1 /*
2 * Copyright (c) 2014, 2019, 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 _refs(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 _old_gen_is_full(false),
60 _num_optional_regions(optional_cset_length),
61 _numa(g1h->numa()),
62 _obj_alloc_stat(NULL)
63 {
64 // We allocate number of young gen regions in the collection set plus one
65 // entries, since entry 0 keeps track of surviving bytes for non-young regions.
66 // We also add a few elements at the beginning and at the end in
67 // an attempt to eliminate cache contention
68 size_t real_length = young_cset_length + 1;
69 size_t array_length = PADDING_ELEM_NUM + real_length + PADDING_ELEM_NUM;
70 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
71 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
72 memset(_surviving_young_words, 0, real_length * sizeof(size_t));
73
74 _plab_allocator = new G1PLABAllocator(_g1h->allocator());
75
76 // The dest for Young is used when the objects are aged enough to
77 // need to be moved to the next space.
78 _dest[G1HeapRegionAttr::Young] = G1HeapRegionAttr::Old;
79 _dest[G1HeapRegionAttr::Old] = G1HeapRegionAttr::Old;
80
81 _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
82
83 _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
84
85 initialize_numa_stats();
86 }
87
88 // Pass locally gathered statistics to global state.
89 void G1ParScanThreadState::flush(size_t* surviving_young_words) {
90 _rdcq.flush();
91 // Update allocation statistics.
92 _plab_allocator->flush_and_retire_stats();
93 _g1h->policy()->record_age_table(&_age_table);
94
95 uint length = _g1h->collection_set()->young_region_length() + 1;
96 for (uint i = 0; i < length; i++) {
97 surviving_young_words[i] += _surviving_young_words[i];
98 }
99 flush_numa_stats();
100 }
101
102 G1ParScanThreadState::~G1ParScanThreadState() {
103 delete _plab_allocator;
104 delete _closures;
105 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
106 delete[] _oops_into_optional_regions;
107 FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
108 }
109
110 size_t G1ParScanThreadState::lab_waste_words() const {
111 return _plab_allocator->waste();
112 }
113
114 size_t G1ParScanThreadState::lab_undo_waste_words() const {
115 return _plab_allocator->undo_waste();
116 }
117
118 #ifdef ASSERT
119 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
120 assert(ref != NULL, "invariant");
121 assert(UseCompressedOops, "sanity");
122 assert(!has_partial_array_mask(ref), "ref=" PTR_FORMAT, p2i(ref));
123 oop p = RawAccess<>::oop_load(ref);
124 assert(_g1h->is_in_g1_reserved(p),
125 "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p));
126 return true;
127 }
128
129 bool G1ParScanThreadState::verify_ref(oop* ref) const {
130 assert(ref != NULL, "invariant");
131 if (has_partial_array_mask(ref)) {
132 // Must be in the collection set--it's already been copied.
133 oop p = clear_partial_array_mask(ref);
134 assert(_g1h->is_in_cset(p),
135 "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p));
136 } else {
137 oop p = RawAccess<>::oop_load(ref);
138 assert(_g1h->is_in_g1_reserved(p),
139 "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p));
140 }
141 return true;
142 }
143
144 bool G1ParScanThreadState::verify_task(StarTask ref) const {
145 if (ref.is_narrow()) {
146 return verify_ref((narrowOop*) ref);
147 } else {
148 return verify_ref((oop*) ref);
149 }
150 }
151 #endif // ASSERT
152
153 void G1ParScanThreadState::trim_queue() {
154 StarTask ref;
155 do {
156 // Fully drain the queue.
157 trim_queue_to_threshold(0);
158 } while (!_refs->is_empty());
159 }
160
161 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
162 size_t word_sz,
163 bool previous_plab_refill_failed,
164 uint node_index) {
165
166 assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
167
168 // Right now we only have two types of regions (young / old) so
169 // let's keep the logic here simple. We can generalize it when necessary.
170 if (dest->is_young()) {
171 bool plab_refill_in_old_failed = false;
172 HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
173 word_sz,
174 &plab_refill_in_old_failed,
175 node_index);
176 // Make sure that we won't attempt to copy any other objects out
177 // of a survivor region (given that apparently we cannot allocate
178 // any new ones) to avoid coming into this slow path again and again.
179 // Only consider failed PLAB refill here: failed inline allocations are
180 // typically large, so not indicative of remaining space.
181 if (previous_plab_refill_failed) {
182 _tenuring_threshold = 0;
183 }
184
185 if (obj_ptr != NULL) {
186 dest->set_old();
187 } else {
188 // We just failed to allocate in old gen. The same idea as explained above
189 // for making survivor gen unavailable for allocation applies for old gen.
190 _old_gen_is_full = plab_refill_in_old_failed;
191 }
192 return obj_ptr;
193 } else {
194 _old_gen_is_full = previous_plab_refill_failed;
195 assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
196 // no other space to try.
197 return NULL;
198 }
199 }
200
201 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
202 if (region_attr.is_young()) {
203 age = !m.has_displaced_mark_helper() ? m.age()
204 : m.displaced_mark_helper().age();
205 if (age < _tenuring_threshold) {
206 return region_attr;
207 }
208 }
209 return dest(region_attr);
210 }
211
212 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
213 oop const old, size_t word_sz, uint age,
214 HeapWord * const obj_ptr, uint node_index) const {
215 PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
216 if (alloc_buf->contains(obj_ptr)) {
217 _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
218 dest_attr.type() == G1HeapRegionAttr::Old,
219 alloc_buf->word_sz() * HeapWordSize);
220 } else {
221 _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
222 dest_attr.type() == G1HeapRegionAttr::Old);
223 }
224 }
225
226 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr const region_attr,
227 oop const old,
228 markWord const old_mark) {
229 const size_t word_sz = old->size();
230
231 uint age = 0;
232 G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
233 // The second clause is to prevent premature evacuation failure in case there
234 // is still space in survivor, but old gen is full.
235 if (_old_gen_is_full && dest_attr.is_old()) {
236 return handle_evacuation_failure_par(old, old_mark);
237 }
238 HeapRegion* const from_region = _g1h->heap_region_containing(old);
239 uint node_index = from_region->node_index();
240
241 HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
242
243 // PLAB allocations should succeed most of the time, so we'll
244 // normally check against NULL once and that's it.
245 if (obj_ptr == NULL) {
246 bool plab_refill_failed = false;
247 obj_ptr = _plab_allocator->allocate_direct_or_new_plab(dest_attr, word_sz, &plab_refill_failed, node_index);
248 if (obj_ptr == NULL) {
249 assert(region_attr.is_in_cset(), "Unexpected region attr type: %s", region_attr.get_type_str());
250 obj_ptr = allocate_in_next_plab(&dest_attr, word_sz, plab_refill_failed, node_index);
251 if (obj_ptr == NULL) {
252 // This will either forward-to-self, or detect that someone else has
253 // installed a forwarding pointer.
254 return handle_evacuation_failure_par(old, old_mark);
255 }
256 }
257 update_numa_stats(node_index);
258
259 if (_g1h->_gc_tracer_stw->should_report_promotion_events()) {
260 // The events are checked individually as part of the actual commit
261 report_promotion_event(dest_attr, old, word_sz, age, obj_ptr, node_index);
262 }
263 }
264
265 assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
266 assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
267
268 #ifndef PRODUCT
269 // Should this evacuation fail?
270 if (_g1h->evacuation_should_fail()) {
271 // Doing this after all the allocation attempts also tests the
272 // undo_allocation() method too.
273 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
274 return handle_evacuation_failure_par(old, old_mark);
275 }
276 #endif // !PRODUCT
277
278 // We're going to allocate linearly, so might as well prefetch ahead.
279 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
280
281 const oop obj = oop(obj_ptr);
282 const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
283 if (forward_ptr == NULL) {
284 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
285
286 const uint young_index = from_region->young_index_in_cset();
287
288 assert((from_region->is_young() && young_index > 0) ||
289 (!from_region->is_young() && young_index == 0), "invariant" );
290
291 if (dest_attr.is_young()) {
292 if (age < markWord::max_age) {
293 age++;
294 }
295 if (old_mark.has_displaced_mark_helper()) {
296 // In this case, we have to install the mark word first,
297 // otherwise obj looks to be forwarded (the old mark word,
298 // which contains the forward pointer, was copied)
299 obj->set_mark_raw(old_mark);
300 markWord new_mark = old_mark.displaced_mark_helper().set_age(age);
301 old_mark.set_displaced_mark_helper(new_mark);
302 } else {
303 obj->set_mark_raw(old_mark.set_age(age));
304 }
305 _age_table.add(age, word_sz);
306 } else {
307 obj->set_mark_raw(old_mark);
308 }
309
310 if (G1StringDedup::is_enabled()) {
311 const bool is_from_young = region_attr.is_young();
312 const bool is_to_young = dest_attr.is_young();
313 assert(is_from_young == from_region->is_young(),
314 "sanity");
315 assert(is_to_young == _g1h->heap_region_containing(obj)->is_young(),
316 "sanity");
317 G1StringDedup::enqueue_from_evacuation(is_from_young,
318 is_to_young,
319 _worker_id,
320 obj);
321 }
322
323 _surviving_young_words[young_index] += word_sz;
324
325 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
326 // We keep track of the next start index in the length field of
327 // the to-space object. The actual length can be found in the
328 // length field of the from-space object.
329 arrayOop(obj)->set_length(0);
330 oop* old_p = set_partial_array_mask(old);
331 do_oop_partial_array(old_p);
332 } else {
333 G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
334 obj->oop_iterate_backwards(&_scanner);
335 }
336 return obj;
337 } else {
338 _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
339 return forward_ptr;
340 }
341 }
342
343 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
344 assert(worker_id < _n_workers, "out of bounds access");
345 if (_states[worker_id] == NULL) {
346 _states[worker_id] =
347 new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length);
348 }
349 return _states[worker_id];
350 }
351
352 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
353 assert(_flushed, "thread local state from the per thread states should have been flushed");
354 return _surviving_young_words_total;
355 }
356
357 void G1ParScanThreadStateSet::flush() {
358 assert(!_flushed, "thread local state from the per thread states should be flushed once");
359
360 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
361 G1ParScanThreadState* pss = _states[worker_index];
362
363 if (pss == NULL) {
364 continue;
365 }
366
367 pss->flush(_surviving_young_words_total);
368 delete pss;
369 _states[worker_index] = NULL;
370 }
371 _flushed = true;
372 }
373
374 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
375 for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
376 G1ParScanThreadState* pss = _states[worker_index];
377
378 if (pss == NULL) {
379 continue;
380 }
381
382 size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
383 _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
384 }
385 }
386
387 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m) {
388 assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
389
390 oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
391 if (forward_ptr == NULL) {
392 // Forward-to-self succeeded. We are the "owner" of the object.
393 HeapRegion* r = _g1h->heap_region_containing(old);
394
395 if (!r->evacuation_failed()) {
396 r->set_evacuation_failed(true);
397 _g1h->hr_printer()->evac_failure(r);
398 }
399
400 _g1h->preserve_mark_during_evac_failure(_worker_id, old, m);
401
402 G1ScanInYoungSetter x(&_scanner, r->is_young());
403 old->oop_iterate_backwards(&_scanner);
404
405 return old;
406 } else {
407 // Forward-to-self failed. Either someone else managed to allocate
408 // space for this object (old != forward_ptr) or they beat us in
409 // self-forwarding it (old == forward_ptr).
410 assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
411 "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
412 "should not be in the CSet",
413 p2i(old), p2i(forward_ptr));
414 return forward_ptr;
415 }
416 }
417 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
418 G1RedirtyCardsQueueSet* rdcqs,
419 uint n_workers,
420 size_t young_cset_length,
421 size_t optional_cset_length) :
422 _g1h(g1h),
423 _rdcqs(rdcqs),
424 _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
425 _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
426 _young_cset_length(young_cset_length),
427 _optional_cset_length(optional_cset_length),
428 _n_workers(n_workers),
429 _flushed(false) {
430 for (uint i = 0; i < n_workers; ++i) {
431 _states[i] = NULL;
432 }
433 memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
434 }
435
436 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
437 assert(_flushed, "thread local state from the per thread states should have been flushed");
438 FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
439 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
440 }