1 /*
2 * Copyright (c) 2014, 2020, 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/g1OopClosures.inline.hpp"
30 #include "gc/g1/g1ParScanThreadState.inline.hpp"
31 #include "gc/g1/g1RootClosures.hpp"
32 #include "gc/g1/g1StringDedup.hpp"
33 #include "gc/g1/g1Trace.hpp"
34 #include "gc/shared/taskqueue.inline.hpp"
35 #include "memory/allocation.inline.hpp"
36 #include "oops/access.inline.hpp"
37 #include "oops/oop.inline.hpp"
38 #include "runtime/prefetch.inline.hpp"
39
40 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
41 G1RedirtyCardsQueueSet* rdcqs,
42 uint worker_id,
43 size_t young_cset_length,
44 size_t optional_cset_length)
45 : _g1h(g1h),
46 _task_queue(g1h->task_queue(worker_id)),
47 _rdcq(rdcqs),
48 _ct(g1h->card_table()),
49 _closures(NULL),
50 _plab_allocator(NULL),
51 _age_table(false),
52 _tenuring_threshold(g1h->policy()->tenuring_threshold()),
53 _scanner(g1h, this),
54 _worker_id(worker_id),
55 _last_enqueued_card(SIZE_MAX),
56 _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
57 _stack_trim_lower_threshold(GCDrainStackTargetSize),
58 _trim_ticks(),
59 _surviving_young_words_base(NULL),
60 _surviving_young_words(NULL),
61 _surviving_words_length(young_cset_length + 1),
62 _old_gen_is_full(false),
63 _num_optional_regions(optional_cset_length),
64 _numa(g1h->numa()),
65 _obj_alloc_stat(NULL)
66 {
67 // We allocate number of young gen regions in the collection set plus one
68 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
69 // We also add a few elements at the beginning and at the end in
70 // an attempt to eliminate cache contention
71 const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
72 size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
73
74 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
75 _surviving_young_words = _surviving_young_words_base + padding_elem_num;
76 memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
77
78 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
79
80 // The dest for Young is used when the objects are aged enough to
81 // need to be moved to the next space.
82 _dest[G1HeapRegionAttr::Young] = G1HeapRegionAttr::Old;
83 _dest[G1HeapRegionAttr::Old] = G1HeapRegionAttr::Old;
84
85 _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
86
87 _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
88
89 initialize_numa_stats();
90 }
91
92 size_t G1ParScanThreadState::flush(size_t* surviving_young_words) {
93 _rdcq.flush();
94 flush_numa_stats();
95 // Update allocation statistics.
96 _plab_allocator->flush_and_retire_stats();
97 _g1h->policy()->record_age_table(&_age_table);
98
99 size_t sum = 0;
100 for (uint i = 0; i < _surviving_words_length; i++) {
101 surviving_young_words[i] += _surviving_young_words[i];
102 sum += _surviving_young_words[i];
103 }
104 return sum;
105 }
106
107 G1ParScanThreadState::~G1ParScanThreadState() {
108 delete _plab_allocator;
109 delete _closures;
110 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
111 delete[] _oops_into_optional_regions;
112 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
113 }
114
115 size_t G1ParScanThreadState::lab_waste_words() const {
116 return _plab_allocator->waste();
117 }
118
119 size_t G1ParScanThreadState::lab_undo_waste_words() const {
120 return _plab_allocator->undo_waste();
121 }
122
123 #ifdef ASSERT
124 void G1ParScanThreadState::verify_task(narrowOop* task) const {
125 assert(task != NULL, "invariant");
126 assert(UseCompressedOops, "sanity");
127 oop p = RawAccess<>::oop_load(task);
128 assert(_g1h->is_in_g1_reserved(p),
129 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
130 }
131
132 void G1ParScanThreadState::verify_task(oop* task) const {
133 assert(task != NULL, "invariant");
134 oop p = RawAccess<>::oop_load(task);
135 assert(_g1h->is_in_g1_reserved(p),
136 "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
137 }
138
139 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
140 // Must be in the collection set--it's already been copied.
141 oop p = task.to_source_array();
142 assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
143 }
144
145 void G1ParScanThreadState::verify_task(ScannerTask task) const {
146 if (task.is_narrow_oop_ptr()) {
147 verify_task(task.to_narrow_oop_ptr());
148 } else if (task.is_oop_ptr()) {
149 verify_task(task.to_oop_ptr());
150 } else if (task.is_partial_array_task()) {
151 verify_task(task.to_partial_array_task());
152 } else {
153 ShouldNotReachHere();
154 }
155 }
156 #endif // ASSERT
157
158 void G1ParScanThreadState::trim_queue() {
159 do {
160 // Fully drain the queue.
161 trim_queue_to_threshold(0);
162 } while (!_task_queue->is_empty());
163 }
164
165 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
166 size_t word_sz,
167 bool previous_plab_refill_failed,
168 uint node_index) {
169
170 assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
171
172 // Right now we only have two types of regions (young / old) so
173 // let's keep the logic here simple. We can generalize it when necessary.
174 if (dest->is_young()) {
175 bool plab_refill_in_old_failed = false;
176 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
177 word_sz,
178 &plab_refill_in_old_failed,
179 node_index);
180 // Make sure that we won't attempt to copy any other objects out
181 // of a survivor region (given that apparently we cannot allocate
182 // any new ones) to avoid coming into this slow path again and again.
183 // Only consider failed PLAB refill here: failed inline allocations are
184 // typically large, so not indicative of remaining space.
185 if (previous_plab_refill_failed) {
186 _tenuring_threshold = 0;
187 }
188
189 if (obj_ptr != NULL) {
190 dest->set_old();
191 } else {
192 // We just failed to allocate in old gen. The same idea as explained above
193 // for making survivor gen unavailable for allocation applies for old gen.
194 _old_gen_is_full = plab_refill_in_old_failed;
195 }
196 return obj_ptr;
197 } else {
198 _old_gen_is_full = previous_plab_refill_failed;
199 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
200 // no other space to try.
201 return NULL;
202 }
203 }
204
205 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
206 if (region_attr.is_young()) {
207 age = !m.has_displaced_mark_helper() ? m.age()
208 : m.displaced_mark_helper().age();
209 if (age < _tenuring_threshold) {
210 return region_attr;
211 }
212 }
213 return dest(region_attr);
214 }
215
216 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
217 oop const old, size_t word_sz, uint age,
218 HeapWord * const obj_ptr, uint node_index) const {
219 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
220 if (alloc_buf->contains(obj_ptr)) {
221 _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
222 dest_attr.type() == G1HeapRegionAttr::Old,
223 alloc_buf->word_sz() * HeapWordSize);
224 } else {
225 _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
226 dest_attr.type() == G1HeapRegionAttr::Old);
227 }
228 }
229
230 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr const region_attr,
231 oop const old,
232 markWord const old_mark) {
233 const size_t word_sz = old->size();
234
235 uint age = 0;
236 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
237 // The second clause is to prevent premature evacuation failure in case there
238 // is still space in survivor, but old gen is full.
239 if (_old_gen_is_full && dest_attr.is_old()) {
240 return handle_evacuation_failure_par(old, old_mark);
241 }
242 HeapRegion* const from_region = _g1h->heap_region_containing(old);
243 uint node_index = from_region->node_index();
244
245 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
246
247 // PLAB allocations should succeed most of the time, so we'll
248 // normally check against NULL once and that's it.
249 if (obj_ptr == NULL) {
250 bool plab_refill_failed = false;
251 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(dest_attr, word_sz, &plab_refill_failed, node_index);
252 if (obj_ptr == NULL) {
253 assert(region_attr.is_in_cset(), "Unexpected region attr type: %s", region_attr.get_type_str());
254 obj_ptr = allocate_in_next_plab(&dest_attr, word_sz, plab_refill_failed, node_index);
255 if (obj_ptr == NULL) {
256 // This will either forward-to-self, or detect that someone else has
257 // installed a forwarding pointer.
258 return handle_evacuation_failure_par(old, old_mark);
259 }
260 }
261 update_numa_stats(node_index);
262
263 if (_g1h->_gc_tracer_stw->should_report_promotion_events()) {
264 // The events are checked individually as part of the actual commit
265 report_promotion_event(dest_attr, old, word_sz, age, obj_ptr, node_index);
266 }
267 }
268
269 assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
270 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
271
272 #ifndef PRODUCT
273 // Should this evacuation fail?
274 if (_g1h->evacuation_should_fail()) {
275 // Doing this after all the allocation attempts also tests the
276 // undo_allocation() method too.
277 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
278 return handle_evacuation_failure_par(old, old_mark);
279 }
280 #endif // !PRODUCT
281
282 // We're going to allocate linearly, so might as well prefetch ahead.
283 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
284
285 const oop obj = oop(obj_ptr);
286 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
287 if (forward_ptr == NULL) {
288 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
289
290 const uint young_index = from_region->young_index_in_cset();
291
292 assert((from_region->is_young() && young_index > 0) ||
293 (!from_region->is_young() && young_index == 0), "invariant" );
294
295 if (dest_attr.is_young()) {
296 if (age < markWord::max_age) {
297 age++;
298 }
299 if (old_mark.has_displaced_mark_helper()) {
300 // In this case, we have to install the mark word first,
301 // otherwise obj looks to be forwarded (the old mark word,
302 // which contains the forward pointer, was copied)
303 obj->set_mark_raw(old_mark);
304 markWord new_mark = old_mark.displaced_mark_helper().set_age(age);
305 old_mark.set_displaced_mark_helper(new_mark);
306 } else {
307 obj->set_mark_raw(old_mark.set_age(age));
308 }
309 _age_table.add(age, word_sz);
310 } else {
311 obj->set_mark_raw(old_mark);
312 }
313
314 if (G1StringDedup::is_enabled()) {
315 const bool is_from_young = region_attr.is_young();
316 const bool is_to_young = dest_attr.is_young();
317 assert(is_from_young == from_region->is_young(),
318 "sanity");
319 assert(is_to_young == _g1h->heap_region_containing(obj)->is_young(),
320 "sanity");
321 G1StringDedup::enqueue_from_evacuation(is_from_young,
322 is_to_young,
323 _worker_id,
324 obj);
325 }
326
327 _surviving_young_words[young_index] += word_sz;
328
329 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
330 // We keep track of the next start index in the length field of
331 // the to-space object. The actual length can be found in the
332 // length field of the from-space object.
333 arrayOop(obj)->set_length(0);
334 do_partial_array(PartialArrayScanTask(old));
335 } else {
336 G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
337 obj->oop_iterate_backwards(&_scanner);
338 }
339 return obj;
340 } else {
341 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
342 return forward_ptr;
343 }
344 }
345
346 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
347 assert(worker_id < _n_workers, "out of bounds access");
348 if (_states[worker_id] == NULL) {
349 _states[worker_id] =
350 new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length);
351 }
352 return _states[worker_id];
353 }
354
355 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
356 assert(_flushed, "thread local state from the per thread states should have been flushed");
357 return _surviving_young_words_total;
358 }
359
360 void G1ParScanThreadStateSet::flush() {
361 assert(!_flushed, "thread local state from the per thread states should be flushed once");
362
363 for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
364 G1ParScanThreadState* pss = _states[worker_id];
365
366 if (pss == NULL) {
367 continue;
368 }
369
370 G1GCPhaseTimes* p = _g1h->phase_times();
371
372 // Need to get the following two before the call to G1ParThreadScanState::flush()
373 // because it resets the PLAB allocator where we get this info from.
374 size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
375 size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
376 size_t copied_bytes = pss->flush(_surviving_young_words_total) * HeapWordSize;
377
378 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
379 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
380 p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
381
382 delete pss;
383 _states[worker_id] = NULL;
384 }
385 _flushed = true;
386 }
387
388 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
389 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
390 G1ParScanThreadState* pss = _states[worker_index];
391
392 if (pss == NULL) {
393 continue;
394 }
395
396 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
397 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
398 }
399 }
400
401 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m) {
402 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
403
404 oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
405 if (forward_ptr == NULL) {
406 // Forward-to-self succeeded. We are the "owner" of the object.
407 HeapRegion* r = _g1h->heap_region_containing(old);
408
409 if (!r->evacuation_failed()) {
410 r->set_evacuation_failed(true);
411 _g1h->hr_printer()->evac_failure(r);
412 }
413
414 _g1h->preserve_mark_during_evac_failure(_worker_id, old, m);
415
416 G1ScanInYoungSetter x(&_scanner, r->is_young());
417 old->oop_iterate_backwards(&_scanner);
418
419 return old;
420 } else {
421 // Forward-to-self failed. Either someone else managed to allocate
422 // space for this object (old != forward_ptr) or they beat us in
423 // self-forwarding it (old == forward_ptr).
424 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
425 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
426 "should not be in the CSet",
427 p2i(old), p2i(forward_ptr));
428 return forward_ptr;
429 }
430 }
431 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
432 G1RedirtyCardsQueueSet* rdcqs,
433 uint n_workers,
434 size_t young_cset_length,
435 size_t optional_cset_length) :
436 _g1h(g1h),
437 _rdcqs(rdcqs),
438 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
439 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
440 _young_cset_length(young_cset_length),
441 _optional_cset_length(optional_cset_length),
442 _n_workers(n_workers),
443 _flushed(false) {
444 for (uint i = 0; i < n_workers; ++i) {
445 _states[i] = NULL;
446 }
447 memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
448 }
449
450 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
451 assert(_flushed, "thread local state from the per thread states should have been flushed");
452 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
453 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
454 }