(optional_cset_length),
87 _numa(g1h->numa()),
88 _obj_alloc_stat(NULL),
89 EVAC_FAILURE_INJECTOR_ONLY(_evac_failure_inject_counter(0) COMMA)
90 _preserved_marks(preserved_marks),
91 _evacuation_failed_info(),
92 _evac_failure_regions(evac_failure_regions)
93 {
94 // We allocate number of young gen regions in the collection set plus one
95 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
96 // We also add a few elements at the beginning and at the end in
97 // an attempt to eliminate cache contention
98 const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
99 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
100
101 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
102 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
103 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
104
105 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
106
107 _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
108
109 _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
110
111 initialize_numa_stats();
112 }
113
114 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
115 _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 #ifdef ASSERT
150 void G1ParScanThreadState::verify_task(narrowOop* task) const {
151 assert(task != NULL, "invariant");
152 assert(UseCompressedOops, "sanity");
153 oop p = RawAccess<>::oop_load(task);
154 assert(_g1h->is_in_reserved(p),
155 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
156 }
157
158 void G1ParScanThreadState::verify_task(oop* task) const {
159 assert(task != NULL, "invariant");
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(PartialArrayScanTask task) const {
166 // Must be in the collection set--it's already been copied.
167 oop p = task.to_source_array();
168 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
169 }
170
171 void G1ParScanThreadState::verify_task(ScannerTask task) const {
172 if (task.is_narrow_oop_ptr()) {
173 verify_task(task.to_narrow_oop_ptr());
174 } else if (task.is_oop_ptr()) {
175 verify_task(task.to_oop_ptr());
176 } else if (task.is_partial_array_task()) {
177 verify_task(task.to_partial_array_task());
178 } else {
179 ShouldNotReachHere();
180 }
181 }
182 #endif // ASSERT
183
184 template <class T>
185 MAYBE_INLINE_EVACUATION
186 void G1ParScanThreadState::do_oop_evac(T* p) {
187 // Reference should not be NULL here as such are never pushed to the task queue.
188 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
189
190 // Although we never intentionally push references outside of the collection
191 // set, due to (benign) races in the claim mechanism during RSet scanning more
192 // than one thread might claim the same card. So the same card may be
193 // processed multiple times, and so we might get references into old gen here.
194 // So we need to redo this check.
195 const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
196 // References pushed onto the work stack should never point to a humongous region
197 // as they are not added to the collection set due to above precondition.
198 assert(!region_attr.is_humongous_candidate(),
199 "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
200 p2i(obj), _g1h->addr_to_region(obj), p2i(p));
201
202 if (!region_attr.is_in_cset()) {
203 // In this case somebody else already did all the work.
204 return;
205 }
206
207 markWord m = obj->mark();
208 if (m.is_marked()) {
209 obj = cast_to_oop(m.decode_pointer());
210 } else {
211 obj = do_copy_to_survivor_space(region_attr, obj, m);
212 }
213 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
214
215 write_ref_field_post(p, obj);
216 }
217
218 MAYBE_INLINE_EVACUATION
219 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
220 oop from_obj = task.to_source_array();
221
222 assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
223 assert(from_obj->is_objArray(), "must be obj array");
224 assert(from_obj->is_forwarded(), "must be forwarded");
225
226 oop to_obj = from_obj->forwardee();
227 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
228 assert(to_obj->is_objArray(), "must be obj array");
229 objArrayOop to_array = objArrayOop(to_obj);
230
231 PartialArrayTaskStepper::Step step
232 = _partial_array_stepper.next(objArrayOop(from_obj),
233 to_array,
234 _partial_objarray_chunk_size);
235 for (uint i = 0; i < step._ncreate; ++i) {
236 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
237 }
238
239 G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
240 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
241 // Process claimed task. The length of to_array is not correct, but
242 // fortunately the iteration ignores the length field and just relies
243 // on start/end.
244 to_array->oop_iterate_range(&_scanner,
245 step._index,
246 step._index + _partial_objarray_chunk_size);
247 }
248
249 MAYBE_INLINE_EVACUATION
250 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
251 oop from_obj,
252 oop to_obj) {
253 assert(from_obj->is_objArray(), "precondition");
254 assert(from_obj->is_forwarded(), "precondition");
255 assert(from_obj->forwardee() == to_obj, "precondition");
256 assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
257 assert(to_obj->is_objArray(), "precondition");
258
259 objArrayOop to_array = objArrayOop(to_obj);
260
261 PartialArrayTaskStepper::Step step
262 = _partial_array_stepper.start(objArrayOop(from_obj),
263 to_array,
264 _partial_objarray_chunk_size);
265
266 // Push any needed partial scan tasks. Pushed before processing the
267 // initial chunk to allow other workers to steal while we're processing.
268 for (uint i = 0; i < step._ncreate; ++i) {
269 push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
270 }
271
272 // Skip the card enqueue iff the object (to_array) is in survivor region.
273 // However, HeapRegion::is_survivor() is too expensive here.
274 // Instead, we use dest_attr.is_young() because the two values are always
275 // equal: successfully allocated young regions must be survivor regions.
276 assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
277 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
278 // Process the initial chunk. No need to process the type in the
279 // klass, as it will already be handled by processing the built-in
280 // module. The length of to_array is not correct, but fortunately
281 // the iteration ignores that length field and relies on start/end.
282 to_array->oop_iterate_range(&_scanner, 0, step._index);
283 }
284
285 MAYBE_INLINE_EVACUATION
286 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
287 verify_task(task);
288 if (task.is_narrow_oop_ptr()) {
289 do_oop_evac(task.to_narrow_oop_ptr());
290 } else if (task.is_oop_ptr()) {
291 do_oop_evac(task.to_oop_ptr());
292 } else {
293 do_partial_array(task.to_partial_array_task());
294 }
295 }
296
297 // Process tasks until overflow queue is empty and local queue
298 // contains no more than threshold entries. NOINLINE to prevent
299 // inlining into steal_and_trim_queue.
300 ATTRIBUTE_FLATTEN NOINLINE
301 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
302 ScannerTask task;
303 do {
304 while (_task_queue->pop_overflow(task)) {
305 if (!_task_queue->try_push_to_taskqueue(task)) {
306 dispatch_task(task);
307 }
308 }
309 while (_task_queue->pop_local(task, threshold)) {
310 dispatch_task(task);
311 }
312 } while (!_task_queue->overflow_empty());
313 }
314
315 ATTRIBUTE_FLATTEN
316 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
317 ScannerTask stolen_task;
318 while (task_queues->steal(_worker_id, stolen_task)) {
319 dispatch_task(stolen_task);
320 // Processing stolen task may have added tasks to our queue.
321 trim_queue();
322 }
323 }
324
325 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
326 size_t word_sz,
327 bool previous_plab_refill_failed,
328 uint node_index) {
329
330 assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
331
332 // Right now we only have two types of regions (young / old) so
333 // let's keep the logic here simple. We can generalize it when necessary.
334 if (dest->is_young()) {
335 bool plab_refill_in_old_failed = false;
336 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
337 word_sz,
338 &plab_refill_in_old_failed,
339 node_index);
340 // Make sure that we won't attempt to copy any other objects out
341 // of a survivor region (given that apparently we cannot allocate
342 // any new ones) to avoid coming into this slow path again and again.
343 // Only consider failed PLAB refill here: failed inline allocations are
344 // typically large, so not indicative of remaining space.
345 if (previous_plab_refill_failed) {
346 _tenuring_threshold = 0;
347 }
348
349 if (obj_ptr != NULL) {
350 dest->set_old();
351 } else {
352 // We just failed to allocate in old gen. The same idea as explained above
353 // for making survivor gen unavailable for allocation applies for old gen.
354 _old_gen_is_full = plab_refill_in_old_failed;
355 }
356 return obj_ptr;
357 } else {
358 _old_gen_is_full = previous_plab_refill_failed;
359 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
360 // no other space to try.
361 return NULL;
362 }
363 }
364
365 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
366 assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
367
368 if (region_attr.is_young()) {
369 age = !m.has_displaced_mark_helper() ? m.age()
370 : m.displaced_mark_helper().age();
371 if (age < _tenuring_threshold) {
372 return region_attr;
373 }
374 }
375 // young-to-old (promotion) or old-to-old; destination is old in both cases.
376 return G1HeapRegionAttr::Old;
377 }
378
379 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
380 oop const old, size_t word_sz, uint age,
381 HeapWord * const obj_ptr, uint node_index) const {
382 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
383 if (alloc_buf->contains(obj_ptr)) {
384 _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
385 dest_attr.type() == G1HeapRegionAttr::Old,
386 alloc_buf->word_sz() * HeapWordSize);
387 } else {
388 _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
389 dest_attr.type() == G1HeapRegionAttr::Old);
390 }
391 }
392
393 NOINLINE
394 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
395 oop old,
396 size_t word_sz,
397 uint age,
398 uint node_index) {
399 HeapWord* obj_ptr = NULL;
400 // Try slow-path allocation unless we're allocating old and old is already full.
401 if (!(dest_attr->is_old() && _old_gen_is_full)) {
402 bool plab_refill_failed = false;
403 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
404 word_sz,
405 &plab_refill_failed,
406 node_index);
407 if (obj_ptr == NULL) {
408 obj_ptr = allocate_in_next_plab(dest_attr,
409 word_sz,
410 plab_refill_failed,
411 node_index);
412 }
413 }
414 if (obj_ptr != NULL) {
415 update_numa_stats(node_index);
416 if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
417 // The events are checked individually as part of the actual commit
418 report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
419 }
420 }
421 return obj_ptr;
422 }
423
424 #if EVAC_FAILURE_INJECTOR
425 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
426 return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
427 }
428 #endif
429
430 NOINLINE
431 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
432 HeapWord* obj_ptr,
433 size_t word_sz,
434 uint node_index) {
435 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
436 }
437
438 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
439 HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
440 HeapRegion* region = _g1h->heap_region_containing(obj_start);
441 region->update_bot_for_obj(obj_start, word_sz);
442 }
443
444 // Private inline function, for direct internal use and providing the
445 // implementation of the public not-inline function.
446 MAYBE_INLINE_EVACUATION
447 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
448 oop const old,
449 markWord const old_mark) {
450 assert(region_attr.is_in_cset(),
451 "Unexpected region attr type: %s", region_attr.get_type_str());
452
453 // Get the klass once. We'll need it again later, and this avoids
454 // re-decoding when it's compressed.
455 Klass* klass = old->klass();
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, 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 update_bot_after_copying(obj, word_sz);
514 }
515
516 // Most objects are not arrays, so do one array check rather than
517 // checking for each array category for each object.
518 if (klass->is_array_klass()) {
519 if (klass->is_objArray_klass()) {
520 start_partial_objarray(dest_attr, old, obj);
521 } else {
522 // Nothing needs to be done for typeArrays. Body doesn't contain
523 // any oops to scan, and the type in the klass will already be handled
524 // by processing the built-in module.
525 assert(klass->is_typeArray_klass(), "invariant");
526 }
527 return obj;
528 }
529
530 ContinuationGCSupport::transform_stack_chunk(obj);
531
532 // Check for deduplicating young Strings.
533 if (G1StringDedup::is_candidate_from_evacuation(klass,
534 region_attr,
535 dest_attr,
536 age)) {
537 // Record old; request adds a new weak reference, which reference
538 // processing expects to refer to a from-space object.
539 _string_dedup_requests.add(old);
540 }
541
542 // Skip the card enqueue iff the object (obj) is in survivor region.
543 // However, HeapRegion::is_survivor() is too expensive here.
544 // Instead, we use dest_attr.is_young() because the two values are always
545 // equal: successfully allocated young regions must be survivor regions.
546 assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
547 G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
548 obj->oop_iterate_backwards(&_scanner, klass);
549 return obj;
550 } else {
551 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
552 return forward_ptr;
553 }
554 }
555
556 // Public not-inline entry point.
557 ATTRIBUTE_FLATTEN
558 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
559 oop old,
560 markWord old_mark) {
561 return do_copy_to_survivor_space(region_attr, old, old_mark);
562 }
563
564 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
565 assert(worker_id < _n_workers, "out of bounds access");
566 if (_states[worker_id] == NULL) {
567 _states[worker_id] =
568 new G1ParScanThreadState(_g1h, rdcqs(),
569 _preserved_marks_set.get(worker_id),
570 worker_id, _n_workers,
571 _young_cset_length, _optional_cset_length,
572 _evac_failure_regions);
573 }
574 return _states[worker_id];
575 }
576
577 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
578 assert(_flushed, "thread local state from the per thread states should have been flushed");
579 return _surviving_young_words_total;
580 }
581
582 void G1ParScanThreadStateSet::flush_stats() {
583 assert(!_flushed, "thread local state from the per thread states should be flushed once");
584
585 for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
586 G1ParScanThreadState* pss = _states[worker_id];
587 assert(pss != nullptr, "must be initialized");
588
589 G1GCPhaseTimes* p = _g1h->phase_times();
590
591 // Need to get the following two before the call to G1ParThreadScanState::flush()
592 // because it resets the PLAB allocator where we get this info from.
593 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
594 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
595 size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _n_workers) * HeapWordSize;
596
597 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
598 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
599 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
600
601 delete pss;
602 _states[worker_id] = NULL;
603 }
604 _flushed = true;
605 }
606
607 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
608 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
609 G1ParScanThreadState* pss = _states[worker_index];
610 assert(pss != nullptr, "must be initialized");
611
612 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
613 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
614 }
615 }
616
617 NOINLINE
618 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
619 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
620
621 oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
622 if (forward_ptr == NULL) {
623 // Forward-to-self succeeded. We are the "owner" of the object.
624 HeapRegion* r = _g1h->heap_region_containing(old);
625
626 if (_evac_failure_regions->record(r->hrm_index())) {
627 _g1h->hr_printer()->evac_failure(r);
628 }
629
630 // Mark the failing object in the marking bitmap and later use the bitmap to handle
631 // evacuation failure recovery.
632 _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
633
634 _preserved_marks->push_if_necessary(old, m);
635
636 ContinuationGCSupport::transform_stack_chunk(old);
637
638 _evacuation_failed_info.register_copy_failure(word_sz);
639
640 // For iterating objects that failed evacuation currently we can reuse the
641 // existing closure to scan evacuated objects because:
642 // - for objects referring into the collection set we do not need to gather
643 // cards at this time. The regions they are in will be unconditionally turned
644 // to old regions without remembered sets.
645 // - since we are iterating from a collection set region (i.e. never a Survivor
646 // region), we always need to gather cards for this case.
647 G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
648 old->oop_iterate_backwards(&_scanner);
649
650 return old;
651 } else {
652 // Forward-to-self failed. Either someone else managed to allocate
653 // space for this object (old != forward_ptr) or they beat us in
654 // self-forwarding it (old == forward_ptr).
655 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
656 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
657 "should not be in the CSet",
658 p2i(old), p2i(forward_ptr));
659 return forward_ptr;
660 }
661 }
662
663 void G1ParScanThreadState::initialize_numa_stats() {
664 if (_numa->is_enabled()) {
665 LogTarget(Info, gc, heap, numa) lt;
666
667 if (lt.is_enabled()) {
668 uint num_nodes = _numa->num_active_nodes();
669 // Record only if there are multiple active nodes.
670 _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
671 memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
672 }
673 }
674 }
675
676 void G1ParScanThreadState::flush_numa_stats() {
677 if (_obj_alloc_stat != NULL) {
678 uint node_index = _numa->index_of_current_thread();
679 _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
680 }
681 }
682
683 void G1ParScanThreadState::update_numa_stats(uint node_index) {
684 if (_obj_alloc_stat != NULL) {
685 _obj_alloc_stat[node_index]++;
686 }
687 }
688
689 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
690 uint n_workers,
691 size_t young_cset_length,
692 size_t optional_cset_length,
693 G1EvacFailureRegions* evac_failure_regions) :
694 _g1h(g1h),
695 _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
696 _preserved_marks_set(true /* in_c_heap */),
697 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
698 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
699 _young_cset_length(young_cset_length),
700 _optional_cset_length(optional_cset_length),
701 _n_workers(n_workers),
702 _flushed(false),
703 _evac_failure_regions(evac_failure_regions) {
704 _preserved_marks_set.init(n_workers);
705 for (uint i = 0; i < n_workers; ++i) {
706 _states[i] = NULL;
707 }
708 memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
709 }
710
711 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
712 assert(_flushed, "thread local state from the per thread states should have been flushed");
713 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
714 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
715 _preserved_marks_set.assert_empty();
716 _preserved_marks_set.reclaim();
717 }